推导过程
A
m
×
n
A_{m \times n}
Am×n 表示待分析变量(anomaly),有m个格点,n个时间片,
A
=
(
a
1
,
a
2
,
.
.
.
,
a
n
)
A = \begin{pmatrix} a_1, a_2, ..., a_n \end{pmatrix}
A=(a1,a2,...,an)
X
=
A
A
T
X = AA^T
X=AAT,
X
m
×
m
X_{m \times m}
Xm×m 表示方差-协方差方阵。对角化如下:
X
m
×
m
=
U
m
×
m
T
S
m
×
m
U
m
×
m
X_{m \times m } = U^T_{m \times m} S_{m \times m} U_{m \times m}\\
Xm×m=Um×mTSm×mUm×m
其中S是对角方阵,
S
λ
S_\lambda
Sλ是对角线上的特征值组成的向量,其中
U
=
(
u
1
,
u
2
,
.
.
.
,
u
m
)
U = \begin{pmatrix} u_1, \\ u_2, \\ ..., \\ u_m \end{pmatrix}
U=
u1,u2,...,um
则有
U
A
(
U
A
)
T
=
S
UA(UA)^T = S
UA(UA)T=S
U
A
UA
UA 可以写作
U
A
=
(
u
a
1
,
u
a
2
,
.
.
.
,
u
a
m
)
u
a
1
=
u
1
A
,
u
a
2
=
u
2
A
,
.
.
.
,
u
a
m
=
u
m
A
,
UA = \begin{pmatrix} ua_1, \\ ua_2, \\ ..., \\ ua_m \end{pmatrix} \\ ua_1 = u_1A, ua_2 = u_2A, ..., ua_m = u_mA,
UA=
ua1,ua2,...,uam
ua1=u1A,ua2=u2A,...,uam=umA,
其中
u
1
,
u
2
,
.
.
.
,
u
m
u_1, u_2, ..., u_m
u1,u2,...,um 是模态,
u
a
1
,
u
a
2
,
.
.
.
,
u
a
m
ua_1, ua_2, ..., ua_m
ua1,ua2,...,uam 是主成分(PC)
通过模态和PC重构A
欲证明
u
1
T
(
u
1
A
)
+
u
2
T
(
u
2
A
)
+
.
.
.
+
u
m
T
(
u
m
A
)
=
A
u_1^T(u_1A) + u_2^T(u_2A) + ... + u_m^T(u_mA) = A
u1T(u1A)+u2T(u2A)+...+umT(umA)=A
对上式左右都取转置,再乘以
A
T
A^T
AT
A
T
u
1
T
u
1
+
A
T
u
2
T
u
2
+
.
.
.
+
A
T
u
m
T
u
m
=
A
T
A
A
T
u
1
T
u
1
+
A
A
T
u
2
T
u
2
+
.
.
.
+
A
A
T
u
m
T
u
m
=
A
A
T
A^Tu_1^Tu_1 + A^Tu_2^Tu_2 + ... + A^Tu_m^Tu_m = A^T\\ AA^Tu_1^Tu_1 + AA^Tu_2^Tu_2 + ... + AA^Tu_m^Tu_m = AA^T
ATu1Tu1+ATu2Tu2+...+ATumTum=ATAATu1Tu1+AATu2Tu2+...+AATumTum=AAT
由于
A
A
T
u
1
T
=
S
1
u
1
T
AA^Tu_1^T = S_1u_1^T
AATu1T=S1u1T,则有
S
1
u
1
T
u
1
+
S
2
u
2
T
u
2
+
.
.
.
+
S
m
u
m
T
u
m
=
A
A
T
U
m
×
m
T
S
m
×
m
U
m
×
m
=
A
A
T
S_1u_1^Tu_1+ S_2u_2^Tu_2 + ... +S_mu_m^Tu_m = AA^T\\ U^T_{m \times m} S_{m \times m} U_{m \times m} = AA^T
S1u1Tu1+S2u2Tu2+...+SmumTum=AATUm×mTSm×mUm×m=AAT
该式成立,原式也成立
备注:
1、matlab中可以通过svd计算EOF,但是不能重构出原始数据
2、每个主成分的解释率和特征值成正比,见Wilks书