Robustness is one of the most important requirements when digital watermarking is applied. Different from the StirMark test and various simulation tests, a mutual information function is defined as a criterion to measure the robustness of watermarking algorithm. Taking the additive spread spectrum watermarking scheme and quantization index modulation (QIM) watermarking scheme as two examples, the calculation formulas of mutual information function are derived to evaluate the robustness of the algorithms. Numerical computation of mutual information is performed with change of watermark noise rate (WNR). In the experiment, spread spectrum watermarking is implemented in discrete cosine transform (DCT) and QIM watermarking is implemented in discrete wavelet transform (DWT). The statistic bit error rate (BER) is derived against Gaussian distribution noise and JPEG compression. Experiment results show that the evaluation conclusion of mutual information method is in accordance with the empirical BER. Mutual information can be selected as a cost function to predict the BER.