Re: QI and MQ Coder: First real-life experiences

"nightlight" <nightlight@xxxxxxxxxxxxxx> wrote in
news:1139080242.233127.152630@xxxxxxxxxxxxxxxxxxxxxxxxxxxx:

I did a test for the 1,000,000 symbol case and based
on what little I could squeeze out of him his coder fails
to compress as well.


My newsreader finally picked it up. So thought I would read his
post. Naturally its mostly hype.

There is no mystery left there at all. Your hardwired _64_ bit
precision AC produces 198121 bytes which is 1584968 bits.
All of your bits are used (since altering any bit in the last byte
breaks the decode to correct input). Also, you're not passing value
N=10^6 (your decoder gets it from OS via file size) or A=3 (it is
hardwired in you encoder & decoder).

Bull. I showed how the length could be added and still it was shorter
than what you claimed was your fixed length. The problem is your trying
to change the game after you failed. I don't blame people for laughing
that you can't get it to work with files. I guess its harder to make
up numbers when people get to use files. You lost your wrong live with it!!

At least your no longer throwing a tamper tantrum about the fact
that a file of all A's still compresses to the same. I wonder what your
code does since appearantly your stuff hand tuned to skip those cases.
Its no wonder you can't get it to work with files. Can't let the masses
actually test it can you?


Under these _same conditions_ for the N and A availability to
decoder, the _32_ bit QI produces output of 1584962.50... bits
(the tapered Huffman code for the top 32 bits of index reaches
on average to within 0.0016890... bits from the exact fraction).
That is only 1.6e-4 bits above the ideal N*log(3) value.

Now, if you want some other coding conditions, e.g. to include
encoding of N and A into the coded output ...

Again I showed how that could be done in a buffer. But the test
was based on what you clearly thought your code would be the best at
Stop trying to change it. Can you write to a file or not so that
others can test it. Be a man (unless your a woman) and either
stop whinning or do it. Or admit you can't. Its really that simple.
No one here belives your numbers there has been to much water
under the bridge for that.


But as explained many times earlier, for any given max codeword
precision of g bits, and requiring that the codewords satisfy Kraft
inequality (decodability condition, cf. eq. (20) in [T3]), the QI's
index .....

Who cares what you EXPLAIN since you don't seem to have a basic
understanding of even simple arihtmetic compression who cares. At
least I see you gave up the garbage about airthmetic compression
not making full use of the output space since it clearly can. Which
is a lot different than your first posts. Maybe you are learning
something yet. There my still be hope.

The QI output size is also mathematically guaranteed to always fit
the size shown (which is determined soleley from N and A via QI
powers table). QI will also code much faster than your 64 bit coder
(even when changed to memory to memory coding), or any other
64 or 32 bit AC (take a look at your multipass ... BLAH BLAH

You wrong about no gaps in an arithmetic your wrong about many
things. Can you code this to files since it the only fair test.
And you cliam its a drop-in replacement for arithmetic. Well if
it is PROVE IT. Since your so way off base on what arithmetic does
it would be foolish to belive any other of your off tangent posts.


In short, smaller output and much faster codinng than AC, as
stated before.


Funny thats not what Thomas got. Your going to have change
your mantra unless you can someday actually get code that does
something.


Of course, this A=3 is the best case for AC. You can try
the test suggested earlier for a much larger A, as described
in the post above (to Jasen). ....

Funny you like A=3 in much earlier posts I wonder why now its
not so hot. Yes I could change my code for higher values but since
you can't understand what happens at A=3 what would be the point.
You can't even get QI to work on this simple example using files
that anyone can test. Why is that?


So please tell what it gets for the extremely simple 5
symbol case.

You can test that with the released copy of QI.exe using
the command line:

QI.exe cr3 i10 n5

which is for g=32 precision the exact enumerative coding (the only
error shown for above command is 1.001e-12 bits which is within
the MSVC floating point error on double for their logs base 2). The
mantissa value for 3^5 is 0xF3=243 (while exponent is 0) i.e. it is
identical to the exact binary code for radix 3 numbers, i.e. 7.9248...


I just love this why am I not surprised. I missed this gem
in the first reading.


What more can I say you could not beat the simple case you
give before what your exact fixed length would be. You have yet
to code even this exteremly simple case to files. It should be
easy after all you rant over and over again how QI is always
better than the best arithmetic. But so far Thomas has showed
its much slower and compresses worse than arithmetic. I modifed
a toy arithmetic compresser to do the case you actually first
brought up. I did make it work to files since that is what
people need to have a true evaluation not your highly controlled
tuned so called test code. Can you code this example to files or
is QI such a poor infexable replacement for an entropy coder that
it would be beyound your capability. Look people here don't mind
mistakes if people are honest about it. I make mistakes I think
you do to. Admit it QI might be hot but you can't make it work
better with files in this case.

In fact if you code it and it works on files and is worse than
arithmetic which I assume you found out since you must have tried
it. After all if you insist you don't have the time people here
would laugh at you. Let the people here help. Show your test code
people will make it better. They might even make it match my code
if they only get the chance. You might even learn something if you
only try and let others help you. In fact many here including me
would be happy to see it beat my code. You do need help don't you?




David A. Scott
--
My Crypto code
http://bijective.dogma.net/crypto/scott19u.zip
http://www.jim.com/jamesd/Kong/scott19u.zip old version
My Compression code http://bijective.dogma.net/
**TO EMAIL ME drop the roman "five" **
Disclaimer:I am in no way responsible for any of the statements
made in the above text. For all I know I might be drugged.
As a famous person once said "any cryptograhic
system is only as strong as its weakest link"

.

Relevant Pages

  • Re: Please, PLEASE, hold your questions/comments/elsewhat til the end. Thank you. :)
    ... compression theory, information theory, and whatever else you'd like to ... does not allow for recursive compression. ... intermediate files had to have a unique MD5 hash. ... that you all could give my current posts a fair treatment regardless. ...
    (comp.compression)
  • Re: Speech processors...
    ... I try to keep my posts short... ... I said that clipping is a form of compression... ... As you know, harmonic distortion can be ... one wants to improve their transmission... ...
    (rec.radio.cb)
  • Re: Shovelhead headscratcher
    ... Do a compression check on both cylinders and compare. ... "Yeah, of course, I was all set to check the compression ... If you read other posts in the thread you'll note that I had a ...
    (rec.motorcycles.harley)
  • Re: Compress Old Files with Disk Cleanup
    ... compression anyway but I always dutifully did it in the past. ... But how does one restore the key? ... I applied the registry edit with the ... "canned" edit suggested in other posts on this subject and no backup was ...
    (microsoft.public.windowsxp.general)
  • Re: QI and MQ Coder: First real-life experiences
    ... >> ported for Unix (thanks Jasen Betts!) ... a static 0-order coder should do. ... > chariot against another chariot. ... My Compression code http://bijective.dogma.net/ ...
    (comp.compression)