求线性预测器的最佳预测系数
我们要在最小
E
[
d
2
]
E[d^2]
E[d2]的条件下,确定一组最佳预测系数(不考虑量化误差)。
E
[
d
2
]
=
E
[
(
S
k
−
S
e
(
k
)
)
2
]
=
E
[
(
S
k
−
∑
i
=
1
N
a
i
S
(
k
−
i
)
)
2
]
(1)
E[d^2]=E\left[ (S_k-S_e(k))^2 \right]=E\left[ \left(S_k-\sum_{i=1}^N a_iS(k-i)\right)^2 \right] \tag{1}
E[d2]=E[(Sk−Se(k))2]=E⎣⎡(Sk−i=1∑NaiS(k−i))2⎦⎤(1)
而要求的最佳预测系数
a
i
a_i
ai满足:
∂
E
[
d
2
]
∂
a
i
=
0
,
i
=
1
,
2
,
⋯
,
N
(2)
\dfrac {\partial E[d^2]}{\partial a_i} = 0,\quad i=1,2,\cdots,N \tag{2}
∂ai∂E[d2]=0,i=1,2,⋯,N(2)
对
i
=
1
,
2
,
⋯
,
N
i=1,2,\cdots,N
i=1,2,⋯,N,代入
(
1
)
(1)
(1)式得:
∂
E
[
(
S
k
−
∑
j
=
1
N
a
j
S
k
−
j
)
2
]
∂
a
i
=
E
[
∂
(
S
k
−
∑
j
=
1
N
a
j
S
k
−
j
)
2
∂
a
i
]
=
E
[
2
(
S
k
−
∑
j
=
1
N
a
j
S
k
−
j
)
(
−
S
k
−
i
)
]
=
E
[
−
2
S
k
S
k
−
i
+
2
∑
j
=
1
N
a
j
S
k
−
j
S
k
−
i
]
=
0
(3)
\begin{aligned} \dfrac {\partial E\left[ \left(S_k-\sum_{j=1}^N a_j S_{k-j}\right)^2 \right]}{\partial a_i} &= E\left[\dfrac {\partial \left(S_k-\sum_{j=1}^N a_j S_{k-j}\right)^2 }{\partial a_i}\right]\\ &= E\left[ 2\left(S_k-\sum_{j=1}^N a_j S_{k-j}\right)\left(-S_{k-i}\right) \right] \\ &= E\left[ -2S_k S_{k-i} + 2\sum_{j=1}^N a_j S_{k-j} S_{k-i}\right] = 0 \end{aligned} \tag{3}
∂ai∂E[(Sk−∑j=1NajSk−j)2]=E⎣⎢⎡∂ai∂(Sk−∑j=1NajSk−j)2⎦⎥⎤=E[2(Sk−j=1∑NajSk−j)(−Sk−i)]=E[−2SkSk−i+2j=1∑NajSk−jSk−i]=0(3)
故
E
[
S
k
S
k
−
i
]
=
E
[
∑
j
=
1
N
a
j
S
k
−
j
S
k
−
i
]
R
(
i
)
=
∑
j
=
1
N
a
j
R
(
i
−
j
)
(4)
\begin{aligned} E\left[S_k S_{k-i} \right] &= E\left[\sum_{j=1}^N a_j S_{k-j} S_{k-i}\right] \\ R(i) &= \sum_{j=1}^N a_j R\left( i-j \right) \end{aligned} \tag{4}
E[SkSk−i]R(i)=E[j=1∑NajSk−jSk−i]=j=1∑NajR(i−j)(4)
其中
R
(
k
)
=
E
[
x
n
x
n
+
k
]
=
R
(
−
k
)
R(k) = E\left[ x_n x_{n+k} \right] = R(-k)
R(k)=E[xnxn+k]=R(−k)为自相关函数。
对
i
=
1
,
2
,
⋯
,
N
i=1,2,\cdots,N
i=1,2,⋯,N,将
(
4
)
(4)
(4)式展开:
R
(
1
)
=
a
1
R
(
0
)
+
a
2
R
(
1
)
+
⋯
+
a
N
R
(
N
−
1
)
R
(
2
)
=
a
1
R
(
1
)
+
a
2
R
(
2
)
+
⋯
+
a
N
R
(
N
−
2
)
⋮
R
(
N
)
=
a
1
R
(
N
−
1
)
+
a
2
R
(
N
−
2
)
+
⋯
+
a
N
R
(
0
)
(5)
\begin{aligned} R(1) &= a_1 R(0) + a_2 R(1) + \cdots +a_N R(N-1) \\ R(2) &= a_1 R(1) + a_2 R(2) + \cdots +a_N R(N-2) \\ \vdots \\ R(N) &= a_1 R(N-1) + a_2 R(N-2) + \cdots +a_N R(0) \\ \end{aligned} \tag{5}
R(1)R(2)⋮R(N)=a1R(0)+a2R(1)+⋯+aNR(N−1)=a1R(1)+a2R(2)+⋯+aNR(N−2)=a1R(N−1)+a2R(N−2)+⋯+aNR(0)(5)
写为矩阵形式即为
R
A
=
P
(6)
\bold R\bold A = \bold P \tag{6}
RA=P(6)
其中
R
=
[
R
(
0
)
R
(
1
)
⋯
R
(
N
−
1
)
R
(
1
)
R
(
2
)
⋯
R
(
N
−
2
)
⋮
⋮
⋱
⋮
R
(
N
−
1
)
R
(
N
−
2
)
⋯
R
(
0
)
]
A
=
[
a
1
a
2
⋮
a
N
]
P
=
[
R
(
1
)
R
(
2
)
⋮
R
(
N
)
]
\begin{aligned} \bold R &= \begin{bmatrix} R(0) & R(1) & \cdots & R(N-1) \\ R(1) & R(2) & \cdots & R(N-2) \\ \vdots & \vdots & \ddots & \vdots \\ R(N-1) & R(N-2) & \cdots & R(0) \\ \end{bmatrix}\\ \bold A &= \begin{bmatrix} a_1 \\ a_2 \\ \vdots \\ a_N \end{bmatrix}\\ \bold P &= \begin{bmatrix} R(1) \\ R(2) \\ \vdots \\ R(N) \end{bmatrix} \end{aligned}
RAP=⎣⎢⎢⎢⎡R(0)R(1)⋮R(N−1)R(1)R(2)⋮R(N−2)⋯⋯⋱⋯R(N−1)R(N−2)⋮R(0)⎦⎥⎥⎥⎤=⎣⎢⎢⎢⎡a1a2⋮aN⎦⎥⎥⎥⎤=⎣⎢⎢⎢⎡R(1)R(2)⋮R(N)⎦⎥⎥⎥⎤
故最佳系数预测向量
A
=
R
−
1
P
(7)
\bold A = \bold R^{-1}\bold P \tag{7}
A=R−1P(7)