Cross-Validation

The core purpose of cross-validation is to estimate a model's generalization error—the error on data not used during coefficient extraction. A model that perfectly memorizes the training waveform but fails on a new modulation scheme is useless for predistortion. Cross-validation provides a realistic performance metric by simulating this unseen data scenario, directly addressing the bias-variance tradeoff.

The most common form of cross-validation divides the captured time-series data into k equal-sized, contiguous folds. The process iterates k times. In each iteration:

• One fold is held out as the validation set.
• The remaining k-1 folds form the training set.
• The model is trained and its error is computed on the held-out fold. The final performance estimate is the average error across all k trials. Typical values for k are 5 or 10.

Standard random shuffling before k-fold splitting is invalid for time-series data like amplifier waveforms. Random shuffling destroys temporal correlations and leaks future information into the past during training. For DPD model validation, folds must be contiguous blocks to preserve the sequential structure of the signal. This is often called blocked cross-validation or time-series split.

Cross-validation is the primary diagnostic for overfitting. A significant divergence between the low training error and the high validation error indicates the model is memorizing noise rather than learning the amplifier's true nonlinear dynamics. This signals the need to reduce model order, apply regularization (like ridge regression), or prune basis functions. The Akaike Information Criterion (AIC) provides a complementary, information-theoretic view of this trade-off.

A simpler alternative to k-fold is the hold-out method, where the dataset is split once into a training set (e.g., 70-80%) and a validation set (20-30%). While computationally cheaper, its performance estimate can have high variance depending on which specific samples fall into the validation set. It is best used for large datasets where a single split is statistically representative, but k-fold is preferred for robust evaluation of extracted PA models.

Beyond error estimation, cross-validation is a powerful tool for model selection. By comparing the average validation error across different model architectures—such as a Memory Polynomial versus a Generalized Memory Polynomial with varying nonlinearity orders and memory depths—an engineer can objectively select the most parsimonious model that achieves the lowest generalization error without manual tuning bias.

A modeling failure where an excessively complex model memorizes measurement noise and specific training data rather than learning the true underlying amplifier behavior. Overfitting produces deceptively low training error but catastrophically high validation error. Key indicators:

• Low training NMSE with high validation NMSE
• Coefficient magnitudes growing without bound
• Model capturing noise floor characteristics
• Poor ACLR on unseen modulation schemes Cross-validation is the primary defense against overfitting in behavioral model extraction.

The fundamental tension in model selection between underfitting (high bias from overly simplistic assumptions) and overfitting (high variance from excessive sensitivity to training data fluctuations). In power amplifier modeling:

• High bias: Memoryless polynomial fails to capture long-term thermal memory effects
• High variance: High-order Volterra series fits noise instead of nonlinear dynamics
• Optimal point: Minimizes total expected prediction error on unseen stimuli Cross-validation empirically estimates this tradeoff by evaluating generalization performance across multiple data partitions.

The process of determining the optimal complexity of a behavioral model by balancing fitting accuracy against the risk of overfitting. Techniques include:

• Akaike Information Criterion (AIC): Penalizes parameter count relative to goodness of fit
• Bayesian Information Criterion (BIC): Applies stronger penalty for larger datasets
• Cross-validated error curves: Plot validation NMSE vs. nonlinearity order and memory depth to identify the elbow point
• Holdout validation: Reserve a portion of captured data exclusively for final model selection Proper order estimation prevents both under-modeling of nonlinear dynamics and over-parameterization.

A technique that adds a penalty term to the cost function during coefficient extraction to prevent overfitting and improve numerical stability. Common forms in DPD model extraction:

• Ridge Regression (L2): Penalizes sum of squared coefficients, shrinking all parameters toward zero
• LASSO (L1): Drives irrelevant coefficients exactly to zero, performing automatic basis function selection
• Elastic Net: Combines L1 and L2 penalties for correlated basis function groups Regularization is essential when the regression matrix is ill-conditioned, which frequently occurs with high-order polynomial basis functions and narrowband training signals.

A carefully designed stimulus signal with specific statistical properties used to excite the power amplifier and capture its full nonlinear dynamic range. Effective training waveforms must:

• Match the peak-to-average power ratio (PAPR) of target modulation schemes
• Span the full amplitude range to characterize compression characteristics
• Include sufficient bandwidth to excite memory effects across the operational range
• Provide persistent excitation so all model parameters are uniquely identifiable Poor waveform design leads to extrapolation failure, where the model performs well on training-like signals but fails on operational waveforms—a failure mode cross-validation with mismatched holdout sets can detect.

A scalar value measuring the sensitivity of matrix inversion to small changes in input data during least-squares coefficient extraction. A high condition number indicates an ill-conditioned regression problem where:

• Coefficient estimates become highly sensitive to measurement noise
• Small data perturbations produce wildly different model parameters
• Numerical rounding errors amplify during pseudoinverse computation Cross-validation reveals the practical consequences: an ill-conditioned model may show acceptable training performance but produce inconsistent, unreliable predictions on validation partitions. Mitigation strategies include regularization, PCA-based dimensionality reduction, and orthogonal basis function design.