Abstract
We propose a decentralized online learning algorithm for distributed randominverse problems over network graphs with online measurements, and unifies thedistributed parameter estimation in Hilbert spaces and the least mean squareproblem in reproducing kernel Hilbert spaces (RKHS-LMS). We transform theconvergence of the algorithm into the asymptotic stability of a class ofinhomogeneous random difference equations in Hilbert spaces with$L_{2}$-bounded martingale difference terms and develop the $L_2$-asymptoticstability theory in Hilbert spaces. We show that if the network graph isconnected and the sequence of forward operators satisfies theinfinite-dimensional spatio-temporal persistence of excitation condition, thenthe estimates of all nodes are mean square and almost surely stronglyconsistent. Moreover, we propose a decentralized online learning algorithm inRKHS based on non-stationary online data streams, and prove that the algorithmis mean square and almost surely strongly consistent if the operators inducedby the random input data satisfy the infinite-dimensional spatio-temporalpersistence of excitation condition.