Deriving Cramer-Rao Bounds and Maximum Likelihood Estimators for Traffic Matrix Inference

摘要

Traffic matrix estimation has caught numerous attentions these days due to its importance on network management tasks such as traffic engineering and capacity planning for Internet Service Providers (ISP). Various estimation models and methods have been proposed to estimate the traffic matrix. However, it is difficult to compare these methods since they adopt different model assumptions. Currently most evaluations are based on some particular realization of data. We propose to use the (Bayesian) Cramer-Rao Bound (CRB) as a benchmark on these estimators. We also derive the maximum likelihood estimator (MLE) for certain models. With coupled mean and variance, our simulations show that the least squares (LS) estimator reaches the CRB asymptotically, while the MLEs are difficult to calculate when the dimension is high.