Autosegmental phonology, feature geometry, and optimality theory

Intuition Beginner

Phonology 51.02.01 showed how a language sorts its speech sounds into contrastive categories. But some sound patterns refuse to stay pinned to a single segment. A tone can drift free of the vowel that first carried it; a vowel-harmony feature can sweep across a whole word; stress can land on whichever syllable is most prominent. Linear, one-segment-at-a-time notation strains to capture these.

Picture a musical score. Melody sits on one staff, rhythm on another, dynamics on a third. They run in parallel and snap together at marked points. Autosegmental phonology gives speech the same treatment. Tones, harmony features, and syllable roles each occupy their own layer, called a tier, linked by association lines to the consonants and vowels underneath.

A second revolution came in the 1990s. Instead of writing rules that rewrite one form into another, list the demands a language places on pronunciation and rank them. Every conceivable pronunciation enters a race. Each demand, a constraint, may be broken, but only the worst offenders lose. The form that breaks only low-ranked demands wins. This is Optimality Theory.

Visual Beginner

A schematic autosegmental representation: tones live on a tier above the segments and attach by association lines. One tone can spread to two vowels; one segment can sit under no tone at all.

The same idea as a tier diagram and an OT tableau:

Tone tier:      H       L       H
               |       |       |
Skeletal tier:  x       x       x        (timing slots)
               |       |       |
Segment tier:   m       b       a

Candidate	NOCODA	MAX-IO
[ap]	*
[a]		*

The lower picture is an Optimality Theory tableau: each candidate collects a mark (*) under each constraint it breaks. Ranking decides the winner.

Worked example Beginner

Optimality Theory picks a pronunciation. A hypothetical language disfavours a consonant at a syllable end (the constraint NOCODA) but also disfavours dropping input material (the constraint MAX-IO). Take the input /ap/. The generator offers two candidate pronunciations, and each collects marks under the constraints it breaks:

Candidate	NOCODA	MAX-IO
[ap]	*
[a]		*

Each candidate collects exactly one mark. Now rank the constraints.

• Ranking NOCODA above MAX-IO: the candidate that breaks the top rule ([ap], which breaks NOCODA) loses. [a] wins. This is a coda-deleting language.
• Ranking MAX-IO above NOCODA: the candidate that breaks the top rule ([a], which breaks MAX-IO) loses. [ap] wins. This is a coda-keeping language.

What this tells us: the same two pressures, ranked in opposite orders, produce opposite phonologies. The mark count (one each) does not decide; the ranking does. That ranking-decides lesson is the heart of Optimality Theory.

Check your understanding Beginner

Formal definition Intermediate+

A linear phonological representation, of the kind standard after Chomsky and Halle (1968), is a single sequence of segments each decorated with a feature bundle. Autosegmental phonology replaces this single sequence by a multitier representation ^[Goldsmith]. Let $T_1,T_2,\dots ,T_n$ be tiers, each a finite sequence of autosegments (tones, features, or skeletal slots). An autosegmental representation is a tuple $(T_1,\dots ,T_n,A)$ where $A$ is a set of association lines joining autosegments on different tiers. A segment tier and a tone tier may share an index, but the tiers vary independently.

A representation is well-formed iff it satisfies Goldsmith's Well-Formedness Conditions: (i) every autosegment is associated to at least one autosegment on every other tier, except at the edges of the domain; (ii) association lines do not cross. Condition (ii), the No-Line-Crossing Constraint, forces the association graph $A$ to be a non-crossing matching between adjacent tiers. Unassociated autosegments at the edge are permitted; fully unassociated autosegments are floating.

Feature geometry (Clements 1985; Clements and Hume 1995) organises the distinctive features of 51.02.01 into a rooted tree rather than a flat bundle ^[Clements]. The root branches into a laryngeal node ([±voice], [±spread glottis], [±constricted glottis]), a Place node (which itself branches into Labial, Coronal, Dorsal, Radical), and manner features ([±continuant], [±nasal], [±lateral]). A phonological process may delink an entire node; placeless consonants, for example, arise from delinking the Place node.

