home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: sci.math.stat
- Path: sparky!uunet!seas.gwu.edu!freytag
- From: freytag@seas.gwu.edu (Richard Freytag)
- Subject: Variance estimation Q.?
- Message-ID: <1992Oct15.203042.19552@seas.gwu.edu>
- Followup-To: freytag@seas.gwu.edu
- Summary: Parameter estimation in non-I.I.D. Gaussian white noise
- Sender: Mark Levin
- Organization: George Washington University
- Date: Thu, 15 Oct 1992 20:30:42 GMT
- Lines: 30
-
- I am posting this for a collegue. Please email me with your responses
- and I will pass them along.
-
- Question follows:
-
- I have a parameter estimation problem which sounds simple but
-
-
- The following question is being posted for a collegue. Email me
- your responses and I will pass them along...
-
- I have a parameter estimation problem which sounds simple but
- seems to have been missed in the literature. We have a Gaussian
- random variable X with mean and variance given by the unknown
- parameters mu_x, and sigma_x. We get observations Yi = X + Ni.
- The Ni are independent, non-identical Gaussian random variables
- with varying but known variances. Our observation Yi consists of
- the observation Yi and the variance of the particular noise at
- time i, what I call sigma_i. I have derived the MLE equations
- for mu_x, and sigma_x and find that they are non-linear and not
- solvable in closed form. While they can be solved for numerically
- we are working in real time and cannot afford a Newton-Raphson type
- numerical solution. I have also worked the problem from a linear
- least squares approach, and while this works OK for estimating mu_x,
- they do not work for estimating sigma_x as sigma_x is an inherently
- non-linear function of the observations. Can anyone direct me to
- any literature regarding this problem.
-
- Thanks in advance for your help,
- Mark Levin
-