How can we use existing cellular infrastructure to locate users?

We are developing AI/ML-based positioning algorithms for realistic urban environments, using NVIDIA Sionna to generate synthetic channel data from site-specific wireless digital twins. Alongside the algorithmic work, we are interested in the fundamental estimation-theoretic limits on what any positioning system can achieve in these settings. The primary focus is FR3 bands. This work is a collaboration with Prof. Tingjun Chen's group at Duke University.

As arrays get larger, the computational complexity gets impossibly expensive. Can we throw most of it away and still get the same performance?

Larger antenna arrays and wider bandwidths promise higher data rates, but at the cost of prohibitive hardware and signal-processing complexity. Beamspace dimensionality reduction offers a way out: transforming the received signal into a spatial-frequency representation concentrates channel energy into a few beams, letting the receiver operate in dramatically reduced dimension. On the fundamentals side, we characterize information-theoretic limits and develop a geometric understanding of interference geometry, validated on measured 28 GHz channels. On the design side, we develop hardware-efficient architectures, including algorithms for severely quantized receivers and tiled layouts that ease packaging constraints. Our LMMSE-based adaptive multiuser detector for tiled beamspace architectures achieves lower BER than conventional LMMSE while reducing training overhead, and remains scalable under realistic power and interconnect constraints. The tiled architecture work is a collaboration with Prof. Zhengya Zhang's group at the University of Michigan.

As we scale sub-THz networks to wider bandwidths and larger arrays, how do we design power-efficient receivers that can handle frequency-dependent beam squint and multiuser interference?

At sub-THz carriers with a large fractional bandwidth, fully digital arrays become impractical: high-speed data converters and downconversion stages dominate power consumption. We study a tiled hybrid beamforming architecture in which each tile is a phased subarray performing analog RF beamforming, followed by digital MU-MIMO processing. Using information-theoretic benchmarks, we compare RF beamforming strategies that mitigate beam squint via quadratic-phase broad beams. The encouraging finding is that simple phase-only RF beamformers paired with broad beams match or outperform more complex strategies requiring joint amplitude and phase control.

How can we use severely quantized, energy-efficient 1-bit ADCs in massive MIMO systems without losing the Angleof Arrival information to severe spatial harmonics?

ADCs dominate the power and cost budget of digital arrays, making 1-bit quantization an appealing design choice. We develop a Fourier analysis of the spatial harmonics produced by the 1-bit nonlinearity. The analysis yields a concrete design payoff: a ramped-phase training sequence that suppresses higher-order harmonics and isolates the fundamental spatial frequency corresponding to each user's true angle of arrival.