skainet-lang-core/sk.ainet.lang.tensor.ops.turboquant/QjlResidual

QjlResidual

QJL (Quantized Johnson-Lindenstrauss) residual stage for TurboQuant.

After scalar quantization, there is a residual error: residual = original_rotated - dequantized

The QJL stage projects this residual onto a random low-dimensional subspace and quantizes the projection. This preserves inner-product accuracy (Johnson-Lindenstrauss property) at the cost of additional storage.

This stage is used only by the TurboQuantPolarQjl variant. The TurboQuantPolar variant omits it for simplicity and speed.

Functions

decode

fun decode(encoded: EncodedResidual, output: FloatArray, seed: Int)

Decode a QJL residual and add it to the base reconstruction.

encode

fun encode(residual: FloatArray, residualBits: Int, seed: Int): EncodedResidual

Encode a residual vector using QJL projection.