Lewis's Triviality Result¶

An impossibility argument showing that a fixed conditional proposition cannot generally have probability equal to conditional probability throughout a nontrivial class closed under relevant conditioning.

Version: v1 · 2026-10-03 · History
Domain-specific #: 13383
Domain group: Humanities
Origin domain: Philosophy
Subdomains: Conditional Logic, Probability of Conditionals → Philosophy
Aliases: Lewis triviality theorem, Triviality of conditional events, Lewis's first and second triviality results

Core Idea¶

Lewis's triviality result is an impossibility argument about trying to identify the probability of an ordinary conditional proposition, “if \(A\), then \(C\),” with the conditional probability \(P(C\mid A)\) in general. Write \(E=A\Rightarrow C\) for a proposed fixed event or proposition and demand the Equation \(P(E)=P(C\mid A)\) whenever \(P(A)>0\). If that same interpretation and equation remain in force across a sufficiently rich class of probability functions closed under the relevant conditioning updates, then a simple decomposition forces \(P(C\mid A)=P(C)\) even when \(A\) and \(C\) are dependent. A nontrivial probability model supplies a counterexample, so the joint demands cannot all be maintained.^[1]^[2]

The proof is exact under its assumptions. Suppose \(P(A\cap C)>0\) and \(P(A\cap\neg C)>0\), and the class contains \(P_C(\cdot)=P(\cdot\mid C)\) and \(P_{\neg C}(\cdot)=P(\cdot\mid\neg C)\). Applying the Equation to those updated measures gives \(P(E\mid C)=P(C\mid A\cap C)=1\) and \(P(E\mid\neg C)=P(C\mid A\cap\neg C)=0\). Because \(E\) is one ordinary event, total probability yields \(P(E)=1P(C)+0P(\neg C)=P(C)\). The prior Equation also gives \(P(E)=P(C\mid A)\). Hence it forces \(P(C\mid A)=P(C)\), an independence condition not generally true.^[1]

Lewis distinguished a first result against a uniform Equation for all probability functions from a second against its application throughout a restricted class of belief functions closed under conditionalization. His 1986 sequel restates that distinction and strengthens the response to objections about which updates are admissible. The result does not say conditional probabilities are inconsistent, that ordinary-language conditionals are meaningless, or that every formal conditional-event calculus is impossible. It says a particular conjunction of eventhood, uniformity, Equation, update closure and nontrivial dependence is untenable.^[1]^[2]

Structural Signature¶

Sig role-phrases: fixed conditional proposition → probability Equation → update-closed class → positive dependent witness → total-probability collapse.

Fixed conditional proposition: \(E=A\Rightarrow C\) must denote one ordinary event before and after conditioning. If the interpretation changes with \(P\), the two-case proof is no longer about one \(E\). Lewis explicitly marks nonuniform interpretations as outside this argument.^[2]
Probability Equation: For each eligible measure \(Q\), require \(Q(E)=Q(C\mid A)\) when \(Q(A)>0\). The theorem targets this stronger proposition-probability claim, not the bare calculation of \(P(C\mid A)\).
Update-closed class: The relevant class must include the conditionals of \(P\) on \(C\) and \(\neg C\) when they are defined. A universal class has this property; a restricted belief-function class needs it as an additional assumption.^[2]
Positive dependent witness: Both \(A\cap C\) and \(A\cap\neg C\) must have positive mass so the conditioned expressions exist, and some admissible pair must satisfy \(P(C\mid A)\ne P(C)\). Without the latter, the conclusion is not a contradiction.
Total-probability collapse: Eventhood licenses partitioning \(P(E)\) by \(C\) and \(\neg C\). The Equation makes the two conditional terms 1 and 0, leaving \(P(E)=P(C)\) and colliding with the prior Equation.^[1]

The role of an ordinary event is not ornamental. If a theory assigns conditional sentences a nonclassical status for which standard event decomposition is unavailable, this particular proof cannot simply be copied; one must inspect that theory's own algebra. Conversely, writing an arrow is not enough to guarantee a new event satisfies the Equation.

