A Schenkerian analysis reduces a melody step by step. Ornamental notes are stripped away until a skeleton of structurally essential tones remains, and every note ends up with an integer depth: the level at which it disappears. Machine parsers for this task now exist. The best of them, AutoSchA, predicts depth correctly about 75% of the time, yet expert theorists still flag two thirds of its analyses as containing "weird or awkward" choices. Depth accuracy is a per-note score, while the theorists are reacting to the overall shape of the tree. No one has a principled way to say what a typical human tree looks like, or how far a parser's trees stray from that.
This proposal borrows that missing yardstick from algebraic combinatorics. Combinatorialists have exact counting formulas for trees and lattice paths classified jointly by shape statistics, such as how many leaves they have or how many leaves sit at the end of unbranched chains. Some of these formulas come with strong guarantees: the counting polynomials have only real roots, which makes the statistic behave like a sum of independent coin flips. The plan is to encode each human Schenkerian analysis as one of these objects. That turns the formulas into two tools. The first is an exact null model that tests whether human analyses, and parser outputs, have unusual shapes. The second is a prior over tree shape with very few parameters and an exactly known normalizer, which we use to rerank parser outputs. The music data are small (145 analyses), and that is the regime where a few interpretable parameters should beat a learned structure model.
Core Claim
We will test whether the shape statistics of human Schenkerian trees deviate from size-matched random trees in a stable, low-dimensional way. If they do, two things follow. First, the errors theorists notice in AutoSchA should show up as deviations in those same statistics. Second, reranking AutoSchA's candidate analyses with a shape prior fit on two or three statistics should cut awkward choices and raise depth accuracy beyond what the parser's own confidence achieves.
The combinatorial input comes from two places. The first is the refined Narayana enumeration of A Combinatorial Proof of a Symmetry for a Refinement of the Narayana Numbers. That paper counts Dyck paths, which are up/down step sequences that never go below the starting level and correspond one-to-one with rooted ordered trees. It counts them by semilength $n$, by the number of peaks (leaves), and by the number of $U^rD$ factors (leaves reached after at least $r$ consecutive up-steps). It gives closed forms, real-rootedness of the polynomials $W_{n,k}(t)$, and a split into symmetric parts $W^+$ and $W^-$ when $n \neq 2k+1$. The paper supplies the method: it lets us compute exact null distributions without sampling.
The second source is Tiered Trees and Theta Operators. It shows that a Theta-operator evaluation, $\Theta_{e_1 e_n e_1}|_{t=1}$, enumerates tiered trees, meaning trees whose vertices carry a level function. Schenkerian depth is exactly a level function on notes. Whether real analyses satisfy the tiered-tree constraints is an empirical question that we check rather than assume. Theta operators come from the line of work in Theta operators, refined delta conjectures, and coinvariants and A proof of the compositional Delta conjecture. They are linear maps on symmetric functions, and their coefficients count decorated paths and trees.
What is new is treating musical reductions as weighted combinatorial objects with known generating functions, rather than as unweighted parse trees. Prior Dyck-language work in machine learning, such as How Can Self-Attention Networks Recognize Dyck-n Languages?, asks only whether a bracket string is well formed and ignores the shape statistics that combinatorics tracks.
Research Program
The first step is the encoding. We start from the heterogeneous-graph representation of A New Dataset, Notation Software, and Representation for Computational Schenkerian Analysis, which shows that an integer depth per note determines the prolongation structure. From each analysis we build a rooted ordered prolongation tree and map it to a Dyck path with the standard bijection. Under this map, peaks are surface ornaments, and $U^rD$ factors are ornaments hanging off chains of $r$ or more nested prolongations. The cases the dataset paper names as hard are converging voices and inner voices with undefined depth. We handle them two ways, either dropping those analyses or using the paper's own heuristics, and report results under both. We also record the fraction of analyses whose depth labeling satisfies the tiered-tree constraints. If that fraction is low, the Theta-operator branch is dropped and the Narayana branch carries the proposal.
The second step is the null-model analysis. Condition on the number of notes and on the number of surface ornaments ($n$ and $k$). The refined Narayana polynomial then gives the exact distribution of the chain statistic $m$ under uniformly random trees. Real-rootedness means this distribution is that of a sum of independent Bernoulli variables, which gives exact means, variances and sharp tail bounds. We compute where each human analysis falls in its conditional null, and do the same for AutoSchA outputs. We then ask two questions. Do human trees show a consistent shift, for example deeper chains than chance? Is the gap between parser and human concentrated in the analyses theorists flagged?
AutoSchA found that rhythmic features predict depth better than pitch features. We test whether the shape deviations correlate with metric position. This is a direct, small-scale probe of the rhythm-versus-pitch tension.
