Frequency-Selective Predistortion

The fundamental mechanism involves decomposing the wideband signal into multiple narrower sub-bands using digital filter banks or frequency-domain processing. Each sub-band is then linearized independently with its own predistorter coefficients before being recombined.

• Filter bank approach: Uses analysis/synthesis filter banks to split and reconstruct signals
• Frequency-domain DPD: Applies predistortion directly in the frequency domain via FFT/IFFT processing
• Sub-band count: Typically 2-8 sub-bands for 5G NR signals with 100-400 MHz bandwidth
• Enables independent optimization of in-band distortion and out-of-band emission per frequency region

Power amplifiers exhibit non-uniform frequency responses due to bias network impedance variations, thermal gradients across the die, and parasitic reactances that vary with carrier frequency. Frequency-selective DPD explicitly models these effects.

• Bias network resonance: Impedance minima at specific frequencies cause envelope-dependent gain variations
• Thermal memory: Different transistor regions heat at different rates, creating frequency-dependent time constants
• Dispersion: Group delay variations across the band cause phase distortion that uniform DPD cannot correct
• Conventional memory polynomials assume frequency-invariant memory depth, which fails for wideband signals

Rather than fitting a single global model, frequency-selective DPD employs piecewise linearization where the predistorter response is a function of both instantaneous signal envelope and frequency location within the carrier bandwidth.

• Each sub-band predistorter uses a localized Volterra or memory polynomial model
• Coefficient interpolation between sub-bands ensures smooth transitions
• Reduces model complexity compared to a single high-order generalized memory polynomial
• Particularly effective for Doherty amplifiers where main and peaking paths have different frequency responses

Frequency-selective DPD provides superior suppression of intermodulation distortion products that fall at specific frequency offsets. By targeting spectral regions individually, it can apply more aggressive correction where needed without over-constraining other regions.

• IMD3 asymmetry: Upper and lower third-order intermodulation products often have different magnitudes due to baseband impedance effects
• Frequency-selective DPD can independently cancel upper and lower sideband distortion
• Reduces ACLR by 3-5 dB beyond uniform DPD for signals exceeding 200 MHz bandwidth
• Critical for meeting 3GPP adjacent channel selectivity requirements in carrier aggregation scenarios

The primary engineering challenge is managing the increased computational burden of multiple parallel predistorters. Implementation strategies balance linearization performance against FPGA/ASIC resource constraints.

• Resource scaling: Complexity grows approximately linearly with the number of sub-bands
• Shared basis functions: Common nonlinear basis terms can be computed once and weighted differently per sub-band
• Decimated processing: Sub-band signals can be processed at lower sample rates after decimation
• Look-up table sharing: LUT-based implementations can share interpolation structures across sub-bands
• Typical implementation requires 1.5-3x the logic resources of a single-band DPD for equivalent throughput

Training frequency-selective DPD requires wideband observation receivers capable of capturing the full predistorted signal bandwidth plus distortion products. The coefficient extraction process must solve a larger optimization problem.

• Indirect learning architecture: Most common approach, comparing PA output to desired linear response per sub-band
• Least squares estimation: Block-based coefficient extraction using sub-band filtered error signals
• Recursive adaptation: LMS or RLS algorithms adapted for multi-band coefficient updates
• Training signal requirements: Wideband OFDM or multi-tone signals that excite all frequency regions simultaneously
• Convergence time typically 2-5x longer than single-band DPD due to larger parameter space

A dynamic nonlinearity where the current output distortion depends on the past amplitude of the input signal envelope. This frequency-dependent behavior, caused by bias network impedance and thermal dynamics, is the primary physical phenomenon that frequency-selective predistortion is designed to compensate for.

• Caused by low-frequency dispersion in the power amplifier
• Manifests as asymmetry in intermodulation distortion sidebands
• Requires models with memory depth to accurately capture

A simplified Volterra series structure that captures both nonlinearity and memory effects using diagonal kernel coefficients. Memory polynomial models serve as the foundational behavioral framework from which frequency-selective predistortion functions are often derived.

• Balances model accuracy with computational efficiency
• Generalized versions include cross-terms between different delay taps
• Widely used for FPGA-based DPD implementations

The maximum signal bandwidth over which a digital predistortion system can effectively suppress nonlinear distortion. Frequency-selective DPD directly targets the expansion of this bandwidth to meet 5G NR and wideband satellite requirements.

• Limited by feedback path sampling rate and aliasing constraints
• Typically requires 3-5x the signal bandwidth for effective cancellation
• Directly impacts adjacent channel leakage ratio compliance

A digital predistortion implementation where the predistorter operates at a higher sampling rate than the baseband signal. This oversampling is essential for frequency-selective DPD to capture and cancel out-of-band distortion products that would otherwise alias back into the signal band.

• Enables cancellation of distortion beyond the Nyquist frequency of the baseband
• Increases computational load on the DPD engine
• Critical for wideband carrier aggregation scenarios

DPD techniques designed to linearize power amplifiers transmitting multiple aggregated component carriers across fragmented spectrum. Frequency-selective predistortion is particularly valuable here, as different carriers experience different amplifier memory effects.

• Addresses cross-modulation between non-contiguous carriers
• Requires wideband observation receivers to capture full composite signal
• Essential for LTE-Advanced Pro and 5G NR multi-carrier deployments

The process of characterizing and compensating for nonlinearities in the DPD observation receiver chain. For frequency-selective predistortion to be effective, the feedback signal must be a faithful copy of the PA output across the entire linearization bandwidth.

• Compensates for ADC nonlinearity and analog filter ripple
• Often uses a separate predistorter for the receiver path
• Ensures coefficient estimation is not corrupted by receiver impairments