In multi-antenna wireless communications, precoding is the mechanism that allows a base station to steer independent data streams to multiple users simultaneously over the same frequency channel. By carefully adjusting the amplitude and phase of the signals transmitted from each antenna, the base station ensures that the signals constructively combine at the intended user and destructively cancel out everywhere else, eliminating multi-user interference.
Traditionally, precoding matrices are designed assuming continuous, infinite-precision signals. However, physical systems must eventually map these continuous values to discrete digital representations. This quantization occurs in two primary places:
- At the hardware level: Low-resolution digital-to-analog converters (DACs) at the transmitter antennas.
- At the network level: Finite-capacity fronthaul links in Cloud Radio Access Networks (C-RANs) or limited feedback channels from users.
The High-SNR Saturation Problem
If you design a precoder ignoring this quantization step and simply round the continuous outputs to the nearest allowed discrete value, you introduce quantization noise.
At low signal-to-noise ratios (SNRs), this quantization noise is negligible because it is masked by dominant background thermal noise. However, as you increase transmit power to achieve higher data rates, a catastrophic phenomenon occurs: the quantization noise scales up proportionally with the signal power.
Because the noise floor rises at the exact same rate as the signal power, the Signal-to-Interference-plus-Noise Ratio (SINR) plateaus. This is known as the rate saturation effect or error floor. Beyond a certain power threshold, increasing transmit power yields zero throughput gains, rendering high-order modulations (like 16-QAM or 64-QAM) entirely impractical.
The Core Insight: Joint Optimization
The breakthrough across this line of research is to stop treating quantization as an unpredictable post-processing error. Instead, we pull the quantization constraint directly inside the precoding optimization loop.
By making the transmitter explicitly aware of the exact discrete alphabet or compression constraints, it can pre-distort the signals or shape the quantization noise. Over the air, this shaped noise interacts with the wireless channel to constructively combine or cancel out, arriving perfectly at the user terminals as a clean, high-resolution signal.
Overcoming Low-Resolution Hardware Constraints
Massive multi-user MIMO (MU-MIMO) systems require hundreds of radio frequency (RF) chains, making high-resolution DACs prohibitively expensive and power-hungry. Moving to ultra-low resolution hardware—such as 1-bit DACs that only output a signal's sign ($\pm 1$)—drastically reduces power consumption but destroys traditional linear precoding performance.
[Continuous Signal] ──> [Linear Precoder] ──> [1-Bit Quantizer] ──> [Channel] ──> (High Error Floor)
[Discrete Alphabet Optimization (SQUID)] ─────────────────────────> [Channel] ──> (Clean Constellation)
Jacobsson et al. demonstrated in Nonlinear 1-Bit Precoding for Massive MU-MIMO with Higher-Order Modulation that this bottleneck is not a fundamental hardware limit, but rather a failure of linear precoding. By replacing linear filters with a nonlinear optimization algorithm called SQUID, the transmitter treats the 1-bit constraint as an integral part of a mean-square error minimization problem.
In Quantized Precoding for Massive MU-MIMO, Jacobsson et al. extended this concept to show that treating signal selection as a combinatorial optimization problem over a discrete alphabet completely removes the high-SNR error floor, allowing 1-bit systems to support higher-order modulations. Castañeda et al. further proved the practical viability of these nonlinear approaches by successfully mapping them to high-throughput VLSI hardware architectures in 1-bit Massive MU-MIMO Precoding in VLSI.
To maximize the geometric efficiency of these discrete signals, Shao et al. proposed a unified framework in A Framework for One-Bit and Constant-Envelope Precoding Over Multiuser Massive MISO Channels that jointly optimizes the precoding weights alongside the inter-point spacing (scaling factors) of the modulation constellation. This allows the received signal geometry itself to adapt dynamically to hardware limitations.
