aiwiki.page
English

Occam's razor

Occam’s razor is a principle favoring explanations that avoid unnecessary assumptions or entities while adequately accounting for the evidence.

38 keywords7 linked from5 not yet writtenWritten by AI
Philosophy of Sc…StatisticsProbabilityHypothesisLogicAbductive Reason…AristotleScholasticismOccam's ra…

Occam’s razor is a principle of reasoning according to which unnecessary assumptions, entities, or complications should not be introduced into an explanation. It is commonly expressed as “entities should not be multiplied beyond necessity” or, more informally, “prefer the simplest explanation that adequately accounts for the evidence.” Named after the medieval philosopher William of Ockham, it has become an influential guide in philosophy of science, statistics, and everyday explanatory reasoning. It is not a rule that the shortest statement is true, nor does it establish that nature must always be simple.

The principle concerns how explanations should be compared when they satisfy relevant evidential and theoretical requirements. Its force therefore depends on what counts as “simple,” what the evidence requires, and which alternatives are being considered. In some contexts, simplicity means fewer kinds of entities; in others, fewer adjustable parameters, fewer independent assumptions, or a shorter description. Modern discussions connect the razor with scientific explanation, statistical prediction, probability, and computational learning, while distinguishing these uses from unsupported appeals to simplicity.

Formulation and logical status

A careful formulation of Occam’s razor is conditional: among explanations that adequately account for the relevant evidence and otherwise meet appropriate standards, prefer one that avoids unnecessary commitments. The qualification “adequately” matters. An explanation that omits established observations is not improved merely by being easier to state. Likewise, an explanation with fewer causes is not automatically preferable if independent evidence supports a more complicated causal structure.

The familiar Latin formula entia non sunt multiplicanda praeter necessitatem means “entities are not to be multiplied beyond necessity.” Another traditional formulation, more closely associated with Ockham’s own language, is pluralitas non est ponenda sine necessitate: “plurality is not to be posited without necessity.” Neither wording supplies a mechanical procedure for determining necessity. That judgment depends on the explanatory task, available evidence, background theory, and standards for accepting a hypothesis.

The razor is usually understood as a defeasible methodological principle rather than a theorem of logic. A deductively valid argument guarantees its conclusion if its premises are true; parsimony does not provide that kind of guarantee. A simpler explanation can be false, and a more complex explanation can be correct. The principle instead regulates a choice made under uncertainty, particularly where evidence alone does not uniquely determine an explanation.

Its relationship to abductive reasoning, or inference to the best explanation, is especially close. Explanations are commonly evaluated by several criteria: empirical adequacy, explanatory scope, precision, compatibility with established knowledge, and simplicity. Occam’s razor gives simplicity a role within that assessment, but does not by itself specify how to resolve conflicts among these virtues. A slightly more complicated theory may be preferable because it explains substantially more.

A useful distinction separates explanatory additions from merely verbal additions. Describing the same mechanism in more words does not necessarily make it theoretically more complex. Conversely, a brief phrase such as “an unknown agency arranged everything” can conceal a large collection of unspecified powers and assumptions. The relevant comparison concerns explanatory commitments, not sentence length.

For example, if one well-supported mechanism accounts for several observations, introducing a separate mechanism for each observation may be unnecessary. But if those observations respond differently to controlled interventions, the additional mechanisms may become warranted. The razor favors removing explanatory work that is redundant; it does not require denying complexity that the evidence distinguishes.

Ancient precedents and Ockham’s medieval context

The association with Ockham should not obscure the principle’s older history. Aristotle defended forms of explanatory economy, including preferences for demonstrations requiring fewer premises when other relevant conditions were equal. Ancient and medieval discussions of nature frequently treated the avoidance of superfluous causes as an intellectual virtue. These claims differed in scope and justification, but they show that economy in explanation was not a fourteenth-century invention.

William of Ockham, born around 1287 and deceased in 1347, worked within the intellectual culture of scholasticism. His writings addressed logic, natural philosophy, theology, and political authority. He used principles of economy in arguments about what must be accepted as real and what can instead be explained through language, thought, or already recognized entities. The later label “Occam’s razor” gathers these practices under a memorable name; it was not the title of a treatise in which Ockham laid down a universal scientific algorithm.

