The Logical Foundations of Scientific Inquiry
Scientific reasoning is commonly thought of as logical, but many practising scientists could not tell you the logic that actually underpins their reasoning or decisions. Deduction, induction, and abduction are three distinct inferential processes, only one of which is deductively valid. A working scientist moves between all three without necessarily being able to name which is which, or even state their basic definitions. This article sets out the basic formalism behind each of the three, and the hypothetico-deductive method that stitches them together into what has come to be known as the scientific method.
A method older than its name
Long before "the scientific method" was named as such, Socrates had already put an elementary form of it into practice, in what has since come to be called the Socratic method. In the fifth century BC, rather than asserting claims outright, he questioned people who believed they understood concepts such as justice, courage, or piety, and pushed them, through successive questions, toward consequences of their own definitions that they could not accept. This isn't the scientific method in any full sense, since it built up no cumulative body of truths the way science does, but it established the idea later scientific enquiry inherits at its core: that a claim is strengthened only by surviving deliberate attempts to break it. The same instinct, preserved in Plato's dialogues, resurfaces almost unchanged in Francis Bacon's push for systematic experiment in the seventeenth century, and again in Popper's twentieth-century insistence that a theory must be falsifiable to count as scientific at all.
Deduction
Deductive inference is the inference form standardly written with the turnstile \( \vdash \): a set of premises \( \Gamma \) deductively entails a conclusion \( \varphi \), written \( \Gamma \vdash \varphi \). This is a formal definition, indifferent to subject matter: it depends solely on the logical form of the premises and conclusion, via the connectives (\( \neg \) for negation, \( \land \) for conjunction, \( \lor \) for disjunction, \( \rightarrow \) for the conditional, \( \leftrightarrow \) for the biconditional) and, once you move to predicate logic, the quantifiers \( \forall \) ("for all") and \( \exists \) ("there exists").
The two inference rules most often cited in scientific argument are modus ponens and modus tollens. Modus ponens states:
\[ P \rightarrow Q, \quad P \ \vdash\ Q. \]
If a theory (together with initial conditions) entails an observation, and the theory-plus-conditions in fact obtain, the observation follows. Modus tollens is the same conditional run in the other direction, and it is the inference actually being used whenever an experiment is described as testing a theory:
\[ P \rightarrow Q, \quad \neg Q \ \vdash\ \neg P. \]
If a theory entails an observation that fails to occur, the theory is false, not "weakened" or "disconfirmed" in some soft sense, but strictly, deductively false, provided the conditional premise \( P \rightarrow Q \) was correct in the first place.
The classical syllogism is the same apparatus one level further into predicate logic. "All ravens are black; this is a raven; therefore this is black" is, formally,
\[ \forall x\,(Rx \rightarrow Bx), \quad Ra \ \vdash\ Ba, \]
and its validity again depends only on form: substitute any predicates for \( R \) and \( B \) and the argument remains valid, even if the premises happen to be false of reality. Validity is not truth: a deductively valid argument with a false premise deductively entails nothing true about reality, which is why deduction alone cannot get scientific inquiry off the ground. Deduction only ever unpacks what is already contained, implicitly, in the premises you started with. This raises the question of where the premises come from.
Validity versus soundness
An argument is valid when its conclusion cannot be false while its premises are true, a matter of form alone. It is sound when it is valid and its premises are true. Modus ponens, \( P \rightarrow Q, P \vdash Q \), is valid in every case, as the truth table below shows:
\[ \begin{array}{cc|c} P & Q & P \rightarrow Q \\ \hline T & T & T \\ T & F & F \\ F & T & T \\ F & F & T \end{array} \]
The only row in which \( P \rightarrow Q \) and \( P \) are both true is the first, and in that row \( Q \) is also true, so there is no row in the table where the premises of modus ponens hold and the conclusion fails. But validity says nothing about whether \( P \rightarrow Q \) or \( P \) are true of the actual world: an argument can be perfectly valid and completely worthless if built on a false premise. "All swans are white; this is a swan; therefore this is white" is exactly as valid as the raven syllogism, and was standardly used as an example of sound scientific reasoning until 1697, when Willem de Vlamingh's expedition to the Swan River in Western Australia found black ones. The syllogism's form was never the problem; its major premise was simply false, and no amount of formal correctness could have caught that. Logic is necessary but not sufficient for scientific knowledge: it guarantees that true premises yield true conclusions, but securing true premises in the first place falls outside deduction, onto induction, abduction, and observation.
