Summary: An individual variable step LMS adaptive algorithm which varies its step size adaptively to the correlation between the input signal and prediction error is proposed. A physical explanation of the step size is presented. The condition of convergence, a matrix difference equation for the second moment of the weight and the misadjustment of the IVLMS algorithm are derived. The noise power of the adaptive weights and the tracking capability of the new algorithm are evidently superior to the conventional LMS algorithm both in stationary and nonstationary inputs. It is shown that the concept of individual variable step size is also available to the normalized LMS and block LMS adaptive algorithms.