Ok, sorry, thought it was another building of a strawman.
Need to read the responses, but basically
- GSM uses 200kHz, kind of 200kbps for 8 timeslots.
- raw data speed could be 200kbps/8 = 25kbps/handset
(without using "higher modulations")
Some of the raw data rate rate is used to
- provide an "instant training signal" in the middle
of the data, time slot, makes it possible to
instantly calculate the channel responce of the
multi path fading,etc, fast connection times.
- some more lost to isolate the time slots (not much,
should maybe have been more)
- most is "lost" for forward error correction, or
channel coding, that convolutional coding.
Basically using only part of all the data which could
be transmitted, difficult to explaine but something
like this: (shannon-ungerboeck-viterbi stuff)
Instead of putting , let's say 10kbps on a "raw data"
10kbps data stream, one just puts 5kbps or 3kbps,
1/2 or 1/3 rate coders. (2/3,4/5 is also possible)
This "actual data" is spread out on the neighboring
data bits according to certain rules.
The receiver, while not being able to decide on the
"raw" data totally correctly, can still figure out
most of the actual data by figuring out which the
actual data probably was encoded into, that
noisy raw data. (garbage)
That is, even if one raw data bit is totally missing, one
can figure out what it was from the neighbors,
as they "contain" parts of that one missing bit.
The really academic explanation is the trick of introducing "redundancy" (like the nagging wife) in the data,
and then calculating that according to the
coding rules this was the most probable thing she
said although I missed half of it.
(additionally their is reed-solomon coding, same used
for CD records to recover the data lost in a scratch,
for mobile enviroment same as a deep fading, but more
like placing the same data in two different places,
if one is lost, one can find it from the other place)
---- bow and accept the shannon thoughts...
According to shannon the ultimate solution would be
to have a telepathic channel back to the sender, to
- detect data bits which are wrong
- telepathically ask the sender to send it once again
but as that would need a higher bandwidth, bitrate
channel in the opposite direction for above operation, it
would be stupid.
The practical solution is a combination of:
- with fairly good, low error rates, do as "whaddyasay"
..- add a checksum for one frame of data, like 1000 bits
....and then 32 as a checksum (1000 bits -> 10032 bits)
..- if the receiver detects the data is wrong, ask for
....it to be sent once more (effectively halving the
....transmission rate for those 1000 bits)
- above convolutional code trick, spread out the same
data over a longer block of data, to help any bit which
has "disappeared", helped by the good ones on both sides.
(true polish socialism)
With low error rates #1 is great, with high error rates it
is useless (consider if every frame is wrong)
The important points are
- the "waddyasay" system wont work if every frame, packet,
block is just garbage and needs to be resent, and is
garbage once again.
- the "waddyasay" system will introduce delays in the data
For the convolutional, shannon, ungerboeck, viterbi,etc coding
this "buddy help" goes "with the data" (forward) so that
the receiver can "figure out those bad data bits anyway".
That is, it doesn't introduce that extra roundtrip delay
"whaddyasay" does, nor does it need that channel
sending that "whaddyasay".
Instead it halves, etc the actual bitrate compared to the
"raw" bitrate.
-- real life is a combination of both:
- convolutional, channel coding, to get the error rate
good enough to get any data block,frame through correctly
- "whaddyasay" error correction to fix those few errors
still slipping through.
The voice-data thing is precisely a question of these
two:
- voice (coders) are OK with every 100-1000 bit wrong,
one still understand the word, sentence, or says "whaddyasay", just like to the wife.
- but "data" demands every bit is finally correct (after all
error correction) as noone can know how important that one
bit might be, buy/sell bit, bankaccount or whatever.
----
This is why QCOM sorts out voice and data on different
bands, to keep the data band "clean", separated from
the real noisy cocktail party for the voice bands.
(also that US operators have narrow bands, poor guys,
cannot get 5MHz)
WCDMA can fall back to 2G for the same, as well as more
"sophisticated" power control, error correction,etc,etc..
-----
Anyway, this is more "sexy" than trying to
predict the stock market, the limited absolute
mathematical impossible equations, far from
realistic implementations, finding the "sweet spot"
inbetween.
Ilmarinen
(enjoying CSPAn and missile defense, another similar
problem) |
