For overcoming the shortage of existing technique, put forward according to inside order to punish the function and the close together and related data segment of the frequency offset estimation. Make use of the variance of the frequency offset estimation to be partial to estimate of namely of the variance with believe the signal-to-noise ratio with the related data segment the relation type of the length. A selection criterion with the metric function is proposed. This proposed criterion utilizes internal penalty function and gradient iteration to maximize the metric function to reduce the variance. Under the conditions of adaptive signal-to-noise ratio changes, it is to realize the control of the autocorrelation complexity and improve the frequency offset estimation of the anti-noise and accuracy.