最近在研究 diffusion model，涉及到了 Time-reversal formula，reverse-time process，以及相关的一些随机过程、随机分析的理论知识，搞得头很晕，特意做了这篇整理。
关注的 SDE 如下
d X t = f ( X t , t ) d t + g ( X t , t ) d W t , (1) dX_t=f(X_t,t)dt+g(X_t,t)dW_t,\tag{1} dXt=f(Xt,t)dt+g(Xt,t)dWt,(1)
我们分别介绍 Forward Kolmogorov 方程和 Backward Kolmogorov 方程（这里就不用正向、倒向来描述了，避免中文带来的直观感觉与后面出现的 reverse-time 产生矛盾）
Forward Kolmogorov 方程为：对
t
⩾
s
t\geqslant s
t⩾s,
−
∂
p
(
x
t
,
t
∣
x
s
,
s
)
∂
t
=
∑
i
∂
∂
x
t
i
[
p
(
x
t
,
t
∣
x
s
,
s
)
f
i
(
x
t
,
t
)
]
−
1
2
∑
i
,
j
,
k
∂
2
[
g
i
k
(
x
t
,
t
)
g
j
k
(
x
t
,
t
)
p
(
x
t
,
t
∣
x
s
,
s
)
]
∂
x
t
i
∂
x
t
j
-\dfrac{\partial p(x_t,t|x_s,s)}{\partial t}=\sum_i\dfrac{\partial}{\partial x_t^i}\left[p(x_t,t|x_s,s)f^i(x_t,t)\right]-\dfrac{1}{2}\sum_{i,j,k}\dfrac{\partial^2\left[g^{ik}(x_t,t)g^{jk}(x_t,t)p(x_t,t|x_s,s)\right]}{\partial x_t^i\partial x_t^j}
−∂t∂p(xt,t∣xs,s)=i∑∂xti∂[p(xt,t∣xs,s)fi(xt,t)]−21i,j,k∑∂xti∂xtj∂2[gik(xt,t)gjk(xt,t)p(xt,t∣xs,s)]
由于偏导数是针对
t
t
t 时刻的状态计算的，因此我们可以对
x
s
,
s
x_s,s
xs,s 条件直接积分，也即方程左右两边同乘
p
(
x
s
,
s
)
p(x_s,s)
p(xs,s) 后对
d
x
s
dx_s
dxs 积分，求偏导和积分运算交换后得到 Fokker-Planck 方程：
−
∂
p
(
x
t
,
t
)
∂
t
=
∑
i
∂
∂
x
t
i
[
p
(
x
t
,
t
)
f
i
(
x
t
,
t
)
]
−
1
2
∑
i
,
j
,
k
∂
2
[
g
i
k
(
x
t
,
t
)
g
j
k
(
x
t
,
t
)
p
(
x
t
,
t
)
]
∂
x
t
i
∂
x
t
j
-\dfrac{\partial p(x_t,t)}{\partial t}=\sum_i\dfrac{\partial}{\partial x_t^i}\left[p(x_t,t)f^i(x_t,t)\right]-\dfrac{1}{2}\sum_{i,j,k}\dfrac{\partial^2\left[g^{ik}(x_t,t)g^{jk}(x_t,t)p(x_t,t)\right]}{\partial x_t^i\partial x_t^j}
−∂t∂p(xt,t)=i∑∂xti∂[p(xt,t)fi(xt,t)]−21i,j,k∑∂xti∂xtj∂2[gik(xt,t)gjk(xt,t)p(xt,t)]
对应的 Backward Kolmogorov 方程为：对
s
⩾
t
s\geqslant t
s⩾t
−
∂
p
(
x
s
,
s
∣
x
t
,
t
)
∂
t
=
∑
i
f
i
(
x
t
,
t
)
∂
p
(
x
s
,
s
∣
x
t
,
t
)
∂
x
t
i
+
1
2
∑
i
,
j
,
k
g
i
k
(
x
t
,
t
)
g
j
k
(
x
t
,
t
)
∂
2
p
(
x
s
,
s
∣
x
t
,
t
)
∂
x
t
i
∂
x
t
j
-\dfrac{\partial p(x_s,s|x_t,t)}{\partial t}=\sum_i f^i(x_t,t)\dfrac{\partial p(x_s,s|x_t,t)}{\partial x_t^i}+\dfrac{1}{2}\sum_{i,j,k}g^{ik}(x_t,t)g^{jk}(x_t,t)\dfrac{\partial^2 p(x_s,s|x_t,t)}{\partial x_t^i\partial x_t^j}
−∂t∂p(xs,s∣xt,t)=i∑fi(xt,t)∂xti∂p(xs,s∣xt,t)+21i,j,k∑gik(xt,t)gjk(xt,t)∂xti∂xtj∂2p(xs,s∣xt,t)
