Similarity / Resemblance¶

Version: v3 · 2026-09-28 · History
Prime #: 1597
Domain group: Humanities
Origin domain: Philosophy
Subdomain: Metaphysics → Philosophy
Also from: Psychology & Behavioral Sciences, Mathematics

Core Idea¶

Similarity or resemblance is the graded relation by which distinguishable entities are alike in selected respects. It supports statements such as an orange being more similar to an apple than to the moon when fruithood and shape count, while allowing orange and moon to resemble one another with respect to roundness alone.

The relation is commonly reflexive and symmetric, but it need not be transitive: an intermediate case can resemble two endpoints that do not closely resemble each other. Respective similarity concerns one declared feature; overall similarity aggregates multiple respects and therefore requires a rule for relevance and weighting.

Similarity is a strict kind of Comparison because every likeness judgment places relata in a shared frame, while many comparisons—earlier/later, larger/smaller, cause/effect—do not assess likeness. Similarity is also weaker than identity and less restrictive than equivalence. It enables categorization and transfer without proving common origin, equal value, or interchangeable behavior.

Three broad formalizations illuminate different parts of the prime. A shared-property account counts or weights features jointly possessed. A metric account embeds cases in a space and makes likeness decrease with distance. A transformation account asks how much structure-preserving change is required to turn one case into another. None is universally privileged: the transformation suitable for musical phrases differs from a genomic substitution model, and a property inventory suitable for fruit differs from a visual embedding. They instantiate the same role pattern only because each declares what correspondence means and how degree is read.

The distinction between respective and overall similarity prevents a recurring error. Respective resemblance can be objective relative to a fixed respect—two rods have equal length within tolerance—even when an all-things-considered judgment is neither specified nor useful. Overall resemblance combines potentially incomparable respects and therefore inherits a salience policy. A well-formed claim may remain plural rather than forcing all respects into one scalar.

How would you explain it like I'm…

Alike in Some Ways

An orange and an apple are alike because they're both fruit and both round. An orange and the moon are alike too — they're both round! Things can be alike in some ways and not in others, and some things are more alike than others.

How Alike, and In What Way?

Similarity / Resemblance is how two different things can be alike in some ways. An orange is more like an apple than like the moon if you care about being fruit, but it's like the moon if you only care about being round. You can compare one feature at a time, like just shape, or add up many features to say how alike things are overall. Being alike doesn't mean being the same, and it doesn't mean they came from the same place. Also, A can be like B, and B can be like C, while A and C aren't very alike at all.

Graded Likeness in Respects

Similarity / Resemblance is the graded relation by which distinct things are alike in certain respects. Usually a thing is similar to itself (reflexive), and if A resembles B then B resembles A (symmetric), but it isn't transitive: an in-between case can resemble two extremes that don't resemble each other, like colors fading gradually across a spectrum. There's a difference between similarity in one declared respect (these two rods are the same length) and overall similarity, which combines many respects and needs a rule for what matters most. Similarity is a special kind of comparison — comparing 'earlier vs later' isn't about likeness — and it's weaker than identity. It lets us group things and transfer knowledge from one case to another, but it doesn't prove shared origin or that things will behave the same way.

Similarity / Resemblance is the graded relation by which distinguishable entities are alike in selected respects. It is typically reflexive and symmetric but not transitive: an intermediate can resemble two endpoints that do not resemble each other closely. Respective similarity concerns one declared feature and can be objective relative to it (two rods equal in length within tolerance); overall similarity aggregates several respects and therefore depends on a relevance and weighting policy, sometimes over respects that are not commensurable. Similarity is a strict subtype of Comparison — every likeness judgment places relata in a shared frame, while comparisons like earlier/later or larger/smaller don't assess likeness — and it is weaker than identity and less restrictive than equivalence. Three formalizations capture different aspects: shared-property accounts (count or weight common features), metric accounts (likeness decreases with distance in an embedding space), and transformation accounts (how much structure-preserving change turns one case into another). None is universally privileged; each fits different domains because each declares what correspondence means and how degree is read. Similarity licenses categorization and transfer without establishing common origin, equal value, or interchangeability, and a well-formed claim may stay plural across respects instead of collapsing to one number.

