Kalman Filtering

The Kalman filter operates as a recursive Bayesian estimator that optimally combines prior state knowledge with new noisy measurements. For DPD coefficient tracking, it maintains a probability distribution over the predistorter parameters, updating the mean and covariance with each new sample. This recursive nature eliminates the need to store historical data batches, making it ideal for sample-by-sample adaptation in real-time systems. The filter propagates the state estimate forward using a process model describing how coefficients are expected to evolve, then corrects this prediction using the measurement model relating coefficients to observed PA output.

DPD coefficient tracking is cast into a state-space representation where:

• State vector: Contains the predistorter coefficients (e.g., memory polynomial taps) to be estimated
• Process equation: Models coefficient evolution over time, typically as a random walk or first-order Markov process to capture thermal drift and aging effects
• Measurement equation: Relates the state to the observed PA output error signal through the basis function matrix

The process noise covariance Q controls how quickly the filter adapts to changing PA characteristics, while the measurement noise covariance R reflects observation path SNR. Tuning these matrices is critical for balancing convergence rate against steady-state misadjustment.

Each iteration executes a predict then update cycle:

Prediction Step:

• Projects the coefficient state estimate forward using the process model
• Propagates the error covariance matrix to reflect increased uncertainty from process noise
• Produces a prior estimate before incorporating the new measurement

Update Step:

• Computes the Kalman gain — the optimal weighting between prediction and measurement based on their relative uncertainties
• Corrects the state estimate using the innovation (difference between actual and predicted measurement)
• Reduces the error covariance to reflect the information gained

This structure naturally handles missing or intermittent feedback by skipping the update step when observation path data is unavailable.

The Kalman filter is the minimum mean squared error (MMSE) estimator when process and measurement noise are zero-mean, white, and Gaussian. Under these conditions, no other linear or nonlinear estimator can achieve lower estimation error. For DPD applications, this optimality translates to the fastest possible convergence rate for a given observation noise level. The filter inherently accounts for varying measurement quality — when the observation path SNR degrades, the Kalman gain automatically decreases, placing more trust in the predicted state. This prevents coefficient drift caused by noisy feedback corrupting the estimate.

When the DPD model or PA behavior introduces nonlinear relationships between coefficients and measurements, the standard linear Kalman filter is extended. The Extended Kalman Filter (EKF) linearizes the nonlinear measurement function around the current state estimate using a first-order Taylor expansion via the Jacobian matrix. This enables tracking of coefficients in:

• Volterra series and memory polynomial predistorters where basis functions are nonlinear in the input signal
• Systems with nonlinear observation paths or IQ modulator impairments
• Architectures requiring joint estimation of PA model parameters and predistorter coefficients

The trade-off is that EKF optimality is only approximate, and poor initialization can cause divergence if the linearization is inaccurate.

In fixed-point FPGA implementations, the standard Kalman filter covariance propagation can suffer from ill-conditioning and loss of positive definiteness due to finite precision arithmetic. Square-root Kalman filtering propagates the matrix square root of the error covariance (e.g., via Cholesky decomposition) rather than the covariance itself, guaranteeing numerical stability. The UD factorization variant stores the covariance as a unit upper triangular matrix and diagonal matrix, reducing computational complexity while preserving the condition number. These techniques are essential for high-order DPD models with dozens of coefficients where covariance matrices become poorly conditioned.

A deterministic adaptive filtering algorithm that recursively finds the coefficients minimizing a weighted linear least squares cost function. Unlike LMS, RLS uses all available input data to compute the gain vector, resulting in an order-of-magnitude faster convergence rate at the expense of higher computational complexity.

• Exhibits zero misadjustment in stationary environments
• Often formulated via the Matrix Inversion Lemma to avoid explicit matrix inversion
• Forms the deterministic foundation that the Kalman filter generalizes to stochastic state-space models

A stochastic gradient descent algorithm that updates filter coefficients using the instantaneous estimate of the mean squared error gradient. Its extreme computational simplicity makes it a baseline for real-time sample-by-sample update implementations.

• Convergence speed is governed by the step-size parameter μ
• Suffers from gradient noise, leading to steady-state misadjustment
• Often used as a lightweight alternative when the Kalman filter's O(N²) complexity is prohibitive

A closed-loop DPD topology where the predistorter coefficients are estimated directly by minimizing the error between the desired input and the actual PA output. The Kalman filter excels in this architecture by treating the time-varying predistorter as a dynamic system state.

• Eliminates the model inversion step required in indirect architectures
• Requires a robust transmit observation receiver for feedback
• Kalman-based DLA adapts seamlessly to thermal memory effects and PA aging

The gradual deviation of predistorter coefficients from their optimal values due to temperature changes, component aging, or numerical instability. Kalman filtering inherently combats this by maintaining a process noise covariance matrix (Q) that injects controlled uncertainty into the state estimate.

• Prevents the estimator from becoming overconfident in stale coefficients
• The Q matrix can be tuned to match the expected rate of physical parameter change
• Critical for maintaining ACPR compliance over long operational periods without recalibration

A numerically stable implementation of the RLS algorithm that uses QR decomposition (via Givens rotations or Householder transforms) to solve the least squares problem directly on the data matrix. This avoids the ill-conditioning and instability that plague the standard RLS covariance matrix update.

• Provides the numerical robustness required for fixed-point FPGA implementation
• Closely related to square-root Kalman filtering formulations
• Essential when the condition number of the input correlation matrix is high

A ridge regression technique that adds a penalty term (λ||w||²) to the least squares cost function to stabilize the solution of ill-posed inverse problems. In the Kalman filter context, this is mathematically equivalent to the initial state covariance matrix (P₀) and the process noise prior.

• Prevents wild coefficient swings during initial training or after signal dropouts
• The regularization parameter λ trades bias for variance in the estimate
• Works synergistically with the Levenberg-Marquardt algorithm for nonlinear PA models