Professional Documents
Culture Documents
Geostatistics Formula Sheet
Geostatistics Formula Sheet
Geostatistics Formula Sheet
ostatisticcs Formu
ula Sheeet O
October 2014
4
Z den
notes a random variable (RVV). z denotess an outcome. FY , Z ,, Z ( y , z1,, z N )
F ( y) 1 N
Z(u) denotes a reggionalized RV at location u.
u The set off Y |Z1,, Z N FZ1 ,, Z N ( z1,, z N )
randoom variables over
o a station
nary domain A A {Z (u ),uA} iss
The covariance and correlation coefficient summarize
know
wn as a random
m function (RF)..
bivarriate dependen
nce between tw
wo random variables:
Uncertainty in a RV is represented by a cumulative e
Cov{ X , Y } C E{[ X m ][Y m ]} E{ XY } m m
distribution functio on (CDF): F(z)=Prob{Z≤z}. The derivative off XY X Y X Y
the C
CDF is the proobability density function (P
PDF): f(z)=F’(z). XY C XYY / X Y
Quan ntiles are z‐values with a probabilistic me
eaning: zp such
h
that FF(zp)=p. The quuantile function is denoted F‐1(p)=zp. d: 2 (h ) E Z (u ) Z (u h )2
The vvariogram for lag h Is defined .
The eexpected value
e operator is written
w z f ( z ) dz .
E{Z } Undeer stationarityy, the variograam, variance aand covariance
E{Z} is denoted m and is also known as the firrst moment orr are reelated by (h)==2‐C(h).
n. The variancce is 2 E{[ Z m ]2 } E{Z 2 } m2 . is the
mean e 2 2 2
h h h
dard deviation. /m is the co
stand oefficient of variation. The sscalar normalizzed distance is h X Y Z
aX aY aZ
A ran mZ)/. E{Y}=0,,
ndom variable Z is standardized by Y=(Z‐m
Clockkwise rotation of X/Y by angle is achieved
d by:
E{Y }==1 and Z=Y + mZ.
2
x1 cos sin x0
Z is uniform in the interval a to b w
when: y sin
ccos y0
1
1/(b a ),
) z[ a ,b ] a b 2
2 (b a )
f ( z) , m and Strattigraphic relativve coordinatess are calculated
d as:
0, otherw
wise 2 12
Z ( x, y ) Zcb ( x, y )
z2
Z reel x, y T
Zct ( x, y ) Zcb ( x, y )
when f ( z ) 1
Z is sttandard normaal or Gaussian w e 2
2
Varioograms are mo uctures: (h ) inst
odeled by stru 0 Ci i (h ) .
Comm
mon standarrdized modelss include the Exponentia
al
3
Exp ( h ) 1exp( 3h / a ) , Spherical Sph ( h ) 1.5( h / a ) 0.5( h / a )
2
if h a ; 1, otherwisee , Gaussian Gaus ( h ) 1 eexp( 3( h / a ) ) .
h
The hhole effect is leess common: h C 1cos
a
The vvolume averagged variogram between v and
d V (gammabar):
)
1
V ,v x y dxdy
|V ||v|V v
The ddispersion variance is given b
by:
The vvariable Z>0 is lognormal witth m and w
V ,V v,v
2
when Y=ln(Z) iss 2
D v ,V E Zv mV 2
mal with mean and variance
norm e . The param
2
meters:
2
lln( m ) / 2
2
ln 1 2 / m2
m2
2
2 2
Variaances add: D v , A D v ,V D V , A
2
v V A
Z e Y 1 Variaance of a linearr combination::
2 /2 2 2
m e 1
2 2
me
1 n 2 n n
x x Var{ x }
V X 1 Cov{ xi , x j }
The m N is defined as:
multivariate disstribution of N RVs Zi,i=1,...,N n i 1 i n n 2 i 1 j 1, i j
Page 1 of 2
The estimation variance is calculated as: Where is the 1xn vector of mean values and is the nxn
n n n matrix of covariances. Conditional distributions defined by
2 2 2 C
E i i , i j Ci , j normal equations (see simple kriging).
i 1 i 1 j 1
LU simulation from a covariance matrix: C=LU; y=Lw.
Minimizing the estimation variance leads to simple kriging
Sequential simulation relies on recursive decomposition of
and minimized estimation variance (kriging variance):
the multivariate distribution:
n 2 2 C n
j Ci , j Ci , i 1,...,n SK i i , P ( A , ..., A ) P ( A | A , ..., A ) P ( A , ..., A )
1 1 N 1 1 N 1
j 1 i 1 N N
L Indicators for continuous variables
Universal kriging: m (u ) al fl (u )
l 0 1, if z (u ) zc
i u ; zc for manycutoffs z
0, otherwise c
n L
j Ci , j l Ci , i 1,...,n
j 1 Indicators for categorical variables
l 0
n 1, if u k
j fl (u j ) fl (u ) l 0,...., L i u ;k for k 1, ..., K
j 1 0, otherwise
Mean and variance of an indicator variable are given by:
External drift considers m (u ) a0 a1 f1 (u)
E i u ;k p
Var i u ;k p 1 pk
k k
2 2
Location dependent variance of SK: Var z*SK SK
Permanence of ratios for combining conditional probabilities:
The cross variogram between variable Zi(u) and Zj(u): n 1
1 P ( A)
2
i, j
(h ) E [ Zi (u ) Zi (u h )][ Z j (u ) Z j (u h )] P ( A | B , i 1, ..., n )
P ( A)
i n 1 n
1 P ( A) 1 P ( A|Bi )
Matrix of cross variograms can be modeled by linear model of
P ( A) i 1 P ( A|Bi )
coregionalization (LMC) i,j=1,...,M:
Stepwise conditional transformation:
K
2 (h ) bik, j k h 1
Prob Z1 z1
i, j Y G
k 0 1
Where each MxM matrix of coefficients (k=0,...,K) must be Y2|1 G 1 Prob Z 2 z2 |Y1 y1
positive definite. Intrinsic model assumes all variograms are
1
Prob Z3 z3|Y2 y2 ,Y1 y1
Y G
proportional. The Markov models assume that the cross 3|2,1
variogram/covariance is proportional to a direct variogram.
Bayesian updating prior and likelihood Gaussian distributions:
C
i, j
( h ) b C ( h ) where b j / i
i ,i i, j yL P2 y 2
P L 2 P2 2
yU U L
Cokriging considers correct covariance between data events. P P L L2
2 2 2 2 2 2 2
P P L L
Z‐data are transformed to be Y‐normal (G(y) is Gaussian CDF) Compositional data could be handled by additive logratios:
with normal score transform:
x exp yi
1 1 y ln i and xi D i 1, ..., D
yG ( F ( z )) and z F
i x D
(G ( y )) exp yi
i 1
The n‐variate multivariate Gaussian distribution is defined:
1 1
exp y μ Σ 1 ( y μ )
T
f (y ) Disclaimer: there may be mistakes on this formula sheet. Any mistakes are
n 1/2 2
2 |Σ| your fault – you should not need a formula sheet anyway.
Copyright © 2014 – Clayton V. Deutsch
Page 2 of 2