Model Order Reduction

A Volterra model simplification technique that separates static nonlinearity from low-order dynamic behavior, drastically reducing the number of parameters needed for weakly nonlinear systems.

• Decomposes the Volterra series into a static nonlinear part and a dynamic linear part
• Assumes higher-order nonlinearities have negligible memory effects
• Reduces parameter count from O(M^K) to O(M*K), where M is memory depth and K is nonlinear order
• Particularly effective for Class-AB power amplifiers operating near their linear region
• Maintains high modeling accuracy for signals with moderate bandwidths (< 100 MHz)

A model pruning approach that applies an L1-norm penalty to force many Volterra coefficients to exactly zero, automatically performing kernel selection during the identification process.

• The LASSO (Least Absolute Shrinkage and Selection Operator) cost function adds λ||w||₁ to the standard least-squares objective
• As the regularization parameter λ increases, more coefficients are driven to zero
• Typically eliminates 70-90% of coefficients while retaining 99% of model fidelity
• Automatically identifies the most significant diagonal and near-diagonal kernel terms
• Requires careful tuning of λ via cross-validation to balance sparsity and accuracy

Mathematical techniques that factorize the high-dimensional Volterra kernel tensor into a sum of lower-rank components, enabling a highly compact representation.

• Canonical Polyadic Decomposition (CPD) expresses each kernel as a sum of rank-one tensors, creating the CP-Volterra model
• Reduces parameters from exponential to linear scaling with respect to nonlinear order
• A 5th-order, 3-memory Volterra with 243 parameters can be compressed to approximately 30-50 parameters
• Tucker decomposition offers an alternative hierarchical factorization for kernels with multi-modal structure
• Trade-off: decomposition introduces a non-convex optimization problem during coefficient estimation

A greedy compressed sensing algorithm that selects the most significant Volterra kernel terms from a large candidate set, building a sparse model one coefficient at a time.

• Starts with an empty model and iteratively adds the kernel term most correlated with the current residual error
• At each step, re-estimates all selected coefficients via least squares
• Terminates when the residual error falls below a threshold or a maximum number of terms is reached
• Guarantees a sparse solution without requiring full matrix inversion of the complete Volterra regressor
• Computationally efficient for initial model structure discovery before fine-tuning with gradient-based methods

A statistical metric that balances model fit against model complexity, used to select the optimal nonlinear order and memory depth for a Volterra series by penalizing over-parameterization.

• AIC = 2k - 2ln(L̂), where k is the number of parameters and L̂ is the maximized likelihood
• The 2k penalty term discourages adding parameters that provide only marginal improvement in fit
• Enables systematic comparison of models with different (nonlinear order, memory depth) configurations
• Corrected AIC (AICc) provides better performance for small sample sizes common in PA characterization
• Often used in conjunction with cross-validation to confirm generalization performance

A structural simplification that retains only diagonal and near-diagonal terms of the Volterra kernels, based on the observation that terms with widely separated time indices contribute minimally to the output.

• The full Volterra kernel hₖ(τ₁,...,τₖ) is sparse, with significant energy concentrated where |τᵢ - τⱼ| is small
• Memory Polynomial retains only the main diagonal (τ₁ = τ₂ = ... = τₖ)
• Generalized Memory Polynomial extends to include lagging envelope cross-terms with one or two index offsets
• Reduces a 5th-order, 4-memory Volterra from 1,024 terms to approximately 40-80 terms
• Physically motivated by the decaying temporal correlation of semiconductor trapping and thermal effects

A Volterra model where the number of active coefficients is minimized using regularization techniques like LASSO. This retains only the most statistically significant terms, directly reducing computational complexity.

• Applies an L1-norm penalty to force coefficients to zero
• Automatically performs model pruning and kernel selection
• Ideal for wideband signals where memory depth creates parameter explosion

A linear regression method that applies an L1-norm penalty to the coefficient estimation process. In Volterra modeling, LASSO forces many kernel coefficients to exactly zero, automatically performing model order reduction.

• Converts a dense Volterra model into a sparse representation
• Balances model fit against the number of non-zero parameters
• The regularization parameter λ controls the sparsity level

A Volterra model simplification technique that separates static nonlinearity from low-order dynamic behavior. It drastically reduces parameters for weakly nonlinear systems by assuming higher-order dynamics contribute negligibly.

• Decomposes the Volterra series into static and dynamic components
• Retains only first-order and second-order dynamic deviations
• Effective for power amplifiers operating below deep compression

A mathematical technique that factorizes the high-dimensional Volterra kernel tensor into a sum of lower-rank components. This dramatically reduces the number of parameters while preserving the model's predictive accuracy.

• Canonical Polyadic Decomposition (CPD) expresses kernels as rank-one tensors
• Enables compact CP-Volterra models
• Reduces a cubic parameter growth to linear scaling with memory depth

A greedy compressed sensing algorithm used to select the most significant Volterra kernel terms from a large candidate set. It builds a sparse model one coefficient at a time by iteratively selecting the term most correlated with the residual error.

• Computationally efficient alternative to exhaustive search
• Stops when the residual falls below a threshold
• Produces a compact model with near-optimal term selection

A statistical metric that balances model fit against model complexity. Used to select the optimal nonlinear order and memory depth for a Volterra series by penalizing over-parameterization.

• AIC = 2k - 2ln(L) where k is the number of parameters
• Lower AIC values indicate a better tradeoff
• Prevents both underfitting and overfitting during model order reduction