Exponential Family Facts Stat 543 Spring 2016 Many standard families of distributions can be treated with a single set of analyses by recognizing them to be of a common “exponential family” form. The following is a list of facts about such families. (See, for example, Schervish Sections 2.2.1 and 2.2.2, or scattered results in Shao or the old books by Lehmann (TSH and TPE ) for details.) De…nition 1 Suppose that P = fP g is de…ned by pdf ’s or pmf ’s f (x) for X. P is called an exponential family if for some h(x) 0, ! k X : dP f (x) = (x) = exp a( ) + . i ( )Ti (x) h(x) 8 d i=1 R P Claim 2 Let = ( 1 ; 2 ; :::; k ); E = f 2 Rk j h(x) exp ( i Ti (x)) d (x) < 1g, and consider also the family of distributions with pdf ’s or pmf ’s of the form ! k X f (x) = C( ) exp , 2E . i Ti (x) h(x) i=1 Call this family P . This set of distributions for X is at least as large as P (and this second parameterization is mathematically nicer than the …rst). E is called the natural parameter space for P . It is a convex subset of Rk . (TSH, page 57.) If E lies in a subspace of dimension less than k, then f (x) (and therefore f (x)) can be written in a form involving fewer than k statistics Ti . We will henceforth assume E to be fully k-dimensional. Note that depending upon the nature of the functions i ( ) and the parameter space , P may be a proper subset of P . That is, de…ning E = f( 1 ( ); 2 ( ); :::; k ( )) 2 Rk j 2 g; E can be a proper subset of E. Claim 3 The "support" of P , de…ned as fxjf (x) > 0g, is clearly fxjh(x) > 0g, which is independent of . Claim 4 From the Factorization Theorem, the statistic T = (T1 ; T2 ; :::; Tk ) is su¢ cient for P. Claim 5 T has distributions derived from the distributions for X, say fP T j 2 exponential family. Claim 6 If E g, which also form an contains an open rectangle in Rk , then T is complete for P. (See pages 142-143 of TSH.) Claim 7 If E contains an open rectangle in Rk (and actually under the much weaker assumptions given on page 44 of TPE) T is minimal su¢ cient for P. Claim 8 If g is any real-valued function such that E jg(X)j < 1, then Z E g(X) = g(x)f (x)d (x) is continuous on E and has continuous partial derivatives of all orders on the interior of E. These can be calculated as Z @ 1+ 2+ + k @ 1+ 2+ + k E g(X) = g(x) f (x)d (x) . @ 11@ 22 @ kk @ 11@ 22 @ kk (See page 59 of TSH.) 1 Claim 9 If for u = (u1 ; u2 ; :::; uk ), both 0 In particular, E 0 Tj (X) = @2 j@ l 0 + u belong to E o + uk Tk (X) = C( 0 ) : C( 0 + u) is in the interior of E, then E 0 (T1 1 (X)T2 2 (X) @ and n exp u1 T1 (X) + u2 T2 (X) + 0 E Further, if 0 ( ln C( )) = @ @ j Tk k (X)) = C( ( ln C( )) = 0) 1+ @ @ 1 1 @ , Var 0 Tj (X) = 0 2+ 2 2 @2 @ 2j + @ k k k 1 C( ) ( ln C( )) = : = 0 , and Cov 0 (Tj (X); Tl (X)) = 0 . 0 Claim 10 If X = (X1 ; X2 ; :::; Xn ) has iid components, Pn with each Xi P , then X generates a k-dimensional exponential family. The k-dimensional statistic i=1 T (Xi ) is su¢ cient for this family. Under the condition that E contains an open rectangle, this statistic is also complete and minimal su¢ cient. 2