Probabilistic Analysis of Foundation Settlement

by Gordon A. Fenton 1, G.M. Paice 2, and D. V. Griffiths 3

Abstract
It is at least intuitively evident that variability in soil properties will have a significant
effect on total and differential settlement of structural foundations. By modeling soils
as spatially random media, whose properties follow certain distributions and spatial
correlation structures, estimates of the reliability of foundations against serviceability
limit state failure, in the form of excessive differential settlements, can in principle be
made. The soil’s property of interest is it’s elastic modulus, , which is represented
here using a lognormal marginal distribution and an isotropic correlation structure.
Prediction of settlement below a foundation can then be made using the finite element
method given a realization of the elastic modulus field underlying the foundation. By
generating and analyzing multiple realizations, the statistics and density functions of
total and differential settlements can be estimated.
This paper estimates probabilistic measures of total settlement under a single spread
footing and of differential settlement under a pair of spread footings using a two-
dimensional model combined with a Monte Carlo simulation. For the cases considered,
total settlement is found to be well represented by a lognormal distribution and simple
relationships are proposed allowing the approximation of the settlement distribution
parameters for a footing founded on a spatially random soil of constant depth and
fixed Poisson’s ratio. A one-parameter exponential distribution is fitted to differential
settlements and found to give reasonable probability estimates, particularly towards
the tail of the distribution. A method of predicting the single parameter is given in
terms of statistics of the elastic modulus field and local averages over the field. An
example is presented to illustrate the proposed methodology for a single footing.
INTRODUCTION
The settlement of structures founded on soil is a subject of considerable interest to
practicing engineers since excessive settlements often lead to serviceability prob-
lems. In particular, unless the total settlements themselves are particularly large, it is
1
M. ASCE, Associate Professor, Department of Applied Mathematics, Technical University of
Nova Scotia, Halifax, Nova Scotia, Canada B3J 2X4
2
Adjunct Assistant Professor, Geomechanics Research Center, Colorado School of Mines, Golden,
Colorado 80401-1887, USA.
3
M. ASCE, Professor, Department of Engineering, Colorado School of Mines, Golden, Colorado
80401-1887, USA.

1 Fenton et al.
actually differential settlements which lead to unsightly cracks in facades and struc-
tural elements, possibly even to structural failure (especially in unreinforced masonry
elements). Existing code requirements limiting differential settlements to satisfy ser-
viceability limit states (see building codes ACI 318-89, 1989, or A23.3-M84, 1984)
specify maximum deflections ranging from

180 to
 480, depending on the type


of supported elements, where is the center-to-center span of the structural element.
In practice, differential settlements between footings are generally controlled, not by
considering the differential settlement itself, but by controlling the total settlement
predicted by analysis using an estimate of the soil elasticity. This approach is largely
based on correlations between total settlements and differential settlements observed
experimentally (see for example D’Appolonia et.al.,1968) and leads to a limitation of
4 to 8 cm in total settlement under a footing as stipulated by the Canadian Foundation
Engineering Manual, Part 2 (1978).
Because of the wide variety of soil types and possible loading conditions, experimental
data on differential settlement of footings founded on soil is limited. With the aid
of modern high-speed computers, it is now possible to probabilistically investigate
differential settlements over a range of loading conditions and geometries. This paper
reports the initial findings of such a study and attempts to provide a relatively simple,
albeit approximate, approach to estimating probabilities associated with settlements.
The paper first considers the case of a single footing, as shown in Figure 1(a), and
estimates the probability density function (PDF) governing total settlement of the
footing as a function of footing width for various input statistics of the underlying soil.
All other parameters are held constant. The footing is assumed to be founded on a
soil layer underlain by bedrock. The results are generalized to allow the estimation
of probabilities associated with total settlement under an isolated footing in many
practical cases. It is emphasized, however, that the results are still preliminary, there
being still many aspects of the problem that need investigation. Thus, the results
presented in this paper should be viewed as providing only ball-park estimates in the
absence of further theoretical and/or empirical developments.
The second part of the paper addresses the issue of differential settlements under a
pair of footings, as shown in Figure 1(b), again for the particular case of footings
founded on a soil layer underlain by bedrock. The mean and standard deviation
of differential settlements are estimated as a function of footing width for various
input statistics of the underlying elastic modulus field. Unfortunately, the probability
density function governing differential settlement is as yet unknown and only rough
estimates of probabilities associated with differential settlement can be made (barring
numerical integration of a joint probability density function). In this paper a simple
one-parameter exponential distribution is fitted to the simulation data. Since such a
simple distribution cannot hope to capture the intricacies of the actual distribution,
the fit is aimed at yielding reasonably accurate probability estimates in the tail of the
distribution for the particular geometry shown in Figure 1(b).
The physical problem is represented using a two-dimensional model. If the footings

extend for a large distance in the out-of-plane direction, , then the 2-D elastic modulus

field is interpreted either as an average over or as having an infinite scale of fluctuation

in the direction. For footings of finite dimension, the 2-D model is admittedly just
an approximation. However, the approximation would be reasonable if the elastic

modulus were suitably averaged in the direction. These issues are not addressed
in here and thus the derived 2-D results must be viewed with caution pending a 3-D
sensitivity study.

