Multi-Rate DPD

The core architectural distinction: the predistorter operates at a sampling rate (Fs_DPD) that is an integer multiple of the baseband signal rate (Fs_BB). This oversampling—typically 3x to 5x—is essential because nonlinear distortion expands signal bandwidth. A 100 MHz baseband signal passing through a power amplifier with 5th-order nonlinearity generates distortion products spanning 500 MHz. Without oversampling, these out-of-band components alias back into the Nyquist zone, making them impossible to cancel. The higher-rate path captures the full nonlinear spectrum, synthesizing a predistorted signal with intentional anti-phase distortion across the expanded bandwidth.

The observation receiver must digitize the PA output with sufficient bandwidth to capture all distortion products the predistorter aims to cancel. This requires an analog-to-digital converter (ADC) sampling at or above the predistorter rate. For a 200 MHz 5G NR carrier with 5th-order DPD, the feedback ADC must sample at 1 GHz or higher. Key challenges include:

• ADC effective number of bits (ENOB) degradation at wide bandwidths
• Anti-aliasing filter design with flat passband and sharp cutoff
• Jitter-induced noise that limits dynamic range at high input frequencies
• Feedback path linearization to prevent the observation receiver itself from introducing uncompensated distortion

Coefficient estimation algorithms in multi-rate DPD must explicitly account for aliased distortion products. When the feedback ADC rate is lower than the full nonlinear bandwidth, out-of-band distortion folds into the observed spectrum. The indirect learning architecture adapts by:

• Band-limiting the model output to match the feedback bandwidth before error calculation
• Spectral extrapolation using the behavioral model to predict unobservable distortion
• Frequency-selective training that weights in-band and adjacent-channel errors differently

Failure to manage aliasing leads to coefficient bias—the predistorter optimizes for the aliased spectrum rather than the true PA output, degrading ACLR performance.

Multi-rate DPD introduces a complexity tradeoff between the oversampling factor and achievable linearization. Higher rates enable cancellation of higher-order intermodulation products but increase:

• Multiply-accumulate operations scaling with the predistorter sample rate
• Coefficient memory for models with memory depth (e.g., memory polynomial with M taps)
• Feedback path power consumption from high-speed ADCs
• Latency through interpolation and decimation filter chains

Practical implementations balance these factors. A 3x oversampled DPD typically captures 3rd and 5th-order distortion adequately for ACLR compliance, while 5x may be necessary for stringent spectral masks or when compensating 7th-order nonlinearities in GaN PAs.

The coefficient estimation path operates at a reduced rate compared to the predistorter to minimize computational load. After capturing the wideband feedback signal, a decimation chain reduces the sample rate before model extraction. This is valid because:

• Coefficient adaptation converges slowly relative to the signal dynamics
• Stochastic gradient methods tolerate subsampled error signals
• Correlation-based estimation (e.g., least squares) can operate on decimated data blocks

The decimation factor must preserve the distortion bandwidth of interest. Aggressive decimation that aliases adjacent-channel energy into the estimation band will produce biased coefficients that fail to suppress spectral regrowth.

The maximum signal bandwidth over which a DPD system can effectively suppress nonlinear distortion. In multi-rate architectures, the linearization bandwidth is determined by the predistorter's higher sampling rate, typically 3-5x the baseband signal bandwidth. This expanded bandwidth captures out-of-band intermodulation products that would otherwise cause adjacent channel interference. For 5G NR signals with 100 MHz carriers, linearization bandwidths of 400-500 MHz are common to suppress third and fifth-order distortion products.

The ratio of the predistorted signal's bandwidth to the original signal's bandwidth. In multi-rate DPD, this factor is intentionally large to accommodate spectral regrowth compensation. The expansion factor directly determines the required digital-to-analog converter (DAC) sampling rate and the computational load on the predistorter. For a 100 MHz OFDM signal with fifth-order nonlinearity compensation, the expansion factor is typically 5, requiring a DAC operating at 500 MHz or higher.

The unwanted appearance of signal energy in adjacent frequency channels caused by intermodulation products generated in a nonlinear power amplifier. Multi-rate DPD specifically targets this regrowth by operating at a higher sampling rate to synthesize anti-phase distortion products that cancel the regrown spectrum. Without adequate predistortion bandwidth, the cancellation signal cannot extend far enough into adjacent channels, leaving residual adjacent channel leakage that violates regulatory emission masks.

A critical metric quantifying the ratio of transmitted power within an assigned channel to the power leaking into adjacent channels. Multi-rate DPD systems are evaluated primarily by their ability to improve ACLR. The higher sampling rate enables cancellation of third-order and fifth-order intermodulation products that fall into the first and second adjacent channels. Typical 5G base station requirements demand ACLR better than -45 dBc, which multi-rate DPD achieves by extending the predistortion bandwidth well beyond the carrier.

A critical impairment in multi-rate DPD systems where the observation receiver's analog-to-digital converter (ADC) samples at an insufficient rate to capture the full bandwidth of the nonlinear PA output. When aliasing occurs, high-frequency distortion products fold back into the Nyquist band, corrupting the feedback signal used for coefficient estimation. Multi-rate architectures mitigate this by employing ADCs with sampling rates matched to the predistorter's expanded bandwidth, ensuring the feedback path captures all relevant distortion products without spectral folding.

The process of characterizing and compensating for nonlinearities in the DPD observation receiver chain to ensure the feedback signal is a faithful copy of the PA output. In multi-rate systems, the feedback path must maintain linearity across the entire expanded bandwidth. Techniques include:

• Post-distortion of the ADC output using a separate inverse model
• Calibration tone injection to measure receiver nonlinearity
• Blind estimation algorithms that jointly identify PA and receiver impairments Without feedback linearization, the DPD algorithm trains on corrupted data, degrading ACLR performance.