Moore-Penrose Pseudoinverse

The pseudoinverse A⁺ is uniquely defined by four axioms: AA⁺A = A, A⁺AA⁺ = A⁺, (AA⁺)ᴴ = AA⁺, and (A⁺A)ᴴ = A⁺A. These conditions guarantee that A⁺ behaves as a true generalized inverse, preserving Hermitian symmetry in the projection operators. In DPD model extraction, satisfying these conditions ensures the coefficient vector minimizes the Euclidean norm of the residual error.

For an overdetermined system Ax ≈ b, the pseudoinverse yields x̂ = A⁺b, which is the unique minimizer of ‖Ax - b‖₂². This is the analytical solution to the normal equations AᴴAx = Aᴴb when A has full column rank. In power amplifier modeling, x̂ represents the optimal DPD coefficients that minimize the sum of squared errors between the linearized and measured output.

The pseudoinverse is computed via A⁺ = VΣ⁺Uᴴ, where A = UΣVᴴ is the SVD. The diagonal matrix Σ⁺ is formed by taking the reciprocal of each non-zero singular value and transposing. Singular values below a tolerance threshold are set to zero rather than inverted, providing a natural mechanism for rank truncation and numerical stabilization in ill-conditioned regression matrices.

When the basis function matrix is rank-deficient—common when using correlated polynomial terms—the pseudoinverse still provides a minimum-norm least-squares solution. Unlike the standard inverse, A⁺ exists for any matrix regardless of rank. The resulting coefficient vector has the smallest L2 norm among all vectors that minimize the residual, acting as an implicit regularizer that suppresses spurious coefficient growth.

Direct computation via A⁺ = (AᴴA)⁻¹Aᴴ is numerically unstable when AᴴA is ill-conditioned. The SVD-based approach avoids forming the normal equations entirely, operating directly on A. By setting a singular value threshold (typically 10⁻⁶ to 10⁻⁸ of the maximum singular value), near-zero singular values are truncated, preventing the amplification of measurement noise in the coefficient estimates.

The truncated SVD pseudoinverse is equivalent to principal component regression, where components associated with small singular values are discarded. This connects directly to ridge regression: adding an L2 penalty λ‖x‖₂² to the least-squares cost yields x̂ = (AᴴA + λI)⁻¹Aᴴb, which can be expressed using a modified pseudoinverse where singular values σᵢ are replaced by σᵢ/(σᵢ² + λ), smoothly shrinking rather than hard-thresholding.

A batch estimation algorithm that finds model coefficients by minimizing the sum of squared errors between predicted and measured outputs. The Moore-Penrose pseudoinverse provides the closed-form solution to the ordinary least squares problem when the design matrix is full column rank. For an overdetermined system Ax ≈ b, the LS solution is x̂ = A⁺b, where A⁺ is the pseudoinverse. This single-computation approach is the workhorse of offline power amplifier model extraction.

A regression scenario where the number of measurement equations exceeds the number of unknown model parameters. In DPD model extraction, thousands of captured I/Q samples form the equations, while the model has far fewer coefficients. No exact solution exists—the pseudoinverse yields the minimum-norm least-squares fit. The residual vector r = b − Ax̂ quantifies the unavoidable modeling error.

A numerical state where the correlation matrix of basis functions is nearly singular, causing the pseudoinverse computation to amplify measurement noise and rounding errors. The condition number κ(A) = σₘₐₓ/σₘᵢₙ quantifies this sensitivity. When κ is large, small perturbations in captured data produce wildly different coefficient estimates. Regularization techniques like ridge regression are essential countermeasures.

A regularized least-squares method that adds an L2 penalty λ‖x‖² to the cost function. The solution becomes x̂ = (AᵀA + λI)⁻¹Aᵀb, which is equivalent to computing the pseudoinverse of an augmented matrix. The regularization parameter λ shrinks coefficient magnitudes, trading a small increase in bias for a large reduction in variance. Critical when basis functions exhibit multicollinearity.

The most numerically stable method for computing the Moore-Penrose pseudoinverse. SVD factorizes the design matrix as A = UΣVᵀ, where Σ contains singular values σᵢ. The pseudoinverse is then A⁺ = VΣ⁺Uᵀ, with Σ⁺ formed by reciprocating non-zero singular values. SVD explicitly reveals the rank and null space, making it the gold standard for ill-conditioned problems in DPD coefficient extraction.

The matrix AᵀA containing pairwise covariances between basis functions. Its structure directly determines whether the pseudoinverse computation is well-behaved. Off-diagonal elements near ±1 indicate highly correlated regressors, leading to a nearly singular matrix. Analyzing the covariance matrix guides basis function selection—removing redundant terms improves conditioning and yields more robust coefficient estimates.