Variogram modeling vs.
Covariance modeling
Parameters and equivalences
Central Idea
Compare
the
methodology
for
Variogram modeling with that of
Covariance modeling, and formulate
(if possible) an equivalence among
their respective parameters
Contents
I. Variogram modeling
a.
b.
c.
d.
Experimental variogram
Parameters of a theoretical variogram
Estimation of the theoretical variogram
Parameter fitting
II. Covariance modeling
a. Parameters of a covariance model
b. Commonly used models
c. Parameter fitting
III. Variogram Covariance relationship
a. Assumptions
b. Formulation
c. Parameters correspondence
Previous considerations
In
this context, a random variable
is
that
defined
at
locations
where , and is the dimension of the input
space
All exemplifications are made supposing the
random variable
is isotropic, that is,
correlation between samples depends only on
the separation distance, no matter their
relative position (direction of the distance
vector)
Variogram modeling
a. Experimental variogram
variogram (also called semi Experimental
variogram) is calculated for each pair of
samples as
Variogram modeling
b. Parameters of a theoretical variogram
model sill
total sill
model
structure
nugget
range
Variogram modeling
c. Common theoretical variogram models
Exponential:
Gaussian:
Nugget:
Where:
(white noise)
is the model variance
is the distance
is the length parameter
Some implementations (e.g., DACE) use as length parameter
Variogram modeling
c. Estimation of the theoretical variogram
1.
2.
3.
Construct the experimental
variogram for the sample (dots)
Estimate the parameters (i.e.,
range, sill, model, nugget) of a
theoretical variogram (red line)
[Optional]
Optimize
the
parameters for a best fit (blue line)
to the experimental variogram
Variogram modeling
d. Parameter fitting
In
almost every implementation, the
fitting is made through weighted least
squares fit to the experimental variogram
There are several criteria to assign the
weight () to each point of the variogram,
the most common are:
Number of averaged points
/squared semi-variogram ratio
/squared distance ratio
.
..
...
Covariance modeling
Since
we are considering the
isotropic case, the distance between
any two samples is a scalar
Covariance modeling
a. Parameters of a covariance model
nugget
total sill model sill
model
structure
range
Covariance modeling
b. Commonly used models
Exponential:
Gaussian:
Where:
is the model variance
is the distance
is the length parameter
Some implementations (e.g., DACE) use as length parameter
Covariance modeling
c. Parameter fitting
This
task is solved by Bayesian inference.
By maximizing the probability of the
posterior for the covariance parameters (),
which is given by:
Where: is the marginal likelihood
is the prior for the parameters
is a normalizing factor
Covariance modeling
c. Parameter fitting
To maximize the probability of the
posterior, it is necessary to maximize
(the probability of observing the
values given the locations and the
parameters ) w.r.t. the parameters .
Covariance modeling
c. Parameter fitting
For
simplicity, it is common
maximize its logarithm, given by:
to
Covariance modeling
c. Parameter fitting
It can be shown1 that maximizing the
marginal
likelihood
w.r.t.
the
parameters is equivalent to maximize
the expression:
Where
is the correlation matrix
is the estimated variance
is the estimated mean
See Jones D.R. (2001) A taxonomy of global optimization methods based on response
surface, J Glob. Opt. 21:345-383
1
Variogram-Covariance
relationship
Where
is the variance of
