Subband DPD

The core signal decomposition engine. A filter bank splits the wideband input into multiple contiguous subbands using a set of bandpass filters. After independent DPD processing, a synthesis filter bank recombines the predistorted sub-signals into a single wideband output.

• Perfect reconstruction: Filter banks are designed to avoid aliasing and amplitude distortion at subband boundaries
• Oversampled architectures: Often employ 2x oversampling per subband to accommodate spectral regrowth from the DPD nonlinearity
• Polyphase implementation: Computationally efficient filter bank structures that reduce redundant operations

Each narrowband sub-signal is linearized by its own dedicated predistorter block. Because each subband occupies a fraction of the total bandwidth, the sampling rate per DPD block is proportionally reduced.

• Model selection flexibility: Different DPD models (memory polynomial, GMP, neural network) can be assigned per subband based on local nonlinear characteristics
• Coefficient isolation: Each predistorter learns only the distortion affecting its specific frequency slice
• Parallel processing: Subband DPD blocks execute concurrently, enabling hardware pipeline optimization

Power amplifiers exhibit frequency-dependent nonlinear behavior — gain compression and phase shift vary across the operating band. Subband DPD inherently addresses this by allowing each subband predistorter to model the local nonlinear response independently.

• Memory effect localization: Long-term memory effects that are frequency-selective are captured within the affected subband
• Thermal memory tracking: Subband-specific thermal time constants can be modeled without cross-band interference
• Improved ACLR: Adjacent channel leakage is suppressed more effectively when distortion is corrected at the subband level

After per-subband DPD, the individually predistorted signals must be recombined into a single wideband signal. This stage requires careful aliasing cancellation because the DPD nonlinearity expands the bandwidth of each subband.

• Guard band allocation: Frequency gaps between subbands absorb spectral regrowth without aliasing into adjacent subbands
• Overlap-add synthesis: Windowing techniques ensure smooth transitions at subband boundaries during recombination
• Digital upconversion: Each predistorted subband is shifted to its original center frequency before summation

The primary motivation for subband DPD is reducing the total multiply-accumulate operations (MACs) required for wideband linearization. By operating at lower per-subband rates, the aggregate computational load is significantly lower than a full-rate wideband DPD.

• Quadratic scaling benefit: DPD complexity often scales quadratically with bandwidth; subband decomposition breaks this relationship
• FPGA resource savings: Lower sample rates reduce DSP slice and block RAM utilization
• Power efficiency: Reduced clock rates translate directly to lower dynamic power consumption in ASIC and FPGA implementations

Subband DPD requires a coordinated training strategy to ensure stable convergence. Coefficients can be adapted independently per subband or jointly using a unified error metric.

• Indirect learning per subband: Each subband post-distorter is trained on the downconverted and filtered PA output for its specific frequency slice
• Joint cost function: A composite error signal combining all subband residuals ensures global linearization optimization
• Cross-subband coupling terms: Optional cross-terms between adjacent subbands capture inter-subband distortion leakage

A linearization paradigm that applies independent processing to different spectral regions of a wideband signal. Unlike subband DPD which decomposes and recombines, frequency-selective DPD often uses filtered basis functions to target specific distortion products.

• Addresses frequency-dependent nonlinearities in GaN amplifiers
• Uses band-limited Volterra kernels for computational efficiency
• Complements subband approaches in hybrid architectures

An architecture where different processing blocks operate at different sampling rates to optimize power consumption while maintaining linearization bandwidth. Subband DPD is a specific implementation of multi-rate processing.

• High-rate paths handle wideband distortion
• Low-rate paths manage narrowband memory effects
• Reduces total computational load by 40-60% compared to full-rate DPD

The overarching challenge of linearizing signals with instantaneous bandwidths exceeding 200 MHz for 5G NR and satellite communications. Subband DPD directly addresses the processing bottleneck that occurs when the DPD sample rate must be 5-7x the signal bandwidth.

• Critical for mmWave and sub-THz systems
• Enables use of lower-cost ADCs and DACs
• Mitigates FPGA resource constraints in massive MIMO

A related but distinct concept where a single PA amplifies multiple carrier signals at different frequencies simultaneously. While subband DPD splits one wideband signal, MB-DPD handles concurrent narrowband carriers with large frequency gaps.

• Addresses cross-band intermodulation products
• Uses 2D memory polynomial models for cross-band terms
• Subband DPD can be applied within each carrier band of an MB-DPD system

Beamforming-aware linearization for antenna arrays with 64 to 256+ elements. Subband DPD is particularly valuable here because the per-antenna DPD complexity must be minimized while handling wideband signals.

• Each antenna path requires independent linearization
• Beam-dependent nonlinearity complicates correction
• Subband processing reduces per-path FPGA utilization by 50-70%

The mathematical techniques for extracting DPD model parameters from observed PA input-output data. Subband DPD requires per-subband coefficient extraction with careful handling of recombination artifacts.

• Least squares (LS) estimation per subband
• Recursive least squares (RLS) for online adaptation
• Must account for filter bank phase alignment during reconstruction