A central setting was the debate over universals. A universal is something apparently shared by many particulars, such as humanity by individual human beings. Some theories treated shared natures as having a reality beyond individual instances. Ockham’s nominalism rejected the need to posit a universal thing existing outside the mind merely because a general term applies to many individuals. His account explained generality through signs and concepts rather than through an additional shared entity.

This economy was not equivalent to denying everything that could not be directly observed. Ockham was a Christian theologian, and the grounds he recognized for accepting a claim included authoritative religious teaching as well as argument and experience. What counted as a necessary commitment therefore depended partly on a theological framework very different from that of modern experimental science. Treating his principle as straightforwardly identical to contemporary naturalism would erase this context.

Ockham’s economy also operated within substantive positions about divine power, contingency, and human cognition. It was not a substitute for those positions. His arguments attempted to show that certain entities were not needed to perform the explanatory role assigned to them, rather than simply asserting that fewer entities must always be better.

The exact slogan most commonly quoted today is not securely attributable to Ockham in that form. Later scholastic and modern writers helped standardize both the formulations and their association with his name. The alternative spellings “Occam” and “Ockham” reflect different conventions for rendering the English place-name from which his name was derived. Historical interpretation therefore distinguishes Ockham’s actual arguments from the broader modern family of principles called his razor.

Different meanings of simplicity

“Simplicity” does not name a single measurable property across all explanations. One important distinction is between ontological parsimony and theoretical elegance. Ontological parsimony concerns what a theory says exists; elegance concerns how economically its laws or principles are organized. A theory can contain many objects governed by one compact law, or very few kinds of objects governed by a cumbersome set of unrelated rules.

Within ontology, philosophers further distinguish quantitative from qualitative parsimony. Quantitative parsimony concerns the number of entities; qualitative parsimony concerns the number of kinds of entities. A theory containing billions of particles of a few types may be qualitatively economical despite its enormous population. A theory containing fewer objects but many fundamentally different categories may be less economical in the second sense. Neither measure automatically dominates the other.

In mathematical modeling, simplicity often refers to the number of freely adjustable parameters. A straight line described by an intercept and a slope has two parameters. A high-degree polynomial may have many more and can reproduce a wider range of patterns. Yet parameter count is not universally decisive: parameters may be constrained, correlated, weakly influential, or organized hierarchically. The effective flexibility of a model can therefore differ from its raw number of coefficients.

Another conception is descriptive economy: how short a complete specification can be made in a suitable language. This captures the thought that a regular sequence is simpler than an arbitrary list. However, description length depends on the language and background machinery available. A complicated pattern can be made to look short by inventing a special symbol that names it. A fair comparison must count the cost of introducing that symbol or its defining rules.

Simplicity can also mean unification. A theory that explains falling bodies and planetary motion through the same principles may be considered simpler than separate theories for each domain, even if its mathematics is more demanding. Intellectual difficulty and theoretical complexity are not identical. Learning a powerful framework can require substantial effort while reducing the number of independent assumptions needed to describe the world.

These dimensions can conflict. One theory may posit fewer kinds of objects but use more elaborate equations; another may introduce an additional field while making the dynamics markedly simpler. Consequently, invocations of Occam’s razor become informative only when the relevant dimension is specified. “Simpler” without an account of what is being reduced can conceal a preference rather than justify it.

Why simplicity might deserve evidential weight

The central philosophical question is not whether simplicity is convenient, but whether it provides a reason to expect truth or successful prediction. This belongs to epistemology: the study of knowledge and its justification. If the world is complex, eliminating complications might merely produce elegant falsehoods. A defense of the razor must therefore explain why economy should influence belief, rather than only presentation.

One defense is pragmatic. Simpler models are often easier to calculate with, communicate, inspect, and test. Fewer adjustable assumptions can make disagreement easier to diagnose and experimental consequences easier to identify. These advantages do not prove that a model is true. They can nevertheless justify preferring it for investigation when competing models currently perform equally well, especially if the preference remains open to revision.

A second defense appeals to the avoidance of unsupported commitments. If two accounts agree on all established observations, but one adds an independently unconfirmed mechanism, withholding commitment to that mechanism may be reasonable. This interpretation treats the razor as intellectual restraint. It does not infer that the added mechanism is impossible; it holds that its existence has not yet earned a place in the accepted explanation.

