home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: sci.math.stat
- Path: sparky!uunet!cs.utexas.edu!sdd.hp.com!mips!news.cs.indiana.edu!uceng.uc.edu!juber
- From: juber@uceng.UC.EDU (james uber)
- Subject: linear covariance estimate for max likelihood
- Message-ID: <1992Aug20.142353.6297@uceng.UC.EDU>
- Keywords: parameter estimation, maximum likelihood, covariance estimate
- Organization: College of Engineering, University of Cincinnati
- Date: Thu, 20 Aug 1992 14:23:53 GMT
- Lines: 65
-
- I obtain parameter estimates via maximum likelihood where
- my model is in the standard reduced form y = f(p), y are the
- data and p are the parameters. I assume that the distribution
- of the model + measurement errors is normal with zero mean
- and known covariance matrix Ve. Thus i am solving the optimization
- problem:
-
- min Tr(y - f(p))Inv(Ve)(y - f(p))
- p
-
- where Tr() is the transpose and Inv() is the inverse. I have a question
- about the mathematics of deriving the standard linear estimate of the
- covariance matrix of the sampling distribution of the parameter estimates, p*.
-
- To form my question, let me go briefly into my interpretation of the
- covariance matrix derivation. First of all, the optimal value function of
- the max. like. problem depends on data, that is p* = p*(y). The definition of
- the covariance matrix of the sampling distribution, Vp, is:
-
- Vp = E[(dp*)Tr(dp*)]
-
- where the perturbation of the optimal parameters, dp* is estimated by the
- linear terms of a Taylor series:
-
- dp* ~ Tr(dp*/dy)dy
-
- where dp*/dy are the partial derivatives of the function p*(y) and are
- evaluated at (y,p*). Substituting this into the definition of Vp gives:
-
- Vp ~ Tr(dp*/dy)E[(dy)Tr(dy)](dp*/dy)
- Vp ~ Tr(dp*/dy)Vy(dp*/dy)
-
-
- Now the rest of the derivation continues with the evaluation of the partial
- derivatives (dp*/dy), often requiring additional approximations based on
- small residuals and other assumptions. The true dp*/dy would certainly
- require second partials of the log likelihood function, whereas certain
- approximations (the most common, i believe) use only first partials. In any
- case, the partials dp*/dy will certainly involve the combined error covariance
- matrix Ve.
-
- My question can now (finally) be stated: Is the covariance matrix Vy above
- one and the same with Ve, the covariance of the model + measurement errors?
- I have been really confused by this for a long time. Just intuitively,
- but perhaps naively, i'd have thought that Vy only includes the *measurement*
- errors, and not the model errors. All of the derivations i've seen, however,
- seem to at some stage treat both Vy and Ve as the same. For example, one
- standard final results is the following:
-
- Vp ~ Inv[Tr(df/dp)Inv(V)(df/dp)]
-
- But, i must confess, i really am unsure how to interpret the covariance
- V in the above equation.
-
- Thank you for your time and assistance in helping me with this problem.
-
- jim uber
- dept. of civil & env. engineering
- univ. of cincinnati
- juber@uceng.uc.edu
-
- --
- --
- james uber
- juber@uceng.uc.edu
-