Convergence Rate

The choice of estimation algorithm fundamentally dictates the speed-accuracy tradeoff.

• Stochastic Gradient Descent (LMS): Slow, linear convergence. Simple but struggles with correlated inputs.
• Recursive Least Squares (RLS): An order of magnitude faster convergence than LMS, as it whitens the input data by inverting the correlation matrix. The cost is O(N²) complexity.
• QR-RLS: Maintains fast convergence with superior numerical stability using Givens rotations, ideal for fixed-point FPGA implementations.

The condition number of the input signal's autocorrelation matrix is the dominant physical factor.

• A high eigenvalue spread (ratio of λ_max to λ_min) indicates a highly correlated input, such as a narrowband LTE signal.
• LMS convergence slows dramatically as the eigenvalue spread increases.
• RLS is theoretically immune to eigenvalue spread because it explicitly estimates the inverse correlation matrix, making it the preferred choice for spectrally colored signals.

In gradient-based algorithms like LMS and NLMS, the step size μ creates a direct tradeoff.

• Large μ: Fast initial convergence but large steady-state misadjustment and potential instability.
• Small μ: Slow convergence but low excess mean squared error.
• Normalized LMS (NLMS) adapts μ inversely to input power, providing stability across fluctuating signal levels without manual tuning.

In RLS-based algorithms, the forgetting factor λ (0 ≪ λ ≤ 1) controls the memory of the estimator.

• λ close to 1: Deep memory, low steady-state noise, but slow response to amplifier characteristic changes due to thermal drift.
• Smaller λ: Fast tracking of time-varying nonlinearities but higher misadjustment.
• Typical values for tracking thermal memory effects range from 0.97 to 0.995.

The number of estimated coefficients directly impacts convergence time.

• A Memory Polynomial with nonlinearity order K=7 and memory depth M=4 requires 28 coefficients.
• A Generalized Memory Polynomial (GMP) can require hundreds of coefficients, increasing the parameter space and slowing convergence.
• Over-parameterization not only slows convergence but increases the risk of overfitting to measurement noise.

The structural approach to coefficient estimation affects convergence behavior.

• Indirect Learning Architecture (ILA): Assumes the postdistorter inverse is equivalent to the predistorter. Convergence is fast but sensitive to PA output noise.
• Direct Learning Architecture (DLA): Closes the loop around the actual PA, converging to the true optimal predistorter. Requires careful loop gain management to prevent instability during adaptation.

A foundational stochastic gradient descent algorithm that updates filter coefficients iteratively based on the instantaneous estimate of the mean squared error gradient.

• Convergence behavior: Linear convergence rate heavily dependent on the eigenvalue spread of the input correlation matrix
• Step-size trade-off: Larger step sizes accelerate convergence but increase steady-state misadjustment
• Computational cost: O(N) per iteration, making it suitable for resource-constrained FPGA implementations
• Practical limit: Convergence slows dramatically for wideband signals with high spectral dynamic range

An adaptive algorithm that recursively updates the inverse of the input correlation matrix to achieve quadratic convergence — significantly faster than gradient-based methods.

• Key advantage: Convergence rate is independent of the input signal's eigenvalue spread
• Cost of speed: O(N²) computational complexity per iteration due to matrix inversion updates
• Forgetting factor λ: Controls the algorithm's memory; smaller values enable faster tracking of non-stationary amplifier characteristics at the expense of increased steady-state variance
• Numerical stability: Standard RLS can suffer from covariance matrix blow-up; QR-RLS variants mitigate this

A variant of LMS that normalizes the step size by the instantaneous input signal power, improving convergence stability in the presence of fluctuating signal levels.

• Mechanism: Step size μ is divided by ||x(n)||² + ε, where ε is a small regularization constant preventing division by zero
• Convergence benefit: Reduces gradient noise amplification during high-power signal bursts, maintaining a more consistent convergence trajectory
• Trade-off: Slightly higher computational cost than standard LMS due to the norm calculation, but still O(N)
• Application: Widely used in DPD systems where signal crest factor varies significantly with traffic load

A scalar parameter in recursive algorithms that exponentially weights recent data more heavily than past data, enabling the algorithm to track time-varying systems.

• Range: Typically 0.95 < λ < 1.0 for DPD applications; λ = 1.0 corresponds to infinite memory
• Effective memory: The asymptotic sample memory is approximately 1/(1-λ) iterations
• Convergence vs. tracking: A smaller λ accelerates convergence and improves tracking of thermal memory effects, but increases steady-state parameter variance
• Design rule: Select λ based on the expected rate of amplifier characteristic drift due to temperature and aging

The normalized difference between the steady-state mean squared error of an adaptive filter and the minimum MSE achievable by the optimal Wiener filter.

• Definition: M = (MSE_∞ - MSE_min) / MSE_min, quantifying the penalty for continuous adaptation
• Convergence trade-off: Faster convergence (larger step size or smaller forgetting factor) directly increases misadjustment
• LMS relationship: Misadjustment is proportional to μ · tr(R), where R is the input correlation matrix
• RLS relationship: Misadjustment is approximately (1-λ)N/2, where N is the number of coefficients
• DPD significance: Excess misadjustment manifests as residual spectral regrowth, directly degrading ACLR performance

The ratio of the largest to smallest singular value of a matrix, quantifying the sensitivity of the solution of a linear system to small perturbations in the input data.

• Impact on convergence: Gradient-based algorithms like LMS converge slowly when the input autocorrelation matrix has a high condition number (large eigenvalue spread)
• RLS immunity: RLS convergence rate is theoretically independent of condition number, making it preferred for ill-conditioned problems
• DPD context: Wideband signals with non-uniform spectral power distribution produce ill-conditioned correlation matrices, challenging LMS convergence
• Mitigation: Regularization parameters and orthogonalizing transforms (DCT, KLT) can reduce effective condition number