Scribd
Upload
77 views1 page

Stat 992: Lecture 01 Gaussian Random Fields.: Moo K. Chung December 11, 2003

1. The document discusses modeling temperature measurements over space and time as a Gaussian random field, where the true temperature is modeled as a signal plus measurement error. 2. It defines Gaussian random fields as random fields where any finite collection of measurements follows a multivariate normal distribution. The covariance function characterizes a Gaussian field. 3. It notes that two Gaussian fields are independent if their cross-covariance is zero for all points, and asks whether there is always a mapping to make two arbitrary Gaussian fields independent.

Uploaded by

Pablo Benitez
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
77 views1 page

Stat 992: Lecture 01 Gaussian Random Fields.: Moo K. Chung December 11, 2003

1. The document discusses modeling temperature measurements over space and time as a Gaussian random field, where the true temperature is modeled as a signal plus measurement error. 2. It defines Gaussian random fields as random fields where any finite collection of measurements follows a multivariate normal distribution. The covariance function characterizes a Gaussian field. 3. It notes that two Gaussian fields are independent if their cross-covariance is zero for all points, and asks whether there is always a mapping to make two arbitrary Gaussian fields independent.

Uploaded by

Pablo Benitez
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 1

Stat 992: Lecture 01

Gaussian Random Fields.


Moo K. Chung [email protected]

December 11, 2003

1. Spatiotemporal model. Suppose we can measure tem- 10


0.6
10
0.3

perature Y at position x and time t in a classroom 0.4 0.2

M ∈ R3 . Since every measurement will be error-


20 20 0.1
0.2

0
0
30 30

prone, we model the temperature as −0.2


−0.1

−0.2
40 40
−0.4
−0.3

Y (x, t) = µ(x, t) + ²(x, t) 50 −0.6 50


−0.4

−0.8
−0.5
60 60

where µ is the real unknown signal and ² is measure-


10 20 30 40 50 60 10 20 30 40 50 60

ment error. The measurement error can be modelled as Figure 1: The first image is the realization of white noise
a random variable. So at each point (x, t) ∈ M ⊗ R+ , which would be unrealistic for modelling room temperature
measurement error ²(x, t) is a random variable. The due to discontinuity. The second image is the realization of
collection of random variables a Gaussian field which show realistic continuous measure-
{²(x, t) : (x, t) ∈ M ⊗ R+ } ments.

is called a stochastic process. For any stochastic pro- mean zero Gaussian fields, e1 and e2 are independent
cess that contains a spatial variable, it is called a ran- if and only if the cross-covariance function
dom field. A formal measure theoretic definition can be
found in Geometry of Random Fields by Adler (1980) R(x, y) = Ee1 (x)e2 (y) = 0
and Introduction to the Theory of Random Processes
by Gikhman and Skorokhod (1969). Functional mag- for all x and y.
netic resonance images (fMRI) is one example spa- Problem 1. For given two arbitrary mean zero Gaus-
tiotemporal modeling is necessary. sian fields, is there mapping that makes them indepen-
dent?
2. Gaussian random fields. A random vector P X =
(X1 , · · · , Xm ) is multivariate normal if i ci Xi is 4. Infinite-dimensional vector space. Let L be a collec-
Gaussian for every possible choice of ci . tion of random fields. Suppose e1 , e2 ∈ L. L forms a
A random fields ²(x) ∈ RN is a Gaussian random field vector space if c1 e1 + c2 e2 ∈ L for any c1 and c2 . Ob-
if ²(x1 ), · · · , ²(xm ) is multivariate normal for any xi ∈ viously the collection of Gaussian fields form a vector
M ⊂ RN . ² is a mean zero Gaussian field if E²(x) = 0 space. Since there is no way to represent every element
for all x. The covariance function of mean zero field is in L with finite basis fields, it is an infinite-dimensional
defined as vector space. However, there exists a finite vector
R(x, y) = E²(x)²(y). space Lp that is the closest to L in the least-squares
sense (Hilbert space theory). Real and Complex Anal-
The variance of field ² at fixed point x is R(x, x).
ysis by Rudin (1986) gives a very nice undergraduate
Mean zero Gaussian field is completely characterized
level introduction of Hilbert space. Functional Analy-
by the covariance function. A Gaussian random vector
sis by Rudin (1991) and A Course in Functional Anal-
field is defined similarly. e(x) = (e1 (x), · · · , en (x))0
ysis by Conway (1997) gives concise graduate level
is a Gaussian vector field if ei are Gaussian fields. We
treatment of the subject matter.
may generalize it further to matrix fields and tensor
fields. Diffusion tensor images (DTI) can be modelled For any linear operator f , f (L) ⊂ L. It can be shown
as a tensor field. Unfortunately, it will not be Gaus- that differentiation of Gaussian fields is again Gaussian
sian tensor fields although it is possible to make images in the mean-square sense.
more Gaussian via kernel smoothing.

3. Independence. Two fields e1 and e2 are independent if


e1 (x) and e2 (y) are independent for every x and y. For

You might also like