The third step is the shape prior and reranking. We define a distribution over trees $T$ with $n$ notes,
$$p_\theta(T) = \frac{x^{k(T)}\, y^{m(T)}}{Z_n(x,y)},$$
where $k(T)$ is the number of peaks, $m(T)$ the number of $UUD$ factors, $x$ and $y$ are positive parameters (collected as $\theta$), and $Z_n(x,y) = \sum_{k,m} w_{n,k,m}\, x^k y^m$ is the exact normalizer from the refined Narayana counts. Fitting $x$ and $y$ by maximum likelihood on training analyses is a convex problem in $\log x$ and $\log y$. We take AutoSchA's top candidate trees, from beam or sampled outputs, and rescore each one by its parser log-probability plus $\lambda \log p_\theta(T)$, tuning $\lambda$ on validation folds. Theorem 5 of the Narayana paper gives closed forms for $U^rD$ factors up to arbitrary $r$, so a richer variant adds the $r=3$ count.
Baselines
The naive reference point assigns depth from metric strength alone: notes on stronger beats get deeper structural levels. AutoSchA's ablations suggest rhythm carries most of the signal, so this baseline is not trivially weak. Any shape prior should beat it clearly.
The strongest existing method is AutoSchA itself, a relational graph network that removes notes level by level and is regularized so a removed note cannot come back. It reports roughly 75% structural depth accuracy on its Bach and Pachelbel fugue-subject benchmark, built from 88 subjects transposed into all 12 keys. In the expert evaluation, it produced awkward choices 66% of the time against 33% for human analysts, and earned a GPA of 3.216 against 3.551 (AutoSchA). Standard graph networks and MLPs are reported only as outperformed, with no numbers found on Lacuna. We retrain AutoSchA on our folds with its published configuration and generate candidate trees from that same model for every reranker, so all rerankers see identical candidates and an identical compute budget. Beating it means a statistically significant gain in depth accuracy across cross-validation folds together with a lower expert-flagged awkwardness rate in a blind re-evaluation.
The key ablation is a learned structured reranker. It is a conditional random field over trees, scored by a dynamic program, and uses many free shape features: depth histograms, branching factors per level, and chain lengths. It gets the same candidates and a matched hyperparameter search. This isolates whether the specific combinatorial statistics and their low parameter count matter, or whether any shape reranker would do. A second ablation sets $x = y = 1$, the uniform prior, which tests whether the fitted parameters carry information.
Why not just…?
Why not just use a dynamic program over trees instead of closed-form counting? This is a real threat, and it is partly correct. A dynamic program can compute the same normalizer for the same features, so the closed form brings convenience and exact tail bounds, not extra modelling power. The claim we actually test is about which features and how many. The learned conditional random field ablation settles it. If a richly featured dynamic-program reranker matches or beats the two-parameter prior on 145 analyses, we report that the contribution reduces to the null-model analysis and the interpretable statistics.
Why not just train a bigger neural parser or fine-tune a music language model? The data are 145 analyses. Can LLMs "Reason" in Music? reports that large language models fail at formal structural reasoning in music. ProGress shows a 3-million-parameter model with hierarchical structure outperforming the 516-million-parameter NotaGen in human ratings. Neither result proves a bigger parser would fail at depth prediction specifically. Our comparison against AutoSchA, the strongest learned parser available, is the direct test.
Why not just compare histograms of tree statistics between parser and human outputs? Raw histograms mix shape with size and ornament count. A permutation test using trees sampled by a dynamic program could fix that, and we will run one as a check. The exact conditional null makes the comparison precise for every $(n,k)$ cell, including cells with only a few analyses, which matters on a dataset this small.
Proposed Experiments
All experiments use the 145-analysis Schenkerian dataset (dataset paper) with folds split by piece, so transposed copies of one subject never cross train and test. Metrics are per-note depth accuracy, tree edit distance to the expert analysis, and the per-analysis tail probability of the chain statistic under the null.
The decisive checks come in three steps. First, human trees must sit off-center in the conditional null in a direction consistent across folds. If they look like random trees, the program stops there.
Second, AutoSchA's deviation from humans in these statistics must predict which analyses experts flag. We test this against the flags from a blind re-evaluation, using the original 5-expert protocol, since per-analysis flags from the original study are not available.
Third, the two-parameter prior must beat the parser-only ranking and match or beat the learned conditional random field. A secondary experiment applies the same encoding to Bach chorale reductions produced by ProGress, to see whether its generated hierarchies inherit realistic shape statistics.
References
- A Combinatorial Proof of a Symmetry for a Refinement of the Narayana Numbers
- Tiered Trees and Theta Operators
- Theta operators, refined delta conjectures, and coinvariants
- A proof of the compositional Delta conjecture
- How Can Self-Attention Networks Recognize Dyck-n Languages?
- A New Dataset, Notation Software, and Representation for Computational Schenkerian Analysis
- AutoSchA: Automatic Hierarchical Music Representations via Multi-Relational Node Isolation
- Can LLMs "Reason" in Music? an Evaluation of LLMs' Capability of Music Understanding and Generation
- ProGress: Structured Music Generation via Graph Diffusion and Hierarchical Music Analysis