Tonal parsimony in chord-sequence analysis: combining modulation cost and tonal vocabulary

François Pachet · 2026

Read Tonal parsimony in chord-sequence analysis: combining modulation cost and tonal vocabulary by François Pachet (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-03459.txt.

Introduction

Pachet studies how to assign local keys to chord sequences with fewer needless changes. The task serves harmonic analysis, composition, and jazz improvisation. Standard dynamic programming cuts modulations but can spawn too many tonal centers. Pachet sets three goals side by side. One counts transitions only, one counts vocabulary only, and one blends both.

The blended goal, tonal parsimony, first cuts modulations, then cuts distinct keys. The joint task is hard in general, but the fixed set of 24 major and minor keys allows exact algorithms. On more than 31,000 sequences from the LMD Chords set, parsimony keeps the best transition count. It also trims the key vocabulary in over half of the cases. With jazz substitution rules added, both mean tonalities and mean modulations fall by clear margins. On annotated jazz standards, chord-scale agreement rises to 95.6 percent, supporting analysis at a professional scale.

Catalog record

CREATORFrançois Pachet
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE49 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.03459

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Tonal parsimony in chord-sequence analysis: combining modulation cost and tonal vocabulary François Pachet Abstract We study the assignment of local tonalities to chord sequences, a task useful for harmonic analysis, composition, and jazz-oriented improvisation. Standard dynamic-programming approaches minimize modulations but can introduce unnecessarily many tonal centers. We compare this transition-only objective with pure minimum-vocabulary analysis and with tonal parsimony, which minimizes lexicographically the number of modulations and then the number of distinct tonalities. Although this joint objective is combinatorially hard in general, we give exact algorithms exploiting the fixed 24-tonality major/minor universe. On 31,032 LMD Chords sequences, tonal parsimony preserves the transition optimum while reducing tonal vocabulary in 55.8% of cases. With weighted jazz-substitution closure, it lowers mean tonalities from 3.802 to 3.206 and modulations from 16.728 to 12.141. On 1,555 annotated jazz standards, it improves compatible chord-scale agreement to 95.6%, supporting tractable professional-scale harmonic analysis. Keywords: tonal parsimony; chord-sequence analysis; jazz harmony; dynamic programming; constraint programming; tonal vocabulary; chord-scale agreement; substitution rules 1 Introduction Given a chord sequence, a harmonic analysis often aims to explain each chord as belonging to one or more local tonal regions. For example, the chord Am7 may be compatible with C major, A minor, G major, and other tonalities depending on the chord vocabulary used. A dominant seventh chord such as D7 naturally points toward G, but may also be interpreted in other contexts. In jazz and related idioms, such an analysis has an immediate practical use: the assigned tonality can be mapped to a scale, providing a compact guide for improvisation, reharmonization, or generative composition. This paper separates the local compatibility layer from the global optimization layer. For each chord position i , we assume that a separate process has already computed a set T_{i} of candidate tonalities. In the baseline experiments reported below, these candidates are drawn from the usual 12 major and 12 minor tonalities using a direct chord-scale compatibility relation. We then study an extension in which this relation is enriched by weighted jazz substitution rules. In both cases, the task studied here is the global one: choose one tonality x_{i}\in T_{i} for each chord while optimizing the sequence-level criteria. Dynamic-programming and optimal-segmentation approaches to local key or tonality analysis have a long history. They commonly combine local evidence with a transition penalty, or equivalently prefer analyses with fewer key changes. Temperley’s reconsideration of the Krumhansl–Schmuckler key-finding algorithm explicitly introduces a modulation penalty between successive segments ( 18 ) . The jazz harmonic analysis of 6 is close to the present setting: harmonic analysis is formulated as an optimal tonality-segmentation problem over a chord sequence. Recent work on joint pitch spelling and local/global key estimation also uses dynamic-programming style optimization over musical sequences ( 3 ) . Probabilistic approaches provide a related inference perspective. 