Wiener-Hopf Equation

A foundational stochastic gradient descent algorithm that iteratively approximates the Wiener-Hopf solution without requiring matrix inversion. Instead of computing the full autocorrelation matrix, LMS uses the instantaneous squared error gradient to update coefficients.

• Update Rule: w(n+1) = w(n) + μe(n)x(n)
• Computational Cost: O(N) per iteration, where N is filter length
• Convergence: Slower than RLS, especially for correlated input signals
• Misadjustment: Proportional to step size μ; larger μ speeds convergence but increases steady-state error

The LMS algorithm trades the Wiener-Hopf equation's batch optimality for continuous adaptability, making it suitable for real-time DPD where amplifier characteristics drift with temperature and aging.

An adaptive algorithm that recursively computes the inverse of the autocorrelation matrix to achieve convergence typically an order of magnitude faster than LMS. RLS directly implements the Wiener-Hopf solution at each iteration with exponentially weighted data.

• Core Mechanism: Uses the matrix inversion lemma to update R⁻¹(n) from R⁻¹(n-1) without full recomputation
• Forgetting Factor λ: Controls memory; λ=0.99 weights recent samples more heavily for tracking time-varying systems
• Computational Cost: O(N²) per iteration, significantly higher than LMS
• Convergence: Independent of input signal eigenvalue spread, unlike LMS

RLS is preferred in DPD applications where rapid adaptation to changing signal statistics is critical, such as wideband 5G systems with dynamic resource block allocation.

The closed-form batch solution to the linear least squares problem that directly yields the Wiener-Hopf equation. It is derived by setting the gradient of the mean squared error cost function to zero.

• Form: w_opt = (X^H X)⁻¹ X^H d, where X is the data matrix and d is the desired signal
• Relationship: The term X^H X is the sample autocorrelation matrix estimate; X^H d is the cross-correlation vector estimate
• Numerical Issues: Direct inversion of X^H X squares the condition number, making the solution sensitive to ill-conditioning
• Practical Use: Forms the basis for offline training of DPD coefficients from captured IQ data

The normal equation provides the theoretical minimum MSE benchmark against which all adaptive algorithms are measured.

A numerically robust implementation of RLS that avoids explicit matrix inversion by maintaining the Cholesky factor (square root) of the inverse correlation matrix. QR-RLS uses Givens rotations to update this factor directly.

• Stability: Maintains positive-definiteness of the correlation matrix even with finite-precision arithmetic
• Implementation: Well-suited for FPGA and fixed-point DSP platforms where numerical precision is limited
• Systolic Array Mapping: The triangular structure of the QR decomposition maps naturally to parallel hardware architectures
• Dynamic Range: Doubles the effective numerical precision compared to conventional RLS

QR-RLS is the algorithm of choice for mission-critical DPD implementations in base station transmitters where numerical stability under high PAPR signals is non-negotiable.

A postdistorter identification method that estimates DPD coefficients by placing a copy of the predistorter after the power amplifier in a training loop. The Wiener-Hopf equation is solved to minimize the error between the postdistorter output and the predistorter input.

• Key Assumption: Assumes the postinverse equals the preinverse, which holds for memoryless nonlinearities but degrades with strong memory effects
• Advantage: Avoids the need for PA model identification; directly estimates predistorter coefficients
• Limitation: The assumption breaks down when the PA exhibits significant nonlinear memory, introducing bias in the Wiener-Hopf solution
• Practical Use: Widely adopted in commercial DPD systems due to implementation simplicity

ILA transforms the DPD coefficient estimation problem into a standard system identification problem solvable by the Wiener-Hopf equation.

The ratio of the largest to smallest singular value of the autocorrelation matrix R, quantifying how sensitive the Wiener-Hopf solution is to small perturbations in the input data. A high condition number indicates an ill-conditioned problem.

• Impact on Wiener-Hopf: The solution error is bounded by κ(R) times the input perturbation; κ(R)=10⁴ means 4 digits of precision are lost
• Causes in DPD: Highly correlated input signals (narrowband modulation) or oversampled wideband signals create near-singular R matrices
• Mitigation: Tikhonov regularization adds δI to R, effectively capping the condition number at (λ_max+δ)/(λ_min+δ)
• Algorithm Sensitivity: LMS convergence slows proportionally to κ(R); RLS convergence is theoretically independent of κ(R)

Understanding condition number is essential for diagnosing why DPD coefficient estimation may fail with certain signal types or sampling rates.