2 Fenton et al.
 a) Wf

H=1 P=1

D=1

b)
P = 1
Wf Wf

P=1
H=1

L=3

Figure 1. Random field/FEM representation of a) a single footing, and b) two foot-

ings founded on a soil layer.

THE RANDOM FIELD/FEM MODEL

As illustrated in Figure 1, the soil mass is discretized into 60 four-noded quadrilateral
elements in the horizontal direction by 20 elements in the vertical direction. The

overall dimensions of the soil model are held fixed at = 3 in width by 
= 1 in
height. Herein, parameters will be expressed without units, it being understood that a
consistent set of units are to be used throughout. The left and right faces of the finite
element model are constrained against horizontal displacement but are free to slide
vertically while the nodes on the bottom boundary are spatially fixed. The footing(s)
are assumed to be rigid, to not undergo any rotations, and to have a rough interface
with the underlying soil (no-slip boundary).
To investigate the effect of the ratio of footing width to soil layer thickness, 

,
 was held constant at 1.0 while the footing width was varied according to Table 1.
In the two footing case, the distance between footing centers was held constant at 1.0,
while the footing widths (assumed equal) were varied. In the latter case, footings of
width greater than 0.5 were not considered since this situation approaches that of a
strip footing (the footings would be joined when  
= 1 0). In all cases, the footing

loads were held constant at 1.0.
The soil has two properties of interest to the settlement problem: these are the elastic
   
modulus, ( ), and Poisson’s ratio, ( ), where is spatial position. At this time for

3 Fenton et al.
simplicity, only the elastic modulus is considered to be a spatially random property,
it being felt that it is the more important variable as far as settlement is concerned.

Poisson’s ratio is held fixed at 0 25 for all analyses over the entire soil mass. The
extension of the results to spatially random Poisson’s ratio is reserved for future work.

Table 1. Input parameters varied in the study while holding = 1, = 1, = 1, 

 
= 1, and = 0 25 constant.


Parameter Values Considered
  0.1, 0.5, 1.0, 2.0, 4.0
0.01, 0.05, 0.1, 0.3, 0.5, 0.7, 1.0, 2.0, 5.0, 10.0

ln
0.1, 0.2, 0.5, 1.0 (single footing)
0.1, 0.3, 0.5 (two footings)
Figure 1 shows, along with the finite element mesh, a grey-scale representation of a
possible realization of the elastic modulus field. Lighter areas denote smaller values

of ( ) so that the elastic modulus field shown in Figure 1(b) corresponds to a higher
elastic modulus under the left footing than under the right – this leads to the substantial
differential settlement indicated by the deformed mesh. This is just one possible
realization of the field; the next realization could just as easily show the opposite
trend, or perhaps something in between.
 
The elastic modulus field is assumed to follow a lognormal distribution so that ln( )
is a Gaussian (normal) random field with mean ln , and variance ln2 . The choice  
of a lognormal distribution is motivated by the fact that the elastic modulus is strictly
positive, as stipulated by the lognormal distribution, while having a simple relationship
with the normal distribution. Note that the normal distribution admits negative values
of with non-zero probability. The spatial dependence is assumed to follow an
isotropic Gauss-Markov correlation function
 2   
 ln  ( ) = exp   (1)
 ln

in which =     
is the vector between spatial points and , and is the absolute   ! 
length of this vector (the lag distance). In this paper, the word ‘correlation’ refers to the
 
correlation coefficient (normalized covariance). The correlation function decay rate
is governed by the so-called scale of fluctuation, ln , which, loosely speaking, is the
distance
 
distance over which elastic moduli are significantly correlated (when the separation
 !  
is greater than ln , the correlation between ( ) and ( ) is less than "
14%).
The assumption of isotropy is, admittedly, somewhat restrictive. Although an ellip-
soidally anisotropic random field can be converted to an isotropic random field by
suitably stretching the coordinate axes, this transformation cannot be performed in a
settlement study since the stress field needs to be preserved. In principal the method-
ology presented in the following is easily extended to anisotropic fields, however, the
accuracy of the proposed distribution parameter estimates would need to be verified.
 
