VINS预积分推导2

协方差矩阵传播

由式子可以看出，预积分噪声是IMU高斯白噪声的线性组合，因此预积分噪声也满足零均值的高斯分布：
$${\eta }_{ij}=\left[\begin{array}{l}\delta {\varphi }_{ij}\\ \delta v_{ij}\\ \delta p_{ij}\end{array}\right]\tilde{\mathcal{N}}\left(0_{9\times 1},{\Sigma }_{ij}\right)$$
同时，IMU离散时间的高斯白噪声为：
$${\eta }_k^d=\left[\begin{array}{l}{\eta }_k^{gd}\\ {\eta }_k^{ad}\end{array}\right]\tilde{\mathcal{N}}\left(0_{6\times 1},{\Sigma }_k^d\right)$$
下面推导协方差矩阵的传播过程。对于姿态噪声项来说：
$$\begin{array}{l}\delta {\varphi }_{ij}=\sum _{k=i}^{j-1}{\tilde{R}}_{k+1,j}^T{J}_{r,k}{\eta }_k^{gd}\Delta t\\ =\sum _{k=i}^{j-2}{\tilde{R}}_{k+1,j}^T{J}_{r,k}{\eta }_k^{gd}\Delta t+J_{r,j-1}{\eta }_{j-1}^{gd}\Delta t\\ =\sum _{k=i}^{j-2}{\left({\tilde{R}}_{k+1,j-1}{\tilde{R}}_{j-1,j}\right)}^T{J}_{r,k}{\eta }_k^{gd}\Delta t+J_{r,j-1}{\eta }_{j-1}^{gd}\Delta t\\ ={\tilde{R}}_{j-1,j}^T\sum _{k=i}^{j-2}{\tilde{R}}_{k+1,j-1}^T{J}_{r,k}{\eta }_k^{gd}\Delta t+J_{r,j-1}{\eta }_{j-1}^{gd}\Delta t\\ ={\tilde{R}}_{j-1,j}^T\delta {\varphi }_{i,j-1}+J_{r,j-1}{\eta }_{j-1}^{gd}\Delta t\end{array}$$
对于速度噪声项来说：
$$\begin{array}{l}\delta v_{ij}=\sum _{k=i}^{j-1}{\tilde{R}}_{ik}{\eta }_k^{ad}\Delta t-{\tilde{R}}_{ik}{\left({\tilde{a}}_k-b_i^a\right)}^{\wedge }\delta {\varphi }_{ik}\Delta t\\ =\sum _{k=i}^{j-2}{\tilde{R}}_{ik}{\eta }_k^{ad}\Delta t-{\tilde{R}}_{ik}{\left({\tilde{a}}_k-b_i^a\right)}^{\wedge }\delta {\varphi }_{ik}\Delta t+{\tilde{R}}_{i,j-1}{\eta }_{j-1}^{ad}\Delta t-{\tilde{R}}_{i,j-1}{\left({\tilde{a}}_{j-1}-b_i^a\right)}^{\wedge }\delta {\varphi }_{i,j-1}\Delta t\\ =\delta v_{i,j-1}+{\tilde{R}}_{i,j-1}{\eta }_{j-1}^{ad}\Delta t-{\tilde{R}}_{i,j-1}{\left({\tilde{a}}_{j-1}-b_i^a\right)}^{\wedge }\delta {\varphi }_{i,j-1}\Delta t\end{array}$$
对于平移噪声项来说：
$$\begin{array}{l}\delta p_{ij}=\sum _{k=i}^{j-1}\delta v_{ik}\Delta t-\frac{1}{2}{\tilde{R}}_{ik}{\left({\tilde{a}}_k-b_i^a\right)}^{\wedge }\delta {\varphi }_{ik}\Delta t^2+\frac{1}{2}{\tilde{R}}_{ik}{\eta }_k^{ad}\Delta t^2\\ =\sum _{k=i}^{j-2}\delta v_{ik}\Delta t-\frac{1}{2}{\tilde{R}}_{ik}{\left({\tilde{a}}_k-b_i^a\right)}^{\wedge }\delta {\varphi }_{ik}\Delta t^2+\frac{1}{2}{\tilde{R}}_{ik}{\eta }_k^{ad}\Delta t^2+\delta v_{i,j-1}\Delta t-\frac{1}{2}{\tilde{R}}_{i,j-1}{\left({\tilde{a}}_k-b_i^a\right)}^{\wedge }\delta {\varphi }_{i,j-1}\Delta t^2+\frac{1}{2}{\tilde{R}}_{i,j-1}{\eta }_{j-1}^{ad}\Delta t^2\\ =\delta p_{i,j-1}+\delta v_{i,j-1}\Delta t-\frac{1}{2}{\tilde{R}}_{i,j-1}{\left({\tilde{a}}_k-b_i^a\right)}^{\wedge }\delta {\varphi }_{i,j-1}\Delta t^2+\frac{1}{2}{\tilde{R}}_{i,j-1}{\eta }_{j-1}^{ad}\Delta t^2\end{array}$$
