Estimation of Random Field Parameters

in Stochastic Analysis and Inverse Modeling

Presented by
Gordon A. Fenton

24
 Introduction

• Our probabilistic models require distributions for the soil

properties.
• Each distribution is characterized by parameters such as the
mean and standard deviation.
• spatial correlation is characterized by a correlation length.
• the mean is generally easy to estimate, mean trends require
somewhat more data.
• the standard deviation requires even more data to estimate –
we often depend on the literature for estimates.
• the correlation length requires huge amounts of data to
estimate – even difficult to find useable estimates in the
literature (design may have to proceed using a worst case
correlation length).
25
 Interpolation vs. Extrapolation

How the statistical estimates are to be used influences how they

are determined;
1. Interpolation: the goal here is to use the data to best
describe the site at which the data was obtained. In this
case all trends should be accounted for. May want to
perform simulations conditioned (i.e. “pinned”) on the
data. Correlation length is now of “residual” variability.
2. Extrapolation: the goal here is to use the data to attempt
to characterize other (similar) sites. In this case, trends
should only be accounted for if they are expected to recur
at the other sites. Mean estimates will be uncertain,
variances will be underestimated. Correlation length will
be larger than estimated (i.e. generally unknown).
26
 Interpolation vs Extrapolation

Interpolation:
• data is used to characterize site at which data was obtained
• estimator errors decrease with increasing correlation
between observations, i.e. the more highly correlated the
site is, the fewer samples required to characterize it.
(Unfortunately, we generally don’t know a-priori how
highly correlated a site is!)
• the data can be assumed known – uncertainty now occurs
between data points, so we need only model this residual
variability.
• Kriging and/or conditional models can be used to
characterize the residual variability.
27
 Interpolation vs. Extrapolation

Extrapolation:
• the data are being used to characterize the soil population
(i.e. to infer the population parameters for use at other
sites).
• estimator error increases with increasing correlation
between observations, i.e. the more highly correlated a site
is, the less representative it is of other sites – you cannot
expect to accurately characterize a neighboring site if all of
your samples are taken from a (highly correlated) stiff soft
clay layer at the current site.
• statistical estimates of population parameters are typically
quite inaccurate (due to correlation), especially estimates of
correlation length.
28
 Interpolation vs. Extrapolation
• Practicing geotechnical engineers are typically interpolating.
That is, they sample with the goal to characterize the site at
which the samples are observed.
• Published research papers and textbooks are extrapolating (or
at least they should be). That is, they are expressing soil
property information that is meant to be useful at sites other
than the single site at which the data were obtained.
• Unfortunately, all too often, research papers will provide soil
property statistics where the locally observed trend has been
removed. This leads to significantly underestimated
variabilities (only useful at sites where similar trends occur
and have been similarly removed).
• In extrapolation trends should be generally considered to be
part of the uncertainty being characterized.
29
 Choosing a Distribution
Once the data have been gathered, we need to decide how to best
represent the “population”. The first step is to decide on a
population distribution. There are several possibilities;
1. Trace driven simulation: use the data directly in a
simulation. This is the least preferable approach since it
can only reproduce the data and not all possibilities. This
approach is most commonly used in earthquake ground
motion simulation.
2. Empirical distribution: the data are used to define an
empirical cumulative distribution function (e.g. P[ X < x ]
is just equal to the number of observed values less than
x). This does not allow for the extremes that often control
design. That is, most samples will not include those
1/1000 extremes that would lead to failure.
30
 Choosing a Distribution

3. Fit a Distribution: A common distribution, such as the

normal or lognormal, is fitted to the data. The advantages
to this approach are that
a) irregularities in the data, due to the natural variability in any
data set, are smoothed out. That is, we don’t end up with a
distribution that is skewed by an outlier in the data set.
b) the known physics of the property can be properly represented
(e.g. properties such as porosity, friction angle, and Poisson’s
ratio have known (or at least almost known) upper and lower
bounds, so a bounded distribution would be appropriate).
c) extremes can also be modeled in a physically reasonable way.

31
 Choosing a Distribution

Extrapolation:
• fit the simplest distribution that you can – you are trying to
model the population, not the specific data set.

Interpolation:
• fit a distribution of reasonable complexity – just remember
that you still need to capture the range of possibilities that
might occur between your observation points (so there is
probably little point in employing a 20 parameter
distribution).

