This dissertation presents techniques to improve the performance of both coherent as well as non-coherent wireless communication systems via optimizing symbol timing, frequency spacing and by making efficient SNR estimations. We show that some of the design choices made in traditional systems are not optimal and demonstrate the gains that may be achieved by making unconventional, but judicious, choices for these parameters. We start in the area of coherent multi-antenna communications where we introduce an offset between the symbol boundaries of the transmitted waveforms from the different antennas and show that this improves performance in comparison to the traditional symbol aligned transmission. For this modified system, we derive various optimal receivers such as maximum likelihood (ML), best linear unbiased estimator (BLUE), minimum mean squared error (MMSE), and zero forcing (ZF) receivers and show that they outperform the equivalent receiver for the system with aligned symbol boundaries. In some system configurations, the performance gain is close to 2dB. The design methodology for a new symbol pulse shape that increases the performance even more is also presented. Next, we extend the study of SNR estimation from the previously published results of a data aided (DA) single antenna system to the non-data aided (NDA) model and also to systems with multiple antennas (MIMO). In both these cases, we have derived the Cramér- Rao lower bound (CRLB) as well as ML estimators that achieve or perform very close to the CRLB. For MIMO systems we define the SNR and then derive the CRLB and the ML estimators for both the DA as well as the NDA data model. We show that previously published results for single antenna systems are a special case of our general solution. The proposed SNR estimation techniques are demonstrated in a patented algorithm to detect the onset of non-linearity in a remote transmitter by dithering the power of transmitted bursts and estimating the difference in the received SNR. For non-coherent systems we show that the performance of multi-tone M-ary frequency shift keying (MT-MFSK) modulation may be significantly improved if, instead of the usual choice of mutually orthogonal tones, non-orthogonal tones are used. In some system configurations, the proposed system can lead to a 4-fold increase in system capacity. The channel capacity, as well as the performance gains of systems using practical receivers such as ML, least squared (LS) error, and compressed sensing (CS) are demonstrated for both flat and frequency selective channels. Many more choices of spectral efficiency are achievable by the non-orthogonal system, thus enabling the system to adapt to changing link SNR and send data at the optimum spectral efficiency. In order to make this practical, we derive the CRLB and ML estimators for SNR estimation for non-orthogonal MT-MFSK in both the DA as well as the NDA data model