LUT Training

LUT training is categorized by when adaptation occurs relative to live signal transmission.

• Offline Training: Coefficients are extracted from captured PA input-output data in a lab environment using least-squares or iterative learning control. The LUT is then frozen during deployment.
• Online Training: Adaptation runs continuously in the field using a feedback path. Algorithms like LMS or RLS update entries in real-time to track temperature drift and aging.
• Hybrid Approach: An offline-trained table serves as the initial seed for online adaptation, ensuring rapid LUT convergence without the risk of startup instability.

The training architecture defines how the error signal is computed to update LUT coefficients.

• Indirect Learning Architecture (ILA): The most common approach. A post-distorter is trained in the feedback path to model the PA's inverse. Once converged, coefficients are copied to the forward predistorter LUT. This avoids instability from amplifier saturation.
• Direct Learning Architecture (DLA): The predistorter itself is trained by minimizing the error between the desired output and the actual PA output. Requires careful gain management to prevent loop oscillation.
• ILA is preferred for its inherent stability, especially during initial LUT initialization when the PA model is unknown.

The Least Mean Square (LMS) algorithm is the workhorse of real-time LUT training due to its computational simplicity.

• Operation: For each LUT entry addressed by the instantaneous signal envelope, the complex coefficient is updated as: w(n+1) = w(n) + μ * e(n) * conj(x(n))
• Step Size (μ): Controls the LUT adaptation rate. A larger μ accelerates convergence but increases steady-state jitter. A smaller μ reduces noise but slows tracking.
• Normalized LMS (NLMS): Scales μ by the input signal power to maintain consistent convergence behavior across varying signal levels, preventing instability during power transients.

Training must account for how coefficients are used during predistortion, not just how they are stored.

• LUT interpolation means a single transmitted sample may blend multiple table entries. Training algorithms must distribute the error correction across adjacent entries proportionally.
• Linear Interpolation: The error is split between the two nearest entries based on the fractional address. This prevents LUT quantization error from dominating residual distortion.
• Higher-order interpolation (cubic, spline) requires distributing updates across more entries, increasing computational load but enabling coarser LUT granularity without sacrificing linearization performance.

The quality and statistical properties of the training signal directly determine LUT accuracy.

• Signal Statistics: The training waveform must exercise the full dynamic range of the PA. Signals with high peak-to-average power ratio (PAPR) ensure all LUT entries receive sufficient updates.
• Memory Depth: For memory-compensating LUTs, the training data must include time-varying envelope trajectories to populate multi-dimensional address spaces.
• Data Length: Sufficient samples are needed for each LUT entry to converge. A rule of thumb is 1000+ updates per entry for LMS-based training to achieve stable LUT convergence.
• Real-world signals (OFDM, 5G NR) are preferred over synthetic tones because they capture the actual statistical distribution the PA will encounter.

Robust training requires mechanisms to detect convergence and prevent divergence.

• Error Vector Magnitude (EVM) and Adjacent Channel Leakage Ratio (ACLR) are monitored in real-time to assess linearization quality.
• Step-size annealing: μ is reduced over time to transition from fast acquisition to low-jitter steady-state operation.
• Coefficient bounding: LUT entries are clamped to physically realizable gain and phase ranges to prevent the adaptation loop from driving the PA into deeper saturation.
• Ping-pong LUT architectures allow training to run on a background table while the foreground table performs predistortion, enabling seamless coefficient updates without glitching the transmitted signal.

The foundational iterative adaptation algorithm for LUT training. It minimizes the mean squared error between the desired linear output and the actual PA output.

• Uses a feedback error signal to compute gradient descent updates.
• Updates only the active LUT entry indexed by the instantaneous signal envelope.
• Convergence rate is controlled by the step size parameter (mu).
• Balances tracking agility against steady-state coefficient jitter.

A mathematical technique for estimating predistortion values between discrete table entries to reduce quantization error.

• Linear interpolation: Simple hardware implementation using adjacent entries.
• Polynomial interpolation: Higher accuracy using multiple surrounding points.
• Critical for achieving smooth correction with limited table sizes.
• Directly impacts adjacent channel leakage ratio (ACLR) performance.

The process of setting initial coefficient values before online adaptation begins. Proper initialization ensures stable convergence and prevents amplifier damage.

• Linear gain model: Sets all entries to a constant complex gain.
• Previously converged values: Restores a known good state from memory.
• Model-based seeding: Uses an offline extracted PA behavioral model to pre-compute inverse coefficients.
• Prevents transient spectral regrowth during startup.

The speed at which LUT coefficients are updated, controlling the trade-off between tracking agility and steady-state noise.

• Fast adaptation: Tracks rapid PA changes due to temperature or voltage drift.
• Slow adaptation: Reduces coefficient jitter and residual NMSE.
• Often implemented with variable step size that decreases over time.
• Critical for Doherty PAs where load modulation changes nonlinear characteristics dynamically.

The computational procedure for deriving optimal predistortion values from measured PA input-output data.

• Direct inversion: Computes the inverse nonlinear function from AM-AM/AM-PM curves.
• Indirect learning: Uses the PA output to iteratively solve for predistorter coefficients.
• Least squares fitting: Batch processes captured data blocks for optimal coefficient sets.
• Enables offline training before deployment in the field.

The state where iterative adaptation has minimized the error signal to a stable residual level. Indicates the LUT accurately models the inverse PA nonlinearity.

• Monitored via error vector magnitude (EVM) or NMSE metrics.
• Convergence time depends on adaptation algorithm and step size.
• False convergence can occur with insufficient excitation signal bandwidth.
• Requires persistent excitation across the full dynamic range for complete table population.