Frequency-Selective DPD

The core mechanism involves splitting a wideband signal into multiple narrowband sub-signals using a filter bank or spectral decomposition technique. Each subband experiences a relatively flat frequency response from the power amplifier, allowing a lower-complexity, narrowband DPD model to be applied independently. This contrasts with wideband DPD, which must model the entire frequency-dependent behavior with a single, high-complexity model. Common decomposition methods include DFT-modulated filter banks and polyphase channelizers.

Wideband power amplifiers exhibit frequency-dependent AM/AM and AM/PM characteristics, where gain and phase distortion vary across the signal bandwidth. A single wideband DPD model often fails to fully correct these variations. Frequency-Selective DPD addresses this by:

• Applying independent memory polynomial models to each subband
• Tailoring the predistortion coefficients to the local nonlinear behavior
• Effectively suppressing spectral regrowth that varies by frequency offset This results in superior Adjacent Channel Leakage Ratio (ACLR) improvement for wideband signals like 5G NR carriers.

A key advantage is the reduction in overall processing complexity. By operating on lower-bandwidth sub-signals, each DPD block can run at a reduced sample rate proportional to the subband bandwidth rather than the full wideband rate. This enables:

• Lower-order Volterra or memory polynomial models per subband
• Parallel processing on FPGA or ASIC hardware
• Reduced total multiply-accumulate operations compared to a single wideband DPD with high nonlinearity order This makes the architecture attractive for power-constrained edge devices and massive MIMO arrays.

After independent linearization, the predistorted subband signals must be recombined into a single wideband signal for transmission. This synthesis filter bank stage is critical and must manage:

• Inter-subband interference caused by nonlinear processing
• Aliasing artifacts from upsampling and recombination
• Phase coherence across subbands to prevent signal distortion Advanced techniques use oversampled filter banks and guard bands between subbands to minimize these artifacts. The synthesis filters are typically matched to the analysis filters to ensure perfect or near-perfect reconstruction.

Training Frequency-Selective DPD requires extracting coefficients for each subband predistorter. This is typically done using an Indirect Learning Architecture (ILA) adapted for the subband structure. The process involves:

• Capturing the PA output and decomposing it into the same subband structure
• Training a post-distorter model for each subband independently
• Copying the trained coefficients to the corresponding forward-path predistorter Joint optimization across subbands can further improve performance by accounting for inter-subband nonlinear interactions.

Frequency-Selective DPD is particularly relevant for 5G NR and future wireless systems due to:

• Carrier aggregation with widely spaced component carriers
• Massive MIMO arrays where per-antenna DPD must be low-complexity
• mmWave systems with extreme bandwidths and severe frequency-dependent effects The technique is often combined with Crest Factor Reduction (CFR) on a per-subband basis to optimize the peak-to-average power ratio before amplification, maximizing power amplifier efficiency.

A closely related architecture that decomposes a wideband signal into multiple narrowband sub-signals using filter banks. Independent DPD is applied to each subband before recombination. The key distinction from Frequency-Selective DPD is the explicit use of analysis and synthesis filter banks to isolate subbands, whereas Frequency-Selective DPD may apply independent processing in the frequency domain without explicit decomposition. This approach reduces the effective sampling rate required for each DPD block, lowering overall computational complexity.

A processing architecture where different DPD blocks operate at different sampling rates. High-rate processing handles wideband distortion correction, while low-rate blocks manage narrowband effects. This is often combined with Frequency-Selective DPD to optimize power consumption in FPGA and ASIC implementations. Key benefits include:

• Reduced clock speeds for memory effect compensation
• Lower dynamic power in coefficient update paths
• Efficient handling of frequency-dependent nonlinearities without oversampling the entire chain

A predistortion model that uses a two-dimensional indexing structure based on the instantaneous magnitudes of two concurrent baseband signals. While Frequency-Selective DPD processes sub-bands of a single wideband carrier, 2D-DPD addresses cross-band distortion between two independent carriers. The 2D approach synthesizes correction signals by indexing into a 2D Look-Up Table (2D-LUT) or evaluating a 2D polynomial function of both envelope magnitudes simultaneously.

A long-term memory phenomenon in multi-band and wideband amplifiers where the nonlinear behavior in one frequency region is influenced by the past envelope history of a signal in a different band. Frequency-Selective DPD must account for these effects when sub-bands interact through the amplifier's thermal and trapping dynamics. Key characteristics:

• Time constants ranging from microseconds to milliseconds
• Strong dependence on sub-band spacing and power distribution
• Requires cross-band memory terms in the predistorter model

An extension of the generalized memory polynomial that incorporates cross-band envelope coupling terms and sample-crossing terms to capture complex nonlinear interactions. This model serves as the mathematical foundation for many Frequency-Selective DPD implementations. The MB-GMP includes:

• In-band memory polynomial terms for each sub-band
• Cross-band envelope terms coupling different frequency regions
• Lagging cross-terms to model frequency-dependent memory effects
• Often pruned using LASSO or OMP algorithms to reduce coefficient count

The signal processing procedure for estimating DPD model parameters from observed input and output waveforms across multiple frequency sub-bands. For Frequency-Selective DPD, this involves:

• Synchronized capture of wideband PA input and output
• Frequency-domain decomposition into sub-band components
• Joint or sequential estimation of per-band and cross-band coefficients
• Common algorithms include Least Squares (LS), Recursive Least Squares (RLS), and Least Mean Squares (LMS) adapted for multi-band formulations