🚀 go-pugleaf

Hello,
I've a problem on autoregressive process. 

A stationary process may be modeled as an autoregressive process (e.g
speech). The reconstruction process involves driving the inverse filter
formed by the polynomial 1 - sum[a_i z^(-i)], summation over the order
of the polynomial. Now usually quantization of the filter coeffts is
not done bec' of stability of the IIR filter. So one quantizes the
reflection coeffts. 
My question is 
1) How can one represent the mean square error of the quantization of the
   reflection coeffts in terms of the quantization of the filter coeffts ?
2) If (1) is possible, how does it translate into a mean square error of
   the reconstructed process ?

My problem is actually a 2-D version where the polynomial in 2-D is not
even separable. But atleast if I can get a handle of the 1-D case, maybe
I can make some assumptions.

My final goal is to get a rate-distortion curve for the quantization of
the reflection coeffts vs the mean square error of the reconstructed process.

Are there any references on this ? I feel this problem has been tackled in
speech coding based on LPC technique. But don't have any references on that
either.

Any help is greatly appreciated since I've to complete my thesis by this
year!!!

Thanks to all in advance.

I can post a summary of responses, if there are enough replies.
Pl. reply to sriram@ipl.rpi.edu or to the net. Either way is fine.

Sriram