Convergence Rate

The choice of adaptive algorithm fundamentally dictates the convergence trajectory. Recursive Least Squares (RLS) offers an order of magnitude faster convergence than Least Mean Squares (LMS) due to its use of the inverse autocorrelation matrix, but at significantly higher computational cost. QR-RLS variants provide superior numerical stability for ill-conditioned signals, preventing divergence during initial acquisition. The trade-off between convergence speed and computational complexity must be evaluated against the coherence time of the PA nonlinearity.

These parameters directly control the stability-speed trade-off. A large step size (μ) in LMS accelerates convergence but increases misadjustment—the excess steady-state error caused by gradient noise. Conversely, a small step size yields precise steady-state performance but sluggish tracking. In RLS, the forgetting factor (λ) weights recent data more heavily; a smaller λ enables faster tracking of time-varying PA behavior but amplifies noise sensitivity and risks coefficient drift during low-excitation periods.

The condition number of the input signal's autocorrelation matrix is the dominant physical limit on convergence speed. A high eigenvalue spread—common in narrowband or highly correlated wideband signals—creates an ill-conditioned problem where different modes converge at vastly different rates. Tikhonov regularization and explicit whitening filters compress the eigenvalue spread, dramatically accelerating convergence. Without proper conditioning, the algorithm may stall before reaching the Wiener solution.

The granularity of coefficient updates creates a latency-accuracy trade-off. Sample-by-sample updates provide the fastest tracking of dynamic distortion but introduce high gradient noise and computational overhead. Block updates average gradient estimates over a data buffer, reducing noise and enabling efficient matrix operations, but introduce a processing delay. Burst training confines adaptation to dedicated intervals, preventing perturbation of active payload transmission but leaving the system vulnerable to inter-burst drift.

The number of predistorter coefficients directly impacts convergence time. Higher-order memory polynomial models with deep memory taps require more iterations to converge due to the expanded parameter space. Excessively complex models risk overfitting to training data noise rather than learning the true PA inverse, degrading generalization on unseen signals. Regularization techniques and cross-validation during offline training constrain model complexity to match the true nonlinear order of the amplifier.

The transmit observation receiver's signal-to-noise ratio and linearity set a fundamental floor on achievable convergence. Quantization noise, local oscillator phase noise, and IQ imbalance in the feedback path inject errors into the error signal used for adaptation. These impairments create a misadjustment floor that no algorithmic tuning can overcome. High-fidelity observation paths with sufficient dynamic range are prerequisites for rapid, accurate coefficient convergence in closed-loop DPD systems.

A foundational stochastic gradient descent algorithm that updates coefficients using the instantaneous gradient estimate. Its convergence rate is directly tied to the step-size parameter (μ).

• Trade-off: Larger μ accelerates convergence but increases steady-state misadjustment.
• Sensitivity: Convergence slows dramatically if the input signal's eigenvalue spread is large.
• Computational Cost: Very low, requiring only O(N) operations per iteration, making it suitable for resource-constrained FPGAs.

An adaptive algorithm that minimizes a weighted linear least squares cost function recursively. It achieves an order-of-magnitude faster convergence than LMS, especially for signals with high eigenvalue spread.

• Mechanism: Replaces the scalar step-size with an inverse correlation matrix estimate.
• Cost: O(N²) complexity per iteration, demanding significantly more hardware resources.
• Instability Risk: Can suffer from coefficient drift due to numerical precision issues in the matrix update; often requires QR decomposition for robust implementation.

A mathematical property of the input signal's autocorrelation matrix that dictates the theoretical convergence speed limit of gradient-based algorithms.

• Definition: The ratio of the largest to smallest eigenvalue (λmax/λmin).
• Impact: A high condition number (ill-conditioned matrix) creates a steep, narrow error surface, forcing LMS to converge slowly.
• Mitigation: Tikhonov regularization or switching to RLS can stabilize convergence when the condition number is poor.

The excess Mean Squared Error (MSE) that persists after convergence, representing the penalty paid for using a stochastic approximation rather than the true gradient.

• Relationship: Proportional to the step-size (μ) and the number of coefficients.
• Trade-off: A faster convergence rate (high μ) inevitably produces a higher misadjustment floor.
• Evaluation: Measured by comparing the converged NMSE against the theoretical Wiener solution minimum.

Two distinct strategies for applying coefficient updates that directly affect perceived convergence speed and computational latency.

• Sample-by-Sample: Updates coefficients every clock cycle. Maximizes tracking speed for fast-changing PA thermal effects but requires high-throughput hardware.
• Block Update: Accumulates a buffer of samples before solving for coefficients. Reduces computational overhead and gradient noise, leading to a smoother convergence trajectory at the cost of update latency.
• Burst Training: A specialized block mode where training occurs only during dedicated preamble slots, freezing coefficients during payload transmission.

The primary metric for quantifying convergence rate and steady-state accuracy in DPD systems. It measures the residual distortion power normalized by the input signal power.

• Convergence Plot: Engineers monitor the NMSE learning curve; the slope indicates convergence rate, and the asymptotic floor indicates misadjustment.
• Target: 5G systems typically require NMSE convergence below -35 dB to meet ACPR specifications.
• Formula: NMSE = 10·log₁₀(E[|y_ideal - y_measured|²] / E[|x|²])