The prosodic hierarchy layers phonological material above the segment:

$$\text{mora}\subset \text{syllable}\subset \text{foot}\subset \text{prosodic word}\subset \text{phonological phrase}\subset \text{intonational phrase}.$$

The Strict Layer Hypothesis requires each category at level $k$ to be exhaustively parsed into categories at level $k+1$. A syllable is structured as $\sigma =(\text{onset})\text{nucleus}(\text{coda})$, with nucleus and coda forming the rhyme. The Sonority Sequencing Principle (SSP) requires sonority to rise from the syllable edge to the nucleus, so onsets rise in sonority toward the vowel and codas fall.

Optimality Theory (Prince and Smolensky 1993/2004) is stated in terms of three functions ^{[PrinceSmolensky]}. The generator $\mathrm{GEN}$ maps an input $I$ to a candidate set $\mathrm{GEN}(I)=\{c_1,c_2,\dots \}$. The constraint set $\mathrm{CON}$ is a finite set of constraints, each assigning a non-negative integer ${\mathrm{C}}_j(c)$ to every candidate $c$. The ranking $\gg$ is a strict total order on $\mathrm{CON}$. The evaluator $\mathrm{EVAL}$ selects the candidate

$$\mathrm{EVAL}(I)=\arg \underset{c\in \mathrm{GEN}(I)}{\min }v(c),$$

where $v(c)$ is the violation vector $({\mathrm{C}}_1(c),{\mathrm{C}}_2(c),\dots ,{\mathrm{C}}_m(c))$ ordered so that ${\mathrm{C}}_1\gg {\mathrm{C}}_2\gg \cdots \gg {\mathrm{C}}_m$, and the minimum is taken lexicographically. Constraints split into markedness constraints (penalising structurally dispreferred output configurations regardless of input) and faithfulness constraints (penalising deviation between input and output, such as MAX-IO against deletion and DEP-IO against insertion).

Counterexamples to common slips

• Tiers are not free decoration. A tier is justified only when its autosegments display behaviour independent of the segment tier: spreading, stability under deletion, or floating. Adding a tier without such evidence is unfalsifiable decoration.
• Feature geometry is not the flat SPE bundle. Grouping matters: a rule that targets the Place node acts on Labial, Coronal, and Dorsal together, which a flat-feature statement cannot express without a disjunction.
• OT rankings are strict total orders, not weights. The lexicographic comparison means one extra violation of a higher-ranked constraint can never be compensated by any number of lower-ranked violations. Weighted variants (Harmonic Grammar) are a different theory.

Key concepts Intermediate+

The Well-Formedness Conditions. Goldsmith's two conditions (full association except at edges; no crossing lines) are the load-bearing claim of autosegmental phonology ^[Goldsmith]. They predict the three hallmark phenomena of tone: spreading (an unassociated autosegment extends an association line to a neighbour), stability (a tone survives the deletion of its host vowel because it lives on its own tier), and floating tones (an unassociated tone that nonetheless causes downstep on a following tone). All three follow from treating tone as an independent tier rather than as a feature attached to a vowel.

The Sonority Sequencing Principle. Sonority is a partial order on sound classes, roughly vowels $>$ glides $>$ liquids $>$ nasals $>$ fricatives $>$ stops. The SSP requires that within a syllable, sonority rises monotonically from the left edge to the nucleus and falls monotonically from the nucleus to the right edge. It explains why English permits /pl/ in play (rising sonority in the onset) but rejects /lp/ as an onset (lpay would fall then rise). Syllable contact and sonority-driven syllabification follow from the same ordering.

Tableau evaluation as lexicographic minimisation. An OT tableau displays, for each candidate, the violation count under each constraint in columns left-to-right by descending rank. The winner is the candidate whose leftmost differing column is smaller. The tableau is a device for reading off the lexicographic minimum that $\mathrm{EVAL}$ computes. Two properties make this nontrivial. First, the ranking is a strict total order, so ties between candidates are broken by the next constraint to the right. Second, markedness and faithfulness are in inherent tension: every alternation (devoicing, epenthesis, deletion, harmony) is a resolution of that tension by ranking.