Necessary and sufficient conditions
A great deal of what a scientific explanation is trying to state can be made precise using the vocabulary of necessary and sufficient conditions, which is simply the conditional \( \rightarrow \) read in two directions. To say \( A \) is a sufficient condition for \( B \) is to assert \( A \rightarrow B \): \( A \)'s obtaining guarantees \( B \)'s obtaining, though \( B \) might arise other ways too. To say \( A \) is a necessary condition for \( B \) is to assert \( B \rightarrow A \), equivalently \( \neg A \rightarrow \neg B \): without \( A \), \( B \) cannot occur, though \( A \) alone might not be enough to produce it. A condition that is both, \( A \leftrightarrow B \), is comparatively rare in real scientific explanation, and its scarcity is itself informative. For example: "a spark in the presence of sufficient oxygen and fuel is sufficient for combustion" is a sufficient-condition claim, not a biconditional, since combustion can also be initiated by sufficient heat alone without a spark; "oxygen is necessary for combustion" is a necessary-condition claim, true but useless as a complete explanation on its own, since the vast majority of oxygen-containing environments are not combusting anything.
Confusing these two directions is one of the most common informal fallacies dressed up as scientific argument, and it has a name once translated into this notation: affirming the consequent, inferring \( A \) from \( B \) and \( A \rightarrow B \), is exactly the mistake of treating a merely sufficient condition as though it were necessary. "Fever is a symptom of infection; this patient has a fever; therefore this patient has an infection" commits exactly this error, since infection being sufficient for fever (roughly) says nothing about what else is sufficient for fever, and the valid inference runs only in the necessary-condition direction: no infection without the possibility of fever does not mean no fever without infection. Mill's methods of agreement and difference, an early systematic attempt (1843) to formalise how experimental variation isolates causes, can be understood as procedures for empirically discriminating necessary from sufficient conditions in this sense: the method of difference holds every factor fixed except one and checks whether the effect disappears when that factor is removed, a direct probe of necessity, while the method of agreement looks for a factor common to every case in which the effect occurs, probing sufficiency. Both methods are inductive, not deductive, in the sense already established above.
Reductio ad absurdum
One more deductive pattern is common enough in both mathematics and empirical science to deserve its own notation. Reductio ad absurdum has the form
\[ \Gamma, \neg\varphi \ \vdash\ \bot \quad \Longrightarrow \quad \Gamma \ \vdash\ \varphi, \]
where \( \bot \) denotes an explicit contradiction (some \( \psi \land \neg\psi \)): assume the negation of what you want to prove, derive a contradiction from that assumption together with your other premises \( \Gamma \), and conclude that the negation must be false, hence \( \varphi \) itself. This isn't a separate logic from modus ponens and modus tollens; it's a derived pattern, valid because a contradiction cannot be true in any model, so anything whose denial entails one must itself hold in every model consistent with \( \Gamma \), which is just soundness again, applied to the special case of a self-defeating assumption.
The Michelson–Morley experiment (1887) is a good example of the same structure applied to a physical hypothesis rather than a mathematical one. The luminiferous ether hypothesis, \( \neg\varphi \) in the schema above where \( \varphi \) is "there is no detectable ether wind," entailed, together with auxiliary assumptions about the interferometer's sensitivity and the earth's orbital motion through the supposed ether, a specific observable fringe shift as the apparatus was rotated. No shift of anything like the predicted magnitude was detected, which is a direct empirical \( \bot \): the ether hypothesis, in conjunction with the auxiliary assumptions, entailed an observation that failed to occur (the modus tollens schema from earlier in this piece), but the negative result was then used constructively, in the reductio pattern, as evidence for the contradictory of the ether hypothesis rather than merely against the conjunction \( T \land A \). The Duhem–Quine complication from before applies here in full force: strictly, the null result only forced \( \neg T \lor \neg A \), and it took decades, and Einstein's alternative account of the relevant phenomena via special relativity, before the disjunction was widely read as falling on the \( \neg T \) side. The reductio pattern establishes that a contradiction has been reached somewhere in \( \Gamma \); it does not, by itself, establish which member of \( \Gamma \) should be discarded, which is the same open question the Duhem–Quine thesis raises for the hypothetico-deductive schema generally, arrived at here by a different formal route.
