51.06.01 · linguistics / historical-comparative

Historical and comparative linguistics — change, families, and reconstruction

shipped3 tiersLean: none

Anchor (Master): Campbell (2013); Hock & Joseph (2009) Language History, Language Change, and Language Relationship (Mouton de Gruyter); Anttila (1989) Historical and Comparative Linguistics (Benjamins)

Intuition Beginner

Languages are alive. Every generation speaks a little differently from the last, and over centuries those small shifts pile up until daughter tongues no longer sound like their parent. English and German were one language fifteen hundred years ago; Latin gave rise to French, Spanish, and Italian; Sanskrit, Greek, and Latin share an ancestor older still. Historical linguistics studies this descent: how languages change, how they split, and how we recover a past no one recorded.

Comparison is the method. When two languages share an ancestor, their words keep matching shapes in a pattern far too regular to be chance. The English word "father", Latin "pater", and Sanskrit "pitár-" line up sound by sound. Such systematic matches — called cognates — are the fingerprints of shared ancestry, and they let us rebuild a parent language no living ear has heard, the way anatomy rebuilds a fossil animal from a few bones.

This unit shows how linguists date, classify, and reverse-engineer language history. We follow sound laws such as Grimm's law, which turned an old "p" into English "f"; reconstruct vanished parent languages from living daughters; map the world's great language families; and weigh two rival pictures of how languages diverge — the branching tree and the spreading wave.

Visual Beginner

A simplified Indo-European family tree: one ancestor (Proto-Indo-European) splits into branches (Germanic, Romance, Celtic, Slavic, Indo-Iranian, Greek, Armenian), with Grimm's law marked on the Germanic branch.

The same ancestor's consonants surface differently in each branch. The table below shows Grimm's law: the first sound shifts one way across a whole branch, not word by word.

Proto-Indo-European	Germanic	Example cognates
*p	f	Latin pater → English father
*t	th	Latin tres → English three
*k	h	Latin canis → English hound
*bʰ	b (d)	Sanskrit bʰrātar → English brother

Worked example Beginner

Trace Grimm's law through the word for "father" across three languages.

Step 1. Start from the shared ancestor sound, Proto-Indo-European *p. In Latin the sound stayed a plain "p": pater. In Sanskrit it also stayed "p": pitár-.

Step 2. In the ancestor of English (Proto-Germanic), Grimm's law shifted every inherited *p to "f". So *pətēr became *faðēr, which later turned into modern "father".

Step 3. Check the pattern holds for the other consonants. The "t" in pater matches the "th" in "three" (Latin tres); the middle "t" of pater became the "th" in "father" before being softened further in modern English.

The lesson: one rule, *p → f, explains dozens of words at once. That breadth — a law acting across the whole vocabulary — is what tells us the match is inheritance, not borrowing and not coincidence.

Check your understanding Beginner

Formal definition Intermediate+

Sound change is a diachronic process in which a phonological category of one stage is realised by a different phone, or merges with another category, in a later stage, applying across the lexicon wherever its phonological environment holds ^{[Campbell Ch. 2]}. A sound change is regular (the Neogrammarians' Ausnahmslosigkeit, "exceptionlessness") when every instance of the input in the defined environment undergoes the same change.

A cognate set is a collection of words, one per related language, descended from a single word of the common ancestor. A correspondence set is the sound-by-sound alignment across a cognate set: the position at which Latin has p, English has f, Sanskrit has p forms the recurring correspondence {p : f : p}.

The comparative method is the procedure that (i) assembles cognate sets, (ii) reads off regular correspondence sets, (iii) assigns each correspondence a proto-phoneme (the ancestor sound), and (iv) writes the resulting proto-forms, the reconstructed lexical items of the ancestor, conventionally prefixed with an asterisk (*pH₂tḗr). The reconstructed ancestor itself is a proto-language.

A language family is the set of languages descended from a single proto-language. Internal reconstruction infers earlier states from alternations within one language (e.g., English keep/kept inferring an older voicing contrast). Typological universals are cross-linguistic generalisations (e.g., if a language has the dual, it has the plural) that constrain but do not by themselves establish genealogy.