Bridge. The autosegmental tier builds toward 51.04.01, where syntax adopts a parallel architecture in which separate modules interact, and appears again in 47.01.01, whose string-rewriting machinery treats tiers and association lines as edges in a planar graph over a finite alphabet. The foundational reason tiers were introduced is that tone and harmony exhibit locality on their own level that the segmental sequence destroys; this is exactly the move that lets a phonological process refer to a class of segments as a geometric node rather than a list; the central insight is that the plane of representation is as explanatory as the rules that act on it. Putting these together, the bridge is that autosegmental structure and OT ranking jointly replace serial rule-ordering by parallel evaluation on a richly structured representation, and the pattern generalises to metrical, prosodic, and morphological domains.

Exercises Intermediate+

Lean formalization Intermediate+

lean_status: none. Autosegmental phonology, feature geometry, and Optimality Theory are empirical linguistic theories whose correctness gate is typological distribution and corpus evidence, not formal proof. The genuinely formal content — multitier representations as planar graphs subject to the No-Line-Crossing Constraint, feature trees as rooted partial orders, OT ranking as a strict total order evaluated by a lexicographic $\mathrm{EVAL}$ function, and factorial typology as the orbit of the ranking permutation group — belongs to the formal-language and order-theory machinery of 47.01.01 rather than to Mathlib's core, and Mathlib currently has no autosegmental-graph, feature-geometry, or OT-tableau formalism.

Advanced results Master

The Well-Formedness Conditions and tier autonomy. Goldsmith's (1976) conditions are not stylistic conventions but empirical hypotheses: they predict that tonal phenomena are insensitive to the segmental composition of their host, a claim confirmed by the cross-linguistic stability of tone under vowel elision, the independence of contour tones from vowel length, and the regularity of tone docking after deletion ^[Goldsmith]. Tier autonomy is the property that makes a tier a tier: the autosegments on $T_i$ obey their own locality, independent of adjacency on $T_j$.

The Obligatory Contour Principle. Leben (1973) and McCarthy (1986) proposed that adjacent identical autosegments on a single tier are prohibited at the level of underlying representation ^[McCarthy]. The OCP predicts that a sequence /H H/ is represented as a single multiply-linked H autosegment, and that a true sequence of like tones is repaired by fusion, dissimilation, or spreading. The OCP functions both as a constraint on representations and, in OT, as a markedness constraint whose ranking explains dissimilatory alternations such as Latin liquid dissimilation and the avoidance of adjacent identical consonants.

Feature geometry and class behaviour. Clements (1985) and Clements and Hume (1995) showed that features are grouped into a rooted tree whose nodes (Place, Laryngeal, Manner, and subnodes) behave as units in phonological processes ^[Clements]. Delinking the Place node produces placeless consonants attested in archiphonemic neutralisation; spreading the Dorsal node produces velar harmony in child phonology and in languages with tongue-root harmony. The geometry explains class behaviour — why features spread, assimilate, and delete in groups — as a structural consequence of node identity rather than as a stipulated disjunction over a flat bundle.

The prosodic hierarchy and Strict Layering. Nespor and Vogel (1986) and Selkirk (1980, 1995) fixed the prosodic hierarchy as a sequence of phonological categories each properly contained in the next. The Strict Layer Hypothesis — that a unit at level $k$ is exhaustively parsed into units at level $k+1$ — predicts the domains of phonological rules: sandhi applies within the phonological phrase, liaison within the prosodic word, and intonational tones align with the intonational phrase. Apparent violations (e.g., Recursive Prosodic Word structure) motivated the move to strict-layering-with-recursion in Selkirk (1995), retaining the hierarchy while admitting limited self-embedding.

