How does digital predistortion work?

DPD operates in a closed-loop system with two paths. The forward path applies a predistortion function to the baseband signal before the DAC and upconverter. This function is designed to be the mathematical inverse of the PA's nonlinear transfer function, so that the cascade of predistorter plus PA produces a linear overall response. The feedback path uses a coupler to sample the PA output, downconverts it, and digitizes it with an ADC running at 3 to 5 times the signal bandwidth (to capture the spectral regrowth products). An adaptive algorithm continuously compares the feedback signal to the original input, computes the PA's current nonlinear behavior, and updates the predistortion coefficients. The model must capture both static nonlinearity (AM-AM gain compression, AM-PM phase distortion) and dynamic effects (memory from thermal, bias network, and matching network time constants). The feedback bandwidth is typically 3 to 5 times the signal bandwidth: for a 100 MHz 5G NR signal, the DPD feedback ADC runs at 300 to 500 MHz sample rate.

What DPD models are used?

The memory polynomial (MP) is the most common model: y(n) = sum over k and m of a(k,m) times x(n-m) times |x(n-m)|^k, where k is the nonlinear order (typically 3 to 9) and m is the memory depth (0 to 5 samples). The generalized memory polynomial (GMP) adds cross-terms between different memory taps, capturing more complex memory effects with better accuracy but more coefficients. For a typical 7th-order, depth-3 MP model, there are 12 complex coefficients. A GMP model might have 50 to 100 coefficients. Neural network DPD uses a small feedforward network (1 to 2 hidden layers, 10 to 50 neurons) trained by backpropagation, capturing arbitrary nonlinear behaviors that polynomial models may miss. The model coefficients are computed using least squares (LS) or recursive least squares (RLS) adaptation, typically updating every 1 to 10 milliseconds to track thermal drift and aging.

How much improvement does DPD provide?

Without DPD, a Doherty PA operating at 3 dB compression typically has ACLR of minus 25 to minus 30 dBc and EVM of 5 to 8 percent, far from the 3GPP requirements of minus 45 dBc ACLR and less than 3 percent EVM. DPD improves ACLR by 15 to 25 dB (to minus 45 to minus 55 dBc) and EVM to less than 1 to 2 percent. This linearization enables the PA to operate 2 to 4 dB closer to compression, increasing average efficiency by 10 to 15 percentage points. For a 5G NR base station with 200 MHz instantaneous bandwidth, the DPD feedback path requires a 600 to 1000 MHz ADC and the predistorter runs at the same rate, consuming 5 to 15 watts of FPGA power, but saving 100 to 300 watts of PA DC power through the efficiency improvement. The net energy saving is substantial, which is why DPD is standard in every modern base station.

Transmitter Linearization

Digital Predistortion

/dij-ih-tul pree-dis-tor-shun/ — DPD

Baseband signal processing that pre-compensates PA nonlinearity by applying an inverse distortion function before the DAC. A feedback ADC captures PA output at 3-5x signal bandwidth. Adaptive algorithms (LMS/RLS) train behavioral models (memory polynomial, GMP, neural network) to characterize AM-AM/AM-PM + memory effects. Improves ACLR by 15-25 dB, enabling Doherty PAs to operate 2-4 dB closer to compression while meeting 3GPP linearity specs.

ACLR gain: 15-25 dB

EVM: <1-2%

FB BW: 3-5x signal

Understanding Digital Predistortion

DPD is the enabling technology that makes efficient Doherty PAs viable for modern cellular systems. Without DPD, a Doherty PA operating near compression produces unacceptable spectral regrowth (ACLR of -25 to -30 dBc) and in-band distortion (EVM of 5-8%). With DPD, the same PA achieves ACLR of -45 to -55 dBc and EVM <2%, meeting or exceeding 3GPP requirements. This linearization costs 5-15 W of FPGA power but saves 100-300 W of PA DC power through the efficiency improvement, a compelling tradeoff.

The key challenge in DPD is accurately modeling the PA's nonlinear behavior, which includes not only static gain compression and phase distortion but also dynamic memory effects. Memory effects arise from thermal time constants (microseconds), bias network impedance variations with frequency (nanoseconds), and matching network bandwidth limitations. A memoryless polynomial model captures the static nonlinearity but fails for wideband signals. Memory polynomial and GMP models add delayed input terms to capture the time-dependent behavior, typically requiring 3-5 memory taps at 3-5x oversampling.

DPD Model Equations

Memory polynomial (MP):
y(n) = Σ_kΣ_m a_k,m x(n−m)|x(n−m)|^k
k = nonlinear order (1,3,5,7,9)
m = memory depth (0 to 4)
Typical: 5 orders × 5 taps = 25 coeff

GMP (generalized):
y = MP + Σ b_k,m,l x(n−m)|x(n−m−l)|^k
+ Σ c_k,m,l x(n−m)|x(n−m+l)|^k
Cross-terms: 50-100 coefficients

Coefficient extraction:
a = (X^HX)⁻¹ X^Hy (LS solution)
Update rate: 1-10 ms (thermal tracking)

Feedback path requirement:
BW_FB = (2K+1) × BW_signal
K=3 (7th order): FB = 7× signal BW
100 MHz 5G: 700 MHz ADC sample rate

DPD Model Comparison

Model	Coefficients	ACLR Improvement	Complexity	Memory	Application
LUT (memoryless)	256-1024	10-15 dB	Low	None	CW/narrow
Memory polynomial	15-30	15-20 dB	Medium	3-5 taps	LTE single
GMP	50-100	20-25 dB	Medium-high	Cross-terms	5G wideband
Neural network	100-500	20-30 dB	High	Arbitrary	Ultra-wideband
Dual-band DPD	100-200	15-20 dB/band	Very high	Cross-band	Multi-band TX

Common Questions

Frequently Asked Questions

How does DPD work?

Forward path: predistortion function (inverse of PA nonlinearity) applied to baseband before DAC. Feedback path: coupler samples PA output, downconverts, ADC at 3-5x signal BW captures spectral regrowth. Adaptive algorithm (LS/RLS) continuously updates model coefficients by comparing feedback to input. Model captures AM-AM, AM-PM, and memory effects. Updates every 1-10 ms for thermal tracking.

What models are used?

Memory polynomial: y(n) = Σa_km × x(n-m)|x(n-m)|^k. 15-30 coefficients, 15-20 dB ACLR improvement. GMP adds cross-terms for 20-25 dB improvement with 50-100 coefficients. Neural network DPD (1-2 hidden layers) captures arbitrary nonlinearities, 20-30 dB improvement. Coefficients extracted via least squares from captured PA input/output data.

How much improvement?

Without DPD: Doherty at 3 dB compression has ACLR -25 to -30 dBc, EVM 5-8%. With DPD: ACLR -45 to -55 dBc, EVM <1-2%. Enables 2-4 dB closer to compression, +10-15% average efficiency. 5-15 W FPGA power cost saves 100-300 W PA DC power. Standard in every modern base station. 100 MHz 5G signal needs ~700 MHz feedback ADC.

Related Terms

← Doherty Amplifier Droop →