根据上式，可以将噪声项的传播过程写成矩阵形式，从而得到预积分协方差矩阵的传播形式：
$$A_{j-1}=\left[\begin{array}{ccc}{\tilde{R}}_{j-1,j}^T& 0& 0\\ -{\tilde{R}}_{i,j-1}{\left({\tilde{a}}_{j-1}-b_i^a\right)}^{\wedge }\Delta t& I& 0\\ -\frac{1}{2}{\tilde{R}}_{i,j-1}{\left({\tilde{a}}_k-b_i^a\right)}^{\wedge }\Delta t^2& \Delta t& I\end{array}\right],B_{j-1}=\left[\begin{array}{cc}{J}_{r,j-1}\Delta t& 0\\ 0& {\tilde{R}}_{i,j-1}\Delta t\\ 0& \frac{1}{2}{\tilde{R}}_{i,j-1}\Delta t^2\end{array}\right]$$$${\eta }_{ij}=A_{j-1}{\eta }_{i,j-1}+B_{j-1}{\eta }_{j-1}^d$$$${\Sigma }_{ij}=A_{j-1}{\Sigma }_{i,j-1}{A}_{j-1}^T+B_{j-1}{\Sigma }_{j-1}^d{B}_{j-1}^T$$
初始时刻，预积分协方差为零矩阵，之后随着预积分的传播，预积分的协方差越来越大，不确定性增大。
所以预积分等式可写为：
$$\begin{array}{l}{R}_{ij}\exp \left(\delta {\varphi }_{ij}\right)={\tilde{R}}_{ij}\\ R_i^T\left(v_j-v_i-g\Delta t_{ij}\right)+\delta v_{ij}={\tilde{v}}_{ij}\\ R_i^T\left(p_j-p_i-v_i\Delta t_{ij}-\frac{1}{2}\Delta t_{ij}^2\right)+\delta p_{ij}={\tilde{p}}_{ij}\end{array}$$
当零偏发生变化时，预积分测量值也会发生相应的变化。设初始零偏为$\bar{b}$，当前零偏为$b$，零偏变化量为$\delta b$，零偏更新过程可写为$b\leftarrow \bar{b}+\delta b$，所以可以根据零偏变化量$\delta b$和一阶雅可比更新预积分量：
$$\begin{array}{l}{\tilde{R}}_{ij}\left(b_i^g\right)={\tilde{R}}_{ij}\left({\bar{b}}_i^g\right)\exp \left(\frac{\partial {\bar{R}}_{ij}}{\partial b^g}\delta b_i^g\right)\\ {\tilde{v}}_{ij}\left(b_i^g,b_i^a\right)={\tilde{v}}_{ij}\left({\bar{b}}_i^g,{\bar{b}}_i^a\right)+\frac{\partial {\bar{v}}_{ij}}{\partial b^g}\delta b_i^g+\frac{\partial {\bar{v}}_{ij}}{\partial b^g}\delta b_i^g\\ {\tilde{p}}_{ij}\left(b_i^g,b_i^a\right)={\tilde{p}}_{ij}\left({\bar{b}}_i^g,{\bar{b}}_i^a\right)+\frac{\partial {\bar{p}}_{ij}}{\partial b^g}\delta b_i^g+\frac{\partial {\bar{p}}_{ij}}{\partial b^g}\delta b_i^g\end{array}$$
在预积分传播过程中，零偏雅可比也会进行传播。将上式中的初始零偏和零偏变化量分离，可以得到：
$$\begin{array}{l}\tilde{R}{\left({\bar{b}}_i^g+\delta b_i^g\right)}_{ij}=\prod _{k=i}^{j-1}\exp \left({\tilde{w}}_k-\left({\bar{b}}_i^g+\delta b_i^g\right)\right)\Delta t\\ =\prod _{k=i}^{j-1}\exp \left(\left({\tilde{w}}_k-{\bar{b}}_i^g\right)\Delta t\right)\exp \left(-J_{r,k}\delta b_i^g\Delta t\right)\\ =\tilde{R}{\left({\bar{b}}_i^g\right)}_{i,i+1}\exp \left(-J_{r,i}\delta b_i^g\Delta