In this thesis, new precoding and equalization techniques for multicarrier systems were proposed and analyzed.First, the error performance of precoded filterbank multicarrier (FBMC) systems was analyzed. It was found out that this performance is highly sensitive to complete subchannel equalization. When there is residual intersymbol interference (ISI) stemming from imperfect subchannel equalization there is a loss of diversity; this loss can be prevented with the adoption of a number of subchannels large enough so that each subchannel suffers flat fading or with the utilization of a subchannel equalizer with sufficient length to compensate the subchannel frequency response.After that, an approximation for the signal to interference-plus-noise ratio (SINR) distribution of SC-FDE systems using linear MMSE equalization was proposed. This approximation uses the lognormal distribution with the smallest Kullback-Leibler distance to the true distribution, and was shown to be precise in the error performance sense; it serves as a system abstraction. With this abstraction, a precise method to obtain the analytical coded error performance of these systems was proposed.Finally, widely linear Tomlinson-Harashima precoders and equalizers (linear and decision-feedback) for SC-FDE systems were proposed. These precoders and equalizers have better error performance when compared to their strictly linear versions if signals coming from an improper constellation are transmitted. Their error performance when decision-feedback equalizers are used is less sensitive to the length of the feedback filter. When widely linear precoders are used, this error performance becomes less sensitive to channel estimation errors.

Dans cette thèse, de nouvelles techniques d'égalisation et de précodage pour des systèmes multiporteuses ont été proposées et analysées. D'abord, la performance d'erreur des systèmes multiporteuses à base de bancs de filtres (FBMC) précodées a été analysée. Il a été découvert que cette performance est très sensible à l'égalisation complète des sous-canaux. Quand il y a de l'interference inter-symbole residuel qui vient de l'égalisation imparfaite du sous-canaux, il y a une perte de diversité; cette diversité peut être recuperée avec l'adoption d'un nombre de sous-canaux assez grand pour que chaque sous-canaux subisse de l'évanouissement plat ou avec l'utilisation d'un égaliseur de sous-canaux avec une longueur assez suffisante pour compenser cette réponse en fréquence. Après, une approximation pour la distribution du rapport signal/bruit-plus-interfèrence (SINR) des systèmes SC-FDE qui utilisent égalisation MMSE linéaire a été proposée. Cette approximation utilise la distribution lognormal avec la plus petit distance de Kullback-Leibler vers la vraie distribution, et il s'est révélé qu'elle est precise dans la performance d'erreur; elle sert aussi comme une abstraction de cette système. Avec cette abstraction, une méthode précise pour obtenir la performance d'erreur analytique codée de ces systèmes a été proposée. Finalement, des précodeurs Tomlinson-Harashima (THP) et égaliseurs (linéaires et à retour de décision) largement linéaires pour des systèmes SC-FDE ont été proposés. Ces précodeurs et égaliseurs ont une performance d'erreur mieux quand comparés avec ses versions strictement linéaires si des signaux de constellations impropres sont transmises. Aussi, la performance d'erreur quand des égaliseurs à retour de décision sont utilisés est moins sensible au longueur du filtre de retour. Quand des précodeurs largement linéaires sont utilisés, cette performance devient moins sensible à des erreurs d'estimation du canaux. / In this thesis, new precoding and equalization techniques for multicarrier systems were proposed and analyzed.First, the error performance of precoded filterbank multicarrier (FBMC) systems was analyzed. It was found out that this performance is highly sensitive to complete subchannel equalization. When there is residual intersymbol interference (ISI) stemming from imperfect subchannel equalization there is a loss of diversity; this loss can be prevented with the adoption of a number of subchannels large enough so that each subchannel suffers flat fading or with the utilization of a subchannel equalizer with sufficient length to compensate the subchannel frequency response.After that, an approximation for the signal to interference-plus-noise ratio (SINR) distribution of SC-FDE systems using linear MMSE equalization was proposed. This approximation uses the lognormal distribution with the smallest Kullback-Leibler distance to the true distribution, and was shown to be precise in the error performance sense; it serves as a system abstraction. With this abstraction, a precise method to obtain the analytical coded error performance of these systems was proposed.Finally, widely linear Tomlinson-Harashima precoders and equalizers (linear and decision-feedback) for SC-FDE systems were proposed. These precoders and equalizers have better error performance when compared to their strictly linear versions if signals coming from an improper constellation are transmitted. Their error performance when decision-feedback equalizers are used is less sensitive to the length of the feedback filter. When widely linear precoders are used, this error performance becomes less sensitive to channel estimation errors.

