Probability Density Reweight
Reweight 是通过将采样样本乘以 reweight 权重,从而将样本从原始密度
P
0
P_0
P0 转移至新密度
P
1
P_1
P1 的方法。
当从原始密度采样样本
x
x
x 时,
x
x
x 的期望为
E
x
∼
P
0
[
x
]
=
∫
P
0
(
x
)
x
d
x
≈
1
N
∑
i
x
i
(1)
E_{x \sim P_0}[x] = \int P_0(x)x dx \approx \frac{1}{N} \sum_i x_i \tag{1}
Ex∼P0[x]=∫P0(x)xdx≈N1i∑xi(1)
其中,
x
i
x_i
xi 为采样点,
N
N
N 为采样个数。
我们的目的是将
x
∼
P
0
x \sim P_0
x∼P0 转换为
x
∼
P
1
x \sim P_1
x∼P1,即求得
x
x
x 在
P
1
P_1
P1 上的期望
E
x
∼
P
1
[
x
]
≈
1
N
∑
i
w
(
x
i
)
⋅
x
i
w
(
x
i
)
=
P
1
(
x
i
)
/
P
0
(
x
i
)
(2)
E_{x \sim P_1}[x] \approx \frac{1}{N} \sum_i w(x_i) · x_i \tag{2} \\ w(x_i) = P_1(x_i) / P_0(x_i)
Ex∼P1[x]≈N1i∑w(xi)⋅xiw(xi)=P1(xi)/P0(xi)(2)
其中,
w
(
x
i
)
w(x_i)
w(xi) 表示 reweight 权重,证明如下所示。
采样时将
x
i
x_i
xi 乘以
w
(
x
i
)
w(x_i)
w(xi),根据式 1 可得
1
N
∑
i
P
1
(
x
i
)
P
0
(
x
i
)
x
i
≈
∫
P
0
(
x
)
P
1
(
x
)
P
0
(
x
)
x
d
x
=
∫
P
1
(
x
)
x
d
x
=
E
x
∼
P
1
[
x
]
(3)
\frac{1}{N}\sum_i \frac{P_1(x_i)}{P_0(x_i)}x_i \approx \int P_0(x) \frac{P_1(x)}{P_0(x)}x dx = \int P_1(x)xdx = E_{x \sim P_1}[x] \tag{3}
N1i∑P0(xi)P1(xi)xi≈∫P0(x)P0(x)P1(x)xdx=∫P1(x)xdx=Ex∼P1[x](3)
要注意,上述所说的概率密度为标准概率密度,即在定义域内积分为 1 。若
P
0
P_0
P0 和
P
1
P_1
P1 为非标准概率密度,需要
E
x
∼
P
0
[
x
]
=
1
∫
P
0
(
x
)
d
x
∫
P
0
(
x
)
x
d
x
≈
1
N
∑
i
x
i
(4)
E_{x \sim P_0}[x] = \frac{1}{\int P_0(x) dx} \int P_0(x)x dx\approx \frac{1}{N} \sum_i x_i \tag{4}
Ex∼P0[x]=∫P0(x)dx1∫P0(x)xdx≈N1i∑xi(4)
x
x
x 在
P
1
P_1
P1 上的期望变为
E
x
∼
P
1
[
x
]
≈
∑
i
w
(
x
i
)
⋅
x
i
∑
i
w
(
x
i
)
(5)
E_{x \sim P_1}[x] \approx \frac{\sum_i w(x_i) · x_i}{\sum_i w(x_i)} \tag{5}
Ex∼P1[x]≈∑iw(xi)∑iw(xi)⋅xi(5)
证明如下
1
N
∑
i
P
1
(
x
i
)
P
0
(
x
i
)
x
i
≈
1
∫
P
0
(
x
)
d
x
∫
P
1
(
x
)
x
d
x
=
∫
P
1
(
x
)
d
x
∫
P
0
(
x
)
d
x
E
x
∼
P
1
[
x
]
≈
∑
i
w
(
x
i
)
E
x
∼
P
1
[
x
]
\frac{1}{N}\sum_i \frac{P_1(x_i)}{P_0(x_i)}x_i \approx \frac{1}{\int P_0(x) dx} \int P_1(x)xdx = \frac{\int P_1(x) dx}{\int P_0(x) dx} E_{x \sim P_1}[x] \approx {\sum_i w(x_i)}E_{x \sim P_1}[x]
N1i∑P0(xi)P1(xi)xi≈∫P0(x)dx1∫P1(x)xdx=∫P0(x)dx∫P1(x)dxEx∼P1[x]≈i∑w(xi)Ex∼P1[x]