2023-07-06 求流形中的偏导数

最新推荐文章于 2026-06-28 22:07:23 发布

原创最新推荐文章于 2026-06-28 22:07:23 发布 · 293 阅读

2 ·

本内容遵循CC 4.0 BY-SA版权协议

GEO检测

标签

#人工智能 #算法

已知流形： $\boxplus b) \boxminus ( c \boxplus d)，a,c \in \mathit{SO(3)}, b,d \in \mathbb{R}^3$ ，求E对b和E对d的偏导数

（1）E对b的偏导数的推导：
$E=(a⊞b)⊟(c⊞d)→E=Log((cExp(d))−1aExp(b))→Exp(E)=(cExp(d))−1aExp(b)→Exp(E)=Exp(−d)c−1aExp(b)→Exp(E+ΔE)=Exp(−d)c−1aExp(b+Δb)→Exp(E)Exp[A(E)TΔE]=Exp(−d)c−1aExp(b)Exp(A(b)TΔb)→Exp(E)Exp[A(E)TΔE]=Exp(E)Exp(A(b)TΔb)→Exp[A(E)TΔE]=Exp(A(b)TΔb)→A(E)TΔE=A(b)TΔb→ΔEΔb=A(E)−TA(b)T→∂E∂b=A(E)−TA(−b)\begin{align} E & =(a \boxplus b) \boxminus ( c \boxplus d) \\ \to E & = Log( (c Exp(d)) ^ {-1} a Exp(b)) \\ \to Exp(E) & = (c Exp(d)) ^ {-1} a Exp(b) \\ \to Exp(E) & = Exp(-d) c ^ {-1} a Exp(b) \\ \to Exp(E+\Delta E) & = Exp(-d) c ^ {-1} a Exp(b+ \Delta b) \\ \to Exp(E)Exp[A(E) ^ T \Delta E] & = Exp(-d) c ^ {-1} a Exp(b) Exp(A(b) ^ T \Delta b) \\ \to Exp(E)Exp[A(E) ^ T \Delta E] & = Exp(E) Exp(A(b) ^ T \Delta b) \\ \to Exp[A(E) ^ T \Delta E] & = Exp(A(b) ^ T \Delta b) \\ \to A(E) ^ T \Delta E & = A(b) ^ T \Delta b \\ \to \frac {\Delta E} {\Delta b} & = A(E) ^ {-T} A(b) ^ T \\ \to \frac {\partial E} {\partial b} & = A(E) ^ {-T} A(-b) \end{align}$
（2）E对d的偏导数的推导：
$E=(a⊞b)⊟(c⊞d)→E=Log((cExp(d))−1aExp(b))→Exp(E)=(cExp(d))−1aExp(b)→Exp(E)=Exp(−d)c−1aExp(b)→Exp(E+ΔE)=Exp(−d−Δd)c−1aExp(b)→Exp(E)Exp[A(E)TΔE]=Exp(−d)Exp[A(−d)T(−Δd)]c−1aExp(b)→Exp[A(E)TΔE]=(c−1aExp(b))−1Exp[A(−d)T(−Δd)]c−1aExp(b)→Exp[A(E)TΔE]=Exp[(c−1aExp(b))−1A(−d)T(−Δd)]→A(E)TΔE=(c−1aExp(b))−1A(−d)T(−Δd)→A(E)TΔE=−(c−1aExp(b))−1A(d)Δd→A(E)TΔE=−Exp(−b)a−1cA(d)Δd→ΔEΔd=−A(E)−TExp(−b)a−1cA(d)→∂E∂d=−A(E)−TExp(−b)a−1cA(d)\begin{align} E & =(a \boxplus b) \boxminus ( c \boxplus d) \\ \to E & = Log( (c Exp(d)) ^ {-1} a Exp(b)) \\ \to Exp(E) & = (c Exp(d)) ^ {-1} a Exp(b) \\ \to Exp(E) & = Exp(-d) c ^ {-1} a Exp(b) \\ \to Exp(E + \Delta E) & = Exp(-d- \Delta d) c ^ {-1} a Exp(b) \\ \to Exp(E) Exp[ A(E) ^ T \Delta E] & = Exp(-d)Exp[A(-d) ^ T (- \Delta d) ] c ^ {-1} a Exp(b) \\ \to Exp[ A(E) ^ T \Delta E] & = {(c ^ {-1} a Exp(b) )} ^ {-1} Exp[A(-d) ^ T (- \Delta d) ] c ^ {-1} a Exp(b) \\ \to Exp[ A(E) ^ T \Delta E] & = Exp[(c ^ {-1} a Exp(b) ) ^ {-1}A(-d) ^ T (- \Delta d) ] \\ \to A(E) ^ T \Delta E & = (c ^ {-1} a Exp(b) ) ^ {-1}A(-d) ^ T (- \Delta d) \\ \to A(E) ^ T \Delta E & = - (c ^ {-1} a Exp(b) ) ^ {-1} A(d) \Delta d \\ \to A(E) ^ T \Delta E & = - Exp(-b) a ^ {-1} c A(d) \Delta d \\ \to \frac {\Delta E} { \Delta d }& = - A(E) ^ {-T} Exp(-b) a ^ {-1} c A(d) \\ \to \frac {\partial E} { \partial d }& = - A(E) ^ {-T} Exp(-b) a ^ {-1} c A(d) \\ \end{align}$