In the meantime, the isotropic case is selected for simplicity.
In practice, one approach to the estimation of ln involves collecting elastic modulus
data from a series of locations in space, estimating the correlations between the log-
data as a function of separation distance, and then fitting Eq. (1) to the estimated
correlations. See, e.g., Degroot and Baecher (1993), de Marsily (July 1985), Asaoka
and Grivas (May 1982), Ravi (1992), Soulié et.al.(1990),and Chiasson et.al.(1995)for
further information on the characterization of spatial variability of soil properties.

4 Fenton et al.
 
Throughout, the mean elastic modulus, , is held fixed at 1.0. Since settlement varies
linearly with the soil elastic modulus, it is always possible to scale the settlement
statistics to the actual mean elastic modulus. The standard deviation of the elastic
modulus is varied from 0.1 to 4.0 to investigate the effects of elastic modulus variability
on settlement variability. The parameters of the transformed ln( ) Gaussian random
field may be obtained from the relations,
 ln2    2   2 )
= ln(1 + (2 ) #
 ln  = ln(  ) 12  ln2  (2 ) $
 
from which it can be seen that the variance of ln( ), ln2 , varies from 0 01 to 2 83  
(note that the mean of ln( ) is not constant).
 
 
To investigate the effect of the scale of fluctuation, ln , on the settlement statistics,
ln is varied from 0.01 (i.e., very much smaller than the soil model size) to 10.0 (i.e.,
substantially bigger than the soil model size). In the limit as ln 0, the elastic
 &%
modulus field becomes a white noise field, with values at any two distinct points
independent. In terms of the finite elements themselves, values of ln smaller than
 
 
the elements results in a set of elements which are largely independent (increasingly
independent as ln decreases). Because of the averaging effect of the details of the  '%
elastic modulus field under a footing, the settlement in the limiting case ln
expected to approach that obtained in the deterministic case, with =
0 is
everywhere, 
and has vanishing variance. By similar reasoning, the differential settlement in this
case (as in Figure 1b) is expected to go to zero. At the other extreme, as ln ,
  % (
the elastic modulus field becomes the same everywhere (different from realization to
realization, according to the lognormal distribution, but spatially constant within any
one realization). In this case, the settlement statistics are expected to approach those
obtained by using a single lognormally distributed random variable, , to model the
 )
)+*-,/.   
soil, ( ) = . That is, if the settlement, , under a footing founded on a soil layer

settlement when = 
with uniform (but random) elastic modulus is given by =
everywhere, then as ln
  % ( ) , for the
the settlement assumes a
)0*1,/.
lognormal distribution with parameters

 ln 2 = ln() *1,/. ) + ln(  )  ln  = ln( ) *1,3. ) + 21  ln2  (3 ) #
 ln 2 =  ln  (3 ) $
where Eq. (2b) was used in Eq. (3a). Also since, in this case, the settlement under the
two footings of Figure 1(b) becomes equal, the differential settlement becomes zero.
Thus, the differential settlement is expected to approach zero both at very small and at
very large scales of fluctuation.

up to a COV =  
Because the variability of the elastic modulus field to be considered can be quite large,
= 4, and because, perhaps more importantly, it is desired
to estimate the entire probability density function (PDF) of settlement, the approach
taken herein is via Monte Carlo simulations. Traditional stochastic finite element
techniques, involving a first or second order perturbation of the random parameters,
cannot be used since they are inaccurate for COV’s in excess of about 20% and
since they do not provide an estimate of the entire PDF. The Monte Carlo approach
adopted here involves the simulation of a realization of the elastic modulus field and
subsequent finite element analysis of that realization to yield a realization of the footing
settlement(s). Repeating the process over an ensemble of realizations generates a set
of possible settlements which can be plotted in the form of a histogram and from
which distribution parameters can be estimated. In this study, 2000 realizations are

5 Fenton et al.
performed for each input parameter set ( , ln , and ). If it can be assumed4   
that log-settlement is approximately normally distributed (which is seen later to be a
reasonable assumption and is consistent with the distribution selected for ), and ln 5 2
6 2
and 2ln are the estimators of the mean and variance of log-settlement, respectively,
 7 89 6 2 ;: < 4= ?
then the standard deviation of these estimors obtained from 2000 realizations are
given by ln ln = 0 022 ln and 2ln 2

2
6 2 2
1 ln = 0 032 ln so that the 8 9 > @BA 6 2  6 2
estimator ‘error’ is negligible compared to the estimated variance.
Realizations of the log-elastic modulus field, ( ), are produced using the two- C ED
dimensional Local Average Subdivision (LAS) technique (Fenton and Vanmarcke,
C  D
1990, Fenton, 1994), where ( ) is the local average of a zero mean, unit variance
Gaussian random field over the domain of the element centered at . The generated 4 D
field correctly reproduces the mean, variance and covariance structure of the 2-D local
average process. The elastic modulus value then assigned to the ’th element is F
 D
( ) = exp G  ln  +  ln  C ( D ) H (4)
Once the field of elastic modulus values is assigned, the settlement(s) are computed
via finite element analysis.
SINGLE FOOTING CASE
A typical histogram of the settlement under a single footing, as estimated by 2000

