On Direction Derivative of Lie Groups
This article is inspired by Henry Taubes’s Differential Geometry
If ∥ A ∥ < 1 \|A\| < 1 ∥A∥<1, then ( I − A ) − 1 = I + A + A 2 + ⋯ + A n + ⋯ (I - A)^{-1}=I + A + A^2 + \cdots + A^n + \cdots (I−A)−1=I+A+A2+⋯+An+⋯
proof. First, the RHS does converge if
∥
A
∥
<
1
\|A\|<1
∥A∥<1, for
∥
∑
n
=
0
∞
A
n
∥
≤
∑
n
=
0
∞
∥
A
n
∥
≤
∑
n
=
0
∞
∥
A
∥
n
=
1
1
−
∥
A
∥
\| \sum_{n=0}^\infty A^n\| \le \sum_{n=0}^\infty \|A^n\| \le \sum_{n=0}^\infty\|A\|^n = \frac 1{1 - \|A\|}
∥n=0∑∞An∥≤n=0∑∞∥An∥≤n=0∑∞∥A∥n=1−∥A∥1
Next, we show that the RHS is indeed the inverse of
I
−
A
I- A
I−A, for
∑
n
=
0
∞
A
n
(
I
−
A
)
=
∑
n
=
0
∞
A
n
−
∑
n
=
1
∞
A
n
=
I
\sum_{n=0}^\infty A^n(I - A) = \sum_{n=0}^\infty A^n - \sum_{n=1}^\infty A^n = I
n=0∑∞An(I−A)=n=0∑∞An−n=1∑∞An=I
Q.E.D.
Using this, we show that the directional derivative of the map
A
↦
A
−
1
A\mapsto A^{-1}
A↦A−1 as the point
A
A
A in the direction
V
V
V is
A
−
1
V
A
−
1
A^{-1} V A^{-1}
A−1VA−1. Choose
V
V
V with small enough matrix norm, then
(
A
+
V
)
−
1
=
(
A
(
I
+
A
−
1
V
)
)
−
1
=
(
I
+
A
−
1
V
)
−
1
A
−
1
(A+V)^{-1} = (A(I + A^{-1}V))^{-1} = (I +A^{-1}V)^{-1} A^{-1}
(A+V)−1=(A(I+A−1V))−1=(I+A−1V)−1A−1
= ∑ n = 0 ∞ ( − 1 ) n A − n V n A − 1 =\sum_{n=0}^\infty(-1)^nA^{-n}V^nA^{-1} =n=0∑∞(−1)nA−nVnA−1
Thus, the 1 t h 1^{th} 1th order term of this equation is A − 1 V A − 1 A^{-1} V A^{-1} A−1VA−1, exactly as we expected.
The next important computation is the directional derivative of det \det det. We show that ∂ i j det m = m i j ⋆ \partial_{ij} \det m = m^\star_{ij} ∂ijdetm=mij⋆, where m i j ⋆ m^\star_{ij} mij⋆ denotes the ( i , j ) (i,j) (i,j)-adjacent matrix of m m m, i.e. to delete the i t h i^{th} ith row and j t h j^{th} jth column of m m m. To me, the prove is quite elegant, for it is telling that the symmetry (instead of finiteness of vector space) of det \det det must let the deduction stop somewhere.
Proposition: If m ∈ M ( n , R ) m\in M(n, \R) m∈M(n,R), then ∂ i j det m = m i j ⋆ \partial_{ij} \det m = m^\star_{ij} ∂ijdetm=mij⋆.
proof.
Using the identity
det
m
=
∑
j
=
1
n
m
i
j
det
m
i
j
⋆
\det m = \sum_{j=1}^n m_{ij} \det m^\star_{ij}
detm=∑j=1nmijdetmij⋆ for any chosen
i
i
i, we have
∂
i
j
det
m
=
det
m
i
j
⋆
+
∑
k
≠
j
m
i
k
∂
i
j
det
m
i
k
⋆
=
det
m
i
j
⋆
+
∑
k
≠
j
m
i
k
∂
i
j
(
∑
l
≠
i
m
l
j
det
m
i
l
,
k
j
⋆
)
\begin{array}{lll} \partial_{ij} \det m &=& \det m^\star_{ij} + \sum_{k\neq j} m_{ik} \partial_{ij} \det m^\star_{ik}\\ &=&\det m^\star_{ij} + \sum_{k\neq j} m_{ik} \partial_{ij}(\sum_{l\neq i} m_{lj} \det m^\star_{il, kj}) \end{array}
∂ijdetm==detmij⋆+∑k=jmik∂ijdetmik⋆detmij⋆+∑k=jmik∂ij(∑l=imljdetmil,kj⋆)
where
m
i
l
,
k
j
⋆
m^\star_{il, kj}
mil,kj⋆ denote subbing matrix of
m
m
m by deleting the
i
,
l
i,l
i,l rows and
k
,
j
k,j
k,j columns. Observing the
m
l
j
m_{lj}
mlj in the above equation cannot be
m
i
j
m_{ij}
mij since
l
≠
i
l\neq i
l=i, we obtain
det
m
i
j
⋆
+
∑
l
≠
i
,
k
≠
j
m
i
k
m
l
j
∂
i
j
det
m
i
l
,
k
j
⋆
\det m^\star_{ij} + \sum_{l\neq i, k\neq j} m_{ik}m_{lj} \partial_{ij} \det m^\star_{il, kj}
detmij⋆+l=i,k=j∑mikmlj∂ijdetmil,kj⋆
But
∂
i
j
det
m
i
l
,
k
j
⋆
=
0
\partial_{ij} \det m^\star_{il, kj} = 0
∂ijdetmil,kj⋆=0 since the
i
t
h
i^{th}
ith row and the
j
t
h
j^{th}
jth column have been deleted, hence the only item left is
det
m
i
j
⋆
\det m^\star_{ij}
detmij⋆, which is exactly what we expected.
Q.E.D.
From this we obtain the differential of
det
\det
det:
d
(
det
)
∣
m
=
det
m
∑
i
,
j
(
m
−
1
)
i
j
d
m
i
j
=
det
m
tr
(
m
−
1
d
m
)
d(\det)|_m = \det m \sum_{i,j} (m^{-1})_{ij} d m_{ij} = \det m \text{ tr}(m^{-1}dm)
d(det)∣m=detmi,j∑(m−1)ijdmij=detm tr(m−1dm)