Metrical phonology. Liberman and Prince (1977) recast stress as relative prominence on a metrical tree rather than as a segmental feature. Each pair of sibling nodes is labelled strong (s) or weak (w); stress patterns follow from the geometry of the tree and from parameter settings (rhythm, boundedness, headedness). This explains the iambic–trochaic typology and the iterative rhythm of stress without diacritic stress marks on vowels.

Correspondence Theory. McCarthy and Prince (1995) generalised input–output faithfulness to a correspondence relation between any two related structures (input–output, base–reduplicant, output–output). Faithfulness constraints (MAX, DEP, IDENT) evaluate pairs of corresponding segments. This recasts reduplication as the optimisation of base-reduplicant identity and unifies a family of opacity phenomena that the original input–output version of OT could not handle.

Factorial typology and Richness of the Base. The set of grammars OT generates is the set of total orders on $\mathrm{CON}$; the typology is the union, over inputs, of the winners those orders select. Richness of the Base holds that the input set is universal: every input is available to every grammar, and cross-linguistic differences arise solely from ranking, never from restricted lexicons. The combination yields strong, falsifiable predictions: if a typology contains a language with final devoicing but none with final voicing under the same constraint set, the constraint set is empirically inadequate.

Synthesis. Autosegmental tiers, feature geometry, the prosodic hierarchy, and OT ranking builds toward a unified parallel-architecture view of phonology, and appears again in 51.04.01 where parallel constraint interaction organises syntax. The foundational reason OT displaced serial derivation is that typology is shaped by conflicting universal pressures rather than by ordered repairs; this is exactly the markedness-versus-faithfulness tension; the central insight is that every surface form is the optimal compromise among ranked, violable constraints acting on a multitier representation. Putting these together, the bridge is that phonology becomes the cleanest laboratory for studying how universal pressures resolve into language-particular grammars, a pattern that generalises to morphology (prosodic templates), to acquisition (the Constraint Demotion Algorithm), and to the formal-language setting of 47.01.01.

Full proof set Master

Proposition (No-Line-Crossing yields a planar association graph). Let $T_S=(s_1,\dots ,s_m)$ and $T_T=(t_1,\dots ,t_n)$ be two tiers of an autosegmental representation, drawn as two parallel horizontal lines with their autosegments in order. Let $A\subseteq \{(i,j):s_i\in T_S,t_j\in T_T\}$ be the set of association lines. If the representation satisfies the No-Line-Crossing Constraint, then no two association lines $(i_1,j_1),(i_2,j_2)\in A$ with $i_1<i_2$ have $j_1>j_2$. Consequently $A$, regarded as a bipartite graph between the index sets $\{1,\dots ,m\}$ and $\{1,\dots ,n\}$, is a non-crossing matching, and the tiered graph is planar.

Proof. Suppose, toward a contradiction, that two association lines cross. Crossing means the line from $s_{i_1}$ to $t_{j_1}$ intersects the line from $s_{i_2}$ to $t_{j_2}$ at an interior point, with the endpoints in opposite orders on the two tiers: without loss of generality $i_1<i_2$ on $T_S$ and $j_1>j_2$ on $T_T$. By the definition of the No-Line-Crossing Constraint no two association lines of a well-formed representation may cross, so this configuration is forbidden; hence no such pair exists. Therefore every pair $(i_1,j_1),(i_2,j_2)\in A$ with $i_1<i_2$ satisfies $j_1\le j_2$, which is precisely the non-crossing matching condition. The two-tier graph embeds in the strip $\mathbb{R}\times [0,1]$ with the two tiers as horizontal boundary lines and association lines as straight segments in the interior; non-crossing segments in a strip do not intersect, so the embedding has no edge crossings and is planar. $\square$

Corollary (Stability under deletion). Deleting an autosegment on one tier removes its incident association lines but leaves every other autosegment and every other line intact. Because the remaining association graph is a subgraph of a non-crossing matching, it is itself non-crossing, so the post-deletion representation remains well-formed. The tonal autosegments of a deleted vowel therefore persist on the tone tier, either reassociating to an adjacent slot or floating — the structural reason for tone stability.

