Condition Number

The condition number of a matrix A is defined as the ratio of its largest to smallest singular value: κ(A) = σ_max / σ_min. For a normal matrix, it equals the ratio of the largest to smallest eigenvalue magnitude. A condition number of 1 indicates a perfectly conditioned (orthogonal) matrix, while a value approaching infinity signals a singular or rank-deficient matrix. In the context of the Normal Equation for least squares, the condition number of the data matrix A^T A is the square of the condition number of A, making the normal equations particularly vulnerable to ill-conditioning.

In Digital Predistortion (DPD) coefficient extraction, a high condition number in the regression matrix amplifies measurement noise and quantization errors, leading to unstable or wildly inaccurate predistorter coefficients. Specifically, the relative error in the solution is bounded by: ||Δx|| / ||x|| ≤ κ(A) · ||Δb|| / ||b||. This means that with a condition number of 10^6, a tiny 0.001% perturbation in measured power amplifier output data can produce a 1000% error in the estimated Volterra kernel coefficients, rendering the linearization ineffective.

Several factors in predistorter modeling lead to ill-conditioned regression matrices:

• Polynomial basis functions: High-order Volterra or memory polynomial terms create near-linear dependencies between columns, especially when using monomial bases like x, x², x³.
• Oversampling and narrowband signals: Highly correlated successive samples in the data matrix cause columns to be nearly collinear.
• Large memory depth: Including many delayed taps increases matrix dimensions without adding independent information.
• Fixed-point arithmetic: In FPGA implementations, limited precision exacerbates the effects of an already moderate condition number.

Engineers employ several techniques to combat ill-conditioning:

• Orthogonal basis functions: Replacing monomials with Chebyshev or Legendre polynomials dramatically reduces column correlation.
• Tikhonov Regularization: Adding a penalty term λI to the correlation matrix improves the condition number of (A^T A + λI) at the cost of a slight bias.
• QR Decomposition with column pivoting: Solving min ||Ax - b|| via QR avoids forming the ill-conditioned normal equations entirely.
• Singular Value Decomposition (SVD): Truncating small singular values provides a robust low-rank approximation, directly addressing the root cause of the high condition number.

The condition number serves as a real-time health metric for adaptive DPD systems:

• Pre-deployment checking: Computing κ(A) on captured data before FPGA synthesis validates that the chosen model structure and signal bandwidth are numerically viable.
• Online monitoring: In Recursive Least Squares (RLS) or QR-RLS implementations, tracking the condition number can trigger a reset or a switch to a more robust algorithm if it exceeds a threshold.
• Precision requirements: The condition number dictates the minimum bit-width needed in fixed-point hardware. A rule of thumb is that the number of bits of numerical precision must exceed log₂(κ(A)) to avoid catastrophic cancellation.

The Singular Value Decomposition (SVD) provides the most direct way to compute and analyze the condition number. Given A = U Σ V^T, the condition number is simply σ_max / σ_min from the diagonal of Σ. The SVD also reveals the effective rank of the matrix by examining the decay of singular values. A sharp drop-off indicates that many basis functions are redundant. By setting a threshold and truncating small singular values, the SVD produces a pseudoinverse that minimizes the norm of the solution while maintaining a user-specified effective condition number, a technique known as Truncated SVD regularization.

A scalar added to the diagonal of the correlation matrix to combat ill-conditioning. In the presence of a high condition number, the normal equation becomes unstable. Adding a regularization parameter (ridge regression) effectively increases the smallest singular values, reducing the condition number and preventing coefficient explosion. This is critical in online training where signal statistics fluctuate.

A numerically stable alternative to solving the normal equations directly. QRD factors the data matrix into an orthogonal matrix Q and upper triangular R. Because it avoids explicitly forming the correlation matrix, QRD squares the condition number only once, making it far more stable than direct matrix inversion for ill-conditioned problems in coefficient estimation algorithms.

A direct consequence of operating with a high condition number. When the data matrix is nearly singular, small noise perturbations cause massive swings in the estimated coefficients. The model fits the noise rather than the underlying power amplifier dynamics. This manifests as excellent training performance but catastrophic generalization to new signals—a classic bias-variance tradeoff failure.

A numerically robust implementation of Recursive Least Squares that uses Givens rotations to directly update the square-root of the inverse correlation matrix. Unlike standard RLS, QR-RLS maintains numerical stability even when the input signal has a high condition number, making it the preferred choice for real-time digital predistortion adaptation in FPGA implementations.

The closed-form least squares solution w = (XᵀX)⁻¹Xᵀy. While mathematically elegant, solving the normal equation directly squares the condition number of the data matrix X. For a matrix with condition number κ, the normal equations have condition number κ², making this approach numerically dangerous for ill-conditioned problems common in wideband system identification.