Generating 1/f Noise Sequences as Constraint Satisfaction: The Voss Constraint

François Pachet, Pierre Roy, Alexandre Papadopoulos, Jason SakellariouIJCAI 2015

Introduction

Natural and human-made phenomena—ranging from the flooding of the Nile to the rhythmic structures of classical music—frequently exhibit 1/f noise, also known as pink noise. Unlike white noise, which is purely random, 1/f noise possesses a "memory" where power spectral density is inversely proportional to frequency. This property makes sequences feel organic and structured to human observers.

While 1/f noise is aesthetically pleasing, existing generation methods are typically "black boxes." For composers or designers who need a sequence to follow 1/f fluctuations while adhering to strict rules (like musical meter, specific keys, or harmony), traditional stochastic algorithms fall short. This paper introduces the Voss Constraint, a novel framework that formalizes 1/f generation as a Constraint Satisfaction Problem (CSP), allowing researchers to inject hard logical rules into natural-sounding sequences.

Methodology: From Stochastic to Deterministic

The authors base their approach on the Voss Algorithm, where a sequence is generated by summing several "dice" that are re-tossed at different frequencies based on a binary counter. To make this compatible with modern solvers, the paper translates this process into an efficient CSP model.

Efficient Ternary Filtering

Instead of a naive model that creates slow, complex n-ary constraints, the authors model the system as a binary tree of ternary sum constraints. By introducing intermediate "partial sum" variables, the problem is transformed into a Berge-acyclic graph (a tree). This structure allows the solver to achieve domain consistency—pruning all impossible values—with high efficiency ($O(N \cdot d^2)$ complexity).

The hierarchical tree structure of the Voss Constraint, showing how partial sums (S) are derived from dice tosses (D).

Evaluation and Results

The method was evaluated using synthetic sequences and musical Markov models trained on the works of composer Michel Legrand. The results demonstrate that the Voss constraint successfully biases sequences toward natural distributions without sacrificing the ability to satisfy other arbitrary constraints.

Spectral Accuracy and Efficiency

Sequences generated with the Voss constraint alone showed a spectral slope ($\alpha$) of approximately 1.03, a near-perfect match for ideal 1/f noise ($\alpha \approx 1$). Even when combined with a Global Cardinality Constraint (GCC)—which usually produces white noise ($\alpha \approx 0.04$)—the system maintained a slope of 1.05.

Spectral analysis of a generated sequence showing a slope of α=0.93, closely approximating ideal 1/f noise.

Performance-wise, the tree-based filtering proved highly effective. For a sequence of 512 notes, the solver found solutions in approximately 0.07 seconds with zero backtracks. Qualitatively, the resulting melodies avoided the aimless feel of pure randomness, sounding more structured and "human."

Conclusion and Future Work

The Voss Constraint demonstrates that well-known random processes can be "unrolled" into deterministic logical relations. By capturing the hierarchical nature of 1/f noise in a tree-based CSP, the authors ensure that global spectral properties are maintained even under local constraints. Future research may explore varying the update patterns to produce other power-law distributions, such as Brownian noise ($\alpha \approx 2$), or investigating different mappings of integers to musical parameters to further refine the aesthetic output.

Summary updated

BibTeX
@inproceedings{pachet2015generating,
  author = {François Pachet and Pierre Roy and Alexandre Papadopoulos and Jason Sakellariou},
  title = {{Generating 1/f Noise Sequences as Constraint Satisfaction: The Voss Constraint}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2015},
}

Loading related content...

Figure