Re: Why kullback-leibler distance?
- From: "David Jones" <dajxxx@xxxxxxxxx>
- Date: Mon, 10 Apr 2006 10:08:40 +0100
kalyan wrote:
Hi,distribution
I am working with calculating the kullback-leibler distance between
two probability distribution. When this distance measure is not
symmetric, I was wondering why is the distance measure almost a
standard to calculate the distances of PDF ( probability
functions).
One of the reasons for importance is its link to the likelihood
function and to maximum likelihood theory, although little of this is
used when treating it as a distance measure I think. In its
max-likelihood background, the non-symmetry is natural.
Or are there any distance measures which overcome this problem of
symmetry. I am calculating distances between multivariate gaussian
distributions.
There are other non-symmetric distances which overcome the problem of
what to do if you having non-matching zero probabilities. Onec of
these is the "continuous ranked probabilty score", although you may
need to move this from a sample-version to a population-version of the
measure.
There are symmetric distance measures which are usually framed as
distances between CDFs rather than PDFs, but should still be good for
what you are doing. For univariate cases, these are the basis of the
Kolmogorov-Smirnov, Cramer-von Mises and Anderson Darling tests ...
you may need to search to find how to convert these ideas to a
multivariate situation. The advantage of CDF-based measures is that
they tend not to be overly inflenced by locations where one or other
density is zero.
David Jones
.
- Follow-Ups:
- Re: Why kullback-leibler distance?
- From: kalyan
- Re: Why kullback-leibler distance?
- References:
- Why kullback-leibler distance?
- From: kalyan
- Why kullback-leibler distance?
- Prev by Date: Reading a input in form of matrix
- Next by Date: Re: Why kullback-leibler distance?
- Previous by thread: Why kullback-leibler distance?
- Next by thread: Re: Why kullback-leibler distance?
- Index(es):
Relevant Pages
|