What It Is Not¶

It is not a denial of conditional probability. \(P(C\mid A)=P(A\cap C)/P(A)\) is well-defined when \(P(A)>0\); the failure concerns identifying it with the unconditional probability of one embeddable proposition under all the relevant measures. Nor is the theorem simply a critique of the material conditional \(\neg A\lor C\). Lewis's argument ranges over proposed uniform interpretations of the arrow, not just that one truth table. A material conditional usually fails the Equation already, but that is a narrower observation.^[1]^[2]

It does not refute Adams's assertability thesis merely by name. Adams related the assertability of indicative conditionals to conditional subjective probability. Lewis targets an attractive explanation that turns the conditional into an ordinary proposition whose probability is exactly that conditional probability. An account can retain the assertability insight while rejecting that explanatory equation.^[3]^[2]

The result is not a declaration that no three-valued, measure-indexed, or otherwise revised conditional system can be studied. Such approaches may reject or alter one premise of this argument; whether they succeed on their own terms requires separate proofs. Similarly, a one-off equality for a particular \(A,C,P\) is not the universal or class-wide Equation Lewis tests.

Scope of Application¶

The result belongs to probability-based theories of indicative conditionals, conditional-event semantics, and formal epistemology. It is useful whenever one proposes to put “if \(A\), then \(C\)” inside a Boolean event algebra and assign it the probability \(P(C\mid A)\), including after beliefs are updated by conditioning. The theorem forces that proposal to declare which commitments it retains and which it relaxes.^[1]^[2]

The same diagnostic applies to a designed reasoning language or knowledge base if it gives conditionals stable event values and uses ordinary probability over compound formulas. The theorem does not say a deployed AI system necessarily uses that semantics, nor does it predict errors in a real system without examining its formal rules. A system that treats \(P(C\mid A)\) as a numerical query, without making an embeddable event \(A\Rightarrow C\), does not trigger Lewis's contradiction.

Lewis's 1986 extension also matters for scope. His earlier second result assumed closure under conditioning broadly; he later considered narrower evidence updates, including a finite partition, and showed a related collapse under additional nontriviality conditions. One should not cite the 1976 proof alone as establishing every imaginable update policy; the exact update closure must be stated.^[2]

Clarity¶

First distinguish three expressions: the numerical \(P(C\mid A)\), the candidate proposition \(E=A\Rightarrow C\), and the numerical \(P(E)\). The first is a conditional measure value; the second must be an event if ordinary Boolean operations and total probability are to apply; the third is its unconditional probability. Their equal sign is a substantive thesis, not probability notation.^[1]

Next state the quantifiers. Does the Equation hold only for one \(P\), for every probability function, or for a specified class? Is the arrow interpretation the same across the class? Is the class closed under conditioning on \(C\) and \(\neg C\)? Are \(P(A\cap C)\) and \(P(A\cap\neg C)\) positive? The contradiction requires all these answers, plus a genuinely dependent witness. Without them, “Lewis proves conditionals cannot be probabilities” is too broad.^[2]

The proof does not infer independence from the definition of a conditional. It derives independence only after substituting the proposed Equation into the total-probability decomposition of one event. That is why changing the treatment of eventhood, update closure, or uniformity can block the inference while leaving ordinary \(P(C\mid A)\) intact.

Manages Complexity¶

The theorem condenses a semantic design problem into a five-part compatibility test. Rather than arguing vaguely over whether conditional sentences “have probabilities,” it asks whether a stable event denotation, conditional-probability Equation, and conditioning-closed credence class can coexist in a model where antecedent and consequent are dependent. A four-cell table or a fair die exposes the contradiction without settling every philosophical theory of “if.”^[1]

It also separates static equality from dynamic stability. A chosen \(E\) might be engineered so that \(P(E)=P(C\mid A)\) for one particular prior. But if the same event must keep that property after learning \(C\) or \(\neg C\), its conditional probabilities become 1 and 0, forcing its prior probability to be \(P(C)\). This is the design pressure: a proposition may be embeddable in ordinary compounds, while an information-relative conditional probability changes its conditioning context. Trying to have both uniformly erases dependence.^[1]^[2]

Abstract Reasoning¶

Specify the probability space, events \(A,C\), the proposed event \(E=A\Rightarrow C\), and a class \(\mathcal K\) of eligible probability functions.
Assert the Equation for \(P\in\mathcal K\) and for any relevant conditioned measures in \(\mathcal K\), never for undefined zero-denominator cases.
Choose \(P(A\cap C)>0\) and \(P(A\cap\neg C)>0\), so both updates permit conditioning on \(A\).
Under \(P_C\), the Equation gives \(P(E\mid C)=1\); under \(P_{\neg C}\), it gives \(P(E\mid\neg C)=0\).
Total probability for the same event gives \(P(E)=P(C)\), while the original Equation gives \(P(E)=P(C\mid A)\).
Exhibit a model with \(P(C\mid A)\ne P(C)\) and identify exactly which assumption an alternative semantics gives up.^[1]^[2]

Knowledge Transfer¶

The proof illustrates a general impossibility-testing move: take a proposed representation equation, apply it under sanctioned transformations, and test whether recombining the transformed cases preserves the original quantity. Here that move is fully formal because conditional probability, event additivity and belief-update closure have precise meanings. The transferable method does not make Lewis's named result itself a prime; outside probabilistic conditional semantics, new assumptions and a new proof are required.

