《Matrix Methods in Data Analysis, Signal Processing, and Machine Learning》MIT Course Number:18.065课程习题答案
LEC6. Singular Value Decomposition(SVD)
Problems of Lecture 6 (from textbook Section I.8)
1 A symmetric matrix
S
=
S
T
S = S^T
S=ST has orthonormal eigenvectors
v
1
v_1
v1 to
v
n
v_n
vn . Then any vector
x
x
x can be written as a combination
x
=
c
1
v
1
+
⋅
⋅
⋅
+
c
n
v
n
x = c_1 v_1 + ··· + c_n v_n
x=c1v1+⋅⋅⋅+cnvn . Explain these two formulas :
x
T
x
=
c
1
2
+
⋅
⋅
⋅
+
c
n
2
x
T
S
x
=
λ
1
c
1
2
+
⋅
⋅
⋅
+
λ
n
c
n
2
.
x^Tx = c^2_1 + ··· + c^2_n \ \ \ \ \ \ \ \ \ \ \ \ \ x^TSx = λ_1 c^2_1 + ··· + λ_n c^2_n .
xTx=c12+⋅⋅⋅+cn2 xTSx=λ1c12+⋅⋅⋅+λncn2.
6 Find the σ’s and v’s and u’s in the SVD for
A
=
[
3
4
0
5
]
.
A=\begin{bmatrix}3&4\\0&5\end{bmatrix}.
A=[3045].Use equation (12).
Solutions to Lecture 6
1
(
c
1
v
1
T
+
⋅
⋅
⋅
+
c
n
v
n
T
)
(
c
1
v
1
+
⋅
⋅
⋅
+
c
n
v
n
)
=
c
1
2
+
⋅
⋅
⋅
+
c
n
2
(c_1 v^T_1 +···+c_n v^T_n )(c_1 v_1 +···+c_n v_n ) = c^2_1 +···+c^2_n
(c1v1T+⋅⋅⋅+cnvnT)(c1v1+⋅⋅⋅+cnvn)=c12+⋅⋅⋅+cn2because the
v
’
s
v’s
v’s are orthonormal.
(
c
1
v
1
T
+
⋅
⋅
⋅
+
c
n
v
n
T
)
S
(
c
1
v
1
+
⋅
⋅
⋅
+
c
n
v
n
)
=
(
c
1
v
1
T
+
⋅
⋅
⋅
+
c
n
v
n
T
)
(
λ
1
c
1
v
1
+
⋅
⋅
⋅
+
λ
n
c
n
v
n
)
=
λ
1
c
1
2
+
⋅
⋅
⋅
+
λ
n
c
n
2
(c_1 v^T_1 +···+c_n v^T_n )S (c_1 v_1 +···+c_n v_n ) = (c_1 v^T_1 +···+c_n v^T_n )(λ_1c_1 v_1 +···+λ_nc_n v_n )=λ_1 c^2_1 + ··· + λ_n c^2_n
(c1v1T+⋅⋅⋅+cnvnT)S(c1v1+⋅⋅⋅+cnvn)=(c1v1T+⋅⋅⋅+cnvnT)(λ1c1v1+⋅⋅⋅+λncnvn)=λ1c12+⋅⋅⋅+λncn2
6 Since we already know the SVD of A = [ 3 0 4 5 ] , A=\begin{bmatrix}3&0\\4&5\end{bmatrix}, A=[3405],all we have to do is to exchange u ′ s u's u′s and v ′ s v's v′s while keeping σ = 45 a n d σ = 5 σ=\sqrt{45}\ and\ σ=\sqrt{5} σ=45 and σ=5.