t\right)\tilde{R}{\left({\bar{b}}_i^g\right)}_{i+1,i+2}\exp \left(-J_{r,i+1}\delta b_i^g\Delta t\right)\cdots \tilde{R}{\left({\bar{b}}_i^g\right)}_{j-1,j}\exp \left(-J_{r,j-1}\delta b_i^g\Delta t\right)\\ =\tilde{R}{\left({\bar{b}}_i^g\right)}_{ij}\prod _{k=i}^{j-1}\exp \left(-\tilde{R}{\left({\bar{b}}_i^g\right)}_{k+1,j}^T{J}_{r,k}\delta b_i^g\Delta t\right)\\ =\tilde{R}{\left({\bar{b}}_i^g\right)}_{ij}\exp \left(-\sum _{k=i}^{j-1}\tilde{R}{\left({\bar{b}}_i^g\right)}_{k+1,j}^T{J}_{r,k}\delta b_i^g\Delta t\right)\end{array}$$$$\begin{array}{l}\tilde{v}{\left({\bar{b}}_i^g+\delta b_i^g,{\bar{b}}_i^a+\delta b_i^a\right)}_{ij}=\sum _{k=i}^{j-1}\tilde{R}{\left({\bar{b}}_i^g+\delta b_i^g\right)}_{ik}\left({\tilde{a}}_k-{\bar{b}}_i^a-\delta b_i^a\right)\Delta t\\ =\sum _{k=i}^{j-1}\tilde{R}{\left({\bar{b}}_i^g\right)}_{ik}\exp \left(\frac{\partial {\bar{R}}_{ik}}{\partial b^g}\delta b_i^g\right)\left({\tilde{a}}_k-{\bar{b}}_i^a-\delta b_i^a\right)\Delta t\\ =\sum _{k=i}^{j-1}\tilde{R}{\left({\bar{b}}_i^g\right)}_{ik}\left(I+{\left(\frac{\partial {\bar{R}}_{ik}}{\partial b^g}\delta b_i^g\right)}^{\wedge }\right)\left({\tilde{a}}_k-{\bar{b}}_i^a-\delta b_i^a\right)\Delta t\\ =\tilde{v}{\left({\bar{b}}_i^g,{\bar{b}}_i^a\right)}_{ij}-\sum _{k=i}^{j-1}\tilde{R}{\left({\bar{b}}_i^g\right)}_{ik}\Delta t\delta b_i^a-\sum _{k=i}^{j-1}\tilde{R}{\left({\bar{b}}_i^g\right)}_{ik}{\left({\tilde{a}}_k-{\bar{b}}_i^a\right)}^{\wedge }\frac{\partial {\bar{R}}_{ik}}{\partial b^g}\Delta t\delta b_i^g\end{array}$$$$\begin{array}{l}\tilde{p}{\left({\bar{b}}_i^g+\delta b_i^g,{\bar{b}}_i^a+\delta b_i^a\right)}_{ij}=\sum _{k=i}^{j-1}\tilde{v}{\left({\bar{b}}_i^g+\delta b_i^g,{\bar{b}}_i^a+\delta b_i^a\right)}_{ij}\Delta t+\frac{1}{2}\tilde{R}{\left({\bar{b}}_i^g+\delta b_i^g\right)}_{ik}\left({\tilde{a}}_k-{\bar{b}}_i^a-\delta b_i^a\right)\Delta t^2\\ =\sum _{k=i}^{j-1}\left(\tilde{v}{\left({\bar{b}}_i^g,{\bar{b}}_i^a\right)}_{ik}+\frac{\partial {\bar{v}}_{ij}}{\partial b^g}\delta b_i^g+\frac{\partial {\bar{v}}_{ij}}{\partial b^a}\delta b_i^a\right)\Delta t+\frac{1}{2}\tilde{R}{\left({\bar{b}}_i^g\right)}_{ik}\left(I-{\left(\frac{\partial {\bar{R}}_{ik}}{\partial b^g}\delta b_i^g\right)}^{\wedge }\right)\left({\tilde{a}}_k-{\bar{b}}_i^a-\delta b_i^a\right)\Delta t^2\\ =\tilde{p}{\left({\bar{b}}_i^g,{\bar{b}}_i^a\right)}_{ij}+\sum _{k=i}^{j-1}\frac{\partial {\bar{v}}_{ij}}{\partial b^a}\Delta t-\frac{1}{2}\tilde{R}{\left({\bar{b}}_i^g\right)}_{ik}\Delta t^2\delta b_i^a+\sum _{k=i}^{j-1}\frac{\partial {\bar{v}}_{ij}}{\partial b^g}\Delta t-\frac{1}{2}\tilde{R}{\left({\bar{b}}_i^g\right)}_{ik}{\left({\tilde{a}}_k-{\bar{b}}_i^a\right)}^{\wedge }\frac{\partial {\bar{R}}_{ik}}{\partial b^g}\Delta t^2\delta b_i^g\end{array}$$
