LUT Step Size

The step size establishes a fundamental engineering trade-off in adaptive LUT systems. Larger step sizes accelerate initial coefficient convergence, allowing the predistorter to rapidly track changes in power amplifier nonlinearity caused by temperature drift or channel switching. However, this agility introduces excess mean squared error at steady state, manifesting as coefficient jitter around the optimal Wiener solution. Smaller step sizes minimize this residual misadjustment noise, improving adjacent channel leakage ratio (ACLR) performance, but slow down the adaptation loop's ability to acquire the inverse PA characteristic during cold start or abrupt environmental changes.

Adaptive LUT algorithms, particularly those based on least mean squares (LMS), exhibit a strict stability condition tied to the step size. The step size must remain below an upper bound determined by the maximum eigenvalue of the input signal's autocorrelation matrix. Exceeding this threshold causes the coefficient error vector to grow exponentially, leading to algorithmic divergence where the predistorter injects distortion rather than canceling it. In practice, the bound is approximated using the total input power to ensure robust operation across varying signal statistics without requiring computationally expensive eigenvalue decomposition.

The effective adaptation rate varies with instantaneous signal envelope power when using a fixed step size. High-power samples, which index into the LUT's gain compression region, generate larger gradient estimates and proportionally larger coefficient updates. This can cause non-uniform convergence across the LUT address space. Normalized step size schemes counteract this by scaling the update term inversely with the instantaneous input power or the local signal activity at each LUT entry, ensuring uniform convergence behavior from the linear region through deep saturation.

Advanced LUT adaptation architectures decouple the step size from a single global parameter, assigning individual step sizes to each table entry. Entries corresponding to frequently visited signal envelope levels can employ smaller step sizes to minimize steady-state jitter, while rarely addressed entries in the compression knee region retain larger step sizes for faster acquisition. This address-dependent step size strategy optimizes the convergence-misadjustment trade-off across the entire dynamic range, improving overall linearization performance compared to a uniform global parameter.

Implementations often employ a time-varying step size schedule to decouple acquisition and tracking phases. During initial LUT training or after a detected fault condition, a large step size forces rapid coefficient convergence. Once the residual error falls below a predefined threshold, the algorithm anneals the step size to a smaller steady-state value. This gear-shifting approach provides the fast initial lock time of an aggressive step size with the low spectral regrowth of a conservative setting, without requiring manual tuning for each operational scenario.

The step size directly influences the achievable ACLR performance of the linearized transmitter. Excessive step size introduces coefficient dithering noise that modulates the predistorted signal, creating wideband noise shoulders in the output spectrum that degrade ACLR. Reducing the step size suppresses this adaptation noise, allowing the residual distortion floor to approach the limits set by LUT granularity and interpolation error rather than algorithmic misadjustment. System designers must verify ACLR compliance under worst-case step size conditions during conformance testing.

The foundational iterative algorithm for coefficient adaptation. It minimizes the mean squared error between the desired linear output and the actual distorted feedback signal. The LUT step size directly controls the convergence behavior of this algorithm.

• Recursively updates coefficients based on instantaneous error
• Low computational complexity for hardware implementation
• Sensitive to step size selection for stability

The frequency and magnitude at which table entries are refreshed. A faster rate improves tracking of thermal memory effects and Doherty amplifier dynamics but introduces steady-state jitter. The step size is the primary tuning knob for this rate.

• Balances tracking agility against noise
• Must be faster than PA characteristic drift
• Often implemented with variable step size scheduling

The stable state where the error signal power is minimized and coefficients cease to change significantly. Convergence time is inversely proportional to the LUT step size—larger steps speed up initial lock but may prevent settling to the true minimum.

• Defined by residual Normalized Mean Squared Error (NMSE)
• False convergence can occur with step size mismatch
• Verification requires monitoring coefficient variance

Distortion arising from representing the continuous predistortion function with finite-resolution coefficients. The step size interacts with quantization: if updates are smaller than the quantization level, the LUT stalls. If too large, it oscillates between levels.

• Determined by fixed-point word length in FPGA
• Adds a noise floor to the linearized output
• Mitigated by dithering techniques in the update loop

Mathematical estimation of correction values between discrete table entries. Linear or polynomial interpolation smooths the staircase effect of a finite LUT granularity. The step size must be coordinated with interpolation order to avoid amplifying interpolation error during adaptation.

• Reduces required table size for a given EVM target
• Linear interpolation is most common in FPGA implementations
• Higher-order interpolation increases computational latency

A post-processing filter applied across adjacent table entries to remove adaptation noise. It prevents spectral regrowth caused by abrupt coefficient discontinuities. Smoothing acts as a low-pass filter on the adapted surface, complementing the step size's role in controlling update granularity.

• Typically a moving average across neighboring bins
• Reduces ACLR degradation from noisy coefficients
• Introduces a slight lag in tracking fast PA changes