Adaptive Filter

The defining feature of an adaptive filter is its ability to automatically modify its transfer function—the mathematical relationship between input and output—without external intervention. This is achieved by an optimization algorithm that iteratively adjusts the filter's coefficients based on an error signal. In DPD applications, this allows the predistorter to continuously track changes in the power amplifier's nonlinear behavior caused by temperature drift, aging, or channel frequency changes.

Adaptation is fundamentally driven by a cost function that quantifies the discrepancy between the desired output and the actual system output. The filter's coefficients are updated to minimize this cost function. Common cost functions include:

• Mean Squared Error (MSE): Minimizes the average of the squared error signal.
• Least Squares (LS): Minimizes the sum of squared errors over a block of data.
• Instantaneous Squared Error: Used in stochastic gradient methods like LMS for sample-by-sample updates.

A fundamental design tension exists between convergence rate and steady-state misadjustment. A large learning rate or aggressive step size enables the filter to rapidly converge to the optimal coefficient set, which is critical for tracking fast-changing PA dynamics. However, this same aggressiveness causes the coefficients to jitter around the optimal solution, introducing excess misadjustment noise that degrades final linearization performance. Conversely, a small step size yields a cleaner steady state but sluggish tracking.

Unlike fixed filters designed for static conditions, adaptive filters excel in non-stationary environments where signal statistics or system characteristics change over time. The forgetting factor in recursive algorithms like RLS is a critical parameter that exponentially discounts older data, giving greater weight to recent observations. This enables the DPD system to track time-varying phenomena such as thermal memory effects in GaN power amplifiers, where the PA's nonlinear profile shifts as the transistor junction temperature changes during transmission bursts.

Adaptive algorithms span a wide spectrum of computational requirements, directly impacting hardware implementation feasibility:

• LMS: O(N) complexity per iteration. Minimal compute, ideal for high-sample-rate FPGA implementation.
• NLMS: O(N) complexity with an additional input power normalization step. Slightly higher cost for improved stability.
• RLS: O(N²) complexity due to explicit matrix inversion or recursive covariance update. Superior convergence but often prohibitive for wideband DPD on resource-constrained hardware.
• QR-RLS: O(N²) complexity with superior numerical stability, achieved through orthogonal triangular decomposition.

When deployed on fixed-point hardware like FPGAs or ASICs, numerical stability becomes a dominant concern. Algorithms that are theoretically sound can diverge due to the accumulation of quantization errors and round-off noise. The correlation matrix in block-based estimators can become ill-conditioned when basis functions are highly correlated, leading to wildly inaccurate coefficient estimates. Techniques such as regularization—adding a small scalar to the matrix diagonal—and QR decomposition are employed to maintain stability in finite-precision arithmetic.

A stochastic gradient descent algorithm that updates filter coefficients to minimize the instantaneous squared error. LMS is prized for its O(N) computational complexity and simple implementation, making it the default choice for resource-constrained FPGA deployments.

• Update rule: w(n+1) = w(n) + μ·e(n)·x(n)
• Converges to the Wiener solution in the mean
• Sensitive to eigenvalue spread of the input correlation matrix
• Trade-off: larger step size μ speeds convergence but increases steady-state misadjustment

An adaptive algorithm that recursively minimizes a weighted linear least squares cost function. RLS achieves an order-of-magnitude faster convergence than LMS by using all available data, at the cost of O(N²) complexity.

• Incorporates a forgetting factor λ to track non-stationary PA behavior
• Converges independently of the input signal's eigenvalue spread
• Standard RLS is numerically sensitive; square-root and lattice variants improve stability
• Ideal for rapid initial acquisition during PA warm-up transients

A variant of LMS that normalizes the step size by the instantaneous power of the input vector. This normalization stabilizes convergence when the input signal exhibits significant power fluctuations, as is common with modern OFDM waveforms.

• Update rule: w(n+1) = w(n) + (μ / ||x(n)||²)·e(n)·x(n)
• Prevents gradient noise amplification during high-power signal bursts
• Adds only a modest computational overhead over standard LMS
• Often preferred over LMS in practical DPD implementations

The instantaneous difference between the desired linear output and the actual observed PA output. This signal is the fundamental driving metric for all adaptive filter coefficient updates in closed-loop DPD.

• Computed as: e(n) = y_desired(n) − y_observed(n)
• Requires precise time alignment between reference and feedback paths
• Any misalignment introduces residual distortion that the filter cannot correct
• The statistical properties of e(n) directly determine the achievable EVM floor

A mathematical function that quantifies the aggregate error the adaptive filter seeks to minimize. In DPD, the cost function typically measures the mean squared error (MSE) between the ideal linear output and the PA's actual output.

• Common forms: MSE, weighted MSE, least squares
• The choice of cost function directly shapes the filter's convergence behavior
• Regularization terms may be added to prevent ill-conditioning
• Minimizing MSE in the time domain correlates strongly with ACLR improvement

The speed at which an adaptive algorithm approaches the optimal coefficient set, measured in iterations or time. Convergence rate is a critical design parameter for DPD systems that must track rapidly changing PA characteristics.

• LMS: slow convergence, highly dependent on input statistics
• RLS: fast convergence, independent of input conditioning
• NLMS: moderate convergence, stabilized against power variations
• Initial convergence during PA cold-start may require a dedicated training sequence before entering background calibration mode