The phase estimation of sinusoidal signal in noise environments is extensively adopted in the field of radar, navigation and DOA estimation, etc. A sinusoidal signal phase estimation method---Singular Value Decomposition (SVD), based on cross-high-order cumulant, is proposed. The signal and noise subspace are obtained using SVD of cross-high-order cumulant matrix. The signal subspace is the optimum solution for signal detection and noise suppression. The high-order cumulant matrix of signal is a conjugated symmetric matrix, its left singular vector is identical with the right one, their amplitudes and frequencies are also the same. The cross-high-order cumulant matrix of sinusoidal signals with different phases is a unconjugated symmetric matrix, its left and right singular vector are different because of the phase difference existing among the harmonic signals. A very important theorem is demonstrated in this article, i.e. the phase angle of the left and the right singular vector inner-product of harmonic signal cross-high-order cumulant matrix is equal to the phase difference of sinusoidal signals. Based on this theorem, the method of SVD for phase estimation of sinusoidal signal is deduced. The availability of the method is verified by simulation.