Normalized Least Mean Squares (NLMS)

The defining characteristic of NLMS is the normalization of the step size by the squared Euclidean norm of the input vector. The coefficient update equation is:

w(n+1) = w(n) + (μ / (||x(n)||² + ε)) * e*(n) * x(n)

• μ (mu): The step-size parameter, bounded between 0 and 2 for stability.
• ||x(n)||²: The instantaneous signal energy, preventing large updates during high-power bursts.
• ε (epsilon): A small regularization constant added to the denominator to prevent division by zero during silent periods.

This normalization effectively decouples the convergence rate from the absolute amplitude of the input signal, solving the gradient noise amplification problem inherent in standard LMS.

NLMS directly addresses the gradient noise amplification flaw of standard LMS. In LMS, a fixed learning rate causes instability if the input power suddenly increases. NLMS provides:

• Consistent Convergence Rate: The effective step size adapts inversely to signal power, maintaining a constant convergence speed regardless of input amplitude fluctuations.
• Improved Stability Margin: For a stationary signal, NLMS converges for any μ between 0 and 2, whereas LMS stability depends on the maximum eigenvalue of the input correlation matrix.
• Reduced Misadjustment: By preventing large coefficient jumps during signal peaks, NLMS typically achieves a lower steady-state Excess Mean Square Error (EMSE) than LMS for the same convergence speed.

For DPD applications with high Peak-to-Average Power Ratio (PAPR) signals, this normalization is not optional—it is essential.

NLMS offers a highly favorable complexity-to-performance ratio, making it suitable for real-time embedded implementation:

• Multiplications per iteration: 2N + 2 (where N is filter length), compared to N² for RLS.
• Divisions per iteration: 1 (the normalization scalar).
• Memory requirements: O(N) for the coefficient vector and input buffer.

The single division operation is the primary hardware bottleneck. In FPGA implementations, this is often handled via:

• Look-up table (LUT)-based reciprocal approximation.
• CORDIC or iterative division pipelines.
• Bit-shift normalization for power-of-two approximations.

This O(N) complexity allows NLMS to scale to the long filter lengths required for wideband Memory Polynomial predistorters.

The regularization parameter ε is a small positive constant added to the denominator ||x(n)||² + ε. Its functions are critical for robust operation:

• Division-by-Zero Prevention: When the input signal is absent (silence or transmission gap), ||x(n)||² approaches zero. Without ε, the update would diverge to infinity.
• Numerical Stability: In fixed-point arithmetic, ε prevents the reciprocal from saturating at maximum value.
• Bias-Variance Trade-off: A larger ε introduces a slight bias toward zero coefficients (leakage), which can improve robustness to ill-conditioned input data at the cost of slower convergence.

Typical values for ε are chosen as a fraction of the expected minimum signal power, often in the range of 10⁻⁴ to 10⁻⁶ relative to the nominal signal level.

NLMS is the workhorse adaptation algorithm for the Indirect Learning Architecture (ILA) in DPD systems. In ILA, a postdistorter is trained in the feedback path, and its coefficients are copied to the predistorter:

• Training Signal: The normalized PA output (attenuated and down-converted) serves as the input to the adaptive filter.
• Desired Signal: The original baseband transmission signal serves as the reference.
• Update Loop: NLMS iteratively minimizes the error between the postdistorter output and the original signal.

The power normalization of NLMS is particularly advantageous here because the feedback receiver captures a PA output whose power varies with the signal envelope and amplifier compression. NLMS automatically compensates for these power fluctuations without requiring a separate AGC loop.

Despite its robustness, NLMS has known limitations that motivate advanced variants:

• Correlated Inputs: Convergence slows significantly when the input signal is highly correlated (colored), common in narrowband or oversampled signals. Affine Projection Algorithm (APA) addresses this by using multiple past input vectors.
• Fixed Normalization: The scalar normalization does not account for the spectral distribution of power. Transform-Domain LMS applies normalization per frequency bin.
• Sparse Systems: NLMS updates all coefficients equally, even when the optimal solution is sparse. Proportionate NLMS (PNLMS) assigns individual step sizes proportional to coefficient magnitude.
• Non-Stationary Tracking: The fixed μ trades off tracking speed against steady-state error. Variable Step-Size NLMS (VSS-NLMS) dynamically adjusts μ based on the error signal's autocorrelation.