注意转移概率描述的还是
X
t
X_t
Xt 到
X
s
X_s
Xs 的转移概率，是正向的转移概率，但是方程中对
t
t
t 和
x
t
x_t
xt 求偏导，描述的是初始状态的变化对转移概率的影响。
至此，我们先做一步总结：Kolmogorov 方程的主体都是正向转移概率的变化，Forward Kolmogorov 描述的是正向转移概率的末端变化，Backward Kolmogorov 描述的是正向转移概率的始端变化。（这都不是什么术语，只是我为了方便区分给出的命名）
我们开始进一步的研究：倒向转移概率，在后面的过程中，注意每一步都要明确时间的先后。我们要计算的是
∂
p
(
x
t
,
t
,
x
s
,
s
)
∂
t
,
s
⩾
t
\dfrac{\partial p(x_t,t,x_s,s)}{\partial t},\quad s\geqslant t
∂t∂p(xt,t,xs,s),s⩾t
这里再强调一下，上面的联合分布
t
,
s
t,s
t,s 的大小关系看起来不太重要，但重要的是，我们研究的是始端状态的变化
∂
t
\partial t
∂t，事实上，如果研究末端变化，直接由 Forward Kolmogorov 方程两侧同乘
p
(
x
s
,
s
)
p(x_s,s)
p(xs,s) 就得到了。而始端变化就稍微复杂些，我们计算
∂
p
(
x
t
,
t
,
x
s
,
s
)
∂
t
=
∂
(
p
(
x
s
,
s
∣
x
t
,
t
)
p
(
x
t
,
t
)
)
∂
t
=
∂
p
(
x
s
,
s
∣
x
t
,
t
)
∂
t
p
(
x
t
,
t
)
+
p
(
x
s
,
s
∣
x
t
,
t
)
∂
p
(
x
t
,
t
)
∂
t
=
−
p
(
x
t
,
t
)
[
∑
i
f
i
(
x
t
,
t
)
∂
p
(
x
s
,
s
∣
x
t
,
t
)
∂
x
t
i
+
1
2
∑
i
,
j
,
k
g
i
k
(
x
t
,
t
)
g
j
k
(
x
t
,
t
)
∂
2
p
(
x
s
,
s
∣
x
t
,
t
)
∂
x
t
i
∂
x
t
j
]
−
p
(
x
s
,
s
∣
x
t
,
t
)
[
∑
i
∂
∂
x
t
i
[
p
(
x
t
,
t
)
f
i
(
x
t
,
t
)
]
−
1
2
∑
i
,
j
,
k
∂
2
[
g
i
k
(
x
t
,
t
)
g
j
k
(
x
t
,
t
)
p
(
x
t
,
t
)
]
∂
x
t
i
∂
x
t
j
]
=
−
∑
i
∂
∂
x
t
i
[
f
i
(
x
t
,
t
)
p
(
x
t
,
t
)
p
(
x
s
,
s
∣
x
t
,
t
)
]
−
1
2
[
p
(
x
t
,
t
)
∑
i
,
j
,
k
g
i
k
(
x
t
,
t
)
g
j
k
(
x
t
,
t
)
∂
2
p
(
x
s
,
s
∣
x
t
,
t
)
∂
x
t
i
∂
x
t
j
+
2
∑
i
,
j
,
k
∂
[
p
(
x
t
,
t
)
g
i
k
(
x
t
,
t
)
g
j
k
(
x
t
,
t
)
]
∂
x
t
i
∂
p
(
x
s
,
s
∣
x
t
,
t
)
∂
x
t
j
+
p
(
x
s
,
s
∣
x
t
,
t
)
∑
i
,
j
,
k
∂
2
[
g
i
k
(
x
t
,
t
)
g
j
k
(
x
t
,
t
)
p
(
x
t
,
t
)
]
∂
x
t
i
∂
x
t
j
]
+
p
(
x
s
,
s
∣
x
t
,
t
)
∑
i
,
j
,
k
∂
2
[
g
i
k
(
x
t
,
t
)
g
j
k
(
x
t
,
t
)
p
(
x
t
,
t
)
]
∂
x
t
i
∂
x
t
j
+
∑
i
,
j
,
k
∂
[
p
(
x
t
,
t
)
g
i
k
(
x
t
,
t
)
g
j
k
(
x
t
,
t
)
]
∂
x
t
i
∂
p
(
x
s
,
s
∣
x
t
,
t
)
∂
x
t
j
=
−
∑
i
∂
∂
x
t
i
[
f
i
(
x
t
,
t
)
p
(
x
t
,
t
,
x
s
,
s
)
]
−
1
2
∑
i
,
j
,
k
∂
2
[
p
(
x
t
,
t
,
x
s
,
s
)
g
i
k
(
x
t
,
t
)
g
j
k
(
x
t
,
t
)
]
∂
x
t
i
∂
x
t
j
+
∑
i
,
j
,
k
∂
∂
x
t
i
[
p
(
x
s
,
s
∣
x
t
,
t
)
∂
[
p
(
x
t
,
t
)
g
i
k
(
x
t
,
t
)
g
j
k
(
x
t
,
t
)
]
∂
x
t
j
]
=
−