Structural Signature¶

Recurring features:

Comparison pair. Two or more distinguishable entities or representations are placed in one frame.
Respect space. Properties, dimensions, relations, transformations, or behaviors define what can match.
Correspondence. Some structure is shared, aligned, or near under the respect space.
Weighting or metric. A qualitative ordering, distance, kernel, or salience rule determines how respects contribute.
Graded result. The comparison returns a degree or ordering of likeness rather than only same/different.
Difference residue. Unmatched structure remains visible so resemblance does not collapse into identity.
Context. Purpose determines which respects are relevant and whether a similarity score licenses any inference.
Collapse condition. Without explicit respects or correspondence, the claim is only association or verbal resemblance.

The roles are jointly necessary for a computable judgment. The pair identifies what can be substituted into the relation; the respect space determines admissible evidence; correspondence supplies the positive matches; the metric or weighting turns matches and mismatches into degree; the residue prevents degree from masquerading as identity; context controls whether the degree matters. Removing one role produces a recognizable neighbor. Remove the metric and one has an unordered list of commonalities. Remove the respect space and one has an impression. Remove the residue and one risks equivalence or identity. Remove context and one cannot tell whether the same ranking is relevant to a medical diagnosis, an aesthetic comparison, or a database search.

Nontransitivity is not a defect to be repaired automatically. Suppose paint chip A is close to B and B close to C under a threshold, while A and C fall outside it. A thresholded similarity graph can therefore contain local chains without making all chain members similar. Forcing transitivity would convert the relation into an equivalence closure and erase meaningful gradation.

What It Is Not¶

Not identity. Numerically distinct entities can resemble one another, even exactly in selected qualitative respects.
Not equivalence. Similarity is often graded and nontransitive; equivalence is normally binary and transitive.
Not correlation. Co-varying measurements need not share represented features or structure.
Not analogy. Analogy maps relational structure to support inference; similarity is the broader likeness relation.
Not proximity alone. Nearness becomes similarity only within a meaningful representation and metric.

Broad Use¶

Perception. Stimuli are grouped by shared shape, color, motion, or learned features.
Biology. Organisms or sequences are compared under morphological, genetic, or functional respects.
Machine learning. Embeddings, kernels, and distances rank records relative to queries or prototypes.
Information retrieval. Documents are ordered by lexical, semantic, or structural resemblance.
Law. Precedents are compared on facts and legal features material to the issue.
Philosophy. Resemblance figures in universals, counterfactual worlds, depiction, and arguments from analogy.

In chemistry, molecular fingerprints rank compounds by shared substructures; a high score can support candidate search without proving equal biological activity. In medicine, patient similarity systems compare phenotypes, histories, or trajectories, but the feature space must not let data availability stand in for clinical relevance. In music, melodic similarity can weight interval pattern, rhythm, contour, and transformation while leaving performance timbre outside the chosen respect. In ecology, community resemblance depends on taxonomic resolution and abundance weighting. In software, clone detection compares tokens, syntax trees, data flow, or behavior, each producing a different and defensible notion of “same-looking code.”

These uses are literal because the comparison roles survive, not because each domain uses a number called a similarity score. Conversely, a score called “similarity” can fail the prime when it is actually a probability of common origin, a causal effect estimate, or an institutional compatibility rating with no likeness interpretation.

Clarity¶

A similarity assertion should answer: which entities, similar in what respect, under which representation, with what weighting, and for what purpose? “These cases are similar” is underdetermined until the comparison frame is supplied. A numerical score can be precise yet conceptually poor when the feature representation omits the differences relevant to the decision.

Respective and overall similarity must also remain distinct. Two objects can closely resemble each other in color while being very dissimilar overall, or be judged similar overall because a task gives one feature exceptional weight. The score records a model of relevance, not a context-free essence.

Several reporting disciplines improve clarity. State whether the result is ordinal or cardinal; identify the zero and maximum conventions; report whether larger numbers mean more likeness or more distance; distinguish learned features from observed features; and name any threshold that turns a graded score into a category. Where a metric is asymmetric—because one case is a prototype, query, or source—call the relation a directed similarity or explain the asymmetry instead of assuming symmetry from the English word.

Similarity can also be multidimensional without one “overall” answer. A policy pair may be fiscally similar but institutionally dissimilar. Two diseases may share symptoms but not mechanism. Preserving a vector of respect-specific results can be clearer than hiding contestable tradeoffs in one average.

Manages Complexity¶

Similarity compresses many correspondences and differences into an ordering or scalar. This makes retrieval, clustering, classification, precedent search, and analogical exploration tractable. The compression is intentionally lossy: it suppresses some features and elevates others.

Exposing the respect space, weights, and difference residue prevents that convenience from becoming false identity. It also explains reversals: changing the task or representation can change which pair is nearest without any underlying object changing.

