In opportunistic spectrum access, where unlicensed secondary users may opportunistically communicate on idle spectral resources, reliable spectrum sensing is the essential technology to minimize harmful interference for the licensed users. A spectrum sensing algorithm is responsible for detecting whether a frequency band is currently used by the licensed primary system or not. It is therefore required to exhibit high detection performance even in low signal-to-noise ratio (SNR) regimes. The class of detectors operating on the eigenvalues of the sample covariance matrix is subsumed under the term eigenvalue-based spectrum sensing. It aims at exploiting correlations in the received signal over time or among multiple cooperating users in the presence of a licensee. Since the receiver noise is typically assumed to be a white random process which is uncorrelated among different receivers, the received signal samples should be free of correlations when the frequency band in question is vacant. Eigenvalue-based spectrum sensing is a prominent detection method since it requires very little knowledge about the signal characteristics of the primary system, while still exhibiting good detection performance. This thesis makes contributions to this field in three areas. Firstly, it explores the potential of reducing detection delays using results from the theory of quickest detection, which is a paradigm to minimize delays in detecting hypothesis changes. Large detection delays are harmful to both the primary and the secondary system in opportunistic spectrum access. This thesis therefore studies whether concepts from quickest detection may be combined with the strengths of eigenvalue-based spectrum sensing in the context of the well-known maximum-minimum eigenvalue (MME) detector.Secondly, performance limits of eigenvalue-based block detectors in the presence of practical model uncertainties are studied. In general, if knowledge about the system model is imperfect, detectors experience an SNR threshold below which reliable detection is impossible irrespective of the number of samples — the so-called SNR wall. In the context of eigenvalue-based spectrum sensing, two questions arise. Can it be shown that well-known detectors suffer from an SNR wall under practical model imperfections? Furthermore, can the location of the SNR threshold be characterized with respect to fundamental system parameters? This thesis answers these questions by investigating the effect of two practical model uncertainties: imperfect noise power calibration, and colored and correlated noise.Finally, this work advances the theoretical analysis of detectors with the help of random matrix theory. A theoretical analysis of the so-called maximum-minus-minimum eigenvalue (MMME) detector is performed in a dual user scenario. Considering that similar theoretical results were obtained for the MME detector, their performances in the presence of noise power uncertainty are compared on the basis of analytical findings.
展开▼