模糊度固定LAMBDA算法详解

最小二乘搜索法

模糊度固定问题可以用一下方式表达：
$$\check{a}=argmin(a-\hat{a}{)}^T{Q}_{\hat{a}}^{-1}(a-\hat{a})\text{\hspace{1em}(1)}$$
若$Q_{\hat{a}}$为对角阵(即模糊度参数之间不相关)，直接对模糊度进行取整便可。因此必须对模糊度进行搜索。直接对模糊度进行搜索，其搜索空间十分狭长，一般采用降相关，再进行搜索。

LAMBDA搜索法采用Z变换的降相关方法：

$$\begin{gathered} z=Z^{T}a \\ \hat{z}=Z^T\hat{a} \\ {Q}_{\hat{z}}=Z^T{Q}_{\hat{a}}Z\text{\hspace{1em}(2)} \end{gathered}$$
因此新的模糊度固定问题可转化为下列式子：
$$\check{z}=argmin(z-\hat{z}{)}^T{Q}_{\hat{z}}^{-1}(z-\hat{z})\text{\hspace{1em}(3)}$$

1、LDL分解

$$Q_{\hat{a}}={\bar{L}}^T\bar{D}\bar{L}\text{\hspace{1em}(4)}$$
$$Q_{\hat{z}}=L^{T}DL\text{\hspace{1em}(5)}$$
$L$矩阵对应的元素为$l$，$\bar{L}$矩阵对应的元素为$\bar{l}$，均为下三角矩阵

将(5)带入(3)可重写为：

$$F(z)=(z-\hat{z}{)}^T{L}^{-T}{D}^{-1}{L}^{-1}(z-\hat{z})\text{\hspace{1em}(6)}$$

2、高斯变换

通过LDL分解的下三角矩阵$L$中的元素并不为零，说明模糊度参数之间存在相关性，给定下三角矩阵$L$，找出幺模矩阵$Z$使得$LZ$尽可能对角化，从而降低模糊度之间的相关性，使得通过$Z$变换的模糊度相关性大大降低.

对$L$下三角的元素$l_{i,j}$,当 $|l_{i,j}|>0.5(j>i)$ 时，对应的高斯变换矩阵为：
$$Z_{i,j}=I_n-[l_{i,j}]\ast e_i{e}_j^T\text{\hspace{1em}(7)}$$
其中，$I_n$为$n$维单位向量，$[\cdot ]$为取整算子，$e_i$和$e_j$分别为$i,j$对应的单位向量:
e.g.
$$e_i=\left[\begin{array}{ccccccc}0& 0& \cdot \cdot \cdot & 1& 0& \cdot \cdot \cdot & 0\end{array}\right]$$
1在第$i$位。

$${\bar{l}}_{k,j}=l_{k,j}-[l_{i,j}]l_{k,i}(k=i,i+1,\cdot \cdot \cdot ,n)\text{\hspace{1em}(8)}$$
举例：

给定单位下三角矩阵$L$如下：

$$L=\left[\begin{array}{cccc}1& 0& 0& 0\\ l_{21}& 1& 0& 0\\ l_{31}& l_{32}& 1& 0\\ l_{41}& l_{42}& l_{43}& 1\end{array}\right]$$
通过右乘高斯变换矩阵的方式，将$L$尽可能对角化，其中[*]为取整操作，$\tilde{0}$表示接近零。
$$L{Z}_{21}=\left[\begin{array}{cccc}1& 0& 0& 0\\ l_{21}& 1& 0& 0\\ l_{31}& l_{32}& 1& 0\\ l_{41}& l_{42}& l_{43}& 1\end{array}\right]\left[\begin{array}{cccc}1& 0& 0& 0\\ -[l_{21}]& 1& 0& 0\\ 0& 0& 1& 0\\ 0& 0& 0& 1\end{array}\right]=\left[\begin{array}{cccc}1& 0& 0& 0\\ l_{21}-[l_{21}]\ast 1& 1& 0& 0\\ l_{31}-[l_{21}]\ast l_{32}& l_{32}& 1& 0\\ l_{41}-[l_{21}]\ast l_{42}& l_{42}& l_{43}& 1\end{array}\right]$$
$$L{Z}_{21}{Z}_{31}=\left[\begin{array}{cccc}1& 0& 0& 0\\ {\tilde{0}}_{21}& 1& 0& 0\\ l_{31}^{(2)}& l_{32}^{(2)}& 1& 0\\ l_{41}^{(2)}& l_{42}^{(2)}& l_{43}^{(2)}& 1\end{array}\right]\left[\begin{array}{cccc}1& 0& 0& 0\\ 0& 1& 0& 0\\ -[l_{31}^{(2)}]& 0& 1& 0\\ 0& 0& 0& 1\end{array}\right]=\left[\begin{array}{cccc}1& 0& 0& 0\\ {\tilde{0}}_{21}& 1& 0& 0\\ l_{31}^{(2)}-[l_{31}^{(2)}]\ast 1& l_{32}^{(2)}& 1& 0\\ l_{41}^{(2)}-[l_{31}^{(2)}]\ast l_{43}^{(2)}& l_{42}^{(2)}& l_{43}^{(2)}& 1\end{array}\right]$$
$$...$$
$$LZ=\left[\begin{array}{cccc}1& 0& 0& 0\\ {\tilde{0}}_{21}& 1& 0& 0\\ {\tilde{0}}_{31}& {\tilde{0}}_{32}& 1& 0\\ {\tilde{0}}_{41}& {\tilde{0}}_{42}& {\tilde{0}}_{43}& 1\end{array}\right],Z=Z_{21}{Z}_{31}{Z}_{32}{Z}_{41}{Z}_{42}{Z}_{43}=\left[\begin{array}{cccc}1& 0& 0& 0\\ -[l_{21}]& 1& 0& 0\\ -[l_{31}^{(2)}]& -[l_{32}^{(3)}]& 1& 0\\ -[l_{41}^{(4)}]& -[l_{42}^{(5)}]& -[l_{43}^{(6)}]& 1\end{array}\right]$$
码表示为：

