LUT Normalization

The primary function of normalization is to scale the input signal envelope so that its maximum value corresponds exactly to the highest LUT address. Without normalization, an input signal exceeding the table's design range causes address overflow, where the indexing logic wraps around or saturates, applying incorrect predistortion coefficients. This results in severe spectral regrowth and potential damage to the power amplifier. The scaling factor is typically derived from the peak-to-average power ratio (PAPR) of the expected signal and the amplifier's compression point.

In hardware implementations on FPGAs or ASICs, normalization must account for fixed-point arithmetic limitations. The scaling operation maps floating-point envelope values to integer addresses within a power-of-two range (e.g., 0 to 255 for an 8-bit index). The normalization factor is chosen to maximize the utilization of available quantization levels without causing overflow. This involves a trade-off: a factor too large wastes addressable range, while one too small risks saturation at the upper index, clipping the correction signal.

In adaptive predistortion systems, the signal's statistical distribution can drift due to changes in modulation scheme, traffic load, or temperature. A static normalization factor becomes suboptimal. Advanced implementations employ adaptive normalization, where the scaling factor is continuously recalculated based on a running estimate of the peak signal envelope. This ensures the LUT indexing remains optimally centered on the active dynamic range, maintaining linearization performance under varying operating conditions.

Normalization directly interacts with LUT granularity. A coarsely spaced table (few entries) requires precise normalization to ensure the limited address space covers the most critical nonlinear region, typically near the amplifier's gain compression point. Fine-granularity tables are more forgiving of slight normalization errors. The normalization scheme often incorporates a deliberate offset or non-linear mapping (e.g., square-root or logarithmic compression) to allocate more addresses to the low-power region where AM-PM distortion varies most rapidly.

For LUTs that incorporate memory depth, normalization extends to multiple dimensions. The instantaneous envelope is normalized for the primary index, while delayed envelope samples (memory taps) must be normalized using the same or a coordinated scaling factor. Inconsistent normalization across dimensions creates a mismatch in the indexing space, distorting the multi-dimensional correction surface. This is critical in memory polynomial predistorters where the LUT address is a vector of normalized envelope values.

Normalization errors are amplified by LUT interpolation. If the normalized address does not accurately represent the true signal envelope position between table entries, linear or polynomial interpolation will compute a correction factor from the wrong neighboring coefficients. This interpolation error manifests as in-band distortion and spectral regrowth. Optimal normalization ensures that the spacing between adjacent addresses corresponds linearly to the spacing between the stored predistortion values, preserving the assumptions of the interpolation algorithm.