The prime manages scale in three ways. First, it reduces a potentially unbounded description of two things to selected comparable features. Second, it permits indexing: candidate pairs too dissimilar under a cheap screen can be excluded before expensive structural comparison. Third, it supports neighborhood reasoning, allowing a new case to borrow hypotheses from nearby cases while keeping verification separate.

Each gain creates a failure mode. Dimensionality can make all points appear similarly distant; sparse representation can overvalue absence; correlated features can be double-counted; learned embeddings can encode historical bias; thresholds can turn tiny numerical changes into categorical discontinuities; and aggregation can hide a decisive mismatch. These failures are not reasons to abandon similarity, but reasons to treat the representation and decision rule as part of the claim.

Abstract Reasoning¶

Identify the relata and keep numerical identity separate from qualitative likeness.
Declare the respect or feature space in which correspondence is possible.
Identify matched structure and the differences excluded or down-weighted.
Choose a qualitative ordering, metric, transformation criterion, or weighting suited to the task.
Check reflexivity, symmetry, and nontransitivity assumptions rather than importing them automatically.
Test robustness under plausible changes of representation and weight.
Ask whether the proposed inference follows from the shared structure or only from surface appearance.
Run a neighbor test. If the rule is transitive and binary, equivalence may be the correct abstraction; if one case is being used to reason about another through mapped relations, analogy may be doing the work; if cases are merely ordered along one dimension, comparison may suffice.
Run a perturbation test. Vary plausible feature weights, normalization, and missing-data treatment. A judgment that flips under every small perturbation is not robust enough for a high-stakes conclusion.
Run a counterexample search. Seek pairs that score highly but fail the intended purpose, and pairs that score poorly but experts regard as substantively alike. These cases reveal absent features or inappropriate aggregation.
Calibrate use. A retrieval task may tolerate approximate neighbors; a safety decision may require an explicit hard exclusion even when overall similarity is high.
Limit the inference. Report what the relation supports—search, grouping, hypothesis transfer, or explanation—and what it does not establish.

One useful decomposition separates representation error from comparison error. Representation error occurs when the feature space omits or distorts relevant structure. Comparison error occurs when the metric or weights misuse an adequate representation. Improving one cannot compensate blindly for the other: a sophisticated metric over the wrong features remains systematically misleading.

Knowledge Transfer¶

Similarity transfers literally whenever the same roles can be filled: relata, respect space, correspondence, weighting rule, graded result, and remaining differences. A visual match, a cosine score, a biological trait comparison, and a precedent analysis can therefore instantiate the same prime without sharing material substrate.

What does not transfer automatically is the metric or inference. Cosine proximity in a language embedding does not establish legal relevance; morphological likeness does not establish homology; similar outcomes do not establish similar causes. Each receiving domain must justify its representation and explain why the matched structure matters.

Geometry to information retrieval. Euclidean distance compares coordinates with a fixed scale; document retrieval often uses cosine similarity so document length does not dominate direction in a term or embedding space. The literal transfer is the pair–space–metric–ranking structure. The coordinate meaning and invariances must be rebuilt.

Biology to engineering. Sequence similarity can suggest homologous ancestry or functional hypotheses, while component similarity in engineering may concern interchangeable dimensions or behavior. A percentage match has different evidential force in the two domains. What transfers is graded correspondence; the inference from correspondence does not.

Perception to law. Visual perception can rely on learned feature salience, and precedent analysis likewise emphasizes some facts as material. Yet legal salience is norm-governed and reason-giving, not merely statistically learned. A similarity engine can retrieve cases, but legal judgment must explain why its respects are doctrinally relevant.

Possible worlds to scenario planning. Philosophical counterfactual accounts order worlds by resemblance to actuality, while planners compare scenarios to a baseline. Both fill the roles, but neither receives a neutral overall metric for free. Background facts held fixed and dimensions allowed to vary must be declared.

Negative transfer. Two interventions can produce similar outcomes through opposite mechanisms; two datasets can have similar summary statistics but different distributions; two faces can look similar without familial relation. In each case, resemblance may be real while the proposed causal or genealogical inference fails. Knowledge transfer should therefore carry a “what matched” statement and an independent validation step.

Examples¶

Formal/abstract¶

Let a respect set contain roundness and fruithood. Orange and apple match on both; orange and moon match only on roundness. With equal weights the first pair ranks as more similar. If the task considers only shape, the distinction can disappear. Mapped back: pair → the compared objects; respect space → the two properties; correspondence → shared properties; weighting → equal or task-specific weights; result → an ordinal ranking; difference residue → unshared qualities.

