In this paper, we introduce a novel technique for adaptive scalar quan
tization, Adaptivity is useful in applications, including image compre
ssion, where the statistics of the source are either not known a prior
i or will change over time, Our algorithm uses previously quantized sa
mples to estimate the distribution of the source, and does not require
that side information be sent in order to adapt to changing source st
atistics, Our quantization scheme is thus backward adaptive, We propos
e that an adaptive quantizer can be separated into two building blocks
, namely, model estimation and quantizer design, The model estimation
produces an estimate of the changing source probability density functi
on, which is then used to redesign the quantizer using standard techni
ques, We introduce nonparametric estimation techniques that only assum
e smoothness of the input distribution, We discuss the various sources
of error in our estimation and argue that, for a wide class of source
s with :! smooth probability density function (pdf), we provide a good
approximation to a ''universal'' quantizer, with the approximation be
coming better as the rate increases, We study the performance of our s
cheme and show how the loss due to adaptivity is minimal in typical sc
enarios, In particular, we provide examples and show how our technique
can achieve signal-to-noise ratios (SNR's) within 0.05 dB of the opti
mal Lloyd-Max quantizer (LMQ) for a memoryless source, while achieving
over 1.5 dB gain over a fixed quantizer for a bimodal source.