Numerical Stability

The condition number κ(A) of the data matrix A quantifies sensitivity to input perturbations. A well-conditioned matrix with κ(A) close to 1 ensures that small rounding errors in finite-precision arithmetic do not amplify into large output errors.

• Ideal range: κ(A) < 10³ for single-precision floating point
• Ill-conditioned threshold: κ(A) > 10⁶ risks catastrophic cancellation
• Mitigation: Orthogonalize basis functions or apply Tikhonov regularization to reduce the effective condition number

Using orthogonal polynomials or orthogonalized Volterra kernels decorrelates the regression matrix columns, dramatically improving numerical conditioning. Orthogonal bases minimize the inner product between distinct basis vectors.

• Chebyshev polynomials: Provide uniform error distribution across the input amplitude range
• Gram-Schmidt orthogonalization: Transforms correlated memory polynomial terms into uncorrelated components
• Benefit: Enables stable coefficient extraction even with high-order nonlinearities and long memory depth

Regularization adds a penalty term λ||w||² to the least-squares cost function, constraining coefficient magnitudes and preventing the inversion of near-singular matrices.

• Tikhonov (L2) regularization: Adds λI to the autocorrelation matrix, ensuring positive definiteness
• LASSO (L1) regularization: Promotes coefficient sparsity while stabilizing the solution
• Elastic Net: Combines L1 and L2 penalties for correlated predictor groups
• Practical impact: A well-chosen λ reduces condition number by orders of magnitude without sacrificing model fidelity

Stable algorithms incorporate iterative refinement to correct residual errors introduced during initial coefficient estimation. Each iteration solves a correction equation using the original factorization.

• Mixed-precision refinement: Compute residuals in higher precision, apply corrections in working precision
• Convergence criterion: Monitor ||r|| / ||b|| to detect stagnation from rounding errors
• Application: Essential for real-time DPD coefficient updates where matrix decompositions are reused across adaptation cycles

The choice of matrix factorization and pivoting strategy directly determines numerical stability. Algorithms that avoid explicit matrix inversion are inherently more stable.

• QR decomposition with column pivoting: Ranks columns by linear independence, isolating near-dependent terms
• Singular Value Decomposition (SVD): Provides the most robust solution for rank-deficient matrices by truncating small singular values
• Cholesky decomposition: Efficient for symmetric positive-definite matrices but fails catastrophically if near-singular
• Avoid: Direct normal equations (AᵀA)⁻¹Aᵀb, which squares the condition number

Preconditioning transforms the original problem into an equivalent system with improved spectral properties. Proper scaling prevents floating-point overflow and underflow during intermediate calculations.

• Column scaling: Normalize basis function amplitudes to unit variance before regression
• Diagonal preconditioning: Apply D⁻¹A where D contains column norms
• Incomplete LU preconditioning: Approximates A⁻¹ for iterative solvers
• Signal normalization: Scale input signals to [-1, 1] range to match basis function domains

A direct measure of a matrix's sensitivity to numerical errors. The condition number of the data matrix quantifies how much a small change in the input (like rounding) can be amplified in the output (the estimated coefficients).

• High Condition Number: Indicates an ill-conditioned problem where solutions are highly sensitive to noise.
• Low Condition Number: Indicates a well-conditioned problem with stable, reliable solutions.
• Practical Impact: A condition number of 10^k can cause a loss of up to k digits of precision during matrix inversion.

The foundational optimization technique for extracting behavioral model coefficients by minimizing the sum of squared residuals. Its numerical stability is directly tied to the structure of the normal equations.

• Normal Equations: Solving A^T A x = A^T b squares the condition number, making the system more sensitive to errors.
• QR Factorization: A more stable alternative that works directly with the data matrix A without forming A^T A.
• Use Case: The primary method for offline extraction of Memory Polynomial and Generalized Memory Polynomial coefficients.

A technique that adds a penalty term to the cost function to constrain coefficient magnitudes, directly improving numerical stability by reducing the effective condition number.

• Ridge Regression (L2): Adds a penalty proportional to the square of coefficient magnitudes, shrinking them toward zero without forcing sparsity.
• LASSO (L1): Adds a penalty proportional to the absolute value of coefficients, promoting coefficient sparsity and automatic model pruning.
• Tikhonov Regularization: A generalized form commonly used to stabilize ill-posed inverse problems in model extraction.

A modeling failure where the extracted model memorizes training data noise instead of learning the true system dynamics. It is a direct consequence of poor numerical stability during estimation.

• Symptom: Excellent performance on training data but poor generalization to new signals, visible in degraded Normalized Mean Square Error (NMSE) on test sets.
• Cause: High model complexity combined with ill-conditioned data matrices leads to wildly oscillating coefficient values.
• Mitigation: Employ cross-validation and regularization to ensure the model captures physical behavior, not measurement artifacts.

A simplified Volterra series model that retains only diagonal terms, offering a computationally efficient structure with inherently better numerical stability than full Volterra models.

• Structure: y(n) = Σ_k Σ_m a_{km} x(n-m) |x(n-m)|^{k-1}
• Stability Advantage: The reduced parameter count and diagonal-only structure lead to better-conditioned data matrices compared to the full Volterra Series.
• Implementation: The most widely adopted model for real-time FPGA-Based DPD Implementation due to its balance of fidelity and stability.

An iterative adaptive filtering algorithm that updates coefficients sample-by-sample, offering inherent robustness to numerical drift in real-time tracking applications.

• Update Rule: w(n+1) = w(n) + μ e*(n) x(n)
• Stability Mechanism: The step-size parameter μ provides direct control over convergence and numerical stability. A smaller μ improves stability at the cost of slower tracking.
• Application: Used in Online Training Algorithms for tracking slow changes in power amplifier behavior due to thermal effects.