Counterexamples to common slips

Similarity is not cognacy. English have and Latin habere look alike but are not cognate (the Latin h- reflects an earlier gʰ, the English h- an inherited k).
Borrowing mimics inheritance. English paternal is borrowed from Latin and so bypasses Grimm's law — that is why it has "p", unlike inherited father.
Sound laws are not universal across families. Grimm's law is a Germanic fact; Romance pater → père follows different (later, weaker) changes.

Key concepts Intermediate+

The regularity hypothesis and Grimm's law. In 1822 Jacob Grimm observed that the Germanic consonant inventory descends from the Proto-Indo-European (PIE) inventory by a systematic rotation: PIE voiceless stops (*p t k) become Germanic voiceless fricatives (f θ h); PIE voiced stops (*b d g) become voiceless stops (p t k); PIE voiced aspirates (*bʰ dʰ gʰ) become voiced stops/fricatives (b d g). This is a sound law: a change applied without exception throughout the lexicon in the defined environment ^[Grimm1822].

Why this matters. A sound law's exceptionlessness is what licenses reconstruction. If the change were sporadic, we could not run it backwards; because it is regular, a single correspondence set {p : f} across Romance vs Germanic lets us reconstruct *p in the ancestor and predict the reflex in any new cognate.

Verner's law (1877). Some words seemed to break Grimm's law: PIE pH₂tḗr should give Germanic faþēr with a voiceless fricative, yet English "father" and Old English fæder show a voiced ð/d. Karl Verner resolved this: the fricatives produced by Grimm's law were voiced when they followed an unstressed syllable in PIE. "Father" had stress on the second syllable (-tḗr), so its *þ voiced to *ð; "three" (tréyes) had initial stress, so its *þ stayed voiceless. Verner's law dissolved the apparent exceptions and confirmed the regularity hypothesis on a deeper level ^{[Campbell Ch. 6]}.

The comparative method in four steps. (1) Collect candidate cognates by meaning and form, rejecting borrowings and look-alikes. (2) Tabulate the recurring sound correspondences. (3) For each correspondence, posit a proto-phoneme whose subsequent changes in each branch yield the observed reflexes. (4) Check that every proto-phoneme so posited is consistent across all cognate sets, then read off proto-forms, morphemes, and eventually proto-phonology and proto-morphology. The method is falsifiable: a single regular counter-correspondence forces a revision.

Bridge. The regularity of sound change builds toward 51.03.01 (morphology), where the same comparative logic reconstructs ancestral affixes and paradigms, and appears again in 31.05.01 (linguistic anthropology), whose study of contact, migration, and prestige explains why the exceptionless ideal is only an approximation. The foundational reason the comparative method works is that phonetic innovation is mechanical rather than free; this is exactly why a single sound law can rotate an entire consonant inventory, and the pattern generalises to morphological and syntactic reconstruction — the central insight uniting historical linguistics with the phonological feature systems developed in 51.02.01.

Exercises Intermediate+

Exercise 8 (hard, symbolic).

Suppose a putative sound law *t → s operates in a language, but you find three words with *t → θ instead. Lay out two distinct hypotheses that preserve the regularity principle, and state what evidence would decide between them.

Hint

Either the environment differs, or the words are not what they seem.

Answer

Hypothesis A: an unrecognised conditioning factor splits the environment (e.g., *t → s before high vowels, *t → θ elsewhere — cf. Verner's resolution of Grimm's law exceptions). Hypothesis B: the *t → θ words are borrowings or analogical reformations that post-date the law. Decide between them by (i) checking whether the *θ words share a definable phonological/morphological environment, and (ii) checking their semantics and distribution for signs of borrowing. A clean environment favours A; foreign-form matches favour B.

Exercise 9 (hard, symbolic).

Present the two competing models of language divergence (tree and wave), state one kind of evidence each explains better, and explain why most modern classifications use both.

Hint

What does a clean nested hierarchy capture that a spreading front does not, and vice versa?