  = 0 2,     
realizations, is shown in Figure 2. This is for the case where the footing has width

= 2, and ln = 0 7. With the requirement that settlement be
non-negative, the shape of the histogram suggests a lognormal distribution, which was
 7 D@  7JI-K  7 D@
adopted in this study (see also Eq. 3) . The histogram itself is computed over 30

 7JI-K
equally spaced intervals between ln( ) and ln( ), in log-space, where and
are the minimum and maximum settlements observed in the sample of 2000. The
histogram is normalized to enclose a unit area and a straight line is drawn between the
interval midpoints.
M
0.2

Frequency Count
Normalized Frequency

µln δ = 1.316, σln δ = 0.7927

M
0.1

L
0

0 5 10 15 20 25 30
Footing Settlement
Figure 2. Typical frequency histogram and fitted lognormal distribution of settlement
under a single footing.
 2  2
Superimposed on the histogram is the fitted lognormal distribution with parameters
given by ln and ln in the line key. At least visually, the fit appears quite reasonable.
In fact a Chi-Square goodness-of-fit test gives a critical p-value of 1 10 8 . The critical N A
p-value may be interpreted as the probability of mistakenly rejecting the lognormal
hypothesis – larger values of p imply a better fit to the data. Unfortunately, the Chi-
Square test is quite sensitive to the ‘smoothness’ of the histogram. Although it would
probably be well worth investigating the Kolmogorov-Smirnov goodness-of-fit test to

6 Fenton et al.
evaluate the fit of the assumed distributions, this was not performed in the current study
because the parameters of the assumed distribution are derived from the data and the
critical statistic is, strictly speaking, unknown under these conditions.
 
Over the entire set of simulations done for each parameter set of interest ( , , and  
ln ) the fraction of critical p-values obtained are listed in Table 2. Since over 30%
have critical p-values in excess of 0.05, and over 70% in excess of 0.0001 (and so are
better fits than that shown in Figure 2) it appears that the lognormal hypothesis is a
reasonable one.
Table 2. Fraction of simulation runs with Chi-Square goodness-of-fit critical p-
value greater than that indicated.
O RP Q D .
S 0 5
Fraction
S 0 1
7%
S 0  05
26%
S 0  01
33%
S 0  0001
49%
71%
Accepting the lognormal hypothesis as a reasonable fit to the simulation results, the

T    
next task is to estimate the parameters of the fitted lognormal distributions as functions
of the input parameters ( , , and ln ). The lognormal distribution,
 
U (   ln 2 2 T
Z [
: 2V  ln 2  exp W
2 )=
1 1
2 X
ln
 ln 2 Y 0 \]_^ ( (5)

has two parameters,  ln 2 and  ln 2 .  2 5 2

varies with  ln  for   = 0  1.
Figure 3 shows how the estimator of ln , ln ,
Similar results were found for the other footing
widths. All scales of fluctuation are drawn in the two plots, but are not individually
labeled since they lie so close together. This observation implies that the mean log-
settlement is largely independent of the scale of fluctuation, ln . This is as expected
 
since the scale of fluctuation does not affect the mean of a local average of a Gaussian
)
process (recall that if is lognormally distributed, then ln( ) is normally distributed). )
Figure 3 suggests that the mean of log-settlement can be estimated by a straight line
of the form
ln = ln( ) + 2 ln2  2 ) *1,/. `  
(6)
where ) *1,3.
is the ‘deterministic’ settlement obtained from a single finite element analy-

sis (or appropriate approximate calculation) of the problem where = everywhere.
For the range of geometries considered in this study, the following relationship can be
) *1,3.
used to approximate ln( ) reasonably accurately
  
ln( ) *1,3. ) = ln( a        b  ; )c
2
) 0 4924 0 6883 ln( ) 0 0964 ln( (7)

which was obtained by regression over the intercepts shown on Figure 3 and over the
d
other footing widths not shown. The 2 coefficient of determination for the above
regression was 0.9999

7 Fenton et al.
 kW / H = 0.10
f

4
mln δ
2

e0
0

f
0.5 g
1
jσ21.5lnf E h
2 f
2.5 i
3

Figure 3. Estimated mean of log-settlement.

