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, it simplifies to the ratio of the largest to smallest singular value: κ(A) = σ_max / σ_min. A condition number of 1 indicates a perfectly conditioned orthogonal matrix, while values approaching infinity signal singularity. In DPD coefficient estimation, the condition number of the data covariance matrix directly determines how much numerical noise is amplified in the solution.

A high condition number in the regression matrix causes ill-conditioning, where small measurement noise or quantization errors in the feedback path are dramatically amplified. This leads to:

• Coefficient drift during online adaptation
• Misadjustment in LMS-based algorithms exceeding theoretical limits
• Slow convergence in iterative solvers like RLS
• Overfitting to noise rather than the true PA nonlinearity

In practice, a condition number above 10⁴ often indicates that direct least squares estimation will produce unreliable predistorter coefficients.

Several factors elevate the condition number in digital predistortion data matrices:

• Highly correlated basis functions in memory polynomial models, especially when memory depth is over-specified
• Narrowband or single-tone test signals that fail to excite all nonlinear modes, leaving the data matrix rank-deficient
• Oversampling without sufficient signal bandwidth, creating near-linear dependencies between adjacent samples
• Numerical precision limits in fixed-point FPGA implementations that compound rounding errors

Wideband signals like 5G NR with high PAPR naturally improve conditioning by providing richer excitation.

To combat ill-conditioning, regularization adds a penalty term to the cost function:

• Tikhonov Regularization (Ridge Regression): Adds λ||w||² to the least squares objective, where λ is the regularization parameter. This effectively adds λ to the diagonal of the covariance matrix, reducing the condition number.
• Levenberg-Marquardt: Interpolates between Gauss-Newton and gradient descent, automatically adjusting damping to handle ill-conditioned Jacobians.
• Truncated SVD: Discards singular values below a threshold, directly reducing the effective condition number.

Optimal λ selection often uses cross-validation or the L-curve criterion.

Real-time condition number monitoring serves as a health metric for DPD adaptation loops. A sudden increase may indicate:

• PA aging or thermal drift changing the nonlinear characteristic
• Feedback path degradation introducing excess noise
• Signal statistics change (e.g., traffic load drop reducing excitation)

Engineers often set a threshold (e.g., κ > 10⁶) to trigger fallback to a safe coefficient table or freeze adaptation. The condition number is computed efficiently using the QR decomposition of the data matrix, where it equals the ratio of the largest to smallest diagonal element of R.

The QR-RLS algorithm is explicitly designed to handle ill-conditioned data matrices that would destabilize conventional RLS. By maintaining the Cholesky factor R of the inverse correlation matrix and updating it via Givens rotations, QR-RLS achieves numerical stability even when the condition number exceeds 10¹⁰. This makes it the preferred adaptive filtering method for DPD applications where:

• Fixed-point arithmetic is used in FPGA implementations
• Training signals may have limited spectral richness
• Long-term continuous adaptation is required without periodic reset

The computational cost of QR-RLS is higher than LMS but justified by its robustness to ill-conditioning.

A stabilization technique that adds a penalty term to the least squares cost function to solve ill-posed inverse problems caused by a high condition number. By introducing a small bias, it reduces the variance of the coefficient estimates, preventing wild oscillations in the predistorter response. This is essential for direct learning architectures where the matrix to be inverted is often ill-conditioned.

A numerically robust implementation of the Recursive Least Squares algorithm that uses QR decomposition to solve the least squares problem directly without explicitly computing the inverse of the correlation matrix. This method is inherently stable against ill-conditioned data matrices, making it a preferred choice for real-time, sample-by-sample DPD coefficient updates where numerical precision is paramount.

The gradual deviation of predistorter coefficients from their optimal values over time. A high condition number in the data covariance matrix makes the system highly sensitive to small perturbations, such as thermal noise or quantization errors. This sensitivity accelerates coefficient drift, leading to a slow degradation of adjacent channel power ratio (ACPR) performance without any change in the physical power amplifier.

A modeling failure where the predistorter learns the specific noise characteristics of the training data rather than the true underlying PA nonlinearity. An ill-conditioned estimation problem exacerbates overfitting, as the model fits the high-frequency noise components that dominate the small singular values of the data matrix. This results in excellent training performance but poor generalization to new signal statistics.

A robust iterative algorithm that interpolates between Gauss-Newton and gradient descent to solve nonlinear least squares problems. It dynamically adjusts a damping parameter to handle the ill-conditioning of the Jacobian matrix. When the condition number is high, the algorithm behaves more like gradient descent to ensure stable convergence during coefficient estimation for complex Volterra-based predistorters.

Offline methods for deriving a power amplifier behavioral model from measured input-output data. The accuracy of the extracted model is directly limited by the condition number of the measurement data matrix. Poorly designed excitation signals that do not sufficiently excite all nonlinear modes lead to an ill-conditioned problem, resulting in an unreliable model that fails to capture the true memory effects of the amplifier.