How does DPD know the inverse of the PA's distortion?

A feedback receiver captures a sample of the PA output, downconverts and digitizes it, then compares it to the original input signal. An adaptive algorithm (typically least-squares or recursive least-squares) continuously updates the predistorter coefficients to minimize the error between the desired linear output and the actual PA output. The model is re-trained every few milliseconds to track PA changes due to temperature drift, aging, and supply voltage variation. The feedback path requires about 40 dB of dynamic range and a bandwidth of 3 to 5 times the signal bandwidth to capture the distortion products.

What is the difference between memoryless and memory DPD?

Memoryless DPD models the PA as a static nonlinearity where the output at time t depends only on the input at time t. This corrects AM-AM and AM-PM compression but ignores memory effects caused by bias networks, thermal time constants, and trapping in GaN devices. Memory DPD (using memory polynomial or generalized memory polynomial models) adds delayed input samples, capturing these dynamic effects. Memory DPD typically provides 5 to 10 dB better ACPR correction than memoryless, and is essential for wideband signals (100+ MHz) where memory effects span multiple nanoseconds.

How much ACPR improvement does DPD provide?

A typical GaN Doherty PA without DPD achieves -33 to -38 dBc ACLR with a 5G NR 100 MHz signal. Memoryless polynomial DPD improves this to -48 to -52 dBc. Memory polynomial DPD reaches -55 to -58 dBc. Generalized memory polynomial (GMP) with cross-terms achieves -58 to -62 dBc. The 3GPP limit is -45 dBc, so even basic DPD provides adequate margin. The real benefit is allowing the PA to operate 2 to 4 dB closer to compression, improving average efficiency from 35% to 45-50% in a Doherty architecture.

Active Components

Digital Predistortion

DPD

A power amplifier compresses at high power levels, generating spectral regrowth that violates emission masks. The brute-force solution is to back off the PA far from compression, but that wastes 60 to 80% of the DC power as heat. DPD takes the opposite approach: it deliberately distorts the digital baseband signal with the mathematical inverse of the PA's nonlinearity before it reaches the DAC. When this pre-distorted signal passes through the compressing PA, the two nonlinearities cancel, producing a linear output from a PA running near saturation. The result is clean spectral emissions at efficiencies that were previously impossible.

Category: Active Components

ACPR Improvement: 15 to 25 dB

Feedback BW: 3 to 5× signal BW

Cancelling Distortion Before It Happens

DPD Model Complexity vs. Performance

DPD Model	Coefficients	ACPR (typical)	EVM	Compute Cost	Best For
No DPD	0	−33 to −38 dBc	5 to 8%	None	Constant-envelope only
LUT (lookup table)	256 to 1024	−45 to −48 dBc	2 to 3%	Very low	Narrowband, low cost
Memory polynomial	30 to 80	−52 to −56 dBc	1 to 2%	Moderate	Sub-6 GHz, 20 to 40 MHz
GMP (with cross-terms)	100 to 300	−58 to −62 dBc	0.5 to 1%	High	Wideband 5G NR (100+ MHz)
Neural network DPD	500+	−60 to −65 dBc	<0.5%	Very high	mmWave, multi-band

The Feedback Loop

DPD adaptation loop:
1. Baseband signal x(n) passes through predistorter: z(n) = f_DPD(x(n))
2. z(n) is upconverted, amplified by PA: y(t) = f_PA(z(t))
3. Feedback receiver samples y(t), digitizes to y(n)
4. Error: e(n) = y(n) − G·x(n) where G is the desired linear gain
5. Coefficient update: minimize ||e(n)||² using LS or RLS algorithm
6. Update f_DPD coefficients every 1 to 10 ms

Feedback receiver requirements:
Bandwidth: 3 to 5× signal BW (captures IMD3 and IMD5)
Dynamic range: ≥40 dB
Sample rate: ≥5× signal BW (for 100 MHz NR: 500 MSPS observation receiver)

Common Questions

Frequently Asked Questions

How does DPD learn the PA's inverse?

A feedback receiver digitizes the PA output and compares it to the original input. An adaptive algorithm (LS/RLS) continuously updates the predistorter coefficients to minimize the error. Re-training occurs every few milliseconds to track temperature, aging, and supply drift.

Memoryless vs. memory DPD?

Memoryless corrects static AM-AM/AM-PM only. Memory DPD adds delayed samples to capture bias network, thermal, and GaN trapping effects. Memory DPD provides 5 to 10 dB better ACPR and is essential for 100+ MHz signals where memory spans multiple nanoseconds.

How much ACPR improvement?

15 to 25 dB depending on model complexity. A GaN Doherty goes from −35 dBc (no DPD) to −58 dBc (GMP DPD). The real benefit: operating 2 to 4 dB closer to compression boosts average efficiency from 35% to 45 to 50%.

Related Terms

← Digital Filter Doherty PA →

Linearization

DPD Coefficient Estimator

Upload your PA AM-AM and AM-PM measurement data and generate memory polynomial coefficients for direct implementation in your FPGA or DSP.

Estimate Coefficients