Re: modifying mahalanobis distance



shaobo hou wrote:
Well, what I am trying to achieve is that in addition to computing the Mahalanobis distance of a data point to a Gaussian, the data point also has some uncertainty (also Gaussian) associated with it, as represented by the matrix C. The greater the uncertainty about the data point, the greater the distance should be.

Another idea I have considered is to compute the integral of the product of the Gaussian and the uncertainty about the data point, but this involves matrix inversion, which will make it too slow for my uses.

Does this describe your situation? You have a population of objects
whose locations in some p-dimensional space have a multivariate normal
distribution with mean vector m and covariance matrix K. The measured
location of any object contains error. That is, every measurement y
can be written as x + e, where x is the true location of an object,
and e is the measurement error. For each measurement, the sampling
distribution of the error is p-variate normal with mean vector 0
and covariance matrix C.

If that's what you have then the sampling distribution of y for a
randomly chosen object is p-variate normal with mean vector m and
covariance matrix K+C, and your post suggests that you want to talk
about (y-m)'K^{-1}(y-m) = (x-m)'K^{-1}(x-m) + 2(x-m)'K^{-1}e +
e'K^{-1}e. The first term on the right is the Mahalanobis distance
of x from m. The expected value of the second term is 0. The third
term is always positive, with expected value trace(CK^{-1}).

.



Relevant Pages

  • Re: Bayesian estimation of structured correlation/covariance
    ... correlation between a pair of paths. ... >> should formulate this as covariance matrix estimation with the ... > large step of using the Wishart/ inverse Wishart distribution for the ... > switch from viewing your problem as one of estimating the covarince ...
    (sci.stat.math)
  • Re: Confidence limits for noise measurement
    ... required to reduce the effect of chance on the measurement result (in ... presense of noise can be expected. ... the noise statistics are stable then you will obtain the same answer ... work he derived the distribution of "t" which allowed confidence ...
    (rec.radio.amateur.antenna)
  • Re: Coffee grinder (burr type) recommendation?
    ... | influencing cup quality, the measurement should be taken many times to ... regular customer at a cafe of the latter type; while the espresso was always ... | size distribution to be unrelated? ...
    (alt.coffee)
  • Re: Sun noise
    ... I have done two series of 250 measurements of audio noise voltage ... estimate of the uncertainty that I referred to in an earlier post. ... distribution, so the variance is 1/12 of the span. ... and that might also be true of measurement of noise voltage ...
    (rec.radio.amateur.antenna)
  • Re: Help me : Transformation of multivariate distribution
    ... K have a spherical distribution not Multivariate Normal. ... while yours are both densities. ... covariance matrix, ... In your case the matrix characterising the ...
    (sci.stat.math)