Syntax and semantics
It matters, for what follows, that there are two independent notions of logical consequence in play, and that a theorem is needed to connect them. The syntactic notion, \( \Gamma \vdash \varphi \), is proof-theoretic: it means there exists a finite derivation of \( \varphi \) from \( \Gamma \) using a fixed set of mechanical inference rules (modus ponens among them), checkable line by line without any reference to meaning or truth at all. In principle, a computer could verify a syntactic derivation without "understanding" a word of it. The semantic notion, written with the double turnstile \( \Gamma \models \varphi \), is model-theoretic: it means that in every possible interpretation (every assignment of truth values to the atomic sentences, or every model in the predicate-logic case) in which every member of \( \Gamma \) comes out true, \( \varphi \) comes out true as well.
Soundness is the direction \( \Gamma \vdash \varphi \Rightarrow \Gamma \models \varphi \): anything the proof system can derive really is a semantic consequence, which is what guarantees the inference rules never manufacture a falsehood from truths. Completeness is the converse, \( \Gamma \models \varphi \Rightarrow \Gamma \vdash \varphi \): anything that is a semantic consequence can in fact be derived by the proof system, which guarantees the rules are not too weak to capture everything that follows. For first-order predicate logic, both directions hold (this is Gödel's 1929 completeness theorem), which is the formal fact behind something scientific practice takes for granted: that a derivation carried out on paper, following syntactic rules alone, faithfully tracks what is really entailed about the models (the possible ways the world could be) consistent with the premises. This completeness theorem is entirely different from Gödel's 1931 incompleteness theorems, which concern what a sufficiently strong formal system can prove about its own arithmetic, not whether first-order logic's proof rules match its semantics.
Aristotelian syllogistic
Modern predicate logic is a nineteenth- and twentieth-century construction (Frege, Peano, Russell and Whitehead), but the study of deductively valid form is much older. I've chosen to touch on it here for its historical priority, and because a great deal of scientific argument outside formal contexts still runs, implicitly, on syllogistic form. Aristotle's syllogistic classifies categorical statements into four types: universal affirmative ("All \( S \) are \( P \)"), universal negative ("No \( S \) are \( P \)"), particular affirmative ("Some \( S \) are \( P \)"), and particular negative ("Some \( S \) are not \( P \)"), traditionally labelled A, E, I, O, and asks which combinations of two premises sharing a middle term validly yield a conclusion relating the two remaining terms.
Barbara, the paradigm case and the one every introductory logic course opens with, is AAA in the first figure: "All \( M \) are \( P \); all \( S \) are \( M \); therefore all \( S \) are \( P \)," which is the raven syllogism above with the middle term \( M \) (raven) mediating between subject \( S \) and predicate \( P \). Celarent is EAE in the same figure: "No \( M \) are \( P \); all \( S \) are \( M \); therefore no \( S \) are \( P \)." Of the 256 syntactically possible combinations of premise-pair and figure, only fifteen are valid, and mediaeval logicians assigned each a mnemonic Latin name (Barbara, Celarent, Darii, Ferio, and so on through Baroco and Bocardo) whose vowels encode which of A, E, I, O occupy the premises and conclusion. The syllogism cannot express relations (it has no device for "taller than" or "causes"), multiple quantification (no way to formalise "every scientist has read some paper"), or the sentential connectives compositionally, all of which is why it was superseded rather than extended; but every syllogistic inference translates into the \( \forall x (Rx \rightarrow Bx) \) notation that is standard today.
Induction
Enumerative induction is the inference from a finite number of observed instances to a general claim covering unobserved cases:
\[ Pa_1 \land Pa_2 \land \cdots \land Pa_n \ \not\vdash\ \forall x\, Px. \]
The \( \not\vdash \) is what we are most interested in here. No matter how large \( n \) gets, the inference to the universal claim is not deductively valid, because it is always logically possible for every observed instance to satisfy \( P \) while some unobserved instance fails to. This is Hume's problem in its sharpest form: there is no way to add a premise that legitimately bridges observed-so-far to all-without-exception without either begging the question (assuming nature is uniform, which is exactly what needs justifying) or smuggling in a further inductive step to justify that premise, which only regresses the same problem one level back.
