We introduce probabilistic neural networks that describe unsupervised synchronous learning on an atomic Hardy space and space of bounded real analytic functions, respectively. For a stationary ergodic vector process, we prove that the probabilistic neural network yields a unique collection of neurons in global optimization without initialization and back-propagation. During learning, we show that all neurons communicate with each other, in the sense of linear combinations, until the learning is finished. Also, we give convergence results for the stability of neurons, estimation methods, and topological statistics to appreciate unsupervised estimation of a probabilistic neural network. As application, we attach numerical experiments on samples drawn by a standing wave.