所以零偏雅可比矩阵为：
$$\begin{array}{l}\frac{\partial \tilde{R}{\left({\bar{b}}_i^g\right)}_{ij}}{\partial b^g}=-\sum _{k=i}^{j-1}\tilde{R}{\left({\bar{b}}_i^g\right)}_{k+1,j}^T{J}_{r,k}\Delta t\\ \frac{\partial \tilde{v}{\left({\bar{b}}_i^g,{\bar{b}}_i^a\right)}_{ij}}{\partial b^g}=-\sum _{k=i}^{j-1}\tilde{R}{\left({\bar{b}}_i^g\right)}_{ik}{\left({\tilde{a}}_k-{\bar{b}}_i^a\right)}^{\wedge }\frac{\partial {\bar{R}}_{ik}}{\partial b^g}\Delta t\delta b\\ \frac{\partial \tilde{v}{\left({\bar{b}}_i^g,{\bar{b}}_i^a\right)}_{ij}}{\partial b^a}=-\sum _{k=i}^{j-1}\tilde{R}{\left({\bar{b}}_i^g\right)}_{ik}\Delta t\\ \frac{\partial \tilde{p}{\left({\bar{b}}_i^g,{\bar{b}}_i^a\right)}_{ij}}{\partial b^a}=\sum _{k=i}^{j-1}\frac{\partial {\bar{v}}_{ij}}{\partial b^a}\Delta t-\frac{1}{2}\tilde{R}{\left({\bar{b}}_i^g\right)}_{ik}\Delta t^2\\ \frac{\partial \tilde{p}{\left({\bar{b}}_i^g,{\bar{b}}_i^a\right)}_{ij}}{\partial b^g}=\sum _{k=i}^{j-1}\frac{\partial {\bar{v}}_{ij}}{\partial b^g}\Delta t-\frac{1}{2}\tilde{R}{\left({\bar{b}}_i^g\right)}_{ik}{\left({\tilde{a}}_k-{\bar{b}}_i^a\right)}^{\wedge }\frac{\partial {\bar{R}}_{ik}}{\partial b^g}\Delta t^2\end{array}$$

预积分雅可比推导

$$\begin{array}{l}{e}_{R_{ij}}\left(R_i\exp \left({\varphi }_i\right)\right)=\log \left({\tilde{R}}_{ij}^T{\left(R_i\exp \left({\varphi }_i\right)\right)}^T{R}_j\right)\\ =\log \left({\tilde{R}}_{ij}^T\exp \left(-{\varphi }_i\right){R_i}^T{R}_j\right)\\ =\log \left({\tilde{R}}_{ij}^T{R_i}^T{R}_j\exp \left(-R_j^T{R}_i{\varphi }_i\right)\right)\\ =e_{R_{ij}}\left(R_i\right)-J_r^{-1}\left(e_{R_{ij}}\left(R_i\right)\right)R_j^T{R}_i{\varphi }_i\end{array}$$
$$\begin{array}{l}{e}_{R_{ij}}\left(R_j\exp \left({\varphi }_j\right)\right)=\log \left({\tilde{R}}_{ij}^T{R}_i^T{R}_j\exp \left({\varphi }_j\right)\right)\\ =e_{R_{ij}}\left(R_j\right)+J_r^{-1}\left(e_{R_{ij}}\left(R_j\right)\right){\varphi }_j\end{array}$$$$\begin{array}{l}{e}_{R_{ij}}\left(b_i^g+\delta b_i^g+\delta {\tilde{b}}_i^g\right)=\log \left({\left(\tilde{R}{\left({\bar{b}}_i^g\right)}_{ij}\exp \left(\frac{\partial \tilde{R}{\left({\bar{b}}_i^g\right)}_{ij}}{\partial b_i^g}\left(\delta b_i^g+\delta {\tilde{b}}_i^g\right)\right)\right)}^T{R_i}^T{R}_j\right)\\ =\log \left({\left(\tilde{R}{\left({\bar{b}}_i^g\right)}_{ij}\exp \left(\frac{\partial \tilde{R}{\left({\bar{b}}_i^g\right)}_{ij}}{\partial b_i^g}\delta b_i^g\right)\exp \left(J_{r,b}\frac{\partial \tilde{R}{\left({\bar{b}}_i^g\right)}_{ij}}{\partial b_i^g}\delta {\tilde{b}}_i^g\right)\right)}^T{R_i}^T{R}_j\right)\\ =\log \exp \left(-J_{r,b}\frac{\partial \tilde{R}{\left({\bar{b}}_i^g\right)}_{ij}}{\partial b_i^g}\delta {\tilde{b}}_i^g\right)\tilde{R}{\left(b_i^g\right)}_{ij}{R_i}^T{R}_j\\ =\log \exp \left(-J_{r,b}\frac{\partial \tilde{R}{\left({\bar{b}}_i^g\right)}_{ij}}{\partial b_i^g}\delta {\tilde{b}}_i^g\right)\exp \left(e_{R_{ij}}\left(b_i^g+\delta b_i^g\right)\right)\\ =\log \exp \left(e_{R_{ij}}\left(b_i^g+\delta b_i^g\right)\right)\exp \left(-\exp {\left(e_{R_{ij}}\left(b_i^g+\delta b_i^g\right)\right)}^T{J}_{r,b}\frac{\partial \tilde{R}{\left({\bar{b}}_i^g\right)}_{ij}}{\partial b_i^g}\delta {\tilde{b}}_i^g\right)\\ =e_{R_{ij}}\left(b_i^g+\delta b_i^g\right)-J_r^{-1}\left(e_{R_{ij}}\left(b_i^g+\delta b_i^g\right)\right)\exp {\left(e_{R_{ij}}\left(b_i^g+\delta b_i^g\right)\right)}^T{J}_{r,b}\frac{\partial \tilde{R}{\left({\bar{b}}_i^g\right)}_{ij}}{\partial b_i^g}\delta {\tilde{b}}_i^g\end{array}$$$$J_r^{-1}\left(e_{R_{ij}}\left(b_i^g+\delta b_i^g\right)\right)=J_r^{-1},J_{r,b}\left(\frac{\partial \tilde{R}{\left({\bar{b}}_i^g\right)}_{ij}}{\partial b_i^g}\delta b_i^g\right)=J_{r,b}$$
所以姿态预积分雅可比为：
$$\begin{array}{l}\frac{\partial e_{R_{ij}}}{\partial \delta {\varphi }_i}=-J_r^{-1}\left(e_{R_{ij}}\left(R_i\right)\right)R_j^T{R}_i\\ \frac{\partial e_{R_{ij}}}{\partial \delta {\varphi }_{\mathrm{j}}}=J_r^{-1}\left(e_{R_{ij}}\left(R_j\right)\right)\\ \frac{\partial e_{R_{ij}}}{\partial \delta b_i^g}=J_r^{-1}\left(e_{R_{ij}}\left(b_i^g+\delta b_i^g\right)\right)\exp {\left(e_{R_{ij}}\left(b_i^g+\delta b_i^g\right)\right)}^T{J}_{r,b}\frac{\partial \tilde{R}{\left({\bar{b}}_i^g\right)}_{ij}}{\partial b_i^g}\end{array}$$
对于速度预积分来说：