Neste texto são propostos algoritmos de filtragem adaptativa de baixo custo computacional para o processamento de sinais lineares no sentido amplo e para beamforming. Novas técnicas de filtragem adaptativa com baixo custo computacional são desenvolvidas para o processamento de sinais lineares no sentido amplo, representados por números complexos ou por quaternions. Os algoritmos propostos evitam a redundância de estatísticas de segunda ordem na matriz de auto correlação, o que é obtido por meio da substituição do vetor de dados original por um vetor de dados real contendo as mesmas informações. Dessa forma, evitam-se muitas operações entre números complexos (ou entre quaternions), que são substituídas por operações entre reais e números complexos (ou entre reais e quaternions), de menor custo computacional. Análises na media e na variância para qualquer algoritmo de quaternions baseados na técnica least-mean squares (LMS) são desenvolvidas. Também é obtido o algoritmo de quaternions baseado no LMS e com vetor de entrada real de mais rápida convergência. Uma nova versão estável e de baixo custo computacional do algoritmo recursive least squares (RLS) amplamente linear também é desenvolvida neste texto. A técnica é modificada para usar o método do dichotomous coordinate descent (DCD), resultando em uma abordagem de custo computacional linear em relação ao comprimento N do vetor de entrada (enquanto o algoritmo original possui custo computacional quadrático em N). Para aplicações em beamforming, são desenvolvidas novas técnicas baseadas no algoritmo adaptive re-weighting homotopy. As novas técnicas são aplicadas para arrays em que o número de fontes é menor do que o número de sensores, tal que a matriz de auto correlação se torna mal-condicionada. O algoritmo DCD é usado para obter uma redução adicional do custo computacional. / In this text, low-cost adaptive filtering techniques are proposed for widely-linear processing and beamforming applications. New reduced-complexity versions of widely-linear adaptive filters are proposed for complex and quaternion processing. The low-cost techniques avoid redundant secondorder statistics in the autocorrelation matrix, which is obtained replacing the original widely-linear data vector by a real vector with the same information. Using this approach, many complex-complex (or quaternion-quaternion) operations are substituted by less costly real-complex (or real-quaternion) computations in the algorithms. An analysis in the mean and in the variance is performed for quaternion-based techniques, suitable for any quaternion least-mean squares (LMS) algorithm. The fastest-converging widely-linear quaternion LMS algorithm with real-valued input is obtained. For complex-valued processing, a low-cost and stable version of the widely-linear recursive least-squares (RLS) algorithm is also developed. The widely-linear RLS technique is modified to apply the dichotomous coordinate descent (DCD) method, which leads to an algorithm with computational complexity linear on the data vector length N (in opposition to the original WL technique, for which the complexity is quadratic in N). New complex-valued techniques based on the adaptive re-weighting homotopy algorithm are developed for beamforming. The algorithms are applied to sensor arrays in which the number of interferer sources is less than the number of sensors, so that the autocorrelation matrix is ill-conditioned. DCD iterations are applied to further reduce the computational complexity.

Neste texto são propostos algoritmos de filtragem adaptativa de baixo custo computacional para o processamento de sinais lineares no sentido amplo e para beamforming. Novas técnicas de filtragem adaptativa com baixo custo computacional são desenvolvidas para o processamento de sinais lineares no sentido amplo, representados por números complexos ou por quaternions. Os algoritmos propostos evitam a redundância de estatísticas de segunda ordem na matriz de auto correlação, o que é obtido por meio da substituição do vetor de dados original por um vetor de dados real contendo as mesmas informações. Dessa forma, evitam-se muitas operações entre números complexos (ou entre quaternions), que são substituídas por operações entre reais e números complexos (ou entre reais e quaternions), de menor custo computacional. Análises na media e na variância para qualquer algoritmo de quaternions baseados na técnica least-mean squares (LMS) são desenvolvidas. Também é obtido o algoritmo de quaternions baseado no LMS e com vetor de entrada real de mais rápida convergência. Uma nova versão estável e de baixo custo computacional do algoritmo recursive least squares (RLS) amplamente linear também é desenvolvida neste texto. A técnica é modificada para usar o método do dichotomous coordinate descent (DCD), resultando em uma abordagem de custo computacional linear em relação ao comprimento N do vetor de entrada (enquanto o algoritmo original possui custo computacional quadrático em N). Para aplicações em beamforming, são desenvolvidas novas técnicas baseadas no algoritmo adaptive re-weighting homotopy. As novas técnicas são aplicadas para arrays em que o número de fontes é menor do que o número de sensores, tal que a matriz de auto correlação se torna mal-condicionada. O algoritmo DCD é usado para obter uma redução adicional do custo computacional. / In this text, low-cost adaptive filtering techniques are proposed for widely-linear processing and beamforming applications. New reduced-complexity versions of widely-linear adaptive filters are proposed for complex and quaternion processing. The low-cost techniques avoid redundant secondorder statistics in the autocorrelation matrix, which is obtained replacing the original widely-linear data vector by a real vector with the same information. Using this approach, many complex-complex (or quaternion-quaternion) operations are substituted by less costly real-complex (or real-quaternion) computations in the algorithms. An analysis in the mean and in the variance is performed for quaternion-based techniques, suitable for any quaternion least-mean squares (LMS) algorithm. The fastest-converging widely-linear quaternion LMS algorithm with real-valued input is obtained. For complex-valued processing, a low-cost and stable version of the widely-linear recursive least-squares (RLS) algorithm is also developed. The widely-linear RLS technique is modified to apply the dichotomous coordinate descent (DCD) method, which leads to an algorithm with computational complexity linear on the data vector length N (in opposition to the original WL technique, for which the complexity is quadratic in N). New complex-valued techniques based on the adaptive re-weighting homotopy algorithm are developed for beamforming. The algorithms are applied to sensor arrays in which the number of interferer sources is less than the number of sensors, so that the autocorrelation matrix is ill-conditioned. DCD iterations are applied to further reduce the computational complexity.