Condition Number

The condition number of a matrix A is defined as the product of its norm and the norm of its inverse: κ(A) = ||A|| · ||A⁻¹||. For the spectral norm, this reduces to the ratio of the largest to smallest singular value: κ(A) = σₘₐₓ / σₘᵢₙ. A condition number of 1 indicates a perfectly orthogonal matrix, while values exceeding 10³ signal significant ill-conditioning. In digital predistortion, the regression matrix formed by basis functions often exhibits high condition numbers due to correlations between polynomial terms and memory taps.

A high condition number amplifies measurement noise and rounding errors during parameter extraction. Specifically, the relative error in the solution vector x is bounded by: ||Δx|| / ||x|| ≤ κ(A) · ||Δb|| / ||b||. This means that with κ = 10⁶, a 0.01% measurement error can produce a 100% error in estimated DPD coefficients. The resulting predistorter becomes unreliable, potentially worsening spectral regrowth rather than mitigating it. This sensitivity is particularly acute in overdetermined systems solved via the normal equations.

Ill-conditioning in DPD model extraction arises from several sources:

• Multicollinearity: High correlation between odd-order polynomial basis functions (e.g., x|x|² and x|x|⁴) when driven by bandlimited signals
• Oversampling: Using a sampling rate far exceeding the signal bandwidth creates near-linear dependencies between adjacent memory taps
• Poor training waveform design: Insufficient peak-to-average power ratio fails to excite the amplifier's full nonlinear range, leaving basis functions underdetermined
• Numerical precision limits: Accumulation of floating-point rounding errors in large covariance matrix constructions

Ridge regression directly addresses high condition numbers by adding a diagonal penalty term to the normal equations: x̂ = (AᵀA + λI)⁻¹Aᵀb. The regularization parameter λ effectively increases all singular values by λ, reducing the condition number to κ_ridge = (σₘₐₓ + λ) / (σₘᵢₙ + λ). This trades a small amount of bias for dramatically reduced variance. In practice, λ is selected via cross-validation or the Akaike Information Criterion to balance model fidelity against numerical stability.

Replacing correlated polynomial terms with orthogonal basis functions fundamentally reduces the condition number. Techniques include:

• Principal Component Analysis (PCA): Transforms the regression matrix into uncorrelated components ordered by variance, discarding dimensions with near-zero singular values
• Orthogonal polynomials: Using Chebyshev or Legendre polynomials instead of standard monomials reduces multicollinearity
• Gram-Schmidt orthogonalization: Iteratively constructs an orthonormal basis from the original regressors, ensuring κ approaches 1 These methods preserve modeling accuracy while eliminating the numerical fragility of direct polynomial fitting.

Real-time DPD systems should continuously monitor the condition number of the regression matrix during online training. Practical guidelines:

• κ < 100: Well-conditioned; coefficient estimates are reliable
• 100 < κ < 1000: Moderately ill-conditioned; monitor solution stability
• κ > 1000: Severely ill-conditioned; apply regularization or switch to recursive least squares with a forgetting factor to reduce the effective data window
• κ > 10⁶: Numerically singular for practical purposes; basis function redesign required Logging κ alongside post-distortion error provides early warning of model degradation.

A numerical state where the covariance matrix of basis functions is nearly singular, causing coefficient estimates to be highly sensitive to measurement noise and computational rounding errors. In power amplifier modeling, ill-conditioning arises when basis functions are highly correlated, making the extracted behavioral model unstable and non-robust. The condition number quantifies the severity of this state.

A generalized matrix inverse used to compute the least-squares solution for model coefficients in overdetermined systems where direct inversion is not possible. When the regression matrix is ill-conditioned, the pseudoinverse computed via Singular Value Decomposition (SVD) provides a numerically stable alternative by discarding or attenuating components associated with small singular values.

A regularized least-squares method that adds an L2 penalty on coefficient magnitudes to the cost function, shrinking parameters to handle multicollinearity in the regression matrix. By adding a constant λ to the diagonal of the Gram matrix, ridge regression directly improves the condition number, trading a small amount of bias for a significant reduction in estimator variance.

A dimensionality reduction technique that transforms correlated basis functions into a smaller set of uncorrelated components, mitigating ill-conditioning in model extraction. By projecting the regression problem onto the principal components with the largest eigenvalues, PCA effectively reduces the condition number and stabilizes coefficient estimation without requiring explicit regularization parameters.

A matrix containing the pairwise covariances between basis functions, used to analyze correlations that lead to ill-conditioning during parameter extraction. The eigenvalue spectrum of the covariance matrix directly determines the condition number: a wide spread between the largest and smallest eigenvalues signals a poorly conditioned problem where minor measurement perturbations can cause large swings in estimated coefficients.

A modeling failure where an excessively complex model memorizes measurement noise and specific training data rather than learning the true underlying amplifier behavior. High condition numbers exacerbate overfitting by amplifying noise components during matrix inversion, producing coefficient vectors with large magnitudes that fit the training data perfectly but generalize poorly to new signals.