$$\begin{array}{l}{e}_{v_{ij}}\left(R_i\exp \left(\delta {\varphi }_i\right)\right)={\left(R_i\exp \left(\delta {\varphi }_i\right)\right)}^T\left(v_j-v_i-g\Delta t_{ij}\right)-{\tilde{v}}_{ij}\\ =\exp {\left(\delta {\varphi }_i\right)}^T{R}_i^T\left(v_j-v_i-g\Delta t_{ij}\right)-{\tilde{v}}_{ij}\\ =\left(I-\delta {\varphi }_i^{\wedge }\right)R_i^T\left(v_j-v_i-g\Delta t_{ij}\right)-{\tilde{v}}_{ij}\\ =e_{v_{ij}}\left(R_i\exp \left(\delta {\varphi }_i\right)\right)+{\left(R_i^T\left(v_j-v_i-g\Delta t_{ij}\right)\right)}^{\wedge }\delta {\varphi }_i-{\tilde{v}}_{ij}\end{array}$$
其他为简单的线形关系，所以速度预积分雅可比为：
$$\begin{array}{l}\frac{\partial e_{{\mathrm{v}}_{ij}}}{\partial \delta {\varphi }_i}={\left(R_i^T\left(v_j-v_i-g\Delta t_{ij}\right)\right)}^{\wedge }\\ \frac{\partial e_{{\mathrm{v}}_{ij}}}{\partial \delta v_i}=-R_i^T\\ \frac{\partial e_{{\mathrm{v}}_{ij}}}{\partial \delta v_j}=R_i^T\\ \frac{\partial e_{{\mathrm{v}}_{ij}}}{\partial b_i^a}=-\frac{\partial \tilde{v}{\left({\bar{b}}_i^g,{\bar{b}}_i^a\right)}_{ij}}{\partial b^a}\\ \frac{\partial e_{{\mathrm{v}}_{ij}}}{\partial b_i^g}=-\frac{\partial \tilde{v}{\left({\bar{b}}_i^g,{\bar{b}}_i^a\right)}_{ij}}{\partial b^g}\end{array}$$
对于位置来说，同理：
$$\begin{array}{l}{e}_{p_{ij}}\left(R_i\exp \left(\delta {\varphi }_i\right)\right)={\left(R_i\exp \left(\delta {\varphi }_i\right)\right)}^T\left(p_j-p_i-v_i\Delta t_{ij}-\frac{1}{2}g\Delta t_{ij}^2\right)-{\tilde{p}}_{ij}\\ =\exp {\left(\delta {\varphi }_i\right)}^T{R}_i^T\left(p_j-p_i-v_i\Delta t_{ij}-\frac{1}{2}g\Delta t_{ij}^2\right)-{\tilde{p}}_{ij}\\ =\left(I-\delta {\varphi }_i^{\wedge }\right)R_i^T\left(p_j-p_i-v_i\Delta t_{ij}-\frac{1}{2}g\Delta t_{ij}^2\right)-{\tilde{p}}_{ij}\\ =e_{p_{ij}}\left(R_i\exp \left(\delta {\varphi }_i\right)\right)+{\left(R_i^T\left(p_j-p_i-v_i\Delta t_{ij}-\frac{1}{2}g\Delta t_{ij}^2\right)\right)}^{\wedge }\delta {\varphi }_i-{\tilde{p}}_{ij}\end{array}$$
$$\begin{array}{l}\frac{\partial e_{{\mathrm{p}}_{ij}}}{\partial \delta {\varphi }_i}={\left(R_i^T\left(p_j-p_i-v_i\Delta t_{ij}-\frac{1}{2}g\Delta t_{ij}^2\right)\right)}^{\wedge }\\ \frac{\partial e_{p_{ij}}}{\partial \delta p_i}=-R_i^T\\ \frac{\partial e_{p_{ij}}}{\partial \delta p_j}=R_i^T\\ \frac{\partial e_{p_{ij}}}{\partial \delta v_i}=-R_i^T\Delta t_{ij}\\ \frac{\partial e_{p_{ij}}}{\partial b_i^a}=-\frac{\partial \tilde{p}{\left({\bar{b}}_i^g,{\bar{b}}_i^a\right)}_{ij}}{\partial b^a}\\ \frac{\partial e_{p_{ij}}}{\partial b_i^g}=-\frac{\partial \tilde{p}{\left({\bar{b}}_i^g,{\bar{b}}_i^a\right)}_{ij}}{\partial b^g}\end{array}$$