Massive MIMO DPD

A linearization architecture that applies predistortion in the beam domain rather than at individual antenna elements. The DPD engine operates on each user's data stream after precoding, compensating for the composite non-linearity experienced by that specific beam.

• Reduces computational complexity from O(N_antennas) to O(N_beams)
• Captures beam-dependent AM-AM and AM-PM distortion
• Requires real-time beamforming weight knowledge
• Example: A 64-element array serving 8 users uses 8 DPD blocks instead of 64

A hybrid architecture that partitions the full antenna array into smaller, independently linearized sub-arrays. Each sub-array has its own DPD block, and the outputs are combined in the far field.

• Balances linearization performance against hardware cost
• Sub-array size is a critical design parameter
• Works well with partially-connected hybrid beamforming architectures
• Mitigates, but does not fully eliminate, beam-dependent distortion
• Common in 5G NR base stations with analog sub-array beamforming

A pragmatic approach that applies a single DPD function to all power amplifiers, assuming identical PA behavior across the array. A single observation receiver captures the combined radiated signal for coefficient adaptation.

• Lowest implementation cost and complexity
• Assumes well-matched PAs with minimal variation
• Vulnerable to beam-dependent distortion in wide-scan scenarios
• Feedback path captures array-level, not per-PA, non-linearity
• Suitable for small arrays with tightly controlled manufacturing tolerances

The most comprehensive architecture where each antenna element has a dedicated DPD block and observation receiver. This brute-force approach linearizes every PA independently before beamforming.

• Guarantees linearity regardless of beam direction
• Computational cost scales linearly with array size
• Requires N_antennas observation receivers and DPD engines
• Impractical for massive MIMO (64+ elements) due to hardware overhead
• Serves as the theoretical performance upper bound for other architectures

A linearization technique that uses a remote observation receiver placed in the far field to capture the radiated signal. The feedback includes the combined effects of PA non-linearity, antenna mutual coupling, and impedance mismatch.

• Compensates for array-level impairments invisible to conducted feedback
• Enables joint DPD and antenna calibration
• Observation receiver placement affects performance
• Critical for active antenna systems with integrated PAs
• Often combined with beam-space DPD for massive MIMO

An advanced architecture using a single neural network conditioned on beamforming weights to generate beam-specific predistortion. The network learns the mapping from beam direction to composite PA non-linearity.

• Input features include beam index, steering angle, or weight vector
• Eliminates need for separate DPD blocks per beam
• Generalized Memory Polynomial often used as baseline comparison
• Enables continuous adaptation as users move and beams change
• Active research area combining deep learning with array theory

The core challenge in Massive MIMO DPD where each formed beam experiences a unique composite distortion characteristic. Unlike single-antenna systems, the effective non-linearity is a function of the complex weighting of individual power amplifier (PA) outputs. The distortion perceived in the far-field changes dynamically as the beamforming precoder updates, requiring a linearization strategy that is aware of the spatial domain. This phenomenon is driven by the interaction between per-antenna PA non-linearity and the coherent combination of signals in the intended direction.

A critical signal metric that quantifies the relationship between peak and average power. High PAPR signals, typical in OFDM systems, force PAs to operate with significant back-off to avoid compression, drastically reducing power-added efficiency (PAE). In Massive MIMO, the effective PAPR can vary across beams. DPD combined with Crest Factor Reduction (CFR) is essential to allow PAs to operate closer to saturation while maintaining signal integrity, directly impacting the energy efficiency of the entire base station.

Physical interactions between antenna elements in a dense array that cause energy from one element to be absorbed by adjacent elements. This mutual coupling alters the effective impedance seen by each PA, changing its non-linear behavior as a function of the array's excitation. A robust Massive MIMO DPD model must account for this spatial interaction, as the distortion is no longer a function of a single PA's input but of the entire array's driving signal. Ignoring this leads to degraded Error Vector Magnitude (EVM) at the beam level.

The application of deep learning to solve the high-dimensional, dynamic non-linearity problem in antenna arrays. Architectures like Convolutional Neural Networks (CNNs) or Graph Neural Networks (GNNs) can learn the spatial-temporal dependencies between array elements. A single neural network can be trained to act as a multi-input, multi-output predistorter, jointly linearizing all PAs while implicitly compensating for beam-dependent effects and mutual coupling. This replaces the need for complex, physics-based Volterra series models that scale poorly with array size.