This isn't a practical limitation awaiting a cleverer sampling method; it's a structural fact about the relationship between finite evidence and universally quantified conclusions. Statistical inference softens the conclusion (from "all ravens are black" to "with 95% confidence, the proportion of black ravens exceeds some bound") but doesn't remove the gap, only relocates it: the inference from sample frequency to population parameter is itself an inductive step, licensed by probabilistic assumptions (independence, representative sampling) that are themselves not deductively certified by the data in hand. Every attempt to underwrite induction with a further principle either turns that principle into an empirical claim needing its own inductive support, or turns it into an a priori assumption whose justification lies outside logic altogether. Both routes leave induction exactly as unproven, in the deductive sense, as Hume found it.
The raven paradox
Carl Hempel noticed in 1945 that even granting some notion of inductive confirmation, without yet worrying about whether it can be fully justified, produces a seemingly nonsensical conclusion. Take the universal claim \( \forall x (Rx \rightarrow Bx) \), "all ravens are black." By ordinary logic this is equivalent to its contrapositive, \( \forall x (\neg Bx \rightarrow \neg Rx) \), "all non-black things are non-ravens." The Nicod criterion says that an instance of the antecedent-and-consequent both holding, here a black raven, confirms a universal conditional, while an instance of the antecedent holding without the consequent would refute it. The equivalence condition says that whatever confirms a statement confirms anything logically equivalent to it, since they say the same thing. Put the two together: a red apple is neither black nor a raven, so it instantiates \( \neg Bx \land \neg Rx \), confirming "all non-black things are non-ravens" by Nicod's criterion, and therefore, by the equivalence condition, confirming "all ravens are black" as well. A red apple, sitting on a table nowhere near a raven, can apparently count as evidence about the colour of ravens.
Hempel's own resolution was to accept the conclusion and explain away the intuitive resistance to it: the apple really does confirm the raven hypothesis, but by an amount so vanishingly small, given how enormously more numerous non-black things are than ravens, that the confirmation is real but practically undetectable, and the intuitive oddness is a failure to notice how weak "some confirmation" is as opposed to "a great deal of confirmation." I. J. Good offered a more detailed diagnosis using the Bayesian tools introduced below: whether an instance confirms a hypothesis at all, not just how much, depends on background assumptions about the relative base rates of ravens and black things, which the syllogistic apparatus alone is silent about and simply does not appear in Hempel's version of the puzzle. Both replies agree on the diagnosis that an adequate theory of confirmation needs more structure than "instances of \( Rx \land Bx \) confirm \( \forall x(Rx \rightarrow Bx) \)" supplies on its own, which is why the probabilistic treatment below was developed.
Probabilistic and Bayesian induction
The most developed formal response to both the raven paradox and the general problem of induction reframes confirmation as a matter of degree, using probability calculus rather than dogmatic categories of syntactic entailment. Bayes' theorem, as a piece of probability theory, states
\[ P(H \mid E) = \frac{P(E \mid H)\, P(H)}{P(E)}, \]
relating the posterior probability of a hypothesis \( H \) given evidence \( E \) to the prior probability \( P(H) \) (credence in \( H \) before \( E \) is taken into account), the likelihood \( P(E \mid H) \) (how probable the evidence would be if \( H \) were true), and \( P(E) \), the unconditional probability of the evidence, which normalises the result. The Bayesian reading of confirmation says \( E \) confirms \( H \) when \( P(H \mid E) > P(H) \), that is, when learning \( E \) raises credence in \( H \); equivalently, using the odds form of the theorem, when the likelihood ratio \( P(E \mid H) / P(E \mid \neg H) \) exceeds 1.
This reframes the raven paradox by making the confirmation Hempel's principles also implied quantitative: a black raven confirms "all ravens are black" strongly because \( P(\text{black raven} \mid H) \) and \( P(\text{black raven} \mid \neg H) \) differ substantially (the observation is much more probable if the generalisation is true), while a non-black non-raven confirms the same generalisation, via the equivalent contrapositive, by an amount governed by how many non-black things there are relative to ravens: a non-zero but minuscule likelihood ratio, matching the intuition that the apple is technically relevant but practically negligible, now derived rather than merely stipulated. It also gives a formal shape to the response to Hume: Bayesian updating never claims to deductively certify \( H \); it's a rule for how credences ought to shift given new evidence, conditional on a prior that itself has to come from somewhere logic cannot reach. Whether the priors themselves can be given a non-arbitrary justification (the frequentist objection to Bayesianism generally) is exactly the sort of question that pushes past pure logic and into the epistemology of science, which the companion piece to this one addresses.