It also teaches a semantic diagnostic for software and formal languages: if an expression is intended to be an ordinary event inside disjunctions, its probability must obey ordinary event algebra. If the expression is instead an information-relative numerical query, it need not be a proposition with truth conditions. Confusing those roles is the exact kind of type mismatch Lewis's result uncovers.^[2]

Examples¶

1. Lewis's fair die. Let a fair die have six equiprobable outcomes. Take \(A\) as “an even number” and \(C\) as “a six.” Then \(P(A\cap C)=1/6\), \(P(A\cap\neg C)=2/6\), \(P(C\mid A)=1/3\), and \(P(C)=1/6\). If a fixed event \(E=A\Rightarrow C\) satisfied the Equation before and after conditioning on \(C\) and \(\neg C\), the two updated values would be 1 and 0. Total probability would force \(P(E)=1/6\), while the prior Equation requires \(P(E)=1/3\). This is Lewis's original nontrivial counterexample.^[1]

Mapped back: The fixed conditional proposition is one event \(E\), not a changing arrow interpretation. The Equation asks for the prior \(1/3\). The update-closed class includes the die distributions conditioned on six and not-six. Both even-and-six and even-and-not-six have positive mass, supplying the dependent witness. The total-probability collapse yields \(1\cdot(1/6)+0\cdot(5/6)=1/6\), contradicting \(1/3\).

2. A schematic probabilistic rule in a knowledge base. Suppose \(A\) means an alarm sounds and \(C\) means a fault is present. Assign joint probabilities \(P(A\cap C)=0.3\), \(P(A\cap\neg C)=0.1\), \(P(\neg A\cap C)=0.1\), and \(P(\neg A\cap\neg C)=0.5\). These numbers are a constructed formal model, not measured device performance. A program that stores “if alarm then fault” as one ordinary Boolean event \(E\), yet requires \(P(E)=P(C\mid A)=0.3/0.4=0.75\) before and after ordinary conditioning updates, will also derive \(P(E)=P(C)=0.4\) by the same partition argument. The contradiction is in the proposed semantics, not the four-cell probabilities.^[1]^[2]

Mapped back: \(E\) remains the same embeddable event across updates. The Equation demands prior probability 0.75. Update closure allows conditioning on fault and no fault, which set \(P(E\mid C)=1\) and \(P(E\mid\neg C)=0\). Both alarm branches are positive, and \(0.75\ne 0.4\) is the nontrivial witness. Recombination produces \(1\cdot0.4+0\cdot0.6=0.4\), so the conjunction of design requirements fails. A system that merely queries \(P(C\mid A)\) without storing an event \(E\) is outside the example's target.

Structural Tensions¶

Embeddable proposition versus conditional credence. A stable proposition supports ordinary Boolean compounds and total probability; an information-relative conditional probability is computed by renormalizing to its antecedent. Equating the two across updates turns that useful embedding into the source of collapse.^[1] Diagnostic: Is the arrow one fixed event under every eligible measure, or a context-dependent numerical operation?

Restricted belief functions versus update closure. Restricting which probability functions represent beliefs might avoid the first universal result, but including both relevant conditioned measures brings back the two-case contradiction. Lewis later questioned how broad a closure premise rational belief needs and developed finer partition-based results.^[2] Diagnostic: Which updates are actually admitted, and do they include the two used in the proof?

Sharp impossibility versus overgeneralized slogan. The theorem gives a decisive contradiction for a package of assumptions in a dependent model. It does not itself choose which premise to sacrifice, or show that every nonclassical conditional account fails. Diagnostic: For a proposed escape, name the exact failed premise and the semantic capabilities retained or lost.

