Adaptive filter-bank tree for power spectrum estimation

We propose a new nonparametric method for power spectrum estimation of a real random process which can be used in situations where the traditional periodogram estimate is the preferred choice over the relatively modern model based approaches, such as the autoregressive (AR), moving average (MA) and other related methods. The proposed technique uses an adaptively grown binary filter bank tree for estimating the spectrum, which results in a trade-off between subband spectral resolution and the variance of the estimate, a feature intuitively very satisfying and absent in the traditional schemes. This is especially valuable in cases where the target spectrum has a combination of relatively flat and non-flat subbands, which intrinsically demand different spectral resolutions. Experimental results are presented to illustrate and support the proposed scheme.