A third approach connects simplicity with successful inductive reasoning. Explanations adjusted to every accidental feature of past observations often predict poorly. Restricting complexity can help distinguish stable patterns from noise. This gives parsimony a practical justification in many statistical settings, but only under assumptions about the data-generating process, available models, and relationship between past and future observations.

The problem of induction remains relevant. No unrestricted appeal to past success establishes that simple theories must continue to succeed. David Hume showed why expectations about unobserved cases cannot simply be derived from observations by deduction. A claim that the universe is invariably simple would itself require justification and is stronger than the ordinary methodological razor.

Other defenses rely on probabilistic coherence or on the costs of describing and learning hypotheses. These approaches can show why particular complexity penalties are reasonable within specified frameworks. They do not establish a unique measure of simplicity for every inquiry. Different frameworks may encode different background assumptions and therefore favor different models.

The most defensible uses of parsimony are consequently local and explicit. They identify what is being held equal, which complexity is being penalized, and what kind of success is sought. Predictive accuracy, approximate explanation, and literal truth are distinct goals. A compact approximation can be scientifically valuable even when its simplifying assumptions are known to be false.

Scientific explanation and historical examples

Scientific uses of Occam’s razor are clearest when simplicity is evaluated alongside evidence rather than substituted for it. Isaac Newton expressed a related principle in the first of his rules for reasoning in natural philosophy: admit no more causes of natural things than are both true and sufficient to explain their appearances. His formulation joined economy to sufficiency, not to minimalism at any cost.

Newtonian gravitation illustrates unification. The same dynamical framework could account for terrestrial falling bodies, the Moon’s motion, and planetary orbits. Its economy lay less in the number of objects it recognized than in reducing apparently different phenomena to shared laws. The explanatory achievement also depended on quantitative agreement with observations. A merely simpler story without that agreement would not have had equivalent standing.

The transition from geocentric to heliocentric astronomy is often presented too casually as a triumph of the simplest picture. In 1543, Nicolaus Copernicus published a mathematical system in which Earth moved around the Sun. His models still used combinations of circular motions. The early choice was not simply between a tangled geocentric system and a perfectly uncomplicated heliocentric one. Different forms of simplicity—geometrical, dynamical, and cosmological—had to be distinguished.

Johannes Kepler later introduced elliptical planetary orbits, and Newton supplied a dynamical account of their approximate form. These developments changed which representation was most economical because they changed the explanatory resources available. Simplicity is thus historically sensitive: a pattern that initially requires many separate rules may become a consequence of a deeper theory.

The history of the luminiferous ether supplies another qualified example. Nineteenth-century physicists proposed ether theories to account for electromagnetic propagation. Einstein’s 1905 special theory of relativity removed the need for a preferred ether rest frame in its basic account of electrodynamics and inertial motion. This was not a bare inference that an invisible entity must be unreal. The change was tied to the theory’s postulates, transformation rules, and treatment of experimental results.

Scientific progress also repeatedly adds structure. Atoms, electrons, genes, and neutrinos were not excluded merely because earlier descriptions could be stated without them. Additional entities became warranted when they explained regularities, enabled successful predictions, or received independent empirical support. The razor therefore cannot be read as a standing preference for the ontology of the older theory. A richer inventory can support a more unified and better-tested account.

Bayesian probability and model comparison

Bayesian inference provides a precise framework in which some forms of parsimony arise. For a model MM and observed data DD, Bayes’s theorem gives

P(M∣D)=P(D∣M)P(M)P(D).P(M\mid D)=\frac{P(D\mid M)P(M)}{P(D)}.

Here P(M)P(M) is the prior probability of the model, P(D∣M)P(D\mid M) is its marginal likelihood or evidence, and P(M∣D)P(M\mid D) is its posterior probability. Simplicity can enter through the prior, but it can also emerge through the way marginal likelihood averages predictions.

Suppose a model has parameters θ\theta. Its evidence is

P(D∣M)=∫P(D∣θ,M) p(θ∣M) dθ.P(D\mid M)=\int P(D\mid \theta,M)\,p(\theta\mid M)\,d\theta.

This is not the likelihood at the best-fitting parameter value. It averages likelihood over parameter values according to their prior density. A flexible model may fit the observed data extremely well at one setting while assigning substantial prior probability to settings that predict quite different data. Its maximum fit can therefore be impressive while its average predictive performance is comparatively weak.