A metric-space example makes the formal boundary sharper. Three points can satisfy d(A,B)<ε and d(B,C)<ε while d(A,C)≥ε. The thresholded “is similar to” relation is reflexive and symmetric but nontransitive even though the underlying distance obeys the triangle inequality. Taking its transitive closure would create a clustering equivalence, a different abstraction. Mapped back: pair → point pairs; respect space → metric coordinates; correspondence → small separation; weighting → coordinate metric; result → thresholded degree; residue → A–C separation.

Applied/industry¶

A case-retrieval system embeds a new legal dispute and candidate precedents using facts material to the legal issue. It ranks cases by the declared feature model, while counsel inspects decisive differences before transfer. Mapped back: pair → current and prior cases; respect space → legally relevant facts; correspondence → matched facts and relations; weighting → issue-specific model; result → retrieval ranking; difference residue → facts requiring doctrinal distinction.

A molecular-search platform encodes compounds with fingerprints and returns structures sharing many subfeatures with a query. Chemists use the ranking to select candidates for assay, not to declare identical activity. A high-scoring pair that differs at a functionally decisive group becomes a hard negative used to improve the representation. Mapped back: pair → query and candidate molecules; respect space → fingerprint substructures; correspondence → shared fragments; weighting → fingerprint coefficient; result → ranked candidates; residue → unmatched functional groups.

Structural Tensions¶

Respective vs. overall likeness. Narrow comparison is reproducible, while overall judgment requires controversial aggregation. Diagnostic: Which respects may legitimately influence this decision?

Stable metric vs. task sensitivity. Fixed weights support comparability, while useful relevance changes with purpose. Diagnostic: Would another legitimate task reverse the ranking?

Transfer aid vs. false inference. Resemblance supports discovery, while superficial likeness can invite unwarranted causal or normative conclusions. Diagnostic: Which conclusion is licensed by the matched structure?

Symmetry vs. directional judgment. Formal similarity is often symmetric, while cognitive or model-based judgments can be asymmetric because one item serves as prototype. Diagnostic: Does changing comparison direction alter the result, and why?

Local continuity vs. categorical action. Similarity varies gradually, while systems often impose a threshold for acceptance, matching, or classification. Nearby scores can then lead to different actions. Diagnostic: Is the threshold calibrated to consequences, and how are borderline cases handled?

Human interpretability vs. learned representation. Expert-defined features expose reasons, while learned embeddings may retrieve better patterns without a transparent account of resemblance. Diagnostic: Must the user defend the matched respects or only achieve predictive performance?

Broad recall vs. precise neighborhood. Loose similarity retrieves diverse potentially useful candidates, while tight similarity reduces false neighbors but can miss novel variants. Diagnostic: Is the task discovery, verification, or automatic action?

Shared structure vs. shared origin. Likeness can arise from common ancestry, convergence, constraint, imitation, or chance. Diagnostic: What independent evidence distinguishes the proposed explanation of resemblance?

Structural–Framed Character¶

Similarity / Resemblance sits at the structural end of the structural–framed spectrum. Its identity rests on a graded relation of likeness between distinguishable entities within a declared respect space, with a weighting rule and a residue of difference that keeps resemblance from collapsing into identity.

Its vocabulary of respects, features, weighting, and degree is generic. It is evaluatively neutral: resemblance does not prove common origin, equal value, or interchangeable behavior. Its origin in philosophy and mathematics is formal rather than institutional. Because purpose decides which respects are relevant, its definition carries some context-dependence, but feature, metric, and transformational accounts can be stated without human practice, so it is only partly practice-bound. Applying it recognizes shared structure under declared respects rather than importing a perspective. An orange resembles an apple more than the moon when fruithood and shape count, yet resembles the moon in roundness alone. The same organization appears when shared features are counted for categorization and when a distance or kernel scores likeness between data points. Taken together, the diagnostics place it on the structural side.

Substrate Independence¶

Replace images with genomes, geometric shapes, legal cases, documents, musical phrases, or possible worlds. In each substrate, two or more relata can occupy a common representational space, share some structure, differ elsewhere, and be ordered by a declared rule. Those literal replacements preserve the prime.

The test fails when only an English word, emotional association, causal link, or coincident outcome is shared. It also fails when “similar” secretly means identical, equivalent, or equally valued. The ability to survive both positive remapping and negative boundary tests supports prime status.

