Error Vector Magnitude (EVM)

EVM is defined as the magnitude of the difference vector between the ideal reference constellation point and the actual measured signal point, normalized by the power of the ideal reference. It is typically expressed as a percentage or in decibels (dB).

• Formula: EVM = |S_measured - S_ideal| / |S_ideal|
• RMS EVM: The root-mean-square average over all symbols in a frame, providing a single quality figure.
• Peak EVM: The maximum instantaneous error, critical for identifying rare but severe distortion events.
• Normalization: Always normalized to the ideal signal power to make it a relative, dimensionless metric.

While both measure distortion, EVM and Adjacent Channel Power Ratio (ACPR) characterize fundamentally different aspects of linearity.

• EVM (In-Band): Quantifies distortion within the intended signal bandwidth, directly impacting bit error rate (BER) and data throughput.
• ACPR (Out-of-Band): Quantifies spectral regrowth into adjacent channels, impacting regulatory compliance and interference.
• Joint Optimization: Modern DPD systems must simultaneously minimize both EVM and ACPR, often using a weighted multi-objective cost function.
• Diagnostic Value: A system with good ACPR but poor EVM suggests a different root cause than one with poor ACPR but good EVM.

Precise EVM measurement requires a calibrated vector signal analyzer (VSA) and a low-noise transmit observation path.

• Reference Generation: The ideal signal must be reconstructed from the demodulated bits or known test patterns.
• Time Alignment: Sub-sample time alignment between the reference and measured waveforms is critical; a misalignment of even a fraction of a sample introduces artificial EVM.
• Equalization: Linear channel impairments (e.g., flat fading) must be equalized out before EVM computation to isolate PA nonlinearity.
• Averaging: RMS EVM is computed over a statistically significant number of frames to ensure a stable, repeatable measurement.

The acceptable EVM threshold is a direct function of the modulation order. Higher-order QAM constellations have tighter spacing between points, demanding far lower EVM.

• QPSK: Tolerates EVM up to ~17.5%.
• 16-QAM: Requires EVM below ~12.5%.
• 64-QAM: Requires EVM below ~8%.
• 256-QAM: Requires EVM below ~3.5%.
• 1024-QAM: Demands EVM below ~1%, pushing the limits of DPD and PA linearity.

In adaptive DPD systems, EVM is often the direct cost function minimized by the coefficient estimation algorithm.

• Stochastic Gradient Descent: Coefficients are updated iteratively to minimize the instantaneous squared error magnitude.
• Least Squares (LS): A block of samples is used to solve for coefficients that minimize the sum of squared errors, equivalent to minimizing the mean squared EVM.
• Regularization: Tikhonov regularization is often added to the LS cost function to prevent coefficient drift and improve numerical stability, trading a slight increase in EVM for robustness.
• Convergence Monitoring: The rate at which EVM decreases during training indicates the convergence speed of the adaptive algorithm.

A critical regulatory metric quantifying spectral regrowth caused by power amplifier nonlinearity. ACPR measures the ratio of power leaking into adjacent frequency channels relative to the main channel power.

• Out-of-band distortion indicator
• Directly linked to ACLR (Adjacent Channel Leakage Ratio)
• Poor ACPR causes interference with neighboring carriers
• DPD aims to suppress ACPR by 15-25 dB in typical deployments

A standard in-band distortion metric that quantifies the deviation between the ideal linear output and the actual linearized signal, normalized by input signal power. NMSE is the primary cost function minimized during DPD coefficient training.

• Expressed in dB, with lower values indicating better linearization
• Typical DPD systems achieve NMSE below -35 dB
• Directly correlates with EVM performance
• Used as the cost function in least squares and LMS-based coefficient estimation

A closed-loop DPD training architecture that directly estimates predistorter coefficients by minimizing the error between the desired input signal and the actual PA output. DLA models the PA forward behavior and mathematically inverts it.

• Requires accurate PA behavioral modeling
• More robust to measurement noise than ILA
• Enables model inversion techniques
• Typically achieves superior EVM compared to indirect methods
• Computationally more intensive due to real-time model extraction

A postdistorter-based DPD training approach where a postdistorter is placed after the PA to identify the inverse transfer function. The trained coefficients are then copied to the predistorter.

• Assumes commutativity of the PA and predistorter
• Simpler implementation than DLA
• Susceptible to noise enhancement in the feedback path
• Widely used in commercial base station DPD systems
• EVM performance may degrade under strong nonlinearity

Two competing performance characteristics of adaptive DPD algorithms. Convergence rate defines how quickly coefficients reach steady-state, while misadjustment quantifies the excess error beyond the theoretical Wiener optimum.

• RLS offers faster convergence than LMS at higher computational cost
• Stochastic gradient noise causes misadjustment in LMS
• Burst training trades convergence speed for stability
• EVM transients occur during coefficient re-convergence
• Tikhonov regularization stabilizes convergence in ill-conditioned scenarios