Analog-to-Digital Converter Clipping

When the input signal amplitude surpasses the ADC's full-scale range (FSR) , the converter output saturates at its maximum or minimum digital code. Unlike soft compression in amplifiers, this is a non-differentiable hard clipping that generates an infinite series of odd-order harmonics. The clipped waveform becomes a square-wave approximation of the original signal, introducing broadband spectral splatter that cannot be removed by downstream filtering. This fundamentally violates the DPD assumption that the feedback path is linear and transparent.

Clipping in the observation path directly corrupts the error signal used in adaptive predistortion algorithms. Key consequences include:

• Model divergence: The indirect learning architecture attempts to invert a distorted version of the PA output, leading to unstable coefficient updates
• ACLR degradation: Predistorter coefficients optimized on clipped feedback data fail to suppress out-of-band emissions, often making spectral regrowth worse than without DPD
• EVM floor: In-band distortion cannot be corrected because the clipping artifacts are indistinguishable from actual PA nonlinearity in the training data

The likelihood of ADC clipping is directly governed by the signal's Peak-to-Average Power Ratio (PAPR) and the ADC's dynamic range allocation. For an OFDM signal with 10 dB PAPR, even with optimal gain staging, the peak-to-average voltage ratio is approximately 3.16:1. If the ADC's full-scale is set to accommodate the average power with 12 dB of headroom, signal peaks exceeding 12 dB above average will clip. The complementary cumulative distribution function (CCDF) quantifies this probability, typically targeting clip rates below 10⁻⁴ for reliable DPD operation.

Proper gain staging in the observation receiver is the primary defense against clipping. The variable gain amplifier (VGA) preceding the ADC must balance two competing requirements:

• Noise floor: Sufficient gain to ensure the ADC's quantization noise does not dominate the feedback SNR
• Headroom margin: Enough back-off to accommodate signal peaks without saturation Typical designs allocate 12-15 dB of peak headroom for wideband signals, though this reduces effective ENOB by 2-2.5 bits. Automatic gain control (AGC) loops must respond slowly to avoid modulating the DPD training signal.

ADC clipping generates a characteristic spectral signature that differs from PA nonlinearity:

• Broadband noise pedestal: Hard clipping produces a flat noise floor extending far beyond the signal bandwidth, unlike the polynomial-shaped spectral regrowth of amplifier distortion
• In-band distortion: The clipped signal's constellation experiences uniform compression rather than the AM-AM/AM-PM pattern of PA nonlinearity
• Harmonic folding: In undersampled or bandpass-sampled architectures, clipping harmonics can alias back into the band of interest, creating spurious tones that mimic actual signal content

Several strategies address ADC clipping in wideband DPD systems:

• Oversampling and digital clipping correction: Sampling at 3-5x the signal bandwidth allows post-hoc reconstruction of clipped peaks using interpolation algorithms
• Predictive peak detection: Analog envelope detectors trigger fast attenuators before the ADC input when signal peaks approach full-scale
• Dual-ADC architectures: A high-gain path captures low-amplitude signals while an attenuated path handles peaks; digital recombination provides extended dynamic range
• Sigma-delta ADCs: Continuous-time oversampling converters with noise shaping inherently resist hard clipping due to their loop filter dynamics

The process of characterizing and compensating for nonlinearities in the DPD observation receiver chain. Before the DPD can correct the power amplifier, the observation path itself must be pristine. This involves modeling the ADC, mixers, and amplifiers in the feedback loop to ensure the digital representation of the PA output is a faithful, undistorted copy. Any uncalibrated nonlinearity here directly translates to coefficient estimation errors in the predistorter.

A critical metric quantifying in-band signal quality degradation. ADC clipping directly increases EVM by distorting the constellation points of the received feedback signal. When the peak-to-average power ratio (PAPR) of the signal causes the ADC to saturate, the resulting amplitude truncation introduces both in-band distortion and spectral regrowth, making it impossible for the DPD algorithm to converge to an optimal solution.

The primary driver of ADC clipping risk. Modern wideband signals like OFDM exhibit high PAPR, meaning the instantaneous signal peaks vastly exceed the average power level. To avoid clipping, the ADC's dynamic range must accommodate these rare peaks, forcing the system to operate with significant back-off. This reduces the effective signal-to-noise ratio and limits the observable linearization bandwidth.

A nonlinear impairment distinct from clipping but often conflated with it. Aliasing occurs when the sampling rate in the DPD feedback path is insufficient to capture the full bandwidth of the nonlinear signal. The intermodulation products generated by the PA fold back into the Nyquist band, corrupting the training data. Unlike clipping, aliasing is a linear phenomenon in the frequency domain but requires strict anti-alias filtering and oversampling to mitigate.

A signal conditioning technique applied before the power amplifier and observation path to reduce PAPR. By intelligently clipping or shaping the peaks of the baseband waveform, CFR reduces the dynamic range required of both the PA and the observation ADC. This allows the feedback receiver to operate at a higher average power level without clipping, maximizing the signal-to-noise ratio and improving DPD coefficient accuracy.

An implementation strategy to combat bandwidth limitations in the observation path. The predistorter operates at a higher sampling rate than the baseband signal to generate out-of-band cancellation products. The feedback ADC, however, may run at a lower rate. This architecture requires careful digital upconversion and filtering to prevent the lower-rate observation data from aliasing the wideband predistorted signal, a challenge directly related to managing the ADC's dynamic range constraints.