[e2e] Re: [Tsvwg] Really End-to-end or CRC vs everything else?

[Jonathan Stone]

In message <5.1.0.14.2.20010611181116.046205a0 at mail.reed.com>,
"David P. Reed" writes:


>Not to belabor this, but just passing the check-function test is not 
>sufficient.  It must also pass a variety of other tests to be accepted in 
>the context of a protocol (correct sequence number, correct address, ...).
>
>So the simple combinatoric argument  you sketch below doesn't quite 
>work.  You can't synthesize just any datagram that passes the checksum.  It 
>also has to meet the other constraints.

Oh, sure Thanks for pointing it out; I never meant it to be taken
that literally.  (In fact the next chapter of my dissertation talks
about this; I went through it in gory detail in in the ATM simulations
I did with Michael Greenwald, Jim Hughes, and Craig Partridge.)

You can interpret the result as applying to a check over just the
payload.  Or you an apply the result to all the bits at a give layer,
by sitting down and counting how many bits of information are check by
the syntactic check at a given layer.  Provided the ``check'' is a
function, not a relation, the proof still holds. (And yes, I did the
work to show how to extend it to a relation, though I didn't address
the specific example of TCP's window.)

The insight that we really cannot rank error-check functions _in
vacuo_, without reference to some error model[*]
(in the sense I use the phrase) is still valid.  

[*] Or with reference to some input distribution, if you
buy the information-maximization story.