A new framework is presented for the analysis of the performance of detection methods, such as AIC and MDL, which are based on the eigenvalues of the covariance matrix. It is shown that theoretical analysis of the probabilities of overestimation and underestimation can be much more conveniently carried out via a proposed, particularly simple, sequence of statistics. Also the breakdown of these detection methods in the presence of model nonidealities is explored by theory, simulations and experimentation with real array data. For example, theoretical arguments are given t o demonstrate the high degree of sensitivity of the detectors to unknown deviations of the noise from whiteness.
|Original language||English (US)|
|Title of host publication||IEEE International Conference on Acoustics, Speech, and Signal Processing, ICASSP 1993|
|Publisher||Institute of Electrical and Electronics Engineers Inc.|
|Number of pages||4|
|State||Published - 1993|
|Event||1993 IEEE International Conference on Acoustics, Speech, and Signal Processing, ICASSP 1993 - Minneapolis, United States|
Duration: Apr 27 1993 → Apr 30 1993
|Name||ICASSP, IEEE International Conference on Acoustics, Speech and Signal Processing - Proceedings|
|Conference||1993 IEEE International Conference on Acoustics, Speech, and Signal Processing, ICASSP 1993|
|Period||4/27/93 → 4/30/93|
Bibliographical noteFunding Information:
Many methods have been suggested for the detection of the number of sources in array processing. The most commonly used detectors are formulated in terms of the eigenvalues of the sample covariance matrix. Typical examples of this kind of eigendecompsition-based detectors are AIC and MDL [l] . These methods are easy to implement and as long as the data comes from the idealized model, MDL gives a strongly consistent estimation of the number of sources independent of the signal-noise-ratio(SNR) . Unfortunately, AIC and MDL do not make effective use of the calibration procedure used in an array , Hence in practice, the estimated number of sources by these detectors may be far from the true number. In very general terms, the deviation of the received signal statistics from the idealized model based on the assumption of Gaussian white noise, causes the failure of these methods. Many nonidealities may be involved in practical arrays. At this point we consider the general design of This work was supported in parts by the National Science Foundation under Grants MIP-8813204 and MIP-9202081 and by the SDIO/IST program managed by the Office of Naval Research under Contract N00014-86-K-0410.
© 1993 IEEE