Further substrate replacements stress different roles. In acoustics, two tones can be similar in spectrum while differing in source. In organizational design, two reporting structures can be graph-isomorphic while their cultures diverge. In mathematics, two objects can be close under one norm and far under another. In art, a depiction can resemble its subject without sharing material properties. In software, two programs can be syntactically dissimilar yet behaviorally similar on a test domain. Each is literal only because the respect and rule are explicit.

The counterfactual independence test asks whether the abstraction survives a wholesale change of material carrier. Replace fruit by legal cases: the same roles remain. Replace visual features by graph relations: the same roles remain. Replace qualitative judgments by a metric: degree remains. Remove graded correspondence entirely and retain only “both were discussed together”: the abstraction disappears. That combination of positive travel and negative collapse establishes substrate independence.

Relationships to Other Abstractions¶

Current abstraction Similarity / Resemblance Prime

Parents (1) — more general patterns this builds on

Similarity / Resemblance is a kind of Comparison Prime

Similarity is a graded kind of comparison whose result depends on declared respects and weighting.

Hierarchy path (1) — routes to 1 parentless root

Similarity / Resemblance → Comparison → Self Checking

Neighborhood in Abstraction Space¶

Similarity / Resemblance sits among the more crowded primes in the catalog (8^th percentile for distinctiveness): several abstractions describe nearly the same structure, so a description that fits it will tend to fit its neighbors too — transporting it usually means disambiguating within this family rather than landing on it exactly.

Family — Unclustered & Miscellaneous (481 primes)

Nearest neighbors

Similarity Measure — 0.84
Comparison — 0.81
Analogy — 0.77
Distance — 0.77
Metric — 0.74

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

Not to Be Confused With¶

Comparison. Comparison is the broader parent; it need not concern likeness. Tell: Does the result grade shared structure?
Identity. Identity concerns sameness, not degree. Tell: Can distinct relata satisfy the claim?
Equivalence. Equivalence normally partitions a domain transitively. Tell: Can A resemble B and B resemble C without A resembling C?
Analogy. Analogy maps relations to support inference. Tell: Is the claim likeness itself or a transfer argument?
Distance. Distance is a separation measure that may induce similarity. Tell: What transformation and representation make small distance meaningful?
Homology. Homology proposes common origin or structural correspondence under a theory. Tell: Is ancestry established, or only likeness measured?
Compatibility. Compatibility asks whether entities can coexist or interoperate. Tell: Can dissimilar things be compatible, or similar things incompatible?
Typicality. Typicality measures relation to a category center. Tell: Is the comparison between any pair or specifically between member and prototype?
Substitutability. Substitutability concerns functional replacement. Tell: Does the matched structure suffice for the required use?

Solution Archetypes¶

No catalogued solution archetypes reference this prime yet.

Notes¶

DAG placement. Comparison remains the strict parent. Every similarity judgment is a comparison, while comparison can order entities by non-likeness relations. No second parent is asserted merely because similarity participates in classification or analogy.

Source boundary. The frozen philosophy survey supports graded likeness, respective versus overall similarity, common formal properties, multiple metaphysical analyses, and the cross-domain applications summarized here. The encyclopedia's typed signature is a structural synthesis, not a claim that one metaphysical theory has won.

Revision trigger. Reconsider the signature if a proposed domain cannot map respects and weighting literally, or if a catalog neighbor is shown to entail the complete role structure.

Operational cautions. Similarity scores require calibration when used as probabilities or thresholds; most are not probabilities. Missing values require a declared policy because “both unknown” can look like a match. Feature duplication can overweight one latent factor. Learned metrics require distribution-shift checks. Cross-group comparisons require examination of whether measurement quality and feature meaning are stable.

Epistemic limit. Similarity is a disciplined way to organize resemblance, not a universal explanation. It can locate a useful precedent or hypothesis and still be silent about cause, mechanism, value, and future behavior. Those conclusions require their own evidence.

References¶

Frozen Wikipedia discovery revision: https://en.wikipedia.org/wiki/Similarity_(philosophy) (revision 1321679412).
Internet Encyclopedia of Philosophy, “Properties”: https://iep.utm.edu/properties/
Stanford Encyclopedia of Philosophy, “Counterfactuals”: https://plato.stanford.edu/entries/counterfactuals/
Stanford Encyclopedia of Philosophy, “Properties”: https://plato.stanford.edu/entries/properties/
Preserved research index: https://philpapers.org/rec/RODRN
Preserved research index: https://philpapers.org/rec/COWR-5

The frozen revision supplies discovery provenance. The structural synthesis is bounded to the documented graded, respect-relative relation and its acknowledged competing analyses.