Model Order Reduction

A complexity reduction technique that removes insignificant kernels from a full Volterra series model based on a significance metric, retaining only the most critical distortion terms.

• Mechanism: Evaluates each kernel's contribution to model accuracy using metrics like the Fisher information matrix or output variance
• Near-diagonal pruning: Retains only kernels where memory tap indices are close together, exploiting the fact that widely separated taps contribute less to PA behavior
• Threshold-based elimination: Removes all kernels whose coefficient magnitude falls below a predefined threshold after initial estimation
• Typical reduction: Can eliminate 50-80% of Volterra kernels while maintaining ACLR within 0.5 dB of the full model

A dimensionality reduction technique applied to the basis function matrix to identify and retain only the most significant principal components, reducing model complexity and improving numerical conditioning.

• Process: Computes the eigenvalue decomposition of the basis function covariance matrix, then projects data onto eigenvectors with the largest eigenvalues
• Variance retention: Typically retains components explaining 99.9% of total variance, often reducing basis dimension by 40-60%
• Numerical benefit: The resulting transformed basis functions are inherently orthogonal, eliminating ill-conditioning in the least squares estimation
• Trade-off: Requires an additional matrix multiplication in the forward path, though this is offset by the reduced number of coefficients

A greedy sparse approximation algorithm that iteratively selects the most correlated basis function from a dictionary to build a compact, low-complexity predistorter model.

• Iteration: At each step, computes the correlation between the residual error and all candidate basis functions, selecting the one with maximum correlation
• Residual update: After each selection, re-estimates all chosen coefficients via least squares and updates the residual
• Stopping criteria: Terminates when the residual power falls below a target threshold or a maximum number of terms is reached
• Sparsity control: Directly produces models with a specified number of active coefficients, making it ideal for hardware-constrained implementations

A regularized least squares estimation technique that adds a penalty on the magnitude of coefficients to the cost function, preventing overfitting and improving the robustness of the extracted DPD model.

• Cost function: Minimizes ||y - Xw||² + λ||w||², where λ is the regularization parameter controlling the bias-variance trade-off
• Shrinkage effect: Drives small, statistically insignificant coefficients toward zero, effectively performing soft pruning without explicit term removal
• Numerical stability: The added diagonal loading (λI) improves the condition number of the normal equations matrix, enabling reliable inversion even with correlated basis functions
• λ selection: Typically chosen via cross-validation or the L-curve method to balance model fidelity against coefficient magnitude

A numerical conditioning process that transforms correlated polynomial basis functions into an orthogonal set to improve the stability and convergence speed of coefficient estimation.

• Gram-Schmidt procedure: Sequentially orthogonalizes basis functions by subtracting projections onto previously processed functions
• QR decomposition alternative: Can be performed via QR decomposition of the basis matrix, yielding an orthogonal Q matrix and upper triangular R matrix
• Coefficient mapping: The orthogonalized coefficients must be back-transformed to the original polynomial basis for implementation, though the forward path complexity remains unchanged
• Convergence benefit: Orthogonal bases enable faster convergence in adaptive algorithms like LMS and RLS by decorrelating the input signals

The systematic selection or pruning of cross-terms in a behavioral model to balance linearization accuracy against the computational complexity of the predistorter.

• Cross-term types: Includes envelope-memory cross-terms (|x(n-m)|ᵏ·x(n-l)) and signal-memory cross-terms (x(n-m)·|x(n-l)|ᵏ) as defined in the Generalized Memory Polynomial
• Lag-lead selection: Restricts cross-terms to a limited range of lag and lead indices around the current sample, exploiting the locality of memory effects
• Symmetry exploitation: Many PA characteristics exhibit symmetry in cross-term coefficients, allowing paired terms to be combined or one eliminated
• Complexity scaling: Each retained cross-term adds one complex multiply-accumulate operation per sample, making aggressive pruning essential for wideband FPGA implementations

A direct complexity reduction technique that systematically removes insignificant Volterra kernels from a full series model. A significance metric, often derived from the kernel's contribution to total output power or its coefficient magnitude, ranks all terms. Kernels falling below a defined threshold are pruned, retaining only the most critical distortion terms. This transforms a dense, computationally prohibitive model into a sparse, efficient structure suitable for real-time FPGA implementation.

A dimensionality reduction technique applied to the basis function matrix (e.g., the MP or GMP regressor matrix). PCA computes the eigenvectors of the data covariance matrix and projects the original, correlated basis functions onto a new coordinate system defined by the principal components. By retaining only components with the largest eigenvalues, the model's parameter count is drastically reduced while preserving the vast majority of the signal variance, improving numerical conditioning and reducing computational load.

A regularized least squares estimation technique that adds an L2 penalty on the coefficient magnitudes to the standard least squares cost function. This penalty, controlled by the regularization parameter (λ), shrinks coefficients toward zero and prevents any single term from dominating. In the context of model order reduction, ridge regression implicitly performs soft pruning by suppressing the influence of less important terms, leading to a more robust, generalizable model that avoids overfitting to measurement noise.

A numerical conditioning process that transforms a set of highly correlated polynomial basis functions (like the standard memory polynomial terms) into a mutually orthogonal set. Techniques like Gram-Schmidt or Cholesky decomposition create new basis functions where each term is uncorrelated with the others. This orthogonalization dramatically improves the stability and convergence speed of coefficient estimation algorithms, making it easier to identify and discard truly redundant model components.

The systematic strategy for selecting or pruning cross-terms in models like the Generalized Memory Polynomial (GMP). Cross-terms, which involve products of the signal and its lagging/leading envelope, capture complex memory effects but explode in number with increased nonlinear order and memory depth. Effective management involves heuristics, sensitivity analysis, or sparse identification to retain only the cross-terms that provide a statistically significant improvement in linearization performance, balancing accuracy against implementation cost.