|Stochastic Systems Group|
Doctoral Student, DSPG, MIT
We consider ``bit stealing'' scenarios which reduce the rate of a data source via requantization. For example, consider sending voice, video, etc., through a multi-hop network. If two packets of rates R_1 and R_2 bits are both destined for a link with capacity C < R_1 + R_2 they can not be simultaneously transmitted. The traditional solution is to drop one packet, but an alternative is to reduce the rate of the sources via further data compression.
If the data sources are encoded in a layered manner, requantization can be accomplished by simply discarding the least significant bits. However, layering is a design constraint. We consider what is possible when this constraint is not imposed and the original encoders are not designed to anticipate requantization.
We study the following fundamental source coding problem: Given that a source is quantized to distortion D by a near-optimal encoder, what bit rate is required to requantize the source to distortion D' > D? We characterize this bit-stealing rate-distortion trade-off, and show that, for the Gaussian-Quadratic case, it is within 1/2 bit of the trade-off for layered systems. We also discuss how information embedding ideas can be used for bit stealing even when the source decoder is unaware of the requantization.
This is joint work with Aaron Cohen, Stark Draper, and Greg Wornell presented in part at DCC 2002 and ISIT 2002.
Problems with this site should be emailed to email@example.com