Digital Pre-Distortion (DPD) Artifacts

DPD algorithms primarily correct static non-linearities, but they struggle to fully compensate for dynamic memory effects caused by thermal and electrical time constants in the PA.

• Thermal Memory: Slow changes in die temperature alter gain, creating a lagging error that a fixed DPD coefficient cannot track.
• Electrical Memory: Bias network resonances and trapping effects cause the current output to depend on previous symbols.
• Signature: The residual error vector magnitude (EVM) exhibits a pattern correlated with the signal's envelope history, not just its instantaneous amplitude.

DPD is implemented in digital hardware with finite precision. The look-up table (LUT) or polynomial coefficients are quantized, introducing a granular noise floor.

• Truncation Error: Rounding coefficients to fit fixed-point arithmetic creates a systematic error pattern.
• Interpolation Artifacts: Linear interpolation between sparse LUT entries fails to capture the smooth, continuous nature of the PA's inverse transfer function.
• Signature: This manifests as a low-level, signal-dependent noise pedestal in the corrected spectrum, distinct from thermal AWGN.

A perfectly linearized PA would confine all power to the allocated channel. DPD imperfections result in residual spectral regrowth—low-level 'shoulders' of power spilling into adjacent channels.

• ACLR Floor: The Adjacent Channel Leakage Ratio (ACLR) cannot be improved beyond a limit set by the DPD's adaptation accuracy and the PA's memory depth.
• Asymmetry: Real PAs often exhibit asymmetric distortion, and an imperfect DPD model will leave one sideband slightly higher than the other.
• Signature: The exact shape and asymmetry of these residual shoulders are unique to the specific PA-DPD pair.

Adaptive DPD systems continuously update coefficients based on a feedback observation path. The adaptation loop itself is a source of artifact.

• Steady-State Dithering: The LMS or RLS adaptation algorithm never perfectly converges; it jitters around the optimal coefficient set, causing a slow, continuous modulation of the distortion.
• Gain Mismatch: A small error in the feedback path's gain calibration causes the DPD to systematically over- or under-compensate.
• Signature: This creates a non-stationary, slowly time-varying component in the residual distortion, unlike a static non-linearity.

DPD is often paired with Crest Factor Reduction (CFR) to manage peak-to-average power ratio (PAPR). The interaction between these two non-linear blocks creates compound artifacts.

• Peak Clipping Residue: CFR clips peaks before the PA, but the DPD expands them again to pre-distort. An imperfect cascade leaves a unique clipping-plus-expansion signature on the peaks.
• Cross-Term Distortion: The DPD's correction for the PA's non-linearity is modulated by the CFR's clipping profile, generating intermodulation products not present in either block alone.
• Signature: The statistical distribution of peak amplitudes in the corrected signal deviates from the ideal in a device-specific manner.

The DPD's adaptation relies on a demodulated feedback signal. Any I/Q imbalance in the observation receiver is learned and compensated for by the DPD, but the correction is imperfect.

• Image Leakage: The DPD pre-distorts the transmit signal to cancel the receiver's own image, creating a faint, inverted image signal in the transmitted spectrum.
• Frequency-Dependent Imbalance: Wideband receivers have I/Q imbalance that varies with frequency, which a simple DPD model cannot fully correct.
• Signature: A weak, frequency-dependent mirror image of the signal appears, with a unique phase and amplitude relationship to the main signal.