Search

All-digital architectures enable taking full advantage of the large number of antennas that can be integrated into mmWave transceivers, with fully flexible beamforming that enables the number of simultaneous users K sharing the band to scale with the number of antennas N. The small carrier wavelength in these bands allows realization of antenna arrays with a large number of elements with a relatively small footprint, opening a path to truly massive Multiple Input Multiple Output (MIMO) systems. However, two key bottlenecks to realizing this potential are the cost/power consumption of Radio Frequency (RF) frontends at high carrier frequencies and the high complexity incurred in the digital baseband processing due to the large number of antennas. In this dissertation, we develop approaches for addressing these bottlenecks by adapting signal processing architectures and algorithms to hardware design considerations, while taking advantage of the unique characteristics of the mmWave band. We first develop an analytical model for the impact of nonlinearities such as RF amplifiers and Analog-to-Digital Converters (ADCs) on the performance of a mmWave massive MIMO uplink. We illustrate the utility of this model in providing specific guidelines for hardware design based on desired system-level perfor- mance. For example, the framework allows specification of the desired 1 dB compression point for RF amplifiers and the desired number of bits of ADC precision in order for the system outage at a target bit error rate to be below 5%. These hardware design prescriptions depend on coarse system-level parameters such as the number of antennas N,the number of simultaneous users K, and the maximum and minimum link distances to be supported. An important conclusion from the analytical framework is that hardware specifications can be substantially relaxed by reducing the load factor, defined as the ratio β = K/N. We then consider the problem of scaling digital signal processing in this regime. For a relatively small number of antennas and users, the Linear Minimum Mean Squared Error (MMSE) approach is a standard technique for handling multiuser interference at reasonable complexity. However, for fixed load factor β, the complexity of computing the LMMSE detector, as well as the complexity of using it for demodulation, grow polynomially with the number of antennas. We propose complexity reduction techniques that substantially improve scaling by taking advantage of the spatial sparsity of the mmWave channel. Specifically, we use a spatial Discrete Fourier transform (DFT) across antennas to create N discrete beams, transforming from antenna space to beamspace. We show that each user’s energy is concentrated in a small number of DFT bins in beamspace. Assuming ideal single path channels, we show that each user can be demodulated reliably using a local LMMSE detector which employs a beamspace window whose size does not scale with the number of antennas. The local LMMSE detector approaches the performance of standard LMMSE detection at substantially reduced complexity, and these performance-complexity tradeoffs become more favorable at lower system load factor β. For larger load factors, the beamspace window required by the local LMMSE detector increases, but we show that it is possible to scale well in such regimes by adding a layer of nonlinear interference cancellation on top of the local LMMSE receiver. Next, drawing on the duality between uplink linear multiuser detection and downlink linear precoding, we demonstrate the efficacy of beamspace techniques for linear precoding on the downlink, in order to reduce the interference seen by a given user due to signals directed at other users by the base station. Finally, we address the problem of simultaneous scaling of bandwidth and number of antennas. As bandwidth and hence symbol rate increase, the signal from a given user impinging on a large antenna array incurs a multi-symbol delay spread across the array, which smears the spatial frequency for each user across the band. We introduce a novel technique that combines DFTs in the spatial and time domains, together with an interpolation technique that limits the dispersion of spatial frequency across the band. We show that this results in significantly reduced complexity in computing LMMSE weights for uplink multiuser detection.

Multiple Input Multiple Output with Orthogonal Frequency Division Multiplexing (MIMOOFDM) has been widely adopted as one of the most promising air interface solutions for future broadband wireless communication systems due to its high rate transmission capability and robustness against multipath fading. However, these MIMO-OFDM advantages cannot be achieved unless the channel state information (CSI) can be obtained accurately and promptly at the receiver to assist coherent detection of data symbols. Channel estimation and training sequence design are, therefore, still open challenges of great interest. In this work, we investigate the Linear Minimum Mean Square Error (LMMSE) channel estimation and design nearly optimal training sequences for MIMO-OFDM systems in spatially correlated fading. We, first, review the LMMSE channel estimation model for MIMO-OFDM in spatially correlated fading channels. We, then, derive a tight theoretical lower bound of the channel estimation Mean Square Error (MSE). By exploiting the information on channel correlation matrices which is available at the transmitter, we design a practical and nearly optimal training sequence for MIMO-OFDM systems . The optimal transmit power allocation for training sequences is found using the Iterative Bisection Procedure (IBP). We also propose an approximate transmit power allocation algorithm which is computationally more efficient than the IBP while maintaining a similar MSE performance. The proposed training sequence design method is also applied to MIMO-OFDM systems where Cyclic Prefixing OFDM (CP-OFDM) is replaced by Zero Padding OFDM - OverLap-Add method (ZP-OFDM-OLA). The simulation results show that the performance of the proposed training sequence is superior to that of all existing training sequences and almost achieves the MSE theoretical lower bound.