输入：下三角矩阵L, 矩阵维度n
输出：变换矩阵Z
初始化：
    check(n >= 2)
    check(L ? unit lowr triangular)
开始：
    for row = 2 : n
        for col= 1 : row - 1
            u = [L(row, col)]
            if (u != 0)
                L(row : n, col) = L(row : n, col) - u * L(row: n, row)
                Z(row, col) = -u
            end
        end
    end

c代码如下

int8_t gauss_transform(matrix_t *L, matrix_t *Z)
{
    if (!lower_matrix_check(L))
    {
        return -1;
    }

    if (L->col <= 2)
    {
        return -1;
    }
    uint32_t col = 0;
    uint32_t row = 0;
    for (row = 1; row < L->row; ++row)
    {
        for (col = 0; col < row; ++col)
        {
            fp64 u = get_integer(L->element[row][col]);
            if (!IS_ZEROS(u))
            {
                Z->element[row][col] = -u;
                uint32_t i = 0;
                for (i = row; i < L->row; ++i)
                {
                    L->element[i][col] -= u * L->element[i][row];
                }
            }
        }
    }

    return 1;

}

3、条件方差排序

满足条件：

$$d_i+{\bar{l}}_{i+1,i}^2{d}_{i+1}⩽d_{i+1}\text{\hspace{1em}(9)}$$
详情后续不上

4、确定模糊度的搜索空间

模糊度的搜索空间无法根据$Q_{\hat{a}}$直接得到，所以我们定义：

$$z-\tilde{z}=L(z-\hat{z})\text{\hspace{1em}(10)}$$
对(10)展开可得：

$$\left[\begin{array}{c}{z}_1-{\tilde{z}}_1\\ ...\\ z_1-{\tilde{z}}_1\end{array}\right]=\left[\begin{array}{cccc}1& 0& 0& ...\\ l_{21}& 1& {\cdot }_{{\cdot }_{\cdot }}& \\ ...& 0& l_{n(n-1)}& 1\end{array}\right]\left(\left[\begin{array}{c}{z}_1\\ ...\\ z_n\end{array}\right]-\left[\begin{array}{c}{\hat{z}}_1\\ ...\\ {\hat{z}}_n\end{array}\right]\right)$$
因此(6)可以更新为：

$$F(z)=(z-\tilde{z}{)}^{T}D(z-\tilde{z})=d_{11}(z_1-{\tilde{z}}_1)+\cdot \cdot \cdot +d_{nn}(z_n-{\tilde{z}}_n)\le {\chi }^2\text{\hspace{1em}(11)}$$
从(11)中可以得到$z-\tilde{z}$的搜索空间：

$$\{\begin{array}{rl}|z_1-{\tilde{z}}_1|& =\sqrt{\frac{{\chi }^2}{d_{11}}}\\ |z_2-{\tilde{z}}_2|& =\sqrt{\frac{{\chi }^2-(z_1-{\tilde{z}}_1{)}^2{d}_{11}}{d_{22}}}\\ & \cdot \cdot \cdot \\ |z_n-{\tilde{z}}_n|& =\sqrt{\frac{{\chi }^2-[(z_1-{\tilde{z}}_1{)}^2{d}_{11}+\cdot \cdot \cdot +(z_{n-1}+{\tilde{z}}_{n-1}{)}^2{d}_{(n-1)(n-1)}]}{d_{nn}}}\end{array}\text{\hspace{1em}(12)}$$
将(10)代入(12)中：

