Stochastic Stability and Statistical Identifiability of Hierarchical Bayesian Inference in Computational Psychiatry
Year: 2026
Type: article
Abstract: Computational psychiatry seeks to connect neurobiological processes, latent cognitive computations, and observable behaviour through generative mathematical models. Adams, Huys, and Roiser (2016) developed an influential conceptual framework connecting hierarchical Bayesian inference, predictive coding, reinforcement learning, dopaminergic signalling, and psychiatric symptoms. However, a unified stochastic treatment of the interaction between hierarchical inference and dopaminergic dynamics, together with rigorous stability and statistical identifiability results, remains an important mathematical direction. We develop a coupled stochastic state-space framework in which latent hierarchical beliefs, prefrontal precision, tonic dopaminergic activity, and phasic dopaminergic activity evolve jointly. The model is formulated as a multidimensional stochastic differential equation with bounded nonlinear coupling and linear dissipative regulation. Our first contribution is a stochastic stability analysis. Under explicit contraction conditions on the coupling matrix, we establish global existence and pathwise uniqueness, exponential contraction of trajectories, existence and uniqueness of an invariant probability measure, and quantitative convergence in Wasserstein distance. We also derive Lyapunov drift bounds and characterise the equilibrium structure of the associated deterministic system.Our second contribution concerns computational regime transitions. We derive a sufficient parameter region in which the deterministic system possesses a unique globally asymptotically stable equilibrium. We then formulate precise Jacobian conditions for detecting loss of stability, saddle-node bifurcations, and Hopf bifurcations outside the contraction region. Our third contribution concerns statistical identifiability. We demonstrate an explicit observational equivalence between prior precision and sensory precision when only posterior means are measured. We show how additional uncertainty measurements and controlled stimulus manipulations can resolve this ambiguity.Finally, under an inverse-Lipschitz identifiability condition and a sub-Gaussian observation model, we derive a finite-sample parameter estimation guarantee.The resulting framework provides a mathematical bridge between stochastic cognitive dynamics, neurocomputational mechanisms, and experimentally distinguishable hypotheses. The model is phenomenological and is not presented as an established biological mechanism. Its neurobiological interpretation requires independent experimental validation.
Source: Figshare
Authors: Rajit Chatterjea, Philip Youn
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7233590939
FWCI:
Cites:
Cited by:
Related to:
