Kriging回归
given set s = [ s 1 , s 2 , … , s m ] T \mathbf{s}=[s_1,s_2,\ldots,s_m]^T s=[s1,s2,…,sm]T, corresponding response g = [ g 1 , g 2 , … , g m ] T \mathbf{g}=[g_1,g_2,\ldots,g_m]^T g=[g1,g2,…,gm]T,
predictor:
g
^
(
t
)
=
f
T
(
t
)
β
+
Z
(
t
)
\hat{g}(t)=f^T(t)\beta+Z(t)
g^(t)=fT(t)β+Z(t)
f
(
t
)
f(t)
f(t) is a regression function, the stochastic process
Z
(
t
)
Z(t)
Z(t) is assumed to have zero mean and covariance is :
E
[
Z
(
W
)
Z
(
Q
)
]
=
σ
2
R
(
θ
;
W
,
Q
)
E[Z(\mathbf{W})Z(\mathbf{Q})]=\sigma^2R(\theta;\mathbf{W},\mathbf{Q})
E[Z(W)Z(Q)]=σ2R(θ;W,Q)
R
R
R is the correlation function defined by parameter
θ
\theta
θ, the most widely used is gaussian correalaton function:
R
(
θ
;
W
,
Q
)
=
∏
j
=
1
n
[
−
θ
(
Q
j
−
W
i
)
2
]
R(\theta;\mathbf{W},\mathbf{Q})=\prod_{j=1}^n\left[-\theta(Q_j-W_i)^2\right]
R(θ;W,Q)=j=1∏n[−θ(Qj−Wi)2]
the response
G
(
u
)
G(u)
G(u) of a given test point
u
\mathbf{u}
u obey a gaussian distribution, denote as
G
(
u
)
∼
N
(
μ
G
(
u
)
,
σ
G
2
(
u
)
)
G(\mathbf{u})\sim N\left(\mu_G(\mathbf{u}),\sigma_G^2(\mathbf{u})\right)
G(u)∼N(μG(u),σG2(u))
the mean and standard devision given as :
μ
G
(
u
)
=
f
(
u
)
T
β
∗
+
r
(
u
)
T
R
−
1
(
g
−
F
β
∗
)
\mu_G(\mathbf{u})=f(\mathbf{u})^T\beta^*+r(\mathbf{u})^T\mathbf{R}^{-1}(\mathbf{g}-\mathbf{F}\beta^*)
μG(u)=f(u)Tβ∗+r(u)TR−1(g−Fβ∗)
σ G 2 ( u ) = σ 2 [ 1 + v T ( F T R T F ) − 1 v − r ( u ) T R − 1 r ( u ) ] \sigma_G^2(\mathbf{u})=\sigma^2\left[1+\mathbf{v}^T(\mathbf{F}^T\mathbf{R}^T\mathbf{F})^{-1}\mathbf{v}-r(\mathbf{u})^T\mathbf{R}^{-1}r(\mathbf{u})\right] σG2(u)=σ2[1+vT(FTRTF)−1v−r(u)TR−1r(u)]
where
v
=
F
T
R
−
1
r
(
u
)
−
f
(
u
)
\mathbf{v}=\mathbf{F}^T\mathbf{R}^{-1}r(\mathbf{u})-f(\mathbf{u})
v=FTR−1r(u)−f(u)
σ 2 = 1 m ( Y − F β ∗ ) T R − 1 ( g − F β ∗ ) \sigma^2=\frac{1}{m}(\mathbf{Y}-\mathbf{F}\beta^*)^T\mathbf{R}^{-1}(\mathbf{g}-\mathbf{F}\beta^*) σ2=m1(Y−Fβ∗)TR−1(g−Fβ∗)
β ∗ = ( F T R − 1 F ) − 1 F T R − 1 g \beta^*=(\mathbf{F}^T\mathbf{R}^{-1}\mathbf{F})^{-1}\mathbf{F}^T\mathbf{R}^{-1}\mathbf{g} β∗=(FTR−1F)−1FTR−1g
r
(
u
)
r(\mathbf{u})
r(u) is correlation vector between
S
\mathbf{S}
S and
u
\mathbf{u}
u
r
(
u
)
=
[
R
(
θ
;
s
1
,
u
)
,
…
,
R
(
θ
;
s
m
,
u
)
]
T
r(\mathbf{u})=[R(\theta;s_1,\mathbf{u}),\ldots,R(\theta;s_m,\mathbf{u})]^T
r(u)=[R(θ;s1,u),…,R(θ;sm,u)]T
F
F
F is the regression matrix
F
=
[
f
(
s
1
)
,
…
,
f
(
s
m
)
]
T
F=[f(s_1),\ldots,f(s_m)]^T
F=[f(s1),…,f(sm)]T
the optimal coefficients
θ
∗
\theta^*
θ∗ of the correlation function solves
θ
∗
=
min
θ
{
∣
R
∣
1
m
σ
2
}
\theta^*=\min_{\theta}\{|\mathbf{R}|^{\frac{1}{m}}\sigma^2\}
θ∗=θmin{∣R∣m1σ2}