Update idea document for zmicro.

notes/zmicro.md

@@ -1,10 +1,31 @@
-ZMicro is a compression algorithm, which was specifically designed for the use in RACC.
+ZMicro is a compression algorithm, which was specifically designed for the use
+in RACC.
 
-ZMicro uses a standard bit-by-bit arithmetic encoder, but has an adaptive model.
+ZMicro uses a standard bit-by-bit arithmetic encoder, but has an adaptive
+model.
 
 The model consists of various submodels:
 
-1. Prediction by partial matching with context order N.
-2. Bit dependency table.
+- Prediction by partial matching with context congruence classes modulo N of order M.
+- Average bit on same congruence class modulo N.
+- Bit dependency table per N bits.
+- Fifty-fifty (identity)
+- Repeat last
 
-These are combined through adaptive context mixing, which works by weighting with the match rate. If model M reads a bit b=1 with prediction of being one P(b=1), then P(b=1) is added to the total sum. If b=0, P(b=0)=1-P(b=1) is added.
+These are combined through adaptive context mixing, which works by weighting
+with the match rate.
+
+If model M reads a bit b=1 with prediction of being one P(b=1), then P(b=1) is
+the error. If b=0, P(b=0)=1-P(b=1) is the error.
+
+Each model gets an accumulated error rate, which consists of a weighted sum of
+the individual errors. For every new error, the sum is multiplied by some
+number K≤1, called the cool-down factor, then the error is added.
+
+Models are compared by ratio. For example model M1 is better than M2 by an
+factor of M1/M2. In other words, a model M is judged on its ratio to the sum of
+all error accumulations.
+
+If a model rates a k'th (for some factor k called the exit factor) of the
+accumulated error sum, it is dropped and it's error accumulation is subtracted
+from the overall sum.