Abduction
There is a third pattern, distinct from both, usually associated with C. S. Peirce, sometimes written schematically as
\[ Q, \quad P \rightarrow Q \ \Longrightarrow\ P \text{ (as a hypothesis worth entertaining)}. \]
Notice the double arrow rather than \( \vdash \): this is not a valid inference in either the deductive or the classical inductive sense. It is affirming the consequent, a deductive fallacy, if read as a claim about what follows necessarily. Read instead as a claim about what is worth investigating, it is the pattern behind essentially all explanatory reasoning in science: an observation \( Q \) is noted, a hypothesis \( P \) is proposed because \( P \rightarrow Q \) and \( P \) would make \( Q \) unsurprising, and \( P \) is provisionally adopted because it is judged the best available explanation among the live alternatives. Abduction generates candidate theories without attempting to prove them.
Ignaz Semmelweis's 1847 investigation of childbed fever at the Vienna General Hospital is a classic historical case of abduction's merit. The observation \( Q \) was a mortality gap: women delivering in the doctors' maternity ward died of puerperal fever at several times the rate of those in the midwives' ward, despite the doctors' ward having, on paper, the more qualified staff. Semmelweis considered and eliminated several candidate explanations for \( Q \) by deducing and checking predictions each would entail (overcrowding, the men's presence during delivery, differences in position during birth) before a colleague's death, from a scalpel wound sustained during an autopsy that produced symptoms matching childbed fever, gave him the abductive leap: doctors performing autopsies before attending deliveries were transferring "cadaverous particles" to women in labour. This hypothesis \( P \) was preferred because it best explained the specific pattern, including the otherwise-puzzling fact that the doctors' ward, uniquely among the two, involved routine autopsy work. Semmelweis then carried out the next step in the hypothetico-deductive process: he deduced a testable prediction (mandatory chlorinated hand-washing between autopsy and delivery should collapse the mortality gap if the hypothesis were correct) and checked it. Ward mortality fell from around 18% to near 1% within a year. The abduction proposed the explanation; the deduction extracted a prediction; the observation checked it.
The hypothetico-deductive method: putting the three together
The method actually used in practice combines all three patterns. Abduction proposes a theory \( T \). Deduction extracts a testable prediction: given \( T \) together with auxiliary assumptions \( A \) (about instruments, background conditions, and everything else the derivation quietly relies on),
\[ (T \land A) \rightarrow O. \]
The prediction \( O \) is then checked. If \( O \) is observed, modus ponens is satisfied but nothing is proven: a true consequent does not deductively certify the antecedent (affirming the consequent again), so confirmation is evidentially suggestive, not logically conclusive. If \( O \) fails to occur, modus tollens applies with full deductive force:
\[ (T \land A) \rightarrow O, \quad \neg O \ \vdash\ \neg(T \land A) \ \equiv\ \neg T \lor \neg A. \]
And there is the catch that Pierre Duhem identified and Willard Quine generalised into the Duhem–Quine thesis: a failed prediction only tells you that something in the conjunction \( T \land A \) is false, and the logic alone does not tell you whether to blame the theory or one of the auxiliary assumptions. Every real test smuggles in instrument calibration, background theory, and ceteris paribus conditions as unstated conjuncts, and a determined defender of \( T \) can always retreat to rejecting some auxiliary \( A \) instead. What actually adjudicates the choice (theoretical conservatism, independent testability of the auxiliary assumptions, track record, simplicity) is epistemological, not logical, which is the territory the companion piece to this one is about.
Falsifiability and the demarcation problem
Karl Popper's proposed solution to the problem of induction was to route around it, such that a theory earns scientific status by being falsifiable: by entailing, via the modus tollens schema above, observations that could in principle turn out false. A theory compatible with every possible observation, entailing no \( O \) whose negation would refute it, is unfalsifiable and, on Popper's criterion, not a scientific theory, regardless of how much confirming evidence its defenders can cite.
The complication, already implied in the Duhem–Quine point, is that falsification is never quite as clean in practice as \( \neg O \vdash \neg T \) makes it look, because the theory under test is never alone in the conjunction being refuted. Popper was aware of this and treated it as a matter of scientific convention and honesty: a live research community that keeps rescuing \( T \) by endlessly adjusting \( A \) rather than risking \( T \) itself is engaged in what he called immunising a theory against refutation, a methodological vice.
Discussion