Proposition (Factorial typology for two constraints). Let $\mathrm{CON}=\{C_1,C_2\}$ be a constraint set, $I$ an input, and $a,b\in \mathrm{GEN}(I)$ two candidates with violation profiles ${\mathrm{C}}_1(a)=0$, ${\mathrm{C}}_2(a)=1$, and ${\mathrm{C}}_1(b)=1$, ${\mathrm{C}}_2(b)=0$. Then the ranking $C_1\gg C_2$ selects $a$ and the ranking $C_2\gg C_1$ selects $b$.

Proof. Under ranking $C_1\gg C_2$ the violation vectors are ordered $({\mathrm{C}}_1,{\mathrm{C}}_2)$, so $v(a)=(0,1)$ and $v(b)=(1,0)$. The lexicographic order compares first components: $0<1$, so $v(a){<}_{\mathrm{lex}}v(b)$ and $\mathrm{EVAL}(I)=a$. Under ranking $C_2\gg C_1$ the vectors are ordered $({\mathrm{C}}_2,{\mathrm{C}}_1)$, so $v(a)=(1,0)$ and $v(b)=(0,1)$. The first components compare $1>0$, so $v(b){<}_{\mathrm{lex}}v(a)$ and $\mathrm{EVAL}(I)=b$. The two total orders on $\mathrm{CON}$ select opposite winners, so the factorial typology of $\mathrm{CON}$ on $I$ is $\{a,b\}$. $\square$

The proposition is the two-constraint kernel of every OT typological argument: reranking two constraints with complementary violation profiles flips the winner. The devoicing typology of Exercise 8 is the instance with $C_1=\ast \mathrm{VOICED}\text{-}\mathrm{CODA}$, $C_2=\mathrm{IDENT}\text{-}\mathrm{IO}[\mathrm{voice}]$, $a=\text{[bat]}$, $b=\text{[bad]}$.

Connections Master

• Phonology foundations 51.02.01. This unit deepens the feature-and-rule apparatus introduced there: the flat SPE feature bundle becomes the feature-geometry tree, the serial rule-based derivation gives way to OT parallel evaluation, and the phoneme-to-phone mapping is recast as the optimisation of a multitier representation against a ranked constraint set.

• Syntax 51.04.01. The parallel, constraint-based architecture developed here is the template for constraint-based syntactic theory: Optimality-Theoretic syntax and the ranked-evaluation model transfer directly to constituency, movement, and well-formedness, and the autosegmental idea of independent tiers reappears as autonomous modules of the syntax–semantics interface.

• Formal languages and automata 47.01.01. Autosegmental tiers and association lines are edges in a planar bipartite graph over a finite alphabet, phonotactic templates are regular constraints on strings, and the OT generator $\mathrm{GEN}$ is a relation over a formal language; the generative power of phonological representations takes its precise measurement in the Chomsky hierarchy.

• Phonetics 51.01.01. The tiers and nodes of this unit are abstractions over the continuous acoustic and articulatory signal described by phonetics: every feature-geometry node has articulatory and acoustic correlates, and the typological predictions of OT are tested against phonetic distributions measured in that signal.

Historical & philosophical context Master

John Goldsmith introduced autosegmental phonology in his 1976 MIT dissertation (Autosegmental Phonology, published in revised form by Basil Blackwell in 1979), where the multitier representation and the Well-Formedness Conditions were proposed to handle African tone languages that the linear SPE framework could not describe cleanly ^{[Goldsmith 1976]}. George Clements extended the autosegmental programme to distinctive features in The Geometry of Phonological Features (1985), arguing that the SPE feature bundle has internal hierarchical structure; the Clements and Hume (1995) synthesis, The Internal Organization of Speech Sounds, fixed the modern tree with Place, Laryngeal, and Manner nodes ^{[Clements 1985]}. Mark Liberman and Alan Prince's (1977) paper on metrical phonology added the metrical tree for stress, completing the representational programme that dominated the 1980s.