Structural–Framed Character¶

Evaluative weight: The contradiction is mathematical once its premises and positive witness are given. Calling a model “trivial” has technical content—relevant dependence cannot be expressed—not a moral or aesthetic condemnation of the model. Whether a proposed semantics is useful is a further assessment.^[1]
Human-practice dependence: Researchers choose whether an indicative conditional is treated as a proposition, which credences count as belief functions, and which updates are admissible. Once those choices fix a formal system, the derivation does not depend on reader opinion. A knowledge-base designer faces the same conditional design choice but does not automatically instantiate the theorem merely by writing an if–then rule.
Institutional origin: Lewis's named result arose in philosophical logic and probability research responding to Adams and Stalnaker. No philosophical school or publisher grants an instance theoremhood; the incompatibility is established by the premises and proof. The name records intellectual history, not a certification rule.^[3]^[1]^[2]
Vocabulary travel: “Lewis triviality” travels literally among conditional semantics, formal epistemology, and probabilistic reasoning systems when the fixed-event Equation and update closure can be tested. Calling any disappointing probability theorem “triviality” does not preserve its specific conditional-event mechanism.
Import versus recognition: One can recognize the Lewis pattern in a newly specified calculus by checking eventhood, uniform Equation, relevant update closure and a dependent witness. A designer may import the theorem's test when building a new conditional representation, but a generic warning that “conditions are tricky” is only analogy.

Its character: A strongly structural impossibility result within a philosophically framed choice of conditional semantics. Its algebraic contradiction is objective given the assumptions, while the decision to regard conditionals as ordinary propositions and the breadth of admissible belief updates are explicit modeling commitments. The wider pattern of testing representation equations under transformations is a future-prime question, not evidence that Lewis's named theorem is itself substrate-independent.

Structural Core vs. Domain Accent¶

Portable skeleton and necessary core: A proposed identity between two representations can fail when the identity must survive transformations and recombination. That broad stress-test is a possible future-prime candidate, not a verified live parent prime. The actual theorem's core is narrower: one fixed conditional event, the Equation, an update-closed probability class, positive antecedent branches, and total-probability recombination.^[1]

Domain accent and variable presentation: Antecedent/consequent notation, whether the witness uses a die or a four-cell table, and whether the eligible class is universal or a specified belief class can vary if the corresponding quantifiers and closure assumptions are stated. Ordinary event probability, conditioning, and a dependent \(A,C\) pair cannot be removed. The first and second Lewis results differ in quantifier scope; they should not be silently collapsed.^[1]^[2]

Why this is not a prime: The result does not apply to any arbitrary equation under any arbitrary update. Its contradiction uses conditional probability's ratio, event additivity, conditioning on \(C\) and \(\neg C\), and a conditional-proposition proposal. A transfer to nonprobabilistic systems requires a newly proved analogue. The live Conditional Probability prime is a necessary operation, represented by the proposed presupposition edge, not a strict genus of this theorem.

This entry presupposes Conditional Probability. Lewis's argument presupposes conditional probability to state and test the proposed Equation.

Relationships to Other Abstractions¶

Current abstraction Lewis's Triviality Result Domain-specific

Parents (1) — more general patterns this builds on

Lewis's Triviality Result presupposes Conditional Probability Prime

Lewis's argument presupposes conditional probability to state and test the proposed Equation.

Hierarchy paths (2) — routes to 2 parentless roots

Lewis's Triviality Result → Conditional Probability → Probability → Measure → Aggregation → Micro Macro Linkage

Show alternative path (1)

Neighborhood in Abstraction Space¶

Lewis's Triviality Result sits in a sparse region of the domain-specific corpus (83^rd percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Argumentation Fallacies & Inference (15 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-10-08

Not to Be Confused With¶

Adams's assertability thesis: Relates conditional assertability to conditional subjective probability; it does not by itself require a single embeddable conditional event whose unconditional probability equals that value.^[3]^[2]
Material implication's mismatch: The material truth function is one particular conditional interpretation. Lewis's reductio applies to any uniform ordinary event interpretation satisfying the Equation across the required measures.
A one-prior matching trick: Choosing an event whose probability happens to equal \(P(C\mid A)\) for one \(P\) does not ensure the same event preserves the Equation after conditioning.
Conditional probability itself: \(P(C\mid A)\) remains valid; the theorem blocks one proposed identification with \(P(A\Rightarrow C)\).
A claim that every conditional semantics is impossible: Changing eventhood, uniformity, Equation, update closure, or admissible model class may avoid this specific proof, subject to separate evaluation.^[2]

References¶

[1] David Lewis, “Probabilities of Conditionals and Conditional Probabilities,” The Philosophical Review 85, no. 3 (1976): 297–315, especially “First Triviality Result,” equations (8)–(12), its fair-die example, and “Second Triviality Result.” Original article scan; journal DOI. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q

[2] David Lewis, “Probabilities of Conditionals and Conditional Probabilities II,” The Philosophical Review 95, no. 4 (1986): 581–589, especially pp. 581–583 reviewing and refining the first and second results. Original article scan. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r

[3] Ernest W. Adams, “The Logic of Conditionals,” Inquiry 8 (1965): 166–197, on conditional assertability. Original article scan. registry ↩a ↩b ↩c