常用表达式的矩阵表示

$X=(x_1,x_2,...,x_m) \in R^ {d\times m}$,$Y=(y_1,y_2,...,y_l)\in R^{d\times l}$,$l \leq m$. $x_i \leftrightarrow y_i$,$i=1,\ldots ,l$. $I$为${l\times l}$单位矩阵。

已知对应点信息的保持：
$\sum_{i=1}^l(x_i-y_i)^2=tr({\left( \begin{matrix}X& Y\\ \end{matrix} \right)}_{d\times(m+l)}{\left( \begin{array}{cc|c}I & 0 & -I \\ 0&0&0\\ \hdashline -I&0&I \end{array} \right) }_{(m+l)\times(m+l)}{\left( \begin{matrix}X^T\\Y^T \end{matrix} \right)}_{(m+l)\times d})$

其中 $\begin{bmatrix}I&0\\0&0\end{bmatrix}$为$m\times m$矩阵，$\begin{bmatrix}I\\0\end{bmatrix}$为$m\times l$矩阵，$\begin{bmatrix}I&0\end{bmatrix}$为$l\times m$矩阵

或者$\sum_{i=1}^l(x_i-y_i)^2=tr({\left( \begin{matrix}X{_{1,...,l}}& Y\\ \end{matrix} \right)}_{d\times(l+l)}{\left( \begin{array}{c|c}I& -I \\ \hdashline -I&I \end{array} \right) }_{(l+l)\times(l+l)}{\left( \begin{matrix}X^T{_{1,...,l}}\\Y^T \end{matrix} \right)}_{(l+l)\times d})$

非对应点尽量远离,第$i$个点$x_i$应与除其对应点$y_i$以外的所有点$y_j$的距离之和尽可能大：

$\sum_{i=1}^l\sum_{j\neq i}(x_i-y_j)^2=\sum_{i=1}^l\sum_{j=1}^l(x_i-y_j)^2-\sum_{i=1}^l(x_i-y_i)^2$

其中$\sum_{i=1}^l\sum_{j=1}^l(x_i-y_j)^2$=$\sum_{i=1}^m\sum_{j=1}^l(x_i-y_j)^2W_{X,Y}^{ij}$ （其中$W_{X,Y}^{ij}=1,i=1,...,l，W_{X,Y}^{ij}=0,i=l+1,...,m$

=trace($\left( \begin{matrix}X&Y\end{matrix} \right)\left( \begin{matrix}\Omega_1&-W_{X,Y}\\-W_{Y,X}&\Omega_2\\ \end{matrix} \right)\left( \begin{matrix}X^T\\Y^T\end{matrix} \right)$)
其中$\Omega_1\in R^{m\times m}$对角矩阵，$\Omega_1^{ii}=\sum{_j}W_{X,Y}^{ij}=\{\begin{matrix}l,i=1,...,l\\0,i=l+1,...,m\end{matrix}$，$\Omega_1=\begin{bmatrix}lI{_{l\times l}}&0\\0&0\end{bmatrix}$
$\Omega_2\in R^{l\times l}$对角矩阵，$\Omega_2^{ii}=\sum{_j}W_{X,Y}^{ij}=l$，$\Omega_2=lI{_{l\times l}}$
所以，原式=trace($\left( \begin{matrix}X&Y\end{matrix} \right)\left( \begin{array}{cc|c}lI_l&0&-E_l \\ 0&0&0\\ \hdashline -E_l&0&lI_l \end{array} \right)\left( \begin{matrix}X^T\\Y^T\end{matrix} \right)$)，其中$E_l$为$l\times l$的元素全为1的矩阵。

保持流形$X$的局部结构：$\sum_{i=1}^m\sum_{j=1}^m(x_i-x_j)^2W_X^{ij}$=2trace($XL_XX^T$)

$L_X=D_X-W_X$，$D_X^{ii}=\sum{_j}W_X^{ij}$

两个流形$X_1\in R^{d\times m_1}$，$X_2\in R^{d\times m_2}$间的对应关系的保持：

$\sum_{i=1}^{m_1}\sum_{j=1}^{m_2}(x_1^i-x_2^j)^2W_{1,2}^{i,j}$

=trace($\left( \begin{matrix}X_1&X_2\\ \end{matrix} \right)\left( \begin{matrix}\Omega_1&-W_{1,2}\\-W_{2,1}&\Omega_2\\ \end{matrix} \right)\left( \begin{matrix}X_1^T\\X_2^T\end{matrix} \right)$)

