Spectral Clustering on Subspace for Parameter Estimation of Jump Linear Models

Abstract: The problem of estimating parameters of a deterministic jump or piecewise linear model is considered. A subspace technique referred to as spectral clustering on subspace (SCS) algorithm is proposed to estimate a set of linear model parameters, the model input, and the set of switching epochs. The SCS algorithm exploits a block diagonal structure of the system input subspace, which partitions the observation space into separate subspaces, each corresponding to one and only one linear submodel. A spectral clustering technique is used to label the noisy observations for each submodel, which generates estimates of switching time epoches. A total least squares technique is used to estimate model parameters and the model input. It is shown that, in the absence of observation noise, the SCS algorithm provides exact parameter identification. At high signal to noise ratios, SCS attains a clairvoyant Cram\'{e}r-Rao bound computed by assuming the labeling of observation samples is perfect.