The slopes of the curves in Figure 3 are almost uniformly 0.5, as predicted for the
settlement by Eq. (3a) in the large and small scale of fluctuation cases. Note that if
the settlement mean is independent of the scale of fluctuation, Eq. (3a) is valid for
any scale. However, there is in fact a slight dependence of the slope, 2 , on ln . The `  
 
second term in the following is a small correction obtained from plots of the slope 2 `
versus 

and ln
  
` 2 = 0 5 + l T 0 041
b ln( ln  ; c 
2
1
 ; exp 4 )+1 (8)

 
Eq. (8) is entirely empirical, but does have the correct limiting forms for large and
small ln . It is unknown at this time if it can be applied for values of outside  
the range investigated. The physical interpretation and analytical verification of the
correction term in the above needs further investigation.
kW /H = 0.10
f
101

nσ /µ
E E = 4.00
10-1

nσ /µ
sln δ

E E = 0.10
10-3

i e i
-5 -4 -3 -2
mln( θ -1 /H ) 0 1 2 3

kW /H = 1.00 ln E

f
101

nσ /µ
E E = 4.00
10-1

nσ /µ
sln δ

E E = 0.10
10-3

i e i
-5 -4 -3 -2
mln( θ -1 /H ) 0 1 2 3
ln E
Figure 4. Estimated standard deviation of log-settlement.
The estimator of the standard deviation of log-settlement, ln , is plotted in Figure 4 6 2
6 2 %  
for the smallest and largest footing widths. Intermediate footing widths give similar
results. In all cases, it can be seen that ln
 
ln for large ln . It is expected

8 Fenton et al.
  
that the reduction in variance as ln decreases is due largely to the local averaging
effect under the footing. That is, if the average of ln( ) is taken over some area under
the footing, then this average is expected to have smaller variance for small scales of
 
fluctuation than for large. This is because there are more ‘independent’ samples in the [ [ [
area when ln is small. Recall that if ¯ is the average of samples 1 2
 o B< , 00 @
then the variance of ¯ is 2 D
if the ’s are mutually independent (the 1 factor   % p<
is the variance reduction) – this case corresponds to the ln
D
hand, if the ’s are fully correlated ( ln
  %q(
0 case. On the other
), then the variance of ¯ is just 2 , so o
that there is no variance reduction. See Vanmarcke (1984) for more details on local
averaging theory. The variance reduction effects are clearly seen in Figure 4.
Following this reasoning, and assuming that local averaging of the area under the
footing accounts for all of the variance reduction seen in Figure 4, the standard
deviation of log-settlement is
 ln 2 = l r (  [  ;  ln  )  ln  (9)
r [
in which (    
; ln ), the so-called variance function (Vanmarcke, 1984), gives the
amount that the variance is reduced when the random field is averaged over a region
of size N
 . Note that the dependence of the averaging region on is apparently 
only valid for the test case considered; if the footing is founded on a much deeper soil
mass, one would not expect to average over the entire depth due to stress distribution
with depth. This issue needs additional study.

v W /H = 0.1
101

t v W /H = 0.1 (pred)
f

v W /H = 1.0 u
5

f
σE/µE = 4.0
v W /H = 1.0 (pred)
f
100

t
5

u σE/µE = 0.1
10-1
sln δ

t
5

v W /H = 0.1
-2

v W /H = 0.1 (pred)
10

t f

v W /H = 1.0
5

v W /H = 1.0 (pred)
f
-3

f
10

-6 -4 -2 s
0 2 4
ln( θln E /H )
Figure 5. Comparison of simulation estimated standard deviation of log-settlement
with theoretical estimate, Eq. (9).
For the isotropic Gauss-Markov correlation function used to represent the ln( ) random
field (Eq. 1), the variance function is closely approximated by
r [ r r r r
( w 1 w 2;  ) = 21 x (w 1) (w 2 w 1) + (w 2) (w 1 w 2)y (10)

9 Fenton et al.
 A A
where,
r D w D { [ r w D {
2 2
3 3

(w ) = z 1 +  ( w D w}| ) = z 1 + ~
3 3

(11 # )
2 2

X Y X |Y
~ | =  z+ + € 1   exp W  w}|  2 Z { (11 $ )
2
2
X 2 Y
in which w D are dimensions of the averaging region. Predictions of  ln 2 using Eq. (9) are

widths and      values considered in this study. The agreement is remarkable.

plotted in Figure 5 against the simulation results for the largest and smallest footing

Intermediate cases show similar, if not better agreement with predictions.

Single Footing Example

Consider a single footing of width 
= 2 0 m to be founded on a soil layer of 
depth 10.0 m and which will support a load = 1000 kN. Suppose also that samples
taken at a nearby location 1 have allowed the estimation of the elastic modulus mean
and standard deviation at the site to be 40 MPa and 40 MPa respectively. Similarly,
 
nearby test results on a regular array have resulted in an estimated scale of fluctuation,

ln = 3 0 m. Assume also that Poisson’s ratio is 0.25.
The results from the previous section can be used to estimate the probability that the
settlement under the footing will not exceed 0.10 m as follows;

everywhere equal to 
1) A deterministic finite element analysis of the given problem with elastic modulus
= 40 MPa gives a deterministic settlement of = )0*1,3.

0 03531 (note that Eq. 7 gives
  ) *1,/.
= 0 03604, a relative difference of 2%). 
2) for   
= 0 2 and ln = 3, use Eq. (8) to compute 2 , `
  
` 
0 041
b ln(3  10) + 1c  
2

: 
1
2 =05+ exp 4 = 0 5907
02

3) compute variance of log-elastic modulus,

 ln2   2 ƒ
= ln 1 + ‚ XY = ln(2) = 0 69315 
 ln  
= 0 83256

4) compute mean of log-settlement,

 
 ln 2 = ln( ) *-,/. ) + ` 2  ln2  = 
3 3437 + 0 5601(0 69315) =   
2 9341

5) compute standard deviation of log-settlement using Eq.’s (9) through (11),

r A A
(  ) = „ 1 + (    ln  )3… 2† 2… 3 = „ 1 + (2  3)3… 2† 2… 3 = 0  74847
1
Note that if elastic modulus measurements were taken at the site itself, then the results presented
in the previous section would not be applicable: When information about the actual site (beyond
statistical information) is known, then the site variability is considerably reduced.

10 Fenton et al.
 r A A
 „
( )= 1+( ‡  ln  )3… 2† 2… 3 = „ 1 + (10  3)3… 2† 2… 3 = 0  27107
~ = 3 zˆ + € 1   exp W    2 Z { = 3  28226
1 2 2 X 2 ln  Y
~ = 3 zˆ + € 1   exp W    2 Z { = 4  69343
2 2 2 X 2 ln  Y
r ~ A 2… 3
( 
"‰ ) = „ 1 + ( 
} 2 )3 … 2 † = 0  84907
r ~ A 2… 3
( Š‹T ) = „ 1 + ( ‡ 1 )3 … 2 † = 0  29261
r [   1r r r r
(   ; ln ) = 2 x (  ) ( Š‹ ) + (  ) (  ‰ )y = 0  22458

 ln 2 = l r ( T [  ;  ln  )  ln  = : 0  22458(0  83256) = 0  39455


Aside: for ln 2 = 2  9341 and  ln 2 = 0  39455, the corresponding settlement mean

and variance can be obtained from the transformations
 2 = exp G  ln 2 + 12  ln2 2 H = 0  0575 m

 2 =  2 l ŒŽ 8 1 = 0  0236 m
2
ln

A trial run of 2000 realizations for this problem gives 5 2 = 0  0582 and 6 2 = 0  0219
for relative differences of 1.2% and 7.7% respectively. The estimated standard error
on 5 2 is approximately 0.0005 for 2000 realizations.


6) compute the desired probability using the lognormal distribution,

ln(0  10)  ln 2
P [)\ 0  10] =   ln 2
X Y
=  (1  6006)
= 0  945
where  ( ‘ ) is the standard normal cumulative distribution, whose table of values
can be found in any good probability textbook.
The simulation run for this problem yielded 1892 samples out of 2000 having settlement
less than 0.10 m. This gives a simulation based estimate of the above probability of

) S 
0.946, a relative difference of only 0.1%. Although this is very good accuracy, one
must a bit cautious since if the probability in question had been P [ 0 10] then the
relative error becomes 1.9%. It is expected that probabilities estimated farther out in
the tail of the distribution may have even larger differences with simulation results
unless the simulation is carried out over very many more realizations, and this is yet
to be verified.

TWO FOOTING CASE

Having established with reasonable confidence the distribution associated with settle-
ment under a single footing founded on a soil layer, attention can now be turned to the
more difficult problem of finding a suitable distribution to model differential settlement
)
between footings. Analytically, if 1 is the settlement under the left footing shown

11 Fenton et al.
 )
in Figure 1 and 2 is the settlement of the right footing, then according to the results
) )
of the previous section, 1 and 2 are joint lognormally distributed random variables
following the bivariate distribution
U [“’   
2 2 ( ) = 2V  ln2 2 dB exp d /„ ”  ln 2 ” K ”–• + ” 2• † 
K 1 ’ (12)
1 2

’
1 2
  2 2
for T—  0, — 0, where ” K = (ln   ln 2 )   ln 2 , ” • = (ln   ln 2 )   ln 2 , and where
 
d 2 = 1 2ln 2 with ln 2 being the correlation coefficient between the log-settlement of
the two footings. It is assumed in the above that ) 1 and ) 2 have the same mean and
variance, which, for the symmetric conditions shown in Figure 1(b), is a reasonable
assumption.
If the differential settlement between footings is defined by
˜ 
= ) 1 ) 2  (13)
then the distribution of
˜ is given by
U™ ( ’4[“’ ’
)=2 šœ› U 2 2 ( + ) w1 2
(14)
0

