Classical multidimensional scaling (CMDS) is a widely used method in manifold learning. It takes in a dissimilarity matrix and outputs a coordinate matrix based on a spectral decomposition. However, there are not yet any statistical results characterizing the performance ofCMDS under randomness, such as perturbation analysis when the objects are sampled from a probabilistic model. In this paper, we present such an analysis given that the objects are sampled from a suitable distribution. In particular, we show that the resulting embedding gives rise to a central limit theorem for noisy dissimilarity measurements, and provide compelling simulation and real data illustration of this CLT for CMDS.