The inverse compositional Gauss-Newton (IC-GN) algorithm and the forward additive Newton-Raphson (FA-NR) algorithm are two mainstream iterative sub-pixel registration algorithms in digital image correlation. This study compares the accuracy and convergence ability of the two algorithms by theoretical analysis and numerical experiments in the speckle images that have been contaminated with artificial Gaussian noise. Based on the derived error model, the systematic errors of the two algorithms are dominated by interpolation-induced error and are insensitive to noise. The random errors are proportional to the noise level. The noise also reduces the convergence radius and rate in the two algorithms. The two algorithms demonstrate equal noise resistance due to their mathematical equivalence. These conclusions are well supported by the experimental study. The recently reported vulnerability of the FA-NR algorithm to noise is not associated with the inherent flaw of the algorithm but with its implementation. If an inappropriate method is employed to estimate the gradients at sub-pixel locations in the FA-NR algorithm, abnormally large errors may be induced. This problem can be eliminated using the method that is proposed in this study, which has an insignificant extra-computation cost.