Unfortunately, this integral cannot be solved analytically insofar as the authors are
aware, although for design purposes it can be estimated using any available reliability
tool, such as first- or second-order reliability methods. Such numerical approximations
˜
to Eq. (14) are being investigated for a future publication. It is not hard to show that
the variance of can be written,  ™ 
= 2(1  2) 2  2
 2  ™2 (15)
where  2 is the correlation coefficient between ) 1 and ) 2.
Figure 6 shows a typical histogram of differential settlement between the two equal
sized footings. Superimposed on the histogram is a trial exponential distribution
U™  
G Ÿ  ™ H
having the form
( ) = 1 exp (16) “ž
with ™ taken as 0.8998 which is the data average. Although this distribution fails the
Chi-Square goodness-of-fit test, it appears to capture the major trends in the histogram,
particularly in the tail.
M
1.5

Frequency Count
Normalized Frequency

µ∆ = 0.8998, σ∆ = 0.9557
1

M
0.5

  ¡ ¢ £4 L ¤
0

0 1 2 3 5 6
Differential Settlement

Figure 6. Frequency histogram and fitted distribution for differential settlement un-
der two equal sized footings.

12 Fenton et al.
 ˜   
¥  ¦ 
Figure 7 shows the estimated mean and standard deviation of as functions of ln
, for   
= 0 5. The other two footing widths considered (0.1 and 0.3) are
,

of similar form. The two plots are nearly identical, suggesting that perhaps it is not
™ ™
unreasonable to take a one-parameter exponential distribution as representative of the
differential settlement (for which = ). A trial function of the form

 2™ = § (1  2 ) 22 (17)
will be investigated (see Eq. 15) to predict the parameter of the exponential distri-  '%
2 %
bution, for some positive constant . Note that when ln   % §(
0, the previous section
predicted 2
settlements, 2 %
0, and when ln , the correlation coefficient between the footing
1. Thus Eq. (17) is in agreement with the observation that differen-
 
tial settlements are expected to go to zero for both very small and very large values of
ln .
ª W /H = 0.50
f
«
103

« σE/µE = 0.10

« σE/µE = 1.00
σE/µE = 4.00
m∆
100

©
10-3

-5 -4 ¨
-3 -2 -1 0 1 2 ¨
3

ª W /H = 0.50 ln( θln E /H )

f
«
103

« σE/µE = 0.10

« σE/µE = 1.00
σE/µE = 4.00
100
s∆

©
10-3

-5 -4 ¨
-3 -2 -1 0 1 2 ¨
3
ln( θln E /H )
Figure 7. Estimated differential settlement mean and standard deviation.
The factor 2
in Eq. (17) can be obtained by first considering the correlation, ln ,  2
between the two local averages of the ln( ) field under the two footings; these local
averages are of dimension in width by 
in height and are separated, center to  


center, by the distance . The correlation, ln , can be found using the variance 2
  
function as
 ln 2 r [ r [ r [
 a 2 ¬ (
 ) (
  )+(
+  ) (
+   ) 2
(
 ) ­ (18)
1 2 2 2
=
2

where the (; ln  ) notation is now dropped 
the variance function is dependent on ln  . Calculating  ln 2 and  ln 2 as in the previous
for convenience, it being understood that

section (Eq’s 6 to 9) then allows the computation of  2 and  22;

 2 = exp G  ln 2  + 12  ln2 2 H (19 # )
Œ 
 22 =  22 ( 8 1) (19 $ )
2
ln

13 Fenton et al.
 2
With these results, the correlation can be found from

 2 = exp G  ln 2 2 ln2 2 H 1
exp G ln 2 H
(20)
1

and ™
can now be estimated using Eq. (17) if a suitable value of is found. §
To test the ability of the assumed distribution to accurately estimate probabilities, the
probability ˜ \]`  ™  Œ0®
P[ ]=1 (21)
`
for varying from 0.5 to 4.0, is plotted against the corresponding probabilities esti-
mated directly from the simulation results. After some trial and error, a value of 2 3 § 
    
was found to give the most accurate probabilities on average over the range of , 

, and ln considered here. Figure 8 shows the predicted probabilities using
  § 
Eq. (17) with = 2 3 alongside the estimated probabilities averaged over all  
and ln parameter values. When results are not averaged over parameter values, the
agreement is typically less good away from the tails (some above, some below) but
generally reasonable in the tails.
1

M
0.9

M
0.8

M
P[ ∆ < αµ∆ ]

± Predicted
0.7

M v W /D = 0.1
v W /D = 0.3
0.6

M v W /D = 0.5
f
0.5

£4
0.4

0.5  
1 ¯
1.5 ¡ ° 2.5¯ 2 3¢ ¯
3.5
α ˜ ™ 
Simulation based estimates of P [ \²`  ], averaged over all ln 
Figure 8.
     cases, compared to that predicted by Eq.’s (17) and (21). and

