Release Notes
----------------------

Release 9.1 is a bug-fix for version 9.0 ; Only two lines in
ppmzesc.c have been changed : if ( escape_prob < 1 ) escape_prob = 1 ;
new 3-2-98 : also fixed (semi-bugs) to handle files that expand more gracefully.

September '97 : PPMZ 9.0 is here.  It's been more than a year since the
last release, and really not much has changed.  I've implemented
a cleaner, better "Z-Escape" modeller (the deterministic escape
modeller I invented for PPMQ, which is the secret to our high
compression) : it now resides in it's own file 'ppmzesc.c'.  This
new escape modeller uses ideas from Malcolm Taylor and the CTW
team to weight contexts.

PPMZ 9.0 also implements an optional LZP pre-pass.  I have not
been able to tune this so that it helps compression, though it
doesn't hurt too much, and it does help speed...

This new version get 2.222 bpb on paper1 , and 1.236 on trans,
both are new records.  These will be roughly on par with the
newest release of RKIVE (not out yet).  The CTW algorithm still
beats us on the books and paper2.  I believe PPMZ should retake
the 'best corpus average' from RKIVE and CTW which had surpassed
it while I was inactive.

All of my big ideas for improving PPMZ will require a major
restructuring of the program, so as to facilitate weighting,
or dynamic LOE, or bit-wise coding, or a seamless merging of
finite and infinite length contexts.

v8.3 fixes another bug, and makes the code a tad cleaner; the LOE
(Local Order Estimation) coders, Coder #9 and up, were using
floating point numbers in their math; this was causing different
files to be made on different architectures.  All math has been
switched to integers, and this problem is gone.

v8.0 is finally available.  This version fixes all known bugs in PPMZ.
Previously PPMZ was doing some probability averaging which occasionally
allowed the total symbol probability to exceed 32768 (which is a no-no).
This bug was only visible on large, skewed files, and has now been
fixed.

v8.1 : 8-8-96 : LRU option works, but is far from perfect. It's killing
compression and it really shouldn't be.

v8.0 notes:

an LRU to limit memory use if provided but IS NOT WORKING.  Hopefully, this will
be finished soon for v8.1

all known bugs fixed.

Version/Revision standardization: files with the same version number are binary-compatible.

Headers (thus files) should now be endian-independent.

Coder-Guess functionality has been removed.  Use the LOE coders instead.

v7.6 notes:

7-20-96 : v7.6 still has bugs.  It works on almost all files, but
occasionally produces errors on decode.  Use at your own risk.
(sorry, guys!)

7-10-96 : v7.6 had a bug in crblib arithc coder : improper termination
of the arithcoder was used that caused problems on some files.  The
ppmz distribution is unchanged, but you need the new crblib to fix
the bug.

Coders 9 through 13 use Local Order Estimation.  These coders perform
consistently better than anything else.  In PPMZ v8 the LOE coders will
supplant all others.

PPMZ v7.6 IS FULLY ANSI COMPATIBLE.  This means that even if your
compiler is not PPMZ will try to make it look like ANSI.
I compile with GCC v2.7.2 with no warnings.  I compile
with Watcom for Windows NT/95 with no warnings.  Visual C sucks, don't
use it.

PPMZ v7.6 runs just fine on Intel chips.

v7.5 includes Recency Scaling, and a preliminary RunLength step (no more
waiting for 'pic' !!) and some fine-tunings for a little more compression

v7.4 adds Automatic Order Guessing (it works on the Corpus, but I'm not too sure about the
rest of the world), and a bug fix which prevented memory freeing.

v7.3 is hopeful little-endian compatible (nope) , eliminates use of timer(),
 and has a better testing interface

v7.2 merges PPMC and PPMQ into PPMZ as special cases of that more
 general method

v7.1 adds 'PPMdet' as the top level, an unbounded-order deterministic-only coder

v7.0 is True ANSI C (no C++ comments) and should work on any UNIX system;
functionality on MS-CRAP based systems is questionable.

v7.0 also is RISC compatible