Least Mean Squares (LMS)

LMS updates filter coefficients using the instantaneous estimate of the gradient of the mean squared error cost function. Unlike batch methods, it processes one sample at a time:

• Update Rule: w(n+1) = w(n) + μ * e(n) * x(n)
• e(n): Instantaneous error between desired and actual output
• x(n): Input signal vector
• μ: Step-size parameter controlling convergence speed and stability This sample-by-sample operation makes it ideal for real-time DPD coefficient tracking.

The step-size parameter μ governs a fundamental performance trade-off:

• Large μ: Faster convergence but larger steady-state misadjustment and potential instability
• Small μ: Lower misadjustment and smoother steady-state behavior but slower convergence
• Stability Bound: 0 < μ < 2 / λ_max, where λ_max is the largest eigenvalue of the input autocorrelation matrix In DPD applications, μ must be carefully tuned to track PA thermal memory effects without introducing excessive gradient noise.

Misadjustment is the excess mean squared error beyond the theoretical Wiener optimum, caused by gradient noise in stochastic updates:

• Source: LMS uses noisy instantaneous gradient estimates instead of true ensemble averages
• Proportional to μ: Misadjustment increases linearly with step-size
• Steady-state MSE: ξ_steady = ξ_min + ξ_excess
• Trade-off: Faster convergence (higher μ) directly increases the noise floor of coefficient estimates In DPD, excessive misadjustment manifests as residual spectral regrowth and degraded ACPR performance.

LMS convergence speed is governed by the eigenvalue spread of the input autocorrelation matrix:

• Eigenvalue Spread: Ratio of λ_max to λ_min determines convergence time
• Highly Correlated Inputs: Signals with large eigenvalue spreads (e.g., narrowband or highly oversampled waveforms) slow convergence dramatically
• Comparison to RLS: Recursive Least Squares achieves faster convergence independent of eigenvalue spread but at O(N²) complexity
• DPD Implication: For wideband signals with flat spectra, LMS converges efficiently; for narrowband training tones, convergence may be unacceptably slow.

The noisy gradient estimates in LMS introduce two related phenomena:

• Gradient Noise: Random perturbations in coefficient updates that prevent exact convergence to the Wiener solution
• Coefficient Drift: Gradual wandering of coefficients from optimal values, exacerbated by:
  • Finite-precision arithmetic in fixed-point FPGA implementations
  • Low-level persistent excitation noise
  • Thermal variations in PA characteristics
• Mitigation: Leakage LMS adds a small leakage factor γ to the update: w(n+1) = γ*w(n) + μ*e(n)*x(n) to bound coefficient magnitudes and prevent drift.

An adaptive filtering algorithm that recursively finds the coefficients minimizing a weighted linear least squares cost function. Unlike LMS, RLS uses all available input data to compute the inverse of the autocorrelation matrix, resulting in faster convergence—typically an order of magnitude quicker than LMS. The trade-off is higher computational complexity: O(N²) versus O(N) for LMS. RLS is particularly effective for tracking time-varying PA nonlinearities in rapidly changing signal conditions.

An iterative optimization method that updates model parameters using the gradient of the loss function computed on a single or small batch of samples. LMS is a specific application of SGD to linear adaptive filtering. Key characteristics include:

• Noisy gradient estimates introduce misadjustment but help escape local minima
• Learning rate critically controls convergence speed vs. steady-state error trade-off
• Forms the foundation for mini-batch training in neural network-based DPD architectures

A measure of the sensitivity of a matrix to numerical errors, defined as the ratio of the largest to smallest singular value. In DPD coefficient estimation, a high condition number in the input signal's autocorrelation matrix indicates ill-conditioning, leading to:

• Amplified numerical noise in coefficient estimates
• Slow convergence or divergence in LMS and RLS
• Sensitivity to finite-precision arithmetic in FPGA implementations

Tikhonov regularization is commonly applied to reduce the effective condition number.

A numerically stable implementation of the RLS algorithm using QR decomposition to solve the least squares problem. By decomposing the data matrix into an orthogonal matrix Q and upper triangular matrix R, QR-RLS avoids explicit computation of the autocorrelation matrix inverse. Benefits include:

• Superior numerical robustness against ill-conditioning
• Well-suited for fixed-point FPGA implementation
• Maintains stability during periods of low input signal power

Often preferred over standard RLS in production DPD systems where numerical precision is constrained.

The excess mean squared error in an adaptive system beyond the theoretical Wiener optimum, caused by gradient noise in stochastic coefficient updates. In LMS-based DPD training:

• Misadjustment is proportional to the step size μ
• Larger μ accelerates convergence but increases steady-state misadjustment
• Represents the fundamental convergence-vs-accuracy trade-off

For DPD applications, misadjustment directly translates to residual nonlinear distortion and degraded ACPR performance.

A robust iterative optimization algorithm that interpolates between Gauss-Newton and gradient descent methods to solve nonlinear least squares curve-fitting problems. For DPD applications:

• Provides quadratic convergence near the optimum for memory polynomial coefficient extraction
• A damping parameter dynamically switches behavior based on error reduction
• Particularly effective for offline model extraction where computational cost is less constrained
• Handles the nonlinear parameter estimation required for Volterra series-based predistorters