CONCLUSIONS
On the basis of this initial simulation study, which is by no means complete, some
tentative observations are made as follows.
It appears that the settlement under a footing founded on a spatially random elastic
modulus field of finite depth overlying bedrock is reasonably well represented by a
 2
 2
2  2
lognormal distribution, if is also lognormally distributed, with parameters ln and
ln . The first parameter, ln , is dependent on the mean and variance of the underlying  
elastic modulus field and may be largely derived by considering limiting values of ln .
Although there is some question as to why the slope correction term appearing in Eq.
(8) is necessary, including it yields quite accurate estimates of the mean log-settlement.
It is significant to note that the second parameter, ln2 , is very well approximated by the  2
 2
variance of a local average of the elastic modulus field in the region directly under the
footing. This gives the prediction of ln2 some generality that could possibly extend
beyond the actual range of simulation results considered herein if a suitable averaging
domain can be defined. Once the statistics of the settlement, ln and ln2 , have been  2  2
14 Fenton et al.
computed using Eq.’s 6 to 9 the estimation of probabilities associated with settlement
involves little more than referring to a standard normal distribution table.
The differential settlement follows a more complicated distribution than that of settle-
ment itself (see Eq. 14). This is seen also in the differential settlement histograms
which tend to be quite erratic with long tails. Clearly the difference between two log-
normally distributed random variables is not exponentially distributed. Neither does
this difference follow a normal distribution, the normal distribution falling off far too
rapidly in the tails. For an accurate estimation of probability relating to differential
settlement where it can be assumed that footing settlement is lognormally distributed,
Eq. (14) should be numerically integrated. This approach was not pursued in this
study since it was observed that the tail of the differential settlement histogram was
reasonably well approximated by an exponential distribution. The results suggest a
relatively simple approach to obtaining ‘ball-park’ estimates of probabilities associ-
ated with differential settlement that involves, again, statistics of the underlying elastic
modulus field and local averages of the field directly under the footings. In that the
statistical parameters of the underlying elastic modulus field are themselves estimates,
that may or may not be very accurate, the proposed probability predictions may be
reasonable.
REFERENCES
American Concrete Institute (1989). ACI 318-89, Building Code Requirements for
Reinforced Concrete, Detroit, Michigan.
Asaoko, A. and Grivas, D.A. (May 1982). “Spatial Variability of the Undrained
Strength of Clays,” ASCE J. Geotech. Eng., 108(5), 743-756.
Canadian Geotechnical Society (1978). Canadian Foundation Engineering Manual,
Montreal, Quebec.
Canadian Standards Association (1984). CAN3-A23.3-M84 Design of Concrete Struc-
tures for Buildings, Toronto, Ontario.
Chiasson, P., Lafleur, J., Soulié, M. and Law, K.T. (1995). “Characterizing spatial
variability of a clay by geostatistics,” Can. Geotech. J., 32, 1–10.
D’Appolonia, D.J., D’Appolonia, E. and Brissette, R.F. (1968). “Settlement of spread
footings on sand,” ASCE J. Soil Mech. Found. Div., 94(SM3), 735–760.
DeGroot, D.J. and Baecher, G.B. (1993). “Estimating autocovariance of in-situ soil
properties,” ASCE J. Geotech. Eng., 119(1), 147–166.
Fenton, G.A. (1994). “Error evaluation of three random field generators,” ASCE J.
Engrg. Mech., 120(12), 2478–2497.
Fenton, G.A. and Vanmarcke, E.H. (1990). “Simulation of Random Fields via Local
Average Subdivision,” ASCE J. Engrg. Mech., 116(8), 1733–1749.
Marsily, G. (July 1985). “Spatial Variability of Properties in Porous Media: A
Stochastic Approach,” in Advances in Transport Phenomena in Porous Media,
NATO Advanced Study Institute on Fundamentals of Transport Phenomena in
Porous Media, Bear, J. and Corapcioglu, M.Y., eds., Dordrecht, Boston, MA,
719-769.
Milovic, D. (1992). Stresses and Displacements for Shallow Foundations, in Devel-
opments in Geotechnical Engineering, 70, Elsevier, Amsterdam.
Ravi, V. (1992). “Statistical modelling of spatial variability of undrained strength,”
Can. Geotech. J., 29, 721–729.
Soulié, M., Montes, P. and Silvestri, V. (1990). “Modelling spatial variability of soil
parameters,” Can. Geotech. J., 27, 617–630.
Vanmarcke, E.H. (1984). Random Fields: Analysis and Synthesis, The MIT Press,
Cambridge, Massachusetts.

15 Fenton et al.