Context Mixing Lite

The ContextMixingLite codec provides a PAQ-style context mixing entropy coder for extreme compression.

Theory

Context mixing coders use multiple independent models (e.g., different historical context lengths or patterns) to predict the next bit or byte. A mixer (often a neural network or logistic regression) combines these predictions dynamically based on which models have been most accurate recently.

This approach achieves state-of-the-art compression ratios but is computationally expensive during both encoding and decoding.

Usage

This codec is opt-in and is not included in the default trial-encode dispatch menu. It is only selected when explicitly named in the codec_menu. It requires a minimum plane size of 4 KiB to amortize its heavy startup costs.

References

• Matt Mahoney, "Data Compression Using Neural Networks" / PAQ algorithm family. Link