Answer

The tree (Stammbaum) model (Schleicher) treats divergence as a sequence of binary splits, each followed by independent innovation; it best explains nested shared innovations (synapomorphies) that define families and subgroups. The wave (Wellentheorie) model (Schmidt) treats change as spreading outward from a centre through a dialect continuum; it best explains graded, overlapping isoglosses and the Sprachbund (areal convergence). Modern classifications use both because genealogy is primary but contact is pervasive: the comparative method recovers the tree from nested innovations, while dialect geography and the wave model account for the diffusion that the tree cannot represent.

Lean formalization Intermediate+

lean_status: none. Historical and comparative linguistics is an empirical science whose reasoning is inductive (correspondence → proto-phoneme) rather than deductive, and whose results are judged by the regularity and breadth of attested correspondences. The formal fragments — glottochronology's exponential decay, Bayesian phylogenetic inference of family trees, stochastic lexical-replacement models — are probabilistic and belong to the statistics curriculum of [26.x] and [45.x], not to Mathlib. Encoding the comparative method itself would be better suited to a domain-specific reconstruction DSL than to a proof assistant.

Advanced results Master

The tree model and the wave model. August Schleicher, drawing on Darwin's recent work, proposed in 1861 that languages speciate like organisms: a parent splits into daughters, each then innovating independently, yielding a branching family tree (Stammbaum) ^{[Schleicher1861]}. Johannes Schmidt replied in 1872 that changes spread as waves through a dialect continuum, so that neighbouring varieties share innovations regardless of genealogy and no clean tree exists. Modern practice holds both as partial truths: shared nested innovations define genetic subgroups (tree), while cross-cutting isoglosses reflect diffusion (wave). The two pictures are complementary projections of a process that has both vertical inheritance and horizontal contact.

Internal reconstruction and the laryngeal triumph. In 1879 Ferdinand de Saussure inferred from vowel-alternation patterns within PIE itself that a class of unattested consonants — coefficients sonantiques, later called laryngeals — must once have conditioned those patterns ^{[Saussure1879]}. Decades later Hittite was deciphered and proved to preserve one such consonant (ḫ), turning a purely structural prediction into attested fact. This is the strongest single argument that internal reconstruction recovers real history, not a convenient fiction.

Typological universals and their limits. Greenbergian implicational universals (e.g., if a language has VSO order, it has prepositions) constrain what reconstructions may posit: a proto-language flouting robust universals is suspect. But universals are statistical, not exceptionless, so they guide rather than decide.

Long-range comparison: a contested frontier. Two positions deserve a hearing. The conservative position (Campbell, Ringe) holds that the comparative method is reliable only at shallow time depths (roughly 6,000–10,000 years), beyond which accumulated change erases the regular correspondences needed for proof; proposed macro-families such as Nostratic or Amerind therefore lack demonstrable cognate sets. The maximalist position (Greenberg's mass lexical comparison) argues that enough residual cognates survive to classify deep groupings by statistical resemblance. The disagreement turns on what counts as a demonstrable correspondence: the conservative camp demands exceptionless sound laws, the maximalist camp accepts multilateral statistical tendencies. The discipline at large sides with the conservative camp because the comparative method's track record (laryngeals, Hittite, Tocharian) rewards exactly the strict standard the maximalists relax.

Synthesis. The foundational reason the field resists a single model of divergence is that real language history mixes descent with contact, and this is exactly the tension between the tree picture and the wave picture; the central insight is that regular sound change licenses reconstruction while diffusion licenses areal grouping, putting these together yields a layered model where the comparative method recovers the family tree and dialect geography recovers the wave, the bridge is the recognition that both are partial projections of one multidimensional history, and the framework generalises to any evolving system whose units replicate by inheritance with occasional horizontal transfer — from manuscripts to genes to software lineages.