15 formulate functional harmonic analysis as inference in a probabilistic model over keys and harmonic functions. More generally, tonal-space representations such as Chew’s Spiral Array provide formal models of tonal centers and tonal distance ( 5 ) , while statistical chord-sequence models study latent harmonic categories and self-emergent grammar ( 19 ) . Grammar-based accounts of jazz harmony, notably the generative grammar for jazz chord sequences of 17 , provide a complementary formal view of substitutional and recursive harmonic structure. Existing dynamic-programming approaches to harmonic analysis penalize modulations or segment transitions, but do not optimize the cardinality of the set of tonal centers used in the analysis. We introduce a tonal-parsimony criterion formulated as a lexicographic optimization of modulation cost and tonal vocabulary size. This captures the musical intuition that key has inertia, while adding a second intuition: among equally smooth analyses, one should prefer a smaller tonal vocabulary. The paper makes three contributions. First, it formulates tonal parsimony as the lexicographic objective \min_{\mathrm{lex}}(C,K), where C is the modulation cost and K is the number of distinct tonalities used. Second, it gives a specialized exact dynamic program for this objective, realizing the combination of a cost-regular path criterion and an NValue criterion over a fixed tonality universe. Third, it introduces a substitution-closure layer that expands local candidate domains with weighted jazz-harmonic reinterpretations before global optimization. The strongest empirical test is the external chord-scale validation: on 1,555 professionally annotated jazz standards ( 9 ) , tonal parsimony reaches 81.1% strict and 95.6% conservative compatible agreement, outperforming both baselines while preserving the minimum modulation cost of the transition-only dynamic program. This indicates that the criterion is not only a more compact objective, but also a better predictor of expert local scale annotations. 2 Problem Statement Let c_{1},\ldots,c_{n} be a chord sequence. For each chord c_{i} , let T_{i}\subseteq\mathcal{T} be its set of candidate tonalities, where in the present work |\mathcal{T}|=24 . An analysis is a sequence x_{1},\ldots,x_{n} such that x_{i}\in T_{i} for all i . We define two quantities: C(x)=\sum_{i=1}^{n-1}[x_{i}\neq x_{i+1}], the number of modulations or adjacent tonality changes, and K(x)=|\{x_{1},\ldots,x_{n}\}|, the number of distinct tonalities used by the analysis. The classical dynamic-programming objective minimizes C . Our proposed objective minimizes (C,K) lexicographically: first minimize the number of modulations, and among all equally smooth analyses choose one using fewer tonalities. 3 Constraint Satisfaction Formulation The assignment problem can be modeled as a finite-domain constraint optimization problem ( 16 ) . There is one decision variable x_{i} per chord position, with finite domain T_{i}\subseteq\mathcal{T} . The hard constraints are the domain constraints x_{i}\in T_{i} ; the musical criteria are expressed as global quantities over the whole sequence. The modulation cost C(x) is a regular sequence cost: it can be represented by a small automaton that reads the selected tonalities and accumulates a unit cost whenever two consecutive labels differ. This places the transition-only objective in the family of Regular and cost-regular global constraints ( 13 ; 7 ) . The number of distinct tonalities K(x) is the value counted by the NValue global constraint ( 2 ) . This follows the broader constraint-programming practice of treating global constraints as structured relations whose graph or automaton structure can be exploited algorithmically ( 1 ) . This constraint-programming view is primarily a modeling statement. A literal off-the-shelf CP model combining a cost-regular path constraint with an NValue objective is unlikely to be tractable at the scale considered here: the solver would still have to coordinate a sequence cost with a global cardinality objective over thousands of instances. The algorithms below are specialized exact algorithms inspired by this model. They exploit the fixed 24-tonality universe directly, instead of relying on generic search over the combined global constraints. As a sanity check, we wrote the direct model in MiniZinc using costRegular and NValue . It reproduces the expected optima on small examples, including the substitution-enriched progression Cmaj7 F#7 F Bb7 Em Am Dm Db7 , for which it returns a single C-major path. Table 1 gives exploratory timings with MiniZinc/Gecode 6.3.0 and the specialized implementation. Instance Chords Tonalities MiniZinc/Gecode Specialized Python Tie-break example, (C,K)=(2,2) 3 3 0.18 s 0.013 ms Substitution example, (C,K)=(0,1) 8 4 0.16 s 0.019 ms Corpus excerpt, (C,K)=(17,5) 32 22 21.9 s 0.26 ms Corpus excerpt, (C,K)=(25,8) 48 22 >2 h 15 min 0.61 ms Table 1: Direct MiniZinc/Gecode model versus our specialized fixed-universe implementation written in Python, measured

Read the complete text