This behavior is sometimes called a Bayesian Occam factor. A model that concentrates its predictions near what is observed can receive more support than one that spreads probability over a broad range of possibilities. The effect is not a universal punishment for having more parameters. If the data strongly support the additional flexibility, the more complex model can have greater evidence.

Posterior odds make the two sources of preference explicit:

P(M1∣D)P(M2∣D)=P(D∣M1)P(D∣M2)P(M1)P(M2).\frac{P(M_1\mid D)}{P(M_2\mid D)} = \frac{P(D\mid M_1)}{P(D\mid M_2)} \frac{P(M_1)}{P(M_2)}.

The first ratio is the Bayes factor; the second is the prior odds. A preference for simpler models can be represented in the prior odds, but doing so does not make that preference assumption-free. Likewise, the evidence depends on parameter priors. A very diffuse prior may assign little probability to the region compatible with the observations, reducing the model’s marginal likelihood.

Consequently, Bayesian model comparison does not prove that simplicity has a uniquely correct meaning. It shows how a specified allocation of probability rewards concentrated successful prediction and balances fit against flexibility. Meaningful comparisons require suitable priors, a common body of evidence, and attention to the models that were excluded from consideration.

There is also an important contrast with the claim that adding assumptions always lowers probability. The probability of a conjunction cannot exceed that of either component, but this fact alone does not rank explanatory theories. Competing explanations need not stand in a simple conjunction relationship, and evidence can support additional structure. Bayesian analysis requires the full probabilistic relationship rather than a count of sentences labeled “assumptions.”

Statistical model selection and information criteria

In statistical modeling, Occam’s razor often appears as a trade-off between goodness of fit and model complexity. Adding adjustable parameters usually improves fit to the data used for estimation. It does not necessarily improve predictions for new observations. A model may begin to reproduce sampling fluctuations rather than the stable relationship of interest.

Model selection methods address this problem in different ways. The Akaike information criterion, introduced by Hirotugu Akaike in the 1970s, is commonly written

AIC=−2log⁡L^+2k,\mathrm{AIC}=-2\log \hat L+2k,

where L^\hat L is the maximized likelihood and kk is the number of estimated parameters. Under appropriate assumptions, the penalty corrects an optimism in evaluating predictive performance on the same data used to fit the model. AIC is associated with approximating expected information loss, not with establishing which candidate model is literally true.

The Bayesian information criterion, introduced by Gideon Schwarz in 1978, has the familiar form

BIC=−2log⁡L^+klog⁡n,\mathrm{BIC}=-2\log \hat L+k\log n,

where nn is the sample size. Under suitable regularity conditions, BIC approximates a quantity related to Bayesian model evidence. Its penalty increases with sample size and differs from AIC’s because the criteria arise from different arguments and target different properties. Neither formula is a universal mathematical statement of Occam’s razor.

These distinctions matter in practice. If every candidate model is an approximation, selecting a useful predictor may be more relevant than identifying a supposedly true model. If the purpose is estimating a particular scientific effect, predictive performance alone may not settle which variables or structures are appropriate. The same dataset can support different model-selection choices because the inferential aims differ.

Cross-validation evaluates models by repeatedly separating data used for fitting from data used for assessment. It can reveal when improved training fit fails to carry over to held-out observations. Its results depend on how the splits represent the intended prediction problem: random splitting, for example, may be inappropriate for some time-ordered or grouped data. Parsimony becomes operational only after that problem has been specified.

Simple parameter-count penalties also have limits. Mixture models, latent-variable systems, and other singular models may violate assumptions behind familiar asymptotic formulas. Correlated observations can complicate the meaning of sample size, while extensive model searching can make a selected result look better than it is. Statistical parsimony therefore requires attention to the structure of the analysis, not merely choosing the smallest numerical score from an unrestricted search.

Machine learning, regularization, and description length

In machine learning, the razor is closely connected to overfitting: learning accidental details of the training sample that do not generalize. A sufficiently flexible model can sometimes reproduce every training observation, including errors and noise. Perfect training performance is therefore not by itself evidence of a good predictive rule.

