This paper proposes Joint Spatial Division and Multiplexing (JSDM) as a method for multiuser MIMO downlink that leverages the structure of channel correlation to enable large antenna arrays at the base station while requiring reduced-dimensional Channel State Information at the Transmitter (CSIT). JSDM achieves this by combining a pre-beamforming matrix, which depends on channel second-order statistics, with a classical multiuser precoder that uses instantaneous knowledge of reduced-dimensional "effective" channels. The paper proves that JSDM incurs no loss of optimality compared to full CSIT in certain conditions. For linear arrays, this condition is closely approached when the number of antennas is large. The paper also extends JSDM to two-dimensional antenna arrays with 3D beamforming, including multiple beams in the elevation angle direction. It provides guidelines for pre-beamforming optimization and calculates system spectral efficiency under proportional fairness and max-min fairness criteria, showing attractive performance. Numerical results are obtained using a deterministic equivalent approximation, which avoids lengthy Monte Carlo simulations and provides accurate results for realistic antenna and user numbers. The paper also discusses the performance of JSDM with joint group processing (JGP) and per-group processing (PGP), showing that JSDM achieves the same sum capacity as full CSIT in certain conditions. The paper also discusses the performance of JSDM with approximate block diagonalization (BD) and provides numerical examples comparing JSDM with deterministic equivalents and Monte Carlo simulations. The paper concludes that JSDM is a promising approach for large MIMO systems, particularly for Frequency Division Duplexing (FDD) systems where channel reciprocity cannot be exploited.This paper proposes Joint Spatial Division and Multiplexing (JSDM) as a method for multiuser MIMO downlink that leverages the structure of channel correlation to enable large antenna arrays at the base station while requiring reduced-dimensional Channel State Information at the Transmitter (CSIT). JSDM achieves this by combining a pre-beamforming matrix, which depends on channel second-order statistics, with a classical multiuser precoder that uses instantaneous knowledge of reduced-dimensional "effective" channels. The paper proves that JSDM incurs no loss of optimality compared to full CSIT in certain conditions. For linear arrays, this condition is closely approached when the number of antennas is large. The paper also extends JSDM to two-dimensional antenna arrays with 3D beamforming, including multiple beams in the elevation angle direction. It provides guidelines for pre-beamforming optimization and calculates system spectral efficiency under proportional fairness and max-min fairness criteria, showing attractive performance. Numerical results are obtained using a deterministic equivalent approximation, which avoids lengthy Monte Carlo simulations and provides accurate results for realistic antenna and user numbers. The paper also discusses the performance of JSDM with joint group processing (JGP) and per-group processing (PGP), showing that JSDM achieves the same sum capacity as full CSIT in certain conditions. The paper also discusses the performance of JSDM with approximate block diagonalization (BD) and provides numerical examples comparing JSDM with deterministic equivalents and Monte Carlo simulations. The paper concludes that JSDM is a promising approach for large MIMO systems, particularly for Frequency Division Duplexing (FDD) systems where channel reciprocity cannot be exploited.