其中$\Omega_1\in R^{m_1\times m_1}$对角矩阵，$\Omega_1^{ii}=\sum{_j}W_{1,2}^{ij}$
$\Omega_2\in R^{m_2\times m_2}$对角矩阵，$\Omega_2^{ii}=\sum{_j}W_{2,1}^{ij}$

保持流形$X_1$，$X_2$的局部结构，并保持流形$X_1$与$X_2$之间的对应关系：

$\frac{1}{2}\mu_2\sum_{i=1}^{m_1}\sum_{j=1}^{m_1}(x_1^i-x_1^j)^2W_1^{ij}$+$\frac{1}{2}\mu_2\sum_{i=1}^{m_2}\sum_{j=1}^{m_2}(x_2^i-x_2^j)^2W_2^{ij}$+$\mu_1\sum_{i=1}^{m_1}\sum_{j=1}^{m_2}(x_1^i-x_2^j)^2W_{1,2}^{i,j}$

=$\mu_2$trace($X_1L_1X_1^T$)+$\mu_2$trace($X_2L_2X_2^T$)+$\mu_1\sum_{i=1}^{m_1}\sum_{j=1}^{m_2}(x_1^i-x_2^j)^2W_{1,2}^{i,j}$
=$\mu_2$trace($\left( \begin{matrix}X_1&X_2\\ \end{matrix} \right)\left( \begin{matrix}L_1&0\\0&0 \end{matrix} \right)\left( \begin{matrix}X_1^T\\X_2^T\\ \end{matrix} \right)$)
+$\mu_2$trace($\left( \begin{matrix}X_1&X_2 \end{matrix} \right)\left( \begin{matrix}0&0\\0&L_2 \end{matrix} \right)\left( \begin{matrix}X_1^T\\X_2^T\\ \end{matrix} \right)$)
+$\mu_1$trace($\left( \begin{matrix}X_1&X_2\\ \end{matrix} \right)\left( \begin{matrix}\Omega_1&-W_{1,2}\\-W_{2,1}&\Omega_2\\ \end{matrix} \right)\left( \begin{matrix}X_1^T\\X_2^T\\ \end{matrix} \right)$)

=trace($\begin{bmatrix}X_1&X_2\end{bmatrix} \begin{bmatrix}\mu_1\Omega_1+\mu_2L_1 & -\mu_1W_{1,2}\\-\mu_1W_{2,1}&\mu_1\Omega_2+\mu_2L_2\\ \end{bmatrix} \begin{bmatrix}X_1^T\\X_2^T\end{bmatrix}$ )

假设有两个流形$X_1\in R^{d\times m_1}$，$X_2\in R^{d\times m_2}$：

$\sum_{i=1}^{m_1}\sum_{j=1}^{m_2}(x_1^i-x_2^j)^2$

=$\sum_{i=1}^{m_1}\sum_{j=1}^{m_2}(x_1^i-x_2^j)^2W_{1,2}^{ij}$ (其中$W_{1,2}^{ij}=1$,$W_{1,2}$为$m_1\times m_2$矩阵，其每个元素都为1)

=trace($\left( \begin{matrix}X_1&X_2\\ \end{matrix} \right)\left( \begin{matrix}\Omega_1&-W_{1,2}\\-W_{2,1}&\Omega_2\\ \end{matrix} \right)\left( \begin{matrix}X_1^T\\X_2^T\end{matrix} \right)$)

其中$\Omega_1\in R^{m_1\times m_1}$对角矩阵，$\Omega_1^{ii}=\sum{_j}W_{1,2}^{ij}=m_2$，$\Omega_1=m_2I{_{m_1\times m_1}}$
$\Omega_2\in R^{m_2\times m_2}$对角矩阵，$\Omega_2^{ii}=\sum{_j}W_{2,1}^{ij}=m_1$，$\Omega_2=m_1I{_{m_2\times m_2}}$

$x_i$是$p\times 1$列向量
$||\sum_{i=1}^nx_i||^2=(\sum_{i=1}^nx_i)^T(\sum_{i=1}^nx_i)=\sum_{i=1}^n\sum_{j=1}^nx_i^Tx_j=tr( XIX^T)$，其中$I$是每个元素都为1的$n\times n$矩阵，$X=[x_1,x_2,...,x_n]$

$x_i$是$p\times 1$列向量，$y_j$是$p\times 1$列向量
$(\sum_{i=1}^nx_i)^T(\sum_{j=1}^my_j)=\sum_{i=1}^n\sum_{j=1}^mx_i^Ty_j=tr(XIY^T)$，其中$I$是每个元素都为1的$n\times m$矩阵，$X=[x_1,x_2,...,x_n]$，$Y=[y_1,y_2,...,y_m]$