Normalized LMS (NLMS)

The core innovation of NLMS is dividing the step size μ by the squared Euclidean norm of the input vector ||u(n)||² plus a small regularization parameter δ.\n\n- Update Equation: w(n+1) = w(n) + [μ / (δ + ||u(n)||²)] · u(n) · e*(n)\n- Effect: This makes the effective step size inversely proportional to the instantaneous signal power, preventing instability during high-amplitude bursts.\n- Regularization: The small constant δ (typically 10⁻⁶ to 10⁻⁴) prevents division by zero during silent periods and ensures numerical stability.

Standard LMS suffers from gradient noise amplification because the correction term scales directly with the input vector magnitude. NLMS solves this by normalizing the correction direction to unit length.\n\n- Problem: In LMS, a large ||u(n)|| produces an excessively large weight update, potentially causing divergence.\n- Solution: NLMS effectively projects the gradient onto a unit-norm direction, making the update magnitude independent of input scaling.\n- Result: The algorithm converges reliably even with signals exhibiting high peak-to-average power ratios (PAPR), such as modern OFDM waveforms.

NLMS exhibits faster convergence than standard LMS for colored or non-stationary inputs while maintaining low computational complexity.\n\n- Convergence Rate: Approaches the optimal Wiener solution in approximately 2N to 4N iterations for white inputs, where N is the filter length.\n- Misadjustment: The steady-state excess MSE is proportional to μ/2, independent of input power, unlike LMS where misadjustment scales with input power.\n- Trade-off: Larger μ accelerates convergence but increases misadjustment. Typical values range from 0.1 to 1.0 for stable operation.

NLMS adds minimal overhead compared to standard LMS, making it suitable for real-time embedded implementation on FPGAs and DSPs.\n\n- Operations per iteration: 2N + 3 multiplications and 2N + 2 additions for a length-N filter.\n- Overhead vs LMS: Only 2 additional multiplications, 1 addition, and 1 division for the norm calculation.\n- Optimization: The squared norm can be computed recursively: ||u(n)||² = ||u(n-1)||² + |u(n)|² - |u(n-N)|², reducing complexity to O(1) per sample.

NLMS provides a deterministic stability bound that is independent of the input signal statistics, unlike LMS which requires knowledge of the input autocorrelation matrix.\n\n- Stability Condition: The algorithm converges in the mean-square sense for any step size satisfying 0 < μ < 2.\n- Robustness: This bound holds for arbitrary input sequences, including non-stationary and chaotic signals.\n- Practical Range: For safety, μ is typically chosen between 0.1 and 1.0 to balance convergence speed with steady-state error.

NLMS is widely used in indirect learning architectures (ILA) for DPD coefficient adaptation due to its robustness to the varying envelope of communication signals.\n\n- Signal Dynamics: Modern 5G and WiFi signals have PAPR exceeding 10 dB, causing standard LMS to become unstable during peaks.\n- NLMS Advantage: The normalization automatically reduces the step size during high-power samples, preventing predistorter coefficient divergence.\n- Implementation: Often combined with delay estimation and IQ imbalance compensation in complete DPD systems for power amplifier linearization.

The foundational stochastic gradient descent algorithm from which NLMS is derived. LMS updates filter coefficients iteratively using the instantaneous estimate of the mean squared error gradient. Its primary limitation is sensitivity to input signal scaling—large input vectors cause gradient noise amplification and potential instability. NLMS directly addresses this by normalizing the step size by the input vector power, making convergence behavior independent of signal level fluctuations common in communication waveforms.

A computationally intensive alternative that recursively updates the inverse of the input correlation matrix to achieve an order of magnitude faster convergence than NLMS. RLS minimizes a weighted least squares cost function with exponential forgetting, making it superior for tracking rapidly time-varying PA characteristics. The trade-off is O(N²) complexity versus NLMS's O(N), limiting RLS deployment to offline model extraction or systems with substantial DSP headroom.

The broader optimization family encompassing both LMS and NLMS. SGD updates parameters using gradients computed from single samples or mini-batches rather than the full dataset, enabling online learning. In DPD contexts, the key distinction is that NLMS employs an adaptive per-sample learning rate (μ / ||x(n)||²), while standard SGD typically uses a fixed or scheduled learning rate. This normalization is critical when input signal statistics vary with modulation scheme and traffic load.

The dominant deployment architecture for NLMS-based DPD. ILA places a postdistorter copy after the PA and trains it to invert the PA's nonlinear response. NLMS adapts the postdistorter coefficients, which are then copied to the predistorter. This avoids the nonlinear feedback path problem of direct architectures. The normalized step size is particularly advantageous in ILA because the postdistorter input (PA output) exhibits power-dependent scaling that would destabilize unnormalized LMS.

A numerically robust RLS implementation using Givens rotations to directly update the square-root of the inverse correlation matrix. QR-RLS avoids the covariance matrix ill-conditioning that plagues standard RLS in fixed-point FPGA implementations. While NLMS remains preferred for low-complexity real-time adaptation, QR-RLS serves as the gold standard for offline PA model extraction where numerical precision and convergence speed outweigh computational cost. The two algorithms represent opposite ends of the complexity-performance Pareto frontier.

A scalar δ added to the input power estimate in the NLMS denominator: μ / (δ + ||x(n)||²). This prevents division by zero during silent intervals and controls the maximum effective step size when input power is low. In DPD applications, proper regularization tuning prevents coefficient divergence during signal nulls in OFDM waveforms. The parameter creates a bias-variance trade-off—too large δ slows convergence, too small risks instability during low-power transmission periods.