Expand description
PPM (Prediction by Partial Matching) byte-level predictor — PPMd variant.
A fundamentally DIFFERENT prediction paradigm from CM:
- CM: hash-based, lossy collisions, bit-level, fixed context orders
- PPM: byte-level, adaptive order with escape/exclusion, checksum-validated
Full-range PPM with orders 0-12. Lower orders (0-3) provide a good fallback on small files. Higher orders (10-12) go BEYOND what any CM model covers, providing unique prediction signal on large files. All orders use checksum validation to avoid hash collisions.
The PPMd Method D escape estimation (Shkarin formula) and full exclusion mechanism provide a different error profile from CM’s bit-level hash models.
CRITICAL: PPM updates at BYTE level, not bit level. Only update after all 8 bits decoded. Byte probabilities are cached and converted to bit predictions on each bit.