$$\{\begin{array}{rl}& |z_1-{\hat{z}}_1|=\sqrt{\frac{{\chi }^2}{d_{11}}}\\ & |l_{21}(z_1-{\hat{z}}_1)+(z_2-{\hat{z}}_2)|=\sqrt{\frac{{\chi }^2-(z_1-{\hat{z}}_1)d_{11}}{d_{22}}}\\ & \cdot \cdot \cdot \\ & |l_{n1}(z_1-{\hat{z}}_1)+\cdot \cdot \cdot +l_{n(n-1)}(z_{n-1}-{\hat{z}}_{n-1})+(z_n-{\hat{z}}_n)|=\sqrt{\frac{{\chi }^2-[(z_1-{\hat{z}}_1{)}^2{d}_{11}+(l_{21}(z_1-{\hat{z}}_1)+(z_2{\hat{z}}_2){)}^2{d}_{22}+\cdot \cdot \cdot ]}{d_{nn}}}\end{array}\text{\hspace{1em}(13)}$$
为了方便表达，我们有如下定义：

$$\{\begin{array}{rl}{\bar{z}}_1& ={\hat{z}}_1\\ {\bar{z}}_2& ={\hat{z}}_2-l_{21}(z_1-{\hat{z}}_1)\\ {\bar{z}}_3& ={\hat{z}}_3-l_{31}(z_1-{\hat{z}}_1)-l_{32}(z_2-{\hat{z}}_2)\\ & \cdot \cdot \cdot \\ {\bar{z}}_n& ={\hat{z}}_n-l_{n1}(z_1-{\hat{z}}_1)-l_{n2}(z_2-{\hat{z}}_2)-\cdot \cdot \cdot -l_{n(n-1)}(z_{n-1}-\widehat{z_{n-1}})\end{array}\text{\hspace{1em}(14)}$$
将(14)代入(13),则有：

$$\{\begin{array}{rl}& |z_1-{\bar{z}}_1|=\sqrt{\frac{{\chi }^2}{d_{11}}}\\ & |(z_2-{\bar{z}}_2)|=\sqrt{\frac{{\chi }^2-(z_1-{\bar{z}}_1)d_{11}}{d_{22}}}\\ & \cdot \cdot \cdot \\ & |(z_n-{\bar{z}}_n)|=\sqrt{\frac{{\chi }^2-[(z_1-{\bar{z}}_1{)}^2{d}_{11}+(z_2-{\bar{z}}_2{)}^2{d}_{22}+\cdot \cdot \cdot +(z_{n-1}-{\bar{z}}_{n-1}{)}^2{d}_{(n-1)(n-1)}]}{d_{nn}}}\end{array}\text{\hspace{1em}(15)}$$
因此可以确定$z$的搜索空间：

$$\{\begin{array}{rl}{\bar{z}}_1-\sqrt{\frac{{\chi }^2}{d_{11}}}& \le z_1\le {\bar{z}}_1+\sqrt{\frac{{\chi }^2}{d_{11}}}\\ {\bar{z}}_2-\sqrt{\frac{{\chi }^2-(z_1-{\bar{z}}_1)d_{11}}{d_{22}}}& \le z_2\le {\bar{z}}_2+\sqrt{\frac{{\chi }^2-(z_1-{\bar{z}}_1)d_{11}}{d_{22}}}\\ & \cdot \cdot \cdot \\ {\bar{z}}_n-\sqrt{\frac{{\chi }^2-[(z_1-{\bar{z}}_1{)}^2{d}_{11}+(z_2-{\bar{z}}_2{)}^2{d}_{22}+\cdot \cdot \cdot +(z_{n-1}-{\bar{z}}_{n-1}{)}^2{d}_{(n-1)(n-1)}]}{d_{nn}}}& \le z_n\le {\bar{z}}_n+\sqrt{\frac{{\chi }^2-[(z_1-{\bar{z}}_1{)}^2{d}_{11}+(z_2-{\bar{z}}_2{)}^2{d}_{22}+\cdot \cdot \cdot +(z_{n-1}-{\bar{z}}_{n-1}{)}^2{d}_{(n-1)(n-1)}]}{d_{nn}}}\end{array}\text{\hspace{1em}(16)}$$

5、恢复模糊度

通过(16)依次搜索的到最优的变换后的模糊度$z$后，根据(2)可得到变换前的模糊度参数$a$的固定解：

$$a=Z^{-T}z\text{\hspace{1em}(17)}$$