保凸运算
透视函数 Perspative Function
P : R n + 1 ↦ R n d o m P = R n ∗ R + + P: R^{n+1} \mapsto R^n \quad dom \; P=R^n*R_{++} P:Rn+1↦RndomP=Rn∗R++
P ( z , t ) = z t z ∈ R n t ∈ R + + P(z,t) = \frac{z}{t} \quad z \in R^n \quad t \in R_{++} P(z,t)=tzz∈Rnt∈R++
降维过程,将前所有元素,除以最后一维,并去掉最后一维。
任何一个函数,通过透视函数之后,还是凸集。
例:考虑 R n + 1 R^{n+1} Rn+1 线段。 x = ( x ~ , x n + 1 ) x ~ ∈ R n x n + 1 ∈ R + + x=(\tilde x,x_{n+1}) \quad \tilde x \in R^n \quad x_{n+1} \in R_{++} x=(x~,xn+1)x~∈Rnxn+1∈R++; y = ( y ~ , y n + 1 ) y ~ ∈ R n y n + 1 ∈ R + + y=(\tilde y,y_{n+1}) \quad \tilde y \in R^n \quad y_{n+1} \in R_{++} y=(y~,yn+1)y~∈Rnyn+1∈R++
θ ≥ 0 \theta \geq 0 θ≥0 线段维 θ x + ( 1 − θ ) y \theta x + (1-\theta) y θx+(1−θ)y
证明:线段经过透视函数P后,仍然时凸集。
x
→
P
P
(
x
)
x \rightarrow ^ P P(x)
x→PP(x)
\quad
y
→
P
P
(
y
)
y \rightarrow ^ P P(y)
y→PP(y)
θ
x
+
(
1
−
θ
)
y
→
P
P
(
θ
x
+
(
1
−
θ
)
y
)
\theta x + (1-\theta) y \rightarrow ^ P P(\theta x + (1-\theta) y)
θx+(1−θ)y→PP(θx+(1−θ)y)
设
μ
=
θ
x
n
+
1
θ
x
n
+
1
+
(
1
−
θ
)
y
n
+
1
\mu =\frac{\theta x_{n+1}}{ \theta x_{n+1} + (1-\theta) y_{n+1} }
μ=θxn+1+(1−θ)yn+1θxn+1
P
(
θ
x
+
(
1
−
θ
)
y
)
=
θ
x
~
+
(
1
−
θ
)
y
~
θ
x
n
+
1
+
(
1
−
θ
)
y
n
+
1
=
θ
x
n
+
1
θ
x
n
+
1
+
(
1
−
θ
)
y
n
+
1
x
~
x
n
+
1
+
(
1
−
θ
)
y
n
+
1
θ
x
n
+
1
+
(
1
−
θ
)
y
n
+
1
y
~
y
n
+
1
=
μ
P
(
x
)
+
(
1
−
μ
)
P
(
y
)
\begin{aligned} P(\theta x + (1-\theta) y) &= \frac{\theta \tilde {x} + (1-\theta) \tilde {y}}{ \theta x_{n+1} + (1-\theta) y_{n+1} } \\ &=\frac{\theta x_{n+1}}{ \theta x_{n+1} + (1-\theta) y_{n+1} } \frac{\tilde {x}}{x_{n+1}} + \frac{(1-\theta) y_{n+1}}{ \theta x_{n+1} + (1-\theta) y_{n+1} } \frac{\tilde {y}}{y_{n+1}} \\ &=\mu P(x) + (1-\mu) P(y) \end{aligned}
P(θx+(1−θ)y)=θxn+1+(1−θ)yn+1θx~+(1−θ)y~=θxn+1+(1−θ)yn+1θxn+1xn+1x~+θxn+1+(1−θ)yn+1(1−θ)yn+1yn+1y~=μP(x)+(1−μ)P(y)
任意凸集的反透视映射仍是凸集。
P − 1 ( C ) = { ( x , t ) ∈ R n + 1 ∣ x t ∈ c , t > 0 } P^{-1}(C)=\lbrace (x,t) \in R^{n+1} | \frac{x}{t} \in c, t > 0 \rbrace P−1(C)={(x,t)∈Rn+1∣tx∈c,t>0}
考虑 ( x , t ) ∈ P − 1 ( C ) ( y , s ) ∈ P − 1 ( C ) 0 ≤ θ ≤ 1 (x,t) \in P^{-1}(C) \quad (y,s) \in P^{-1}(C) \quad 0 \leq \theta \leq 1 (x,t)∈P−1(C)(y,s)∈P−1(C)0≤θ≤1
θ x + ( 1 − θ ) y θ t + ( 1 − θ ) s ∈ ? C \frac{\theta x+(1-\theta)y}{\theta t+(1-\theta)s} \in^? C θt+(1−θ)sθx+(1−θ)y∈?C
θ
t
θ
t
+
(
1
−
θ
)
s
x
t
+
(
1
−
θ
t
θ
t
+
(
1
−
θ
)
s
)
y
s
\begin{aligned} \frac{\theta t}{\theta t+(1-\theta)s}\frac{x}{t}+(1-\frac{\theta t}{\theta t+(1-\theta)s})\frac{y}{s} \end{aligned}
θt+(1−θ)sθttx+(1−θt+(1−θ)sθt)sy
x
t
\frac{x}{t}
tx 与
y
s
\frac{y}{s}
sy 都是C中的一点,C是凸集。所以结论成立。
线性分数函数
g
:
R
n
↦
R
m
+
1
g:R^n \mapsto R^{m+1}
g:Rn↦Rm+1 为伪仿射函数
g
(
x
)
=
[
A
C
T
]
x
+
[
b
d
]
A
∈
R
m
+
n
b
∈
R
m
C
∈
R
n
d
∈
R
g(x)=\begin{bmatrix} A \\ C^T \\ \end{bmatrix} x + \begin{bmatrix} b \\ d \\ \end{bmatrix} \quad A \in R^{m+n} \quad b\in R^{m} \quad C \in R^{n} \quad d \in R
g(x)=[ACT]x+[bd]A∈Rm+nb∈RmC∈Rnd∈R
P
:
R
m
+
1
↦
R
m
P:R^{m+1} \mapsto R^{m}
P:Rm+1↦Rm
定义:
f
:
R
n
↦
R
m
=
P
∘
g
f:R^n \mapsto R^m = P \circ g
f:Rn↦Rm=P∘g 先经过
g
g
g,在经过
P
P
P。
本质是仿射变换+透视变换。
f ( x ) = A x + b C T + d d o m f { x ∣ C T x + d > 0 } f(x) = \frac{Ax+b}{C^T+d} \quad dom \;f \lbrace x | C^Tx + d > 0 \rbrace f(x)=CT+dAx+bdomf{x∣CTx+d>0}
例:两个随机变量的联合概率 → \rightarrow → 条件概率
u ∈ { 1.. n } , v ∈ { 1... m } u \in \lbrace {1..n} \rbrace,v \in \lbrace {1...m} \rbrace u∈{1..n},v∈{1...m}
P
i
j
=
P
(
u
=
i
,
v
=
j
)
P_{ij}=P(u=i,v=j)
Pij=P(u=i,v=j)
f
i
j
=
P
(
u
=
i
∣
v
=
j
)
f_{ij}=P(u=i|v=j)
fij=P(u=i∣v=j)
f i j = P i j ∑ k = 1 n P k j f_{ij}= \frac{P_{ij}}{\sum_{k=1}^n P_{kj}} fij=∑k=1nPkjPij