We develop a realization theory for a class of multiscale stochastic proces
ses having white-noise driven, scale-recursive dynamics that are indexed by
the nodes of a tree. Given the correlation structure of a 1-D or 2-D rando
m process, our methods provide a systematic way to realize the given correl
ation as the finest scale of a multiscale process. Motivated by Akaike's us
e of canonical correlation analysis to develop both exact and reduced-order
model for time-series, we too harness this tool to develop multiscale mode
ls. We apply our realization scheme to build reduced-order multiscale model
s for two applications, namely linear least-squares estimation and generati
on of random-field sample paths. For the numerical examples considered, lea
st-squares estimates are obtained having nearly optimal mean-square errors,
even with multiscale models of low order. Although both field estimates an
d field sample paths exhibit a visually distracting blockiness, this blocki
ness is not an important issue in many applications. For such applications,
our approach to multiscale stochastic realization holds promise as a valua
ble, general tool.