💻 computer science

Frequency Range of Fourier Phase Space as a Candidate Code for Musical Pitch Structure: Evidence from Artificial Neural Network Models of Iranian Classical Music

This study demonstrates that artificial neural networks trained on the micro-tonal pitch structure of Iranian classical music spontaneously develop hidden unit representations dominated by Fourier phase spaces, confirming that this frequency range serves as a universal candidate code for musical pitch structure beyond Western tonal systems.

Original authors: Yousef Aftabi, Hamid Farvaresh

Published 2026-11-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yousef Aftabi, Hamid Farvaresh

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Music is a universal human trait, found in every known society, yet the way our brains encode the building blocks of melody remains one of the great mysteries of cognitive science. We know that the brain processes sound through complex networks of neurons, but we do not know the specific language these neurons use to represent a musical note or a chord. Is the brain storing music as a simple list of notes, or does it use a more abstract, mathematical code? For decades, researchers have suspected that the brain might organize pitch using a system based on frequency, similar to how a prism separates light into a rainbow of colors. This idea, rooted in the mathematics of the Fourier transform, suggests that the brain breaks down musical structures into their underlying wave components. Until now, this theory has been tested almost exclusively using Western music, with its familiar twelve-note scale. But if this frequency-based code is a fundamental part of human musical cognition, it should appear in any musical system, even those with completely different rules and structures.

To test this, researchers at the University of Kurdistan turned to Iranian classical music, a tradition that operates on a much finer grid of pitch than Western music. While Western music divides an octave into twelve equal steps, Iranian classical music recognizes seventeen distinct pitch classes within the same space, including subtle quarter-tone shifts that have no direct equivalent in the standard Western scale. The team built artificial neural networks—computer programs designed to mimic the learning process of the human brain—and trained them to solve specific musical problems using this Iranian system. They asked the networks to identify musical intervals and specific four-note chord structures, known as Dangs, purely by analyzing the relationships between the notes. Crucially, the networks were never told to use any mathematical formulas or frequency analysis; they were simply given the task of learning to categorize the music and allowed to figure out the best way to do it on their own.

The results were striking. As the networks learned to solve these classification tasks, their internal connections spontaneously organized themselves into patterns that perfectly matched the mathematical structure of the Fourier phase space. This means that without any instruction, the computer models discovered that the most efficient way to represent these complex musical relationships was to map them onto a set of frequency-based waves. This finding confirms that the frequency-based code is not just a quirk of Western music theory but a general solution that learning systems can discover when faced with musical structure. However, the study also revealed a fascinating difference between the two musical worlds. Because the Iranian system uses a finer grid of pitches, the networks trained on it recruited a much wider range of frequencies to do their work. While the networks trained on Western music utilized a range of six distinct frequency bands, the networks trained on Iranian music utilized the full twelve bands permitted by their more detailed pitch grid.

This difference is not merely a matter of scale; it points to a deeper principle about how musical cultures might be encoded. The researchers found that about thirty-one percent of the units in the Iranian networks were using high-frequency bands that simply cannot exist in the Western twelve-tone system. These units were capturing nuances of the music that the Western system is mathematically incapable of representing. The study rigorously ruled out the possibility that this was a trick of the computer setup or a result of how the data was fed into the system. When the researchers scrambled the musical labels or randomized the input order, the pattern disappeared, proving that the networks only developed this specific code because it was the most effective solution for the musical task at hand.

The implications of this work extend beyond computer science. It suggests that while the fundamental code for representing music might be shared across human cultures—a frequency-based system that the brain naturally gravitates toward—the specific range of frequencies used might vary depending on the musical tradition. Just as a language uses the same basic sounds but arranges them differently to create distinct words, different musical cultures might be using different slices of the same universal frequency spectrum to capture their unique sonic identities. The study does not claim to have solved the mystery of human musical cognition, nor does it prove that human brains work exactly like these computer models. Instead, it offers a precise, testable hypothesis: that the diversity of the world's musical traditions might be explained by how different cultures utilize different ranges of a shared, underlying frequency code. By showing that this code emerges naturally in a system as different as Iranian classical music, the research provides a new lens through which to view the universal and the unique in human musical experience.

Technical Summary: Frequency Range of Fourier Phase Space as a Candidate Code for Musical Pitch Structure

Problem Statement
The central question addressed is how musical pitch structure is represented across different cultures. While previous research using Artificial Neural Networks (ANNs) demonstrated that networks trained on Western tonal classification problems spontaneously develop connection weights correlated with discrete Fourier phase spaces of the 12-tone system, it remains unclear whether this representational format generalizes to structurally distinct musical systems. Specifically, it is unknown if Fourier phase space coding is a universal candidate code for musical cognition or merely an artifact of the Western 12-tone equal temperament (12-TET) structure. The study tests this by examining Iranian classical music, a microtonal system with 17 pitch classes per octave based on a 24-tone quarter-tone lattice, which is neither a subset nor a transposition of the Western chromatic scale.

Methodology
The study employed a comparative computational modeling approach using Gaussian value-unit networks trained on classification tasks derived from Iranian and Western music theories.

Stimuli and Tasks:
- Iranian System: Defined on a 24-tone lattice (50-cent steps) containing 17 active pitch classes. The study defined nine classification tasks: two based on melodic intervals (153 unordered pitch-class pairs) and seven based on Dangs (tetrachordal scale segments). Tasks varied in how dissonant combinations were handled (collapsed or excluded).
- Western System: Constructed using an identical rule set but substituting the 12 pitch classes of 12-TET. This yielded two interval tasks (78 unordered pairs). No Dang tasks were created for the Western system due to the lack of a direct theoretical counterpart.
Network Architecture:
- Networks utilized Gaussian ("value unit") activation functions.
- Input layers were binary vectors representing the presence (1) or absence (0) of specific pitch classes (17 dimensions for Iranian, 12 for Western).
- Hidden layers were sized to be "minimal-sufficient" (7 units for interval tasks; 4–12 units for Dang tasks).
- Twenty-five networks were trained per task using the generalized delta rule until perfect classification was achieved.
Fourier Phase Space Construction:
- Phase spaces were defined over the respective pitch-class lattices. For the 24-tone Iranian lattice, this yielded 12 distinct phase spaces (frequencies $k=1$ to $12$). For the 12-tone Western lattice, 6 spaces ( $k=1$ to $6$).
- Each phase space $k$ corresponds to the span of $\{\cos(2\pi k g_n/G), \sin(2\pi k g_n/G)\}$ , where $g_n$ are the lattice indices and $G$ is the lattice size.
Analysis and Controls:
- Fit Analysis: Hidden unit connection weights were regressed against each phase space. The best-fitting phase space was identified via the multiple correlation coefficient ( $R_k$ ).
- Statistical Controls: Significance was assessed using a permutation null (randomizing weights) with Benjamini–Hochberg false-discovery-rate (FDR) control. Non-Fourier baselines included pitch-height ramps and random low-pass ceilings.
- Task-Driven Controls:
  - Label Permutation: Networks were retrained on randomly permuted output labels to test if Fourier structure persists without task structure.
  - Input-Coordinate Permutation: The mapping of pitch classes to input units was scrambled before training to test if the structure relied on input adjacency rather than pitch identity.
- Task-Matched Comparison: To isolate the effect of lattice cardinality from task type, a specific comparison was restricted to the interval tasks, which were defined identically in both systems.

Key Results

Dominance of Single Phase Spaces: Across all tasks and systems, hidden units were reliably dominated by a single Fourier phase space. This dominance exceeded chance levels (FDR-controlled) and surpassed non-Fourier baselines. The spectral-concentration index confirmed that units were not split across multiple frequencies but focused on a single dominant frequency.
Frequency Range Recruitment:
- Western Networks: Recruited the full range of their lattice's Nyquist limit (frequencies 1–6).
- Iranian Networks: Recruited the full range of their lattice's Nyquist limit (frequencies 1–12).
The High-Frequency Distinction:
- The Nyquist frequency ( $k=12$ ) on the 24-tone lattice is mathematically identical to a binary partition between chromatic and microtonal (Koron/Sori) pitches. Consequently, units fitting this frequency may act as simple feature detectors rather than periodic codes.
- Excluding $k=12$ : In the band of frequencies 7–11 (which a 12-tone lattice cannot represent), 31% of significant Iranian units (86 out of 279 in the task-matched interval comparison) found their best fit. In contrast, 0% of Western units occupied this band.
- This difference persisted even when controlling for task type (comparing only interval tasks) and was not driven by the Dang tasks.
Task-Driven Nature:
- When networks were trained on random labels, the Fourier fits collapsed to chance levels, confirming the structure is driven by the musical task, not the input coding or model flexibility.
- Scrambling input coordinates did not degrade the fits, indicating the structure depends on pitch-class identity and task structure, not the spatial arrangement of input units.

Significance and Claims
The paper claims to extend the finding that Fourier phase spaces are a spontaneously discovered representational format in ANNs from Western tonal music to a structurally distinct microtonal system. The primary contribution is the demonstration that networks trained on Iranian classical music recruit higher-frequency phase spaces (7–11) that are mathematically inaccessible to networks trained on 12-TET.

The authors propose a modest, falsifiable hypothesis: musical cultures may not possess qualitatively different representational codes, but rather share a single Fourier-based code where the effective frequency range is determined by the cardinality and structure of the underlying pitch-class lattice. The study suggests that while the format (Fourier phase space) may be universal, the range of frequencies utilized is culture-specific and constrained by the lattice's Nyquist limit.

Limitations Acknowledged by the Authors

Theoretical vs. Performance: The model relies on the modified Vaziri temperament (a theoretical 24-TET system) rather than the flexible, context-dependent intonation of actual Persian performance practice.
Nyquist Ambiguity: The exclusion of the $k=12$ frequency is necessary because it degenerates into a binary chromatic/microtonal detector on this specific lattice, making it impossible to distinguish from simple feature detection.
Scope of Generality: The study uses only two musical systems. The authors note that Arabic and Turkish traditions are not independent enough (sharing historical roots) to serve as a definitive test of cross-cultural generality. They suggest Indian and Chinese systems as future candidates but note the difficulty in defining their pitch lattices quantitatively without imposing external frameworks.
Architecture: The results are based on small, minimal-sufficient networks. While robust across layer sizes, the authors do not claim that larger or different architectures (e.g., deep learning) would necessarily converge on the same code, though they cite convergent evidence from variational autoencoders.

The paper concludes that Fourier phase space is a candidate code for musical cognition beyond Western tonality, with the specific frequency range recruited serving as a potential marker of the underlying pitch-class system's structural properties.

More like this