Sub-Array DPD

Elements are grouped based on shared nonlinear behavioral characteristics rather than physical proximity alone. The clustering algorithm analyzes AM-AM and AM-PM distortion profiles to identify power amplifiers with nearly identical compression curves and memory effects. This ensures that a single predistorter model accurately linearizes all elements in the sub-array without introducing residual distortion from mismatched characteristics.

• Groups formed via k-means clustering on extracted PA behavioral parameters
• Accounts for thermal coupling and shared heat dissipation paths
• Re-clustering triggered when beamforming weights change significantly

A single set of DPD basis function coefficients is computed for the entire sub-array and applied identically to each element within the cluster. This coefficient sharing eliminates redundant computation across elements with similar nonlinear behavior. The shared predistorter is trained using a representative element or an averaged feedback signal from multiple elements in the sub-array.

• Reduces FPGA DSP slice utilization proportionally to cluster size
• Enables real-time adaptation with limited hardware resources
• Compatible with both memory polynomial and neural network predistorter structures

Each sub-array designates a primary element whose nonlinear response serves as the training target for the shared DPD engine. Selection criteria include median distortion characteristics within the cluster, central physical location to capture average thermal conditions, and accessibility for feedback receiver connection. The representative element's predistorter is then broadcast to all sibling elements.

• Selection updated periodically to track thermal drift
• Feedback receiver multiplexed across representative elements only
• Outlier elements with divergent behavior may be reassigned to different clusters

Sub-array boundaries are dynamically adjusted based on the active beamforming configuration. As beam-steering angles change, the active impedance seen by each power amplifier shifts, altering its nonlinear behavior. The clustering algorithm incorporates beamforming weight vectors to predict which elements will experience similar impedance trajectories and should remain grouped together.

• Clustering updated at slot-level granularity in 5G NR systems
• Pre-computed cluster maps stored in look-up tables for rapid switching
• Integrates with hybrid beamforming architectures where analog and digital domains interact

Sub-array DPD introduces a controlled linearization penalty in exchange for substantial hardware savings. The trade-off is quantified by measuring adjacent channel leakage ratio (ACLR) degradation as cluster size increases. System designers select the maximum acceptable ACLR loss—typically 0.5 to 2 dB—and configure sub-array sizes accordingly.

• 2-element clusters: near-identical performance to per-element DPD
• 8-element clusters: ~1 dB ACLR degradation, 8x resource savings
• 16-element clusters: ~2 dB degradation, suitable for less stringent requirements
• Adaptive sizing balances spectral mask compliance with cost constraints

Sub-array DPD pairs naturally with single-feedback receiver architectures where one observation receiver sequentially samples outputs. Instead of sampling every element, the feedback receiver captures only the representative element from each sub-array, further reducing switching complexity and training latency. This combination enables cost-effective linearization for arrays with 64 or more elements.

• Feedback sampling rate reduced by factor equal to sub-array size
• Training convergence time shortened due to fewer elements to characterize
• Enables over-the-air feedback capture from representative elements only

A resource-efficient technique where a common set of DPD basis function coefficients is applied across multiple antenna branches with similar nonlinear behavior. This is the foundational principle behind sub-array DPD.

• Reduces total coefficient storage by grouping elements with correlated distortion
• Leverages spatial correlation of PA characteristics across adjacent array elements
• Typically combined with cluster identification algorithms to determine optimal groupings

A dimensionality reduction technique that identifies and compensates for the dominant spatial modes of nonlinear distortion in massive MIMO arrays.

• Decomposes the array's distortion into orthogonal spatial components
• Retains only the most significant eigenmodes for linearization
• Can reduce the effective number of DPD engines from hundreds to fewer than ten
• Pairs naturally with sub-array architectures by identifying which spatial modes dominate each cluster

The variation in the impedance seen by an individual power amplifier due to beam steering and mutual coupling, causing the amplifier's nonlinear behavior to change dynamically.

• The primary physical mechanism that complicates sub-array DPD grouping
• Elements at array edges experience different impedance variations than center elements
• Sub-array clusters must account for load-dependent AM/AM and AM/PM distortion
• Beam-dependent clustering may be required for wide steering ranges

A complexity-reduction technique that identifies and selects only the most significant basis functions from a large candidate set to build an efficient array predistorter.

• Uses algorithms like LASSO or orthogonal matching pursuit for basis selection
• Can reduce the number of active coefficients by 70-90% with minimal linearization loss
• Complements sub-array DPD by further reducing per-cluster computational load
• Particularly effective when combined with memory polynomial model pruning

A digital predistortion technique that accounts for the dynamic changes in PA nonlinearity caused by varying beamforming weights in a phased array.

• Recognizes that sub-array clustering may need to adapt as beam angles change
• Integrates beamforming coefficient knowledge into the DPD coefficient update loop
• Can pre-compute cluster assignments for different beam configurations
• Essential when sub-array groupings based on static characteristics prove insufficient

A cost-effective array linearization architecture that uses a single observation receiver to sequentially sample the output of multiple PAs for DPD training.

• Critical enabler for practical sub-array DPD implementation
• Reduces hardware cost by eliminating per-element feedback chains
• Introduces time-multiplexed training constraints that affect coefficient update rates
• Sub-array architectures can share one feedback path per cluster for faster convergence