Bayesian Analysis High-dimensional Problems推荐.pdfVIP

9 High-dimensional Problem s Rather than begin by defining what is meant by high-dimensional, we begin with a couple of examples . Example 9.1. (Steins example) Let N{fipxi^ ^pxp = o-^Ipxp) be a p-variate normal population and Xi = (Xn^..., X^p), i = 1 , . . . , n be n i.i.d. p-variate samples. Because U = cr^/, we may alternatively think of th e dat a as p in- dependent samples of size n from p univariate normal population s N{/j.j, cr^), J = 1,.. . ,p . Th e parameter s of interest are th e /ijs . For convenience, we ini- tially assume cr^ is known . Usually, th e number of parameters , p , is large and the sample size n is small compared with p. These have been called problems with large p , small n . Note that n in Stein s example is th e sample size, if we think of th e dat a as a p-variate sample of size n . However, we could also think of th e dat a as univariate samples of size n from each of p univariate populations . Then th e tota l sample size would be np. Th e second interpreta - tion leads t o a class of similar examples . Note that th e observations are not exchangeable except in subgroups , in thi s sense one may call them partially exchangeable. Example 9.2. Let f{x\fij)^j = 1,.. . ,p , denote th e densities for p populations , and Xij^i = l,...,n,j i = l , . . . , p denote p samples of size n from these p populations . As in Example 9.1 , f{x\fij) may contain additional common parameters . Th e object is t o make inference about th e /ijs .