∑
i
∂
∂
x
t
i
[
p
(
x
t
,
t
,
x
s
,
s
)
⋅
(
f
i
(
x
t
,
t
)
−
1
p
(
x
t
,
t
)
∑
j
,
k
∂
[
p
(
x
t
,
t
)
g
i
k
(
x
t
,
t
)
g
j
k
(
x
t
,
t
)
]
∂
x
t
j
)
]
−
1
2
∑
i
,
j
,
k
∂
2
[
p
(
x
t
,
t
,
x
s
,
s
)
g
i
k
(
x
t
,
t
)
g
j
k
(
x
t
,
t
)
]
∂
x
t
i
∂
x
t
j
\begin{align*} &\quad \dfrac{\partial p(x_t,t,x_s,s)}{\partial t}\\ &=\dfrac{\partial \left(p(x_s,s|x_t,t)p(x_t,t)\right)}{\partial t}\\ &=\dfrac{\partial p(x_s,s|x_t,t)}{\partial t}p(x_t,t)+p(x_s,s|x_t,t)\dfrac{\partial p(x_t,t)}{\partial t}\\ &=-p(x_t,t)\left[\sum_if^i(x_t,t)\dfrac{\partial p(x_s,s|x_t,t)}{\partial x_t^i}+\dfrac{1}{2}\sum_{i,j,k}g^{ik}(x_t,t)g^{jk}(x_t,t)\dfrac{\partial^2 p(x_s,s|x_t,t)}{\partial x_t^i\partial x_t^j}\right]\\ &\quad-p(x_s,s|x_t,t)\left[\sum_i\dfrac{\partial}{\partial x_t^i}\left[p(x_t,t)f^i(x_t,t)\right]-\dfrac{1}{2}\sum_{i,j,k}\dfrac{\partial^2\left[g^{ik}(x_t,t)g^{jk}(x_t,t)p(x_t,t)\right]}{\partial x_t^i\partial x_t^j}\right]\\ &=-\sum_i\dfrac{\partial}{\partial x_t^i}\left[f^i(x_t,t)p(x_t,t)p(x_s,s|x_t,t)\right]\\ &\quad -\dfrac{1}{2}\bigg[p(x_t,t)\sum_{i,j,k}g^{ik}(x_t,t)g^{jk}(x_t,t)\dfrac{\partial^2 p(x_s,s|x_t,t)}{\partial x_t^i\partial x_t^j}+2\sum_{i,j,k}\dfrac{\partial\left[p(x_t,t)g^{ik}(x_t,t)g^{jk}(x_t,t)\right]}{\partial x_t^i}\dfrac{\partial p(x_s,s|x_t,t)}{\partial x_t^j}\\ &\qquad\quad+p(x_s,s|x_t,t)\sum_{i,j,k}\dfrac{\partial^2\left[g^{ik}(x_t,t)g^{jk}(x_t,t)p(x_t,t)\right]}{\partial x_t^i\partial x_t^j}\bigg]\\ &\quad +p(x_s,s|x_t,t)\sum_{i,j,k}\dfrac{\partial^2\left[g^{ik}(x_t,t)g^{jk}(x_t,t)p(x_t,t)\right]}{\partial x_t^i\partial x_t^j}+\sum_{i,j,k}\dfrac{\partial\left[p(x_t,t)g^{ik}(x_t,t)g^{jk}(x_t,t)\right]}{\partial x_t^i}\dfrac{\partial p(x_s,s|x_t,t)}{\partial x_t^j}\\ &=-\sum_i\dfrac{\partial}{\partial x_t^i}\left[f^i(x_t,t)p(x_t,t,x_s,s)\right]\\ &\quad -\dfrac{1}{2}\sum_{i,j,k}\dfrac{\partial^2\left[p(x_t,t,x_s,s)g^{ik}(x_t,t)g^{jk}(x_t,t)\right]}{\partial x_t^i\partial x_t^j}\\ &\quad +\sum_{i,j,k}\dfrac{\partial}{\partial x_t^i}\left[p(x_s,s|x_t,t)\dfrac{\partial\left[p(x_t,t)g^{ik}(x_t,t)g^{jk}(x_t,t)\right]}{\partial