Levenberg-Marquardt

The algorithm's defining feature is its damping parameter (λ), which dynamically interpolates between gradient descent and Gauss-Newton behavior. When λ is large, the algorithm behaves like gradient descent, taking cautious steps downhill—ideal when far from the minimum or when the cost surface is highly non-convex. When λ is small, it approximates the Gauss-Newton method, exploiting quadratic curvature for rapid terminal convergence. The damping parameter is adjusted heuristically at each iteration: if a step reduces the cost function, λ is decreased to accelerate convergence; if a step increases the error, λ is increased to improve stability. This self-tuning mechanism makes LM particularly effective for power amplifier model extraction where the error surface can exhibit narrow valleys and flat plateaus simultaneously.

Unlike pure Gauss-Newton, which can diverge catastrophically if initial coefficient estimates are far from the optimum, Levenberg-Marquardt maintains stability through its damping mechanism. This robustness is critical in digital predistortion coefficient estimation where the initial predistorter may be a simple linear gain or a previously calibrated model that has drifted due to temperature changes. The algorithm can recover from poor starting points without requiring precise model order estimation beforehand. In practice, this means an engineer can initialize a memory polynomial model with all coefficients set to zero, and LM will reliably converge to the correct nonlinear parameters, whereas Gauss-Newton might oscillate or produce invalid solutions due to ill-conditioning in the early iterations.

Behavioral model extraction frequently encounters ill-conditioned regression matrices due to high correlation between basis functions, especially in memory polynomial models where delayed terms and their nonlinear expansions create near-linear dependencies. The LM algorithm addresses this by augmenting the diagonal of the approximate Hessian matrix (JᵀJ) with the damping term λI. This diagonal loading is mathematically equivalent to ridge regression or Tikhonov regularization applied at each iteration. The effect is twofold:

• Numerical stability: The modified normal equations remain solvable even when JᵀJ is nearly singular
• Implicit regularization: Large λ values shrink coefficient estimates, preventing overfitting to measurement noise This property makes LM particularly suitable for extracting Volterra series and generalized memory polynomial models where basis function correlation is unavoidable.

Levenberg-Marquardt operates as a batch optimization algorithm, processing the entire training dataset in each iteration to compute the Jacobian matrix and the resulting parameter update. This contrasts with recursive methods like RLS or LMS that update coefficients sample-by-sample. For offline model extraction in laboratory settings, this batch nature is advantageous:

• Full data utilization: Every captured sample contributes to each parameter update, maximizing information extraction from expensive measurement campaigns
• Deterministic convergence: Given a suitable training waveform, LM produces repeatable, verifiable coefficient sets without the stochastic behavior of sample-by-sample methods
• Diagnostic capability: The full Jacobian enables analysis of parameter sensitivity and confidence intervals However, the computational cost scales with O(N·M²) where N is the number of data points and M is the number of parameters, making it impractical for real-time online training without significant hardware acceleration.

When the current parameter estimate approaches the true minimum and the cost function becomes approximately quadratic, the LM algorithm with a small damping parameter closely approximates the Gauss-Newton method, achieving quadratic convergence. This means the error decreases proportionally to the square of the previous error in each iteration, dramatically accelerating final refinement. In practical power amplifier modeling, this translates to:

• Rapid reduction of normalized mean squared error (NMSE) from -35 dB to below -45 dB in just 2-3 final iterations
• Efficient extraction of high-order nonlinear kernels that have subtle but important contributions to spectral regrowth prediction
• Minimal iteration count when warm-starting from a previously extracted model at an adjacent frequency or power level This fast terminal convergence is a key advantage over pure gradient descent methods, which slow to linear convergence near the minimum and may require hundreds of additional iterations to achieve the same final accuracy.

The LM algorithm requires explicit computation of the Jacobian matrix—the matrix of partial derivatives of each residual with respect to each model parameter. For behavioral models, this demands:

• Analytical derivative expressions for each basis function with respect to its coefficient. For a memory polynomial, these derivatives are straightforward: the derivative of a delayed nonlinear term with respect to its coefficient is simply the basis function itself
• Structured computation: The Jacobian has a predictable block structure corresponding to memory taps and nonlinear orders, enabling efficient construction without numerical differentiation
• Memory considerations: For a capture of 100,000 samples and a model with 50 coefficients, the Jacobian requires storage of 5 million elements, which is manageable on modern workstations but must be considered for wideband signal linearization with very long captures The need for the Jacobian distinguishes LM from derivative-free optimization methods and ties it closely to models where analytical derivatives are available. Fortunately, most Volterra series variants and neural network architectures used in DPD provide straightforward derivative calculations.