Alan Prince and Paul Smolensky's Optimality Theory: Constraint Interaction in Generative Grammar, circulated as a manuscript in 1993 and published by Blackwell in 2004, replaced serial rule ordering with parallel evaluation of ranked violable constraints ^{[PrinceSmolensky 1993]}. John McCarthy's work on the Obligatory Contour Principle (1986) and prosodic morphology (McCarthy and Prince 1986–1995) supplied the empirical cases that established the new framework, and Bruce Hayes's Introductory Phonology (2009) codified the OT-first pedagogy. The 1990s debate between rule-based serialists and OT parallelists concerned the nature of grammatical computation itself — whether a grammar is a sequence of operations or a ranking of conditions.

Bibliography Master

@phdthesis{Goldsmith1976,
  author = {Goldsmith, John A.},
  title  = {Autosegmental Phonology},
  school = {Massachusetts Institute of Technology},
  year   = {1976},
}

@book{Goldsmith1979,
  author    = {Goldsmith, John A.},
  title     = {Autosegmental Phonology},
  publisher = {Basil Blackwell},
  year      = {1979},
}

@incollection{Clements1985,
  author    = {Clements, George N.},
  title     = {The Geometry of Phonological Features},
  booktitle = {Advances in Nonlinear Phonology},
  editor    = {van der Hulst, Harry and Smith, Norval},
  publisher = {Foris},
  year      = {1985},
}

@incollection{ClementsHume1995,
  author    = {Clements, George N. and Hume, Elizabeth V.},
  title     = {The Internal Organization of Speech Sounds},
  booktitle = {The Handbook of Phonological Theory},
  editor    = {Goldsmith, John A.},
  publisher = {Blackwell},
  year      = {1995},
}

@article{LibermanPrince1977,
  author  = {Liberman, Mark and Prince, Alan},
  title   = {On Stress and Linguistic Rhythm},
  journal = {Linguistic Inquiry},
  volume  = {8},
  number  = {2},
  pages   = {249--336},
  year    = {1977},
}

@unpublished{PrinceSmolensky1993,
  author = {Prince, Alan and Smolensky, Paul},
  title  = {Optimality Theory: Constraint Interaction in Generative Grammar},
  note   = {Rutgers University Center for Cognitive Science Technical Report 2},
  year   = {1993},
}

@book{PrinceSmolensky2004,
  author    = {Prince, Alan and Smolensky, Paul},
  title     = {Optimality Theory: Constraint Interaction in Generative Grammar},
  publisher = {Blackwell},
  year      = {2004},
}

@article{McCarthy1986,
  author  = {McCarthy, John J.},
  title   = {OCP Effects: Gemination and Antigemination},
  journal = {Linguistic Inquiry},
  volume  = {17},
  number  = {2},
  pages   = {207--263},
  year    = {1986},
}

@article{McCarthyPrince1995,
  author  = {McCarthy, John J. and Prince, Alan},
  title   = {Faithfulness and Reduplicative Identity},
  journal = {University of Massachusetts Occasional Papers in Linguistics},
  volume  = {18},
  pages   = {249--384},
  year    = {1995},
}

@book{NesporVogel1986,
  author    = {Nespor, Marina and Vogel, Irene},
  title     = {Prosodic Phonology},
  publisher = {Foris},
  year      = {1986},
}

@book{TesarSmolensky2000,
  author    = {Tesar, Bruce and Smolensky, Paul},
  title     = {Learnability in Optimality Theory},
  publisher = {MIT Press},
  year      = {2000},
}

@book{McCarthy2008,
  author    = {McCarthy, John J.},
  title     = {Doing Optimality Theory: Applying Theory to Data},
  publisher = {Blackwell},
  year      = {2008},
}

@book{Hayes2009,
  author    = {Hayes, Bruce},
  title     = {Introductory Phonology},
  publisher = {Wiley-Blackwell},
  year      = {2009},
}