Normalized Least Mean Squares (NLMS)

The defining feature of NLMS is the step-size normalization by the Euclidean norm of the input vector. The update equation is:

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

• ||x(n)||²: Squared norm of the input signal vector
• ε: Small regularization constant preventing division by zero
• μ: Step-size parameter (0 < μ < 2 for stability)

This normalization decouples convergence behavior from input signal scaling, making the algorithm inherently resistant to gradient noise amplification during loud signal bursts.

Standard LMS suffers from a gradient noise amplification problem: when the input signal power is large, the instantaneous gradient estimate becomes proportionally large, causing coefficient overshoot.

NLMS solves this by:

• Normalizing the correction term at each iteration
• Ensuring the a posteriori error is zero after each update in the noise-free case
• Providing the solution that exhibits minimum disturbance to the existing coefficient vector

This makes NLMS the normalized solution to the constrained optimization problem of minimizing coefficient change while achieving zero a posteriori error.

Unlike LMS, where convergence speed depends on the eigenvalue spread of the input autocorrelation matrix, NLMS convergence is largely independent of input signal statistics.

Key properties:

• Fast initial convergence even with highly colored input signals
• Uniform convergence rate across all filter modes
• μ = 1 provides the fastest convergence in stationary environments
• Convergence guaranteed for 0 < μ < 2 regardless of input power

This statistical independence makes NLMS the preferred choice for applications with non-stationary or unpredictable signal characteristics, such as DPD coefficient tracking in wireless transmitters.

NLMS adds minimal overhead compared to standard LMS while delivering significantly improved robustness:

• LMS complexity: 2N + 1 multiplications per iteration (N = filter length)
• NLMS complexity: 3N + 1 multiplications per iteration
• The additional cost is computing ||x(n)||², which can be updated recursively: ||x(n)||² = ||x(n-1)||² + x²(n) - x²(n-N)

This O(N) complexity makes NLMS suitable for real-time FPGA and DSP implementations in DPD systems, where sample rates exceed hundreds of MHz.

The regularization constant ε prevents numerical instability when the input signal power approaches zero:

• Typical value: ε = 0.01 to 0.001 relative to expected signal power
• Acts as a soft threshold preventing infinite step-size magnification
• Introduces a slight bias-variance tradeoff: larger ε improves stability but slows convergence
• In DPD applications, ε must be tuned to handle signal dropouts during transmission gaps

Without ε, silent periods in the input signal would cause division by near-zero values, leading to coefficient divergence and potential amplifier damage.

NLMS can be understood as a special case of the Affine Projection Algorithm (APA) with projection order P = 1:

• APA reuses past input vectors to accelerate convergence with colored signals
• NLMS uses only the current input vector for the gradient estimate
• As P increases, APA approaches RLS performance at higher computational cost

This hierarchical relationship allows DPD designers to scale algorithm complexity based on available hardware resources, starting with NLMS and upgrading to APA if faster convergence is required for wideband signals.

The foundational stochastic gradient descent algorithm from which NLMS is derived. LMS adapts filter coefficients sample-by-sample using a fixed step size, minimizing instantaneous squared error. Its primary vulnerability is gradient noise amplification when input signal power fluctuates, leading to slow convergence or instability. NLMS directly addresses this by normalizing the step size.

An adaptive algorithm offering order-of-magnitude faster convergence than NLMS by minimizing a weighted least-squares cost function. RLS recursively computes the inverse of the input autocorrelation matrix, achieving insensitivity to input signal conditioning. The trade-off is quadratic computational complexity (O(N²)) versus NLMS's linear complexity (O(N)), making RLS prohibitive for many real-time embedded systems.

The core mechanism distinguishing NLMS from LMS. The fixed step size μ is divided by the squared Euclidean norm of the input vector plus a small regularization constant ε:

• Prevents divergence during high-power signal bursts
• Accelerates convergence during low-power periods
• The regularization parameter ε prevents division by zero during silent intervals
• Optimal normalized step size typically lies in the range 0.1 to 0.9 for stable operation

A numerical pathology where the condition number of the input correlation matrix becomes extremely large. In such scenarios, the error surface forms steep, narrow valleys that cause gradient-based algorithms like NLMS to oscillate or stall. Ill-conditioning commonly arises from:

• Highly correlated basis functions in behavioral models
• Narrowband or tonal training signals
• Insufficient spectral excitation during model extraction

A scalar parameter β ∈ (0,1] that exponentially discounts past data in recursive estimation. While native NLMS does not use a forgetting factor, its leaky NLMS variant incorporates one to prevent coefficient windup in fixed-point implementations. The effective memory of the algorithm is approximately 1/(1-β) samples. Typical values range from 0.95 to 0.999 for tracking slowly varying power amplifier characteristics.