i have problem @ hand feel there should rather elegant solution it, @ point have problems finding right search terms or getting first step towards right direction.
basics: have high dimensional data space d = 19 , have 100 points in space (100 measurements). pca , dimensionality estimation algorithms, confirmed latent space on points lie on relatively low dimensional (max 5 dimensions or so). therefore, think in general not impossible asking.
the problem: now, based on uncomplete measurements of new point, estimate missing values. problem not know values missing. combinations of missing values (somewhat) likely. -> have 1 missing value, 19 missing values or in between. in perfect world, algorithm looking not gives estimate of missing values, error measure.
to further illustrate, attach 1 image raw data. x-axis shows 19 individual measured parameters , y axis gives values of parameters. can see measurements highly correlated. if specify 1 measurement/dimension should able give reliable estimation of rest. 
does of have pointers me? thoughts or advice helpful! thanks, thomas
the right way (tm) handle missing data average (i.e., integrate) on missing variables, given values of known variables. bayesian belief network formalization of idea. if can more variables are, can more how go building suitable belief network.
