Vastola & Rajan's Markov model reveals hidden timescales in animal behavior

In a new paper accepted to the Cognitive Computational Neuroscience conference (CCN 2026), John Vastola and Kanaka Rajan tackle a fundamental theoretical question: if spontaneous animal behavior can be accurately modeled as a sequence of neurally defined 'syllables' using Markov models, what do these models really tell us? The authors use hierarchical Markov models to quantitatively formulate questions about the sequence-like nature, irreversibility, predictability, and effective dimensionality of behavior. A key insight is that the eigenvalues and eigenvectors of model-associated matrices furnish interpretable time scales—e.g., how fast certain behavioral patterns decay or persist—and reveal how behavior modifications occur on those time scales. The analysis uses both toy examples and Markov models fit to real behavioral data to illustrate these points, providing a rigorous mathematical foundation for quantifying complex animal behavior.

The work has significant implications for neuroscience and quantitative biology. By connecting the eigenvalues of transition matrices to physically meaningful timescales, researchers can now decompose behavioral sequences into components that evolve at different rates, offering a principled way to link neural activity patterns to behavioral dynamics. The paper also clarifies the theoretical meaning of progress in quantifying behavior, ensuring that as toolkits for behavioral analysis mature, they rest on solid mathematical ground. This approach could eventually help scientists understand how the brain generates structured, reversible and irreversible behavioral sequences, and how to predict future actions from past ones. For the AI and neuroscience communities, it bridges high-dimensional neural data with low-dimensional behavioral descriptions.