Non-linear Optimization
Least square method
1. Generalized pseudo-inverse
Simply Define :
A
x
=
b
,
A
∈
R
m
×
n
Ax=b, A \in \mathbb{R}^{m\times n}
Ax=b,A∈Rm×n
The solution can be generally defined as :
x
=
A
+
b
x=A^+b
x=A+b
Assume
r
a
n
k
(
A
)
=
r
rank(A) = r
rank(A)=r, then deploy BC decomposition method:
A
m
×
n
=
B
m
×
r
C
r
×
n
A_{m\times n} = B_{m\times r}C_{r\times n}
Am×n=Bm×rCr×n
The generalized pseudo-inverse of
A
A
A is:
A
+
=
C
T
(
C
C
T
)
−
1
(
B
T
B
)
−
1
B
T
A^{+} = C^T(CC^T)^{-1}(B^TB)^{-1}B^T
A+=CT(CCT)−1(BTB)−1BT
If
r
=
m
r = m
r=m(row full rank)
A
=
E
A
A = EA
A=EA
If r = n r = n r=n(column full rank)
A
=
A
E
A=AE
A=AE
Then, we can deduce that if matrix
A
A
A is row full rank, the generalized pseudo-inverse of
A
A
A is:
A
+
=
A
T
(
A
A
T
)
−
1
A^+=A ^T(AA^ T)^ {−1}
A+=AT(AAT)−1
If
A
A
A is column full rank,
A
+
A^+
A+ is:
A
+
=
(
A
T
A
)
−
1
A
T
A^ + =(A ^T A) ^{−1} A^ T
A+=(ATA)−1AT
2. Singular value decomposition (SVD)
Define:
A
=
U
Σ
V
T
,
A
∈
R
m
×
n
A=U\Sigma V^T, A \in \mathbb{R}^{m\times n}
A=UΣVT,A∈Rm×n
The least square problem is:
min
∥
A
x
−
b
∥
2
2
\min \| Ax-b \| ^2 _2
min∥Ax−b∥22
When
m
>
n
=
r
a
n
k
(
A
)
m>n=rank(A)
m>n=rank(A), it is overdetermined equations:
A
=
(
U
1
,
U
2
)
(
Σ
1
0
)
V
T
A = (U_1,U_2) (\begin{matrix} \Sigma_1 \\ 0 \end{matrix}) V^T
A=(U1,U2)(Σ10)VT
U
1
∈
R
m
×
r
U_1 \in \mathbb{R}^{m\times r}
U1∈Rm×r ,
x
x
x can be described as:
x = V Σ 1 − 1 U 1 T b x = V\Sigma_1^{-1}U_1^T\bm b x=VΣ1−1U1Tb
When
n
>
m
=
r
a
n
k
(
A
)
n>m=rank(A)
n>m=rank(A), then:
A
=
U
T
(
Σ
1
0
)
(
V
1
T
V
2
T
)
A = U^T (\begin{matrix} \Sigma_1 & 0 \end{matrix}) (\begin{matrix} V_1^T \\ V_2^T \end{matrix})
A=UT(Σ10)(V1TV2T)
V
1
T
∈
R
r
×
n
V_1^T \in \mathbb{R}^{r\times n}
V1T∈Rr×n, thus:
x
=
V
1
Σ
1
−
1
U
T
b
x = V_1\Sigma_1^{-1}U^T\bm b
x=V1Σ1−1UTb