Even when perfect channel knowledge is unavailable, quantization-aware designs remain highly resilient. Chu et al. showed in Robust Precoding Design for Coarsely Quantized MU-MIMO Under Channel Uncertainties-V0 that reformulating the joint problem into a tractable convex optimization space yields a negligible performance gap compared to ideal, unquantized systems. Similarly, Yapici et al. incorporated the statistical properties of 1-bit quantization distortion directly into a Signal-to-Leakage-plus-Noise Ratio (SLNR) objective function in SLNR Based Precoding for One-Bit Quantized Massive MIMO in mmWave Communications, successfully preventing high-SNR performance collapse by treating quantization as a predictable part of the signal model.
Managing Network Compression and Feedback Limits
The exact same mathematical tension arises when the bottleneck is a digital link within the network architecture rather than the antenna hardware. In Cloud Radio Access Networks (C-RANs), a Central Unit (CU) computes precoding vectors centrally but must compress the resulting signals before transmitting them over capacity-constrained fronthaul links to distributed Radio Units (RUs).
In Multivariate Fronthaul Quantization for Downlink C-RAN, Shamai introduced Multivariate Quantization (MQ), demonstrating that jointly optimizing the precoding matrix and the quantization codebooks across multiple RUs yields massive spectral efficiency gains compared to independent quantization. Instead of treating quantization noise as an unresolvable penalty, Simeone et al. explained in Cloud Radio Access Network: Virtualizing Wireless Access for Dense Heterogeneous Systems that a centralized controller can shape this quantization noise so that it acts as over-the-air interference cancellation, maintaining high signal quality despite limited fronthaul capacity.
This joint optimization problem is notoriously non-convex and mathematically challenging because compression noise directly couples with and degrades the precoded signals, as detailed by Shamai in Fronthaul Compression and Precoding Design for C-RANs Over Ergodic Fading Channels. However, Fan et al. provided a major theoretical breakthrough in Efficiently and Globally Solving Joint Beamforming and Compression Problem in the Cooperative Cellular Network Via Lagrangian Duality, proving that Semidefinite Relaxation (SDR) is tight for this problem. This allows the system to find globally optimal solutions efficiently, perfectly balancing user SINR against compression noise.
[Central Unit] ──(Joint Precoding & MQ)──> [Fronthaul Link] ──> [Radio Units] ──> [Over-the-Air Noise Cancellation]
Beyond fronthaul compression, quantization also plagues the feedback loop where users report channel state information (CSI) back to the base station. Zhou et al. showed in Robust Precoding Designs for Multiuser MIMO Systems With Limited Feedback that integrating tight mathematical approximations of feedback quantization errors directly into the matrix inversion process of an MMSE framework prevents rate saturation.
To bypass manual statistical modeling entirely, Jang et al. proposed an end-to-end deep learning framework in Deep Learning for Multi-User MIMO Systems: Joint Design of Pilot, Limited Feedback, and Precoding. By training a modular neural network to jointly optimize pilots, feedback quantization, and precoding simultaneously, the system automatically learns to make the precoder aware of the specific noise characteristics of the feedback loop, avoiding performance plateaus at high SNRs.
Turning Distortion into a Feature
Remarkably, once you have full control over the statistical properties of quantization noise, you can even weaponize it for security.
In Fronthaul quantization as artificial noise for enhanced secret communication in C-RAN, Shamai demonstrated that by jointly optimizing the precoding matrix and the quantization noise covariance matrix, the system can shape the quantization noise to fall harmlessly into the null spaces of legitimate users while steering it directly toward potential eavesdroppers. This turns a fundamental hardware and network limitation into an artificial noise source for physical-layer security, proving that quantization noise, when properly managed, can be an asset rather than a liability.
Key Foundational Work
Cloud Radio Access Network: Virtualizing Wireless Access for Dense Heterogeneous Systems (2015) — contributed 2 concept elements to this direction.
Multivariate Fronthaul Quantization for Downlink C-RAN (2016) — contributed 2 concept elements to this direction.
Nonlinear 1-Bit Precoding for Massive MU-MIMO with Higher-Order Modulation (2016) — contributed 2 concept elements to this direction.