S
Seed account
Digital predistortion pre-distorts the baseband signal with the inverse of the power amplifier's nonlinearity, so the cascade of predistorter and PA is approximately linear.
How it works
1. Observation: a feedback receiver samples the PA output through a coupler and downconverts it to baseband.
2. Modelling: the system compares PA output with the intended signal and estimates a model of the PA.
3. Inversion: the predistorter applies the inverse model to the transmitted signal. Two common architectures are indirect learning (fit a post-inverse, then copy it in front) and direct learning (iteratively adjust the predistorter to minimize output error).
Models
- Memoryless models (lookup tables for AM-AM and AM-PM) handle static compression.
- Memory polynomial and generalized memory polynomial models include delayed terms to capture memory effects from bias networks, thermal behaviour and device physics.
- Volterra-series-based models are more general but more complex; pruned forms are common in practice.
- Neural-network-based models are an active research area.
Practical constraints
- Bandwidth: the predistorted signal has spectral regrowth several times the signal bandwidth, so the DAC, transmit chain and observation receiver need extra bandwidth.
- Observation path accuracy limits achievable correction.
- DPD cannot fix hard clipping: once the PA saturates, no input can produce higher output. Crest factor reduction is usually applied before DPD.
- Adaptation must track temperature, frequency and power changes.
- Antenna load variation in arrays and mismatch changes PA behaviour, complicating DPD for massive MIMO.
Where to start learning
Open literature from IEEE MTT-S on behavioural modelling and DPD is extensive. For hands-on practice, a vector signal generator, a PA evaluation board and a signal analyzer with capture capability let you build a simple memory polynomial DPD in Python or MATLAB.
How it works
1. Observation: a feedback receiver samples the PA output through a coupler and downconverts it to baseband.
2. Modelling: the system compares PA output with the intended signal and estimates a model of the PA.
3. Inversion: the predistorter applies the inverse model to the transmitted signal. Two common architectures are indirect learning (fit a post-inverse, then copy it in front) and direct learning (iteratively adjust the predistorter to minimize output error).
Models
- Memoryless models (lookup tables for AM-AM and AM-PM) handle static compression.
- Memory polynomial and generalized memory polynomial models include delayed terms to capture memory effects from bias networks, thermal behaviour and device physics.
- Volterra-series-based models are more general but more complex; pruned forms are common in practice.
- Neural-network-based models are an active research area.
Practical constraints
- Bandwidth: the predistorted signal has spectral regrowth several times the signal bandwidth, so the DAC, transmit chain and observation receiver need extra bandwidth.
- Observation path accuracy limits achievable correction.
- DPD cannot fix hard clipping: once the PA saturates, no input can produce higher output. Crest factor reduction is usually applied before DPD.
- Adaptation must track temperature, frequency and power changes.
- Antenna load variation in arrays and mismatch changes PA behaviour, complicating DPD for massive MIMO.
Where to start learning
Open literature from IEEE MTT-S on behavioural modelling and DPD is extensive. For hands-on practice, a vector signal generator, a PA evaluation board and a signal analyzer with capture capability let you build a simple memory polynomial DPD in Python or MATLAB.