Model Order Reduction

Applies orthogonal linear transformation to the DPD basis function matrix, projecting high-dimensional coefficient vectors onto a lower-dimensional subspace that captures maximum variance. The technique identifies the most significant eigen-directions in the behavioral model, discarding components that contribute minimally to linearization. Typical compression ratios range from 40-70% with less than 0.5 dB ACLR degradation. PCA is particularly effective for memory polynomial models where adjacent tap coefficients exhibit high correlation due to oversampling.

An iterative sparse recovery algorithm that selects the most correlated basis functions from an overcomplete DPD dictionary one at a time. At each iteration, OMP identifies the regressor that best explains the current residual error, then orthogonalizes the remaining candidates. This produces a sparse predistorter where only 10-30% of the original coefficients are non-zero. OMP is preferred when the power amplifier exhibits localized non-linearity that can be captured by a small subset of basis functions rather than a dense polynomial expansion.

A post-training compression method that ranks DPD coefficients by their absolute magnitude and removes those falling below a threshold. The underlying assumption is that small-magnitude coefficients contribute negligibly to the predistorter output. After pruning, the remaining coefficients may be fine-tuned through additional training iterations to recover any lost performance. This technique is hardware-friendly because it produces unstructured sparsity that maps efficiently to general-purpose DSP cores without requiring specialized sparse matrix accelerators.

Constructs a reduced set of latent variables that maximize the covariance between the DPD basis matrix and the desired predistorter output. Unlike PCA, which focuses solely on input variance, PLS simultaneously considers the input-output relationship, ensuring that the retained components are directly relevant to linearization performance. PLS is particularly advantageous for indirect learning architectures where the post-distorter error signal guides dimensionality reduction, yielding models with 30-50% fewer coefficients and minimal NMSE degradation.

Trains a compact student network to mimic the linearization behavior of a larger, high-performance teacher model. The student is optimized using a composite loss function that combines the standard predistortion error with a soft-target loss matching the teacher's output distribution. This transfers the teacher's generalization capability to a model with 5-10x fewer parameters. Distillation is especially effective for RVTDNN architectures, where wide hidden layers in the teacher can be compressed into narrow student networks suitable for FPGA deployment.

Applies CANDECOMP/PARAFAC (CP) or Tucker decomposition to the multi-dimensional coefficient tensors of Volterra-based DPD models. By factorizing the full tensor into a sum of rank-one components or a core tensor with factor matrices, the number of parameters scales linearly with the model order rather than exponentially. For a 5th-order Volterra model with memory depth 3, tensor decomposition can reduce coefficients from thousands to hundreds while maintaining -50 dBc ACLR performance in Doherty amplifier linearization.

A sparsification technique that identifies and removes low-magnitude coefficients from a DPD model that contribute negligibly to linearization. By setting a threshold based on the coefficient's absolute value relative to the dominant terms, the model's computational footprint is reduced without significant degradation in Adjacent Channel Leakage Ratio (ACLR).

• Magnitude-based pruning removes coefficients below a fixed threshold
• Iterative pruning retrains the model between pruning steps to recover performance
• Can reduce the number of coefficients in a Generalized Memory Polynomial (GMP) by 50-80%

A statistical dimensionality reduction method that transforms the correlated basis functions of a DPD model into a smaller set of uncorrelated principal components. By retaining only the components that capture the majority of the variance in the predistorter's behavior, the effective number of parameters is dramatically reduced.

• Projects high-dimensional Volterra series kernels onto a lower-dimensional subspace
• The number of retained components is determined by a variance explained threshold (e.g., 99.9%)
• Particularly effective for models with strong correlations between adjacent memory taps

A supervised dimensionality reduction technique that projects both the input basis functions and the desired predistorter output into a latent variable space that maximizes their covariance. Unlike PCA, PLS explicitly considers the target signal when constructing the reduced model, leading to more efficient compression for a given linearization target.

• Constructs latent variables that are optimal for predicting the inverse PA non-linearity
• Often achieves equivalent EVM performance with 60-70% fewer latent variables than PCA
• Well-suited for Direct Learning Architecture (DLA) implementations

An iterative selection algorithm that builds a reduced DPD model by sequentially adding the basis function that most improves linearization performance. Starting from an empty model, each step evaluates all remaining candidate terms and selects the one that minimizes the residual Error Vector Magnitude (EVM).

• Uses orthogonal matching pursuit or similar greedy optimization
• Produces a compact model with explicitly ranked basis function importance
• Allows a direct trade-off between model complexity and ACLR improvement

A compression paradigm where a compact student model is trained to mimic the linearization behavior of a large, high-performance teacher model. The student learns not just from the desired output but from the teacher's full predistortion signal, capturing nuanced compensation patterns with far fewer parameters.

• The student model can use a simpler architecture (e.g., a reduced-tap RVTDNN)
• Training minimizes both the output error and the Kullback-Leibler divergence between teacher and student predistortion distributions
• Enables deployment of complex DPD on resource-constrained FPGAs

A multi-linear algebra technique that factorizes the high-dimensional coefficient tensor of a DPD model into a set of low-rank component matrices. For models like the Generalized Memory Polynomial, the coefficient tensor can be decomposed using CP (CANDECOMP/PARAFAC) or Tucker decomposition, dramatically reducing the parameter count.

• Exploits the inherent low-rank structure of PA non-linearity across memory and non-linearity dimensions
• Can reduce a GMP model's parameters by an order of magnitude
• The decomposed structure maps efficiently to parallel hardware implementations