x_t^j}\right]\\ &=-\sum_i\dfrac{\partial}{\partial x_t^i}\left[p(x_t,t,x_s,s)\cdot\left(f^i(x_t,t)-\dfrac{1}{p(x_t,t)}\sum_{j,k}\dfrac{\partial\left[p(x_t,t)g^{ik}(x_t,t)g^{jk}(x_t,t)\right]}{\partial x_t^j}\right)\right]\\ &\quad -\dfrac{1}{2}\sum_{i,j,k}\dfrac{\partial^2\left[p(x_t,t,x_s,s)g^{ik}(x_t,t)g^{jk}(x_t,t)\right]}{\partial x_t^i\partial x_t^j} \end{align*}
∂t∂p(xt,t,xs,s)=∂t∂(p(xs,s∣xt,t)p(xt,t))=∂t∂p(xs,s∣xt,t)p(xt,t)+p(xs,s∣xt,t)∂t∂p(xt,t)=−p(xt,t)
i∑fi(xt,t)∂xti∂p(xs,s∣xt,t)+21i,j,k∑gik(xt,t)gjk(xt,t)∂xti∂xtj∂2p(xs,s∣xt,t)
−p(xs,s∣xt,t)
i∑∂xti∂[p(xt,t)fi(xt,t)]−21i,j,k∑∂xti∂xtj∂2[gik(xt,t)gjk(xt,t)p(xt,t)]
=−i∑∂xti∂[fi(xt,t)p(xt,t)p(xs,s∣xt,t)]−21[p(xt,t)i,j,k∑gik(xt,t)gjk(xt,t)∂xti∂xtj∂2p(xs,s∣xt,t)+2i,j,k∑∂xti∂[p(xt,t)gik(xt,t)gjk(xt,t)]∂xtj∂p(xs,s∣xt,t)+p(xs,s∣xt,t)i,j,k∑∂xti∂xtj∂2[gik(xt,t)gjk(xt,t)p(xt,t)]]+p(xs,s∣xt,t)i,j,k∑∂xti∂xtj∂2[gik(xt,t)gjk(xt,t)p(xt,t)]+i,j,k∑∂xti∂[p(xt,t)gik(xt,t)gjk(xt,t)]∂xtj∂p(xs,s∣xt,t)=−i∑∂xti∂[fi(xt,t)p(xt,t,xs,s)]−21i,j,k∑∂xti∂xtj∂2[p(xt,t,xs,s)gik(xt,t)gjk(xt,t)]+i,j,k∑∂xti∂[p(xs,s∣xt,t)∂xtj∂[p(xt,t)gik(xt,t)gjk(xt,t)]]=−i∑∂xti∂
p(xt,t,xs,s)⋅
fi(xt,t)−p(xt,t)1j,k∑∂xtj∂[p(xt,t)gik(xt,t)gjk(xt,t)]
−21i,j,k∑∂xti∂xtj∂2[p(xt,t,xs,s)gik(xt,t)gjk(xt,t)]
如果整理一下系数，重新给定记号
f
ˉ
i
(
x
t
,
t
)
=
f
i
(
x
t
,
t
)
−
1
p
(
x
t
,
t
)
∑
j
,
k
∂
∂
x
t
j
[
p
(
x
t
,
t
)
g
i
k
(
x
t
,
t
)
g
j
k
(
x
t
,
t
)
]
\bar f^i(x_t,t)=f^i(x_t,t)-\dfrac{1}{p(x_t,t)}\sum_{j,k}\dfrac{\partial}{\partial x_t^j}\left[p(x_t,t)g^{ik}(x_t,t)g^{jk}(x_t,t)\right]
fˉi(xt,t)=fi(xt,t)−p(xt,t)1j,k∑∂xtj∂[p(xt,t)gik(xt,t)gjk(xt,t)]
则有
t
⩽
s
t\leqslant s
t⩽s,
−
∂
p
(
x
t
,
t
,
x
s
,
s
)
∂
t
=
∑
i
∂
∂
x
t
i
[
f
ˉ
i
(
x
t
,
t
)
p
(
x
t
,
t
,
x
s
,
s
)
]
+
1
2
∑
i
,
j
,
k
∂
2
[
p
(
x
t
,
t
,
x
s
,
s
)
g
i
k
(
x
t
,
t
)
g
j
k
(
x
t
,
t
)
]
∂
x
t
i
∂
x
t
j
-\dfrac{\partial p(x_t,t,x_s,s)}{\partial t}=\sum_i\dfrac{\partial}{\partial x_t^i}\left[\bar f^i(x_t,t)p(x_t,t,x_s,s)\right]+\dfrac{1}{2}\sum_{i,j,k}\dfrac{\partial^2\left[p(x_t,t,x_s,s)g^{ik}(x_t,t)g^{jk}(x_t,t)\right]}{\partial x_t^i\partial x_t^j}
−∂t∂p(xt,t,xs,s)=i∑∂xti∂[fˉi(xt,t)