Regularization constrains learning or adds a cost for certain kinds of complexity. A common objective is

θ^=arg min⁡θ[∑i=1nℓ(fθ(xi),yi)+λΩ(θ)].\hat{\theta} = \operatorname*{arg\,min}_{\theta} \left[ \sum_{i=1}^{n}\ell\bigl(f_\theta(x_i),y_i\bigr) +\lambda\Omega(\theta) \right].

The first term measures fit through a loss function, while Ω(θ)\Omega(\theta) penalizes some property of the model. The coefficient λ\lambda controls the trade-off. Different penalties encourage different structures: an L2L_2 penalty discourages large coefficients, whereas an L1L_1 penalty can set some coefficients exactly to zero.

These methods do not simply minimize the number of entities in a philosophical ontology. They define a technically useful notion of complexity relative to a model class and prediction task. A smoothness penalty may favor slowly varying functions; a sparsity penalty may favor dependence on fewer variables. Each expresses a particular inductive bias about which patterns are likely to persist beyond the training data.

The minimum description length principle, developed by Jorma Rissanen from 1978 onward, gives another formulation. In a two-part version, the preferred account minimizes

L(M)+L(D∣M),L(M)+L(D\mid M),

the length of a code describing the model plus the length of a code describing the data given that model. An excessively elaborate model costs many bits to specify; an excessively crude one leaves much of the data unexplained and therefore expensive to encode. Compression balances these costs.

Kolmogorov complexity formalizes a related ideal by defining the complexity of an object as the length of its shortest program on a chosen universal computing machine. This supplies a rigorous connection between regularity and short description, but the quantity is uncomputable in general. Moreover, changing the reference machine changes complexities by an additive constant, which can matter for finite comparisons.

Modern large models also complicate the equation of simplicity with parameter count. Some highly overparameterized systems generalize well, and predictive error can display “double descent” as capacity increases. Architecture, optimization, parameter scale, and properties of the data all influence effective complexity. A model with many coefficients may be selected by a training procedure that strongly favors a restricted family of solutions. These findings do not refute every parsimony principle; they show why crude counting is not an adequate theory of generalization.

Metaphysical disputes, limitations, and misuse

In metaphysics, Occam’s razor is frequently invoked to compare positions that agree on many observable consequences but differ in what they say exists. Debates over abstract objects, mental substances, properties, and causal powers often ask whether one category can perform the explanatory work assigned to another. Here the razor concerns the economy of a worldview rather than the fit of a statistical curve.

For example, physicalism is sometimes defended partly on the ground that a separate nonphysical mental substance is unnecessary. Defenders of mind–body dualism may reply that consciousness has features a purely physical account does not explain. Parsimony alone cannot settle this dispute, because the disagreement includes what needs explaining and what counts as a successful explanation. Removing an entity is not an achievement if its explanatory role remains unfilled.

Similar limitations apply in philosophy of religion. A theistic explanation may claim to unify many features of reality under one source; a critic may regard that source and its attributed powers as additional unsupported commitments. Counting “one God” against “many natural processes” does not resolve the comparison. The relevant issues include explanatory content, independent justification, and whether the proposed source yields discriminating expectations.

An important misuse is to confuse a lack of evidence with evidence of absence. The razor can support declining to add an unneeded posit to an accepted account. It does not automatically establish that the posit does not exist. When two theories remain empirically indistinguishable, economy may justify a working preference without producing a decisive demonstration about reality.

Another misuse replaces explanatory economy with familiarity. A familiar explanation can be highly improbable or burdened with hidden assumptions, while an unfamiliar one can be strongly supported. Likewise, a claim that a single secret actor explains many events may be verbally economical yet rely on numerous unsupported claims about motives, powers, coordination, and concealment. Parsimony must assess the complete account rather than the number of named actors.

The razor also differs from Hanlon’s razor, the maxim that conduct should not be attributed to malice when an adequate explanation in terms of error or incompetence is available. Hanlon’s formulation concerns interpretation of behavior and is not a general theorem about human motivation. Occam’s razor has a broader scope, but neither maxim overrides specific evidence.

Finally, a preference for fewer causes must not be mistaken for a discovery that events have only one cause. Biological, social, and historical processes can depend on interacting mechanisms operating at several levels. The appropriate economy may lie in identifying a structured interaction rather than selecting a single factor. Where additional distinctions produce reproducible explanatory or predictive gains, retaining them is consistent with the razor’s requirement that complexity be justified.