A cosine-function-based approximation method is proposed to reconstruct atrial fibrillatory wave in QRST intervals for estimating dominant frequency of atrial fibrillation (AF) using single-lead ECG. It is hypothesized that atrial fibrillatory waves in two adjacent QRST intervals can be approached by a cosine function and direct component. As harmonics of dominant frequency are neglected, the determination of harmonics' number and Tikhonov regularization are avoided compared with previous methods. Initially, we build a learning data set composed of at most 3 TQ intervals neighboring to the two QRST intervals. Next, the optimum frequency of cosine is determined in an enumeration way using the learning data set, and the corresponding amplitude, phase, and direct component are estimated in a least square method. Finally, the atrial fibrillatory waves in the two adjacent QRST intervals are calculated based on the cosine function and direct component. This method is simple and convenient, and the atrial fibrillatory waves in ectopic QRST intervals can also be easily estimated. Through evaluation with artificial and clinical AF ECG, we proved that frequency analysis of atrial fibrillation based on this method can obtain a dominant frequency profile with less standard deviation than that based on ABS method.