Digital Pre-Distortion (DPD)

DPD functions by creating a complementary distortion profile that is the exact mathematical inverse of the Power Amplifier (PA) transfer function. When the pre-distorted signal passes through the PA, the non-linearities cancel out, resulting in a linear output.

• AM-AM Distortion: Corrects amplitude-dependent gain compression.
• AM-PM Distortion: Corrects amplitude-dependent phase rotation.
• Memory Effects: Modern DPD models (like Volterra series) account for thermal and electrical memory, where the current output depends on previous inputs.

PA non-linearity causes spectral regrowth, where the transmitted signal spills energy into adjacent frequency channels, violating regulatory masks (e.g., 3GPP). DPD is the primary mechanism to suppress this interference.

• Adjacent Channel Leakage Ratio (ACLR): DPD typically improves ACLR by 15-25 dB.
• Mask Compliance: Essential for meeting strict spectral emission requirements without resorting to expensive, bulky analog cavity filters.
• Efficiency Trade-off: DPD allows the PA to operate closer to its saturation point (compression), where efficiency is highest but linearity is worst.

PA behavior drifts over time due to temperature changes, voltage fluctuations, and component aging. A static DPD lookup table is insufficient; the system must adapt in real-time.

• Indirect Learning Architecture (ILA): The most common adaptation structure. It compares the attenuated PA output with the desired linear signal to update the pre-distorter coefficients.
• Direct Learning Architecture (DLA): Minimizes the error between the desired signal and the actual PA output directly, often requiring more complex optimization.
• Coefficient Update Rate: Typically operates on a millisecond scale to track thermal transients.

Traditional polynomial-based models (e.g., Generalized Memory Polynomial) struggle with strong non-linearities in wideband scenarios. Neural networks offer superior modeling capability for complex PA behaviors.

• Real-Valued Time-Delay Neural Networks (RVTDNN): A classic architecture that processes IQ components as separate real inputs.
• Augmented Real-Valued Time-Delay Neural Networks (ARVTDNN): Incorporates envelope-dependent terms to better capture AM-AM/AM-PM characteristics.
• Complex-Valued Neural Networks (CVNN): Process IQ data natively in the complex domain, preserving phase relationships and often requiring fewer parameters than real-valued equivalents.

DPD is often integrated with Crest Factor Reduction (CFR) to manage the signal's dynamic range before the PA. High PAPR forces the PA to operate with a large back-off, killing efficiency.

• Clipping and Filtering: A basic CFR technique that clips signal peaks, but DPD must then linearize the resulting in-band distortion.
• Pulse Injection: DPD can be combined with techniques that inject cancellation pulses to reduce peaks without excessive distortion.
• Joint Optimization: Modern systems jointly optimize CFR and DPD coefficients to maximize PA efficiency while maintaining EVM and ACLR targets.

The choice of adaptation architecture fundamentally impacts DPD performance and stability. The Indirect Learning Architecture (ILA) is favored for its robustness, while the Direct Learning Architecture (DLA) offers theoretical optimality.

• ILA Process: 1) Identify the post-distorter (inverse of PA) using PA output/input. 2) Copy coefficients to the pre-distorter. This avoids the need for a PA model.
• DLA Process: 1) Model the PA. 2) Backpropagate the error through the PA model to update the pre-distorter. Highly dependent on PA model accuracy.
• Stability: ILA is inherently more stable because it solves a linear-in-parameters problem, whereas DLA can suffer from convergence issues if the PA model is inaccurate.

The physical root cause requiring DPD. As a Power Amplifier (PA) approaches saturation to maximize efficiency, its gain compresses, creating AM-AM distortion (amplitude-dependent gain) and AM-PM distortion (amplitude-dependent phase shift). This non-linear transfer function generates spectral regrowth in adjacent channels and in-band distortion that degrades Error Vector Magnitude (EVM). DPD applies an inverse non-linearity to pre-distort the signal, such that the cascade of the pre-distorter and the PA results in a linear output.

A metric defining the ratio of a signal's peak power to its average power, expressed in dB. Modern modulation schemes like OFDM exhibit high PAPR (10-13 dB), forcing the PA to operate with a large back-off from its saturation point to avoid clipping and distortion. This back-off drastically reduces power efficiency. Crest Factor Reduction (CFR) is a complementary technique to DPD that intelligently clips signal peaks before the PA, reducing PAPR and allowing the amplifier to operate closer to saturation, maximizing efficiency while DPD handles the residual linearization.

A regulatory compliance metric measuring the ratio of transmitted power within the assigned channel to the power leaking into adjacent frequency channels. PA non-linearity causes spectral regrowth, which broadens the signal's bandwidth and creates interference in neighboring channels. DPD is the primary technique to suppress this regrowth and meet strict ACLR requirements (typically -45 dBc or better). Failure to meet ACLR limits results in regulatory non-compliance and denial of certification by bodies like the FCC.

The dominant adaptive DPD parameter identification method. Instead of directly modeling the PA's inverse, ILA places a copy of the pre-distorter in a feedback path. The error signal between the PA's normalized output and the pre-distorter copy's output drives an adaptive algorithm (e.g., LMS or RLS) to update the pre-distorter coefficients. This architecture avoids the need to explicitly compute the inverse of a non-linear model, which is mathematically ill-posed, and is robust to PA characteristic drift due to temperature and aging.