Full proof set Master

Proposition (the comparative method reconstructs proto-phonemes uniquely up to featural decomposition). Let $L = {L_{1}, \dots, L_{n}}$ be $n$ related languages, and let $C = {C_{1}, \dots, C_{m}}$ be a set of cognate sets covering the same semantic slots, each regular in the Neogrammarian sense (every cognate in the same environment exhibits the same reflex). For each position $j$ in a cognate set, define the correspondence $κ_{j} = (r_{1 j}, \dots, r_{nj})$ where $r_{ij}$ is the reflex in language $L_{i}$ . Then (i) correspondences partition into equivalence classes by the relation "occurs in the same set of cognate positions," and (ii) assigning one proto-phoneme $\astp_{κ}$ per equivalence class is well-defined: the assignment is invariant under reordering or relabelling the daughter languages, and any reconstruction assigning two distinct proto-phonemes to the same regular correspondence is inconsistent with the data.

Proof. (Well-definedness.) Define two correspondences $κ, κ^{'}$ equivalent, $κ \sim κ^{'}$ , iff they occur at the same set of cognate positions. Reflexivity, symmetry, and transitivity are immediate, so $\sim$ partitions the space of correspondences into equivalence classes $[κ]$ . Assign to each class a single proto-phoneme $\astp_{[κ]}$ . This assignment depends only on the partition, which depends only on which reflex-vectors occur together, not on the names or order of the languages; relabelling $L_{i}$ permutes coordinates of every $κ$ identically and so permutes classes without merging or splitting them. Hence the reconstruction is label-invariant.

(Injectivity on classes.) Suppose a reconstruction $R$ assigns two distinct proto-phonemes $\astp \neq = \astp^{'}$ to the same class $[κ]$ . Then some single cognate position $j$ has correspondence $κ_{j} = κ$ , yet $R$ must derive the attested reflex-vector $(r_{1 j}, \dots, r_{nj})$ from $\astp$ under one set of changes and from $\astp^{'}$ under another. But the changes posited for each language are functions of environment; within the fixed environment defining $κ$ , each $L_{i}$ has exactly one attested reflex $r_{ij}$ , so each $L_{i}$ admits exactly one change from a given proto-phoneme to $r_{ij}$ . Two distinct proto-phonemes would then require, in at least one language, two distinct changes mapping to the same reflex in the same environment — a contradiction of regularity (the Neogrammarians' Ausnahmslosigkeit). Therefore $\astp = \astp^{'}$ , and the assignment is injective on classes, hence unique up to the featural decomposition chosen for the proto-phoneme inventory. ∎

Remark. The proof formalises why Verner's law was a genuine discovery rather than a labelling trick: it splits one apparent class (the "exceptions" to Grimm's law) into two by environment (stressed vs unstressed), and the split is forced by the data because within each environment the correspondence is exceptionless.

Connections Master

Phonology 51.02.01. Sound change operates on phonological categories and their features; the phonological inventories and alternations developed there are the substrate that the comparative method reads backwards in time.
Morphology 51.03.01. Once proto-phonemes are fixed, the same comparative logic reconstructs ancestral morphemes and paradigms; analogical levelling, treated under morphology, is a chief disturber of the regularity the comparative method assumes.
Linguistic anthropology 31.05.01. Migration, prestige, and contact — the subject of linguistic anthropology — explain the diffusion the wave model captures and motivate the borrowing-versus-inheritance test on which sound-law reasoning depends.
Probability and statistical phylogenetics 45.07.01. Bayesian tree-inference and glottochronology quantify the comparative method's qualitative judgements, turning cognate sets into dated phylogenies with uncertainty estimates.

Historical & philosophical context Master

Modern historical linguistics begins with Sir William Jones's 1786 observation that Sanskrit, Greek, and Latin share a common source "no philologer could examine ... without believing them to have sprung from some common source." Franz Bopp's Über das Conjugationssystem (1816) extended the comparison to morphology, founding Indo-European studies as a discipline ^{[Campbell Ch. 20]}. Jacob Grimm codified the systematic consonant shift in Deutsche Grammatik (1822), turning scattered resemblances into a law-governed science ^[Grimm1822].

August Schleicher, an explicit Darwinian, introduced the Stammbaum (family-tree) model in Compendium (1861) and published a reconstructed Indo-European fable (Schleicher's fable) to demonstrate that a proto-language could be spoken on paper ^{[Schleicher1861]}. The Neogrammarians (Osthoff, Brugmann, 1878) fixed the regularity hypothesis as dogma: sound change operates without exceptions ("every sound change, in so far as it occurs mechanically, takes place according to laws that admit no exception"). Karl Verner's 1877 law vindicated that dogma by dissolving Grimm's law's apparent exceptions through the discovery of PIE stress conditioning.

Ferdinand de Saussure's Mémoire (1879) performed the field's most celebrated structural prediction: from vowel alternations alone he inferred a class of unattested consonants whose reality was confirmed half a century later by Hittite ḫ ^{[Saussure1879]}. Johannes Schmidt's wave model (1872), Joseph Greenberg's mass-comparison programme (mid-20th century), and the recent Bayesian phylogenetics of Gray and Atkinson continue the same debate Schleicher and Schmidt opened: how deep, and how branching, is the history a method can honestly recover.

Bibliography Master

@book{Campbell2013,
  author = {Campbell, Lyle},
  title = {Historical Linguistics: An Introduction},
  edition = {3},
  publisher = {Edinburgh University Press},
  year = {2013},
}

@book{Trask1996,
  author = {Trask, Robert L.},
  title = {Historical Linguistics},
  publisher = {Arnold},
  year = {1996},
}

@book{HockJoseph2009,
  author = {Hock, Hans Henrich and Joseph, Brian D.},
  title = {Language History, Language Change, and Language Relationship},
  publisher = {Mouton de Gruyter},
  year = {2009},
}

@book{Anttila1989,
  author = {Anttila, Raimo},
  title = {Historical and Comparative Linguistics},
  edition = {2},
  publisher = {John Benjamins},
  year = {1989},
}

@book{Grimm1822,
  author = {Grimm, Jacob},
  title = {Deutsche Grammatik},
  publisher = {Dieterich, Göttingen},
  year = {1822},
}

@book{Schleicher1861,
  author = {Schleicher, August},
  title = {Compendium der vergleichenden Grammatik der indogermanischen Sprachen},
  publisher = {Böhlau, Weimar},
  year = {1861},
}

@phdthesis{Saussure1879,
  author = {de Saussure, Ferdinand},
  title = {Mémoire sur le système primitif des voyelles dans les langues indo-européennes},
  publisher = {Teubner, Leipzig},
  year = {1879},
}

Prerequisites

51.02.01

Tier anchors

beginner: Campbell (2013) Historical Linguistics: An Introduction (Edinburgh UP) Ch. 1; Trask (1996) Historical Linguistics (Arnold) Ch. 1, in plain English
intermediate: Campbell (2013) Historical Linguistics (Edinburgh UP) Chs. 2–6; Trask (1996) Historical Linguistics (Arnold) Chs. 5–8, the comparative method and sound change
master: Campbell (2013); Hock & Joseph (2009) Language History, Language Change, and Language Relationship (Mouton de Gruyter); Anttila (1989) Historical and Comparative Linguistics (Benjamins)

References

Campbell — Historical Linguistics: An Introduction · Edinburgh University Press (2013). Chs. 2–6: sound change, the comparative method, internal reconstruction, language families; Ch. 21 on models of diversification (tree vs wave).
Trask — Historical Linguistics · Arnold (1996). Ch. 5 the comparative method; Ch. 7 phonological change; Ch. 15 on the regularity hypothesis and the Neogrammarians.
Hock & Joseph — Language History, Language Change, and Language Relationship · Mouton de Gruyter (2009). Chs. 2–4: sound change and reconstruction; Ch. 17 on diffusion and the wave model.
Grimm — Deutsche Grammatik · Göttingen (1822). First systematic statement of the Germanic consonant shift (Grimm's law): PIE voiceless stops *p *t *k surface as Germanic fricatives f θ h.
Schleicher — Compendium der vergleichenden Grammatik der indogermanischen Sprachen · Weimar (1861). The Stammbaum (family-tree) model and the first reconstructed Indo-European fable.
Saussure — Mémoire sur le système primitif des voyelles dans les langues indo-européennes · Leipzig (1879). Abstract coefficient sonantique (later identified as laryngeals): structural reasoning about proto-forms from internal patterning alone.

Estimated time

beginner: 15m
intermediate: 40m
master: 85m