motion sensor signal compression/prediction

Hi folks.

I'm currently chewing on the following problem:

I have a bunch of huge motion sensor signals to lossy compress. Since I
need random access to the samples block based compression schemes are
out of question.


The current approach works like this:

I store every 8th sample, quantized from 32 bit down to 8 bits and use
this as a rough approximation of the signal. Along with this prediction
data I store the errors, agan quantized and with a variable bit-width
per 8 sample block.

I know that this is a very crude approach but unfortunatley the system
where I have to decompress the data on is not as powerful as I'd like it
to be. I have to decompress several hundrets of streams at nearly an
instant. I have fast RAM and a slow CPU. Fortunately compression time is
not *that* much of an issue.

In the current approach I see one thing where I can improve a lot: My
prediction. Taking every n'th sample is far from optimal. I'm sure I can
do much better here.

I thought about doing a linear least square fit to get my error down.
That will help, but what I really want is a prediction that minimizes
the the number of significant bits of the error per block.

Does some function-fitting like this exist? Links to (free available)
papers or keywords for a good google session are welcome..

Thanks, guys..

  Nils