Dr. Abdelhak M Zoubir, Fellow of the IEEE and IEEE Distinguished Lecturer
Technische Universität Darmstadt, Germany
The use of more accurate models in signal processing applications such as communications, radar, sonar, biomedical engineering, speech and image processing and machine learning has become a fundamental requirement. With an improved accuracy the models have become more complex and inferential statistical signal processing required in parameter estimation and signal detection and classification, for example, has become intractable. The signal processing practitioner requires a simple but accurate method for assessing errors of estimates and answering inferential questions. Asymptotic approximations are useful only when enough data is available, which is not always possible due to time constraints, the nature of the signal or the measurement setting. In place of the formulae and tables of parametric and non parametric procedures based on complicated mathematics and asymptotic approximations, tools such as the Bootstrap have revolutionized statistics in the last decade and have become powerful for solving complex engineering problems. It is the method of an engineer's choice for solving inferential signal processing problems, such as signal detection, confidence limits estimation and model selection, to mention a few. In this talk, we first give a brief overview of and the basic principle underlying the bootstrap methodology. We then discuss bootstrap techniques for dependent data. Frequency domain bootstrap methods are discussed along with a combined time-domain and frequency-domain bootstrap technique for spectral analysis.