A novel hybrid stochastic gradient adaptation apparatus and method for calibrating the gain of an RF or non-RF digitally controlled oscillator (DCO). The adaptation algorithm determines a true stochastic gradient between a forcing function and its corresponding system measure to estimate the system parameters being adapted. A momentum term is generated and injected into the adaptation algorithm in order to stabilize the algorithm by adding inertia against any large transient variations in the input data. In the case of adaptation of DCO gain KDCO, the algorithm determines the stochastic gradient between time varying calibration or actual modulation data and the raw phase error accumulated in an all digital phase locked loop (ADPLL). Two filters preprocess the observable data to limit the bandwidth of the computed stochastic gradient providing a trade-off between sensitivity and settling time.