32
 Choosing a Distribution

The normal distribution is a very popular choice, especially if the

soil property is a random field (since we then only need to know
the mean and covariance structure).
The one major disadvantage to the normal distribution is that its
range is from −∞ to +∞, so for many non-negative soil properties
it is not physically possible.
However, if the probability of obtaining a negative value is
sufficiently small, the normal distribution might be a reasonable
approximation. What is meant by “sufficiently small” depends on
the acceptable probabilities of the extremes that might lead to
failure.

33
 Choosing a Distribution

Probability that X < 0 for coefficients of variation v = 0.3 and 1.0

34
 Goodness-of-Fit

Once a distribution has been selected and then fit, by estimating

its parameters using the collected data, the fit must be assessed.
There are two common approaches to measuring how well the
assumed distribution fits the data;
1. Frequency comparisons and probability plots, and
2. Goodness-of-Fit tests.

35
 Frequency Comparison

Example:
Suppose that, just after construction, a series of 50 randomly selected one-
kilometre long sections of highway through a hilly region were selected to
evaluate the annual probability of slope failure under the existing design code.
The number of years until an observable slope failure occurred within each
one-kilometre length, ti , was recorded, with the following results
3, 2, 8, 9, 10, 4, 4, 2, 7, 7, 1, 14, 2, 1, 8, 3, 4, 5, 4, 2, 10,
2, 1, 7, 8, 4, 3, 3, 21, 1, 3, 9, 1, 4, 5, 1, 4, 1, 4, 3, 5, 3, 1,
9, 1, 6, 3, 5, 12, 11
A previous analysis of similar data suggested that the annual probability of
observable slope failure in each one-kilometre section of highway is 0.2.
Assuming that sections fail independently and that each year constitutes an
independent trial, how reasonable does the hypothesis that the annual
probability of slope failure per km is 0.2 appear to be on the basis of this data?

36
 Frequency Comparison

If sections fail independently, and each year is also independent, then we have
50 independent observations of the ‘number of trials’ (i.e. years) to first failure
of a 1-km section. Under the given assumptions, the ‘number of trials to first
failure’ follows a geometric distribution.
The estimate of the annual probability of slope failure per km is just one (year)
over the average time to slope failure;

1
=pˆ = 0.199
( 3 + 2 +  + 11) / 50
which is very close to the hypothesized annual probability.

The following page compares the frequency histogram with that predicted by
theory along with the empirical and fitted cumulative distribution functions.

37
Frequency Histogram

Frequency-density plot of annual

failure probability and fitted
geometric distribution. The lower
plot compares the empirical cumulative
distribution with the fitted cumulative
geometric distribution.

38
 Parameter Estimation

There are many ways to estimate distribution parameters – some

are better than others. The common criteria used to compare
estimators are
1. unbiasedness: E[estimator] = parameter?
2. consistency: lim estimator = parameter?
n→∞
3. efficiency: Var[estimator] small?
4. sufficiency: utilizes all pertinent information?

39
 Classical Estimators
1 n
Sample Mean: µˆ X= x= ∑
n i =1
xi estimates the true mean μX

1 n
Sample Variance: σˆ = s= 2
∑ i
2

n − 1 i =1
X ( x − x ) 2

is an estimate of the true variance σ X2

n− j
1
=
Sample Correlation: ρˆ X ( j ∆x) ∑
σˆ X (n − j − 1) i=1
2
( xi − x )( xi + j − x )

is an estimate of the true correlation ρX

40
 Estimation in the Presence of Correlation

Friction angle measured at regular locations along a 10 km line.

• if we measured only from 0 – 0.75 km, our estimate of the global average
would be very poor. In fact, most 1 km segments would give poor results.
• best to sample at widely spaced points (>> θ).
41
• the variance estimated over 0 – 0.75 km is far less than the site variance.
 Estimation Without Correlation

Friction angles measured along a 10 km line where soil properties are largely
spatially independent.
• in this case, both the estimated mean and variance obtained over
0 – 0.75 km are quite representative of the entire 10 km.
42
 1 n
Classical Estimate of the Mean: µˆ X= x= ∑
n i =1
xi

43
Introduction of Correlation Between Samples

Recall from last slide:

44
 1 n
=
Classical Estimate of the Variance: s 2
∑
n − 1 i =1
( xi − x ) 2

45
Introduction of Correlation Between Samples

Recall from last slide:

46
 Case 1: Data are Gathered over the Design Site

• we will know the soil properties at the data site locations and
will not be attempting to extrapolate beyond the site borders,
• estimates for μX , σX , and θX are “local” and can be be
considered to be reasonably accurate
• best estimates for the value and variability of the random field
between observation points can be obtained using Best Linear
Unbiased Estimation (BLUE) or Kriging.
• probability estimates should be obtained using a random field
conditioned on the data (possibly via conditional simulation)

47
 Case 2: Data are Gathered at a Similar Site

• data gathered at another similar site are used to characterize

the design site (extrapolation) – this might occur during
preliminary design, e.g. before the site has been cleared.
• much greater uncertainty in applying the resulting statistics to
the design site. The sample mean should be viewed with
caution and the sample variance should be assumed to be
underestimated.
• BLUE and Kriging are not options because no data available at
site.
• treatment of trends needs special care – are they likely to recur
at the design site?

48
 Characterization of Trends?

• locally, trends should be accounted for in design

• globally, trends should only be accounted for if
expected to continue (or repeat) offsite

49
 Estimating the Mean

1 n
Classical sample mean: µˆ X = ∑ X i
n i =1
1 n  1 n
E [ µˆ X ] E=
=  n ∑ X i  n=
∑ E[ Xi ] µX (unbiased)
=  i 1=  i 1

1 n n 1 n n  2
Var [ µˆ X ] =2 ∑∑
Cov 
 X i , X 
j  2 ∑∑ ij  X
ρ σ  γ ( T ) σ 2
X
n =i 1 =j 1  =i 1 =j 1 
n

where T is the domain over which the samples are gathered.

For highly correlated samples, γ(T ) ≈ 1.0, and the sample mean
could be quite variable (i.e. large estimator error).
50
 Effect of Correlation on Statistical Estimates

- mean and variance estimates are locally accurate but globally

poor
- this is one of the reasons why geotechnical engineering is difficult
51
to codify
 Estimating the Variance

Classical Maximum Likelihood Estimator:

1 n
= σX ∑ ( Xi − µX )
2 2
ˆ ˆ
n i =1

E σ
=ˆ X2  σ X2 1 − γ (T ) 

Since γ(T ) lies between 0 and 1, E σˆ X2  ≤ σ X2 (unconservative)

For strongly correlated fields, E σˆ X2  → 0 (highly biased)

52
 Estimating the Covariance Structure

Consider a sequence of observations X1, X2, …, Xn, each

separated by distance Δx. Then for τj = jΔx, j = 0, 1, …, n – j – 1
we have n− j
Cˆ (=
τj) ( X i − µˆ X ) ( X i+ j − µˆ X )
1
∑
n − j − 1 i =1
Cˆ (τ j )
= ρˆ (τ j ) = (where σ ˆ 2
Cˆ (0))
σˆ X2 X

53
 Estimating the Covariance Structure
2  n − j +1 
Bias: E C (τ j )   σ X 
 ˆ    ρ (τ j ) − γ ( D ) 
 n 

 n − j + 1   ρ (τ j ) − γ ( D ) 
E  ρˆ (τ j )     
 n   1− γ ( D) 
 
Note that in a strongly correlated field, ρˆ (τ j ) will become
negative, often at about the field midpoint.

54
 Estimating the Covariance Structure

Correlation function estimates from a finite-scale process (θ = 3)

55
 Estimating the Covariance Structure

Correlation function estimates from a fractal process (H = 0.95)

56
 The Sample Semivariogram

The semivariogram gives essentially the same information as the

correlation function since they are related according to
V (τ j ) = E ( X i + j − X i )  =σ X2 1 − ρ (τ j ) 
1  2

2  
Its estimator is
n− j
V (τ j ) = ( X i+ j − X i ) ,
1
∑
2
ˆ j=0,1, , n − 1
2(n − j ) i =1
This estimator does not depend on µˆ X , which is a significant
advantage. In particular, it means it is unbiased,

V (τ j )  ( )
1  2

E=

ˆ E
 2  X i + j − X i 
57
 The Sample Semivariogram

Semivariogram estimates from a finite-scale process (θ = 3)

58
 The Sample Semivariogram

Semivariogram estimates from a fractal process (H = 0.95)

59
 Conclusions

• the mean is relatively easy to estimate at a site, the variance

less easy, and the correlation length is very hard to estimate.
• interpolation is generally more accurate than extrapolation due
to correlation between observations
• use caution when using statistics from the literature – these are
generally unconservative due to correlation
• account for trends when interpolating, but not usually when
extrapolating
• when estimating the correlation structure
• correlation function estimates can be highly biased
• the variogram is approximately unbiased.

60