By Stéphane Rivaud, François Pachet, and Pierre Roy.
Journées Francophones sur les Réseaux Bayésiens et les Modèles Graphiques Probabilistes (JFRB 2016), Clermont-Ferrand, France, June 2016.
About the paper
Repeating a musical pattern can require notes at distant positions to be equal. This paper studies what happens when such equality constraints are imposed on sequences sampled from a Markov model.
Abstract summary
The authors establish a general hardness result for exact sampling under binary equality constraints. They also show experimentally that adding constraint filtering to a random walk can produce valid sequences while changing their probabilities substantially. Specific configurations of equalities remain tractable, and the paper illustrates one with chord-sequence generation.
Related work
See Markov Constraints for the broader research. Exact Sampling for Regular and Markov Constraints with Belief Propagation describes exact sampling for a class of regular constraints. Hidden Biases in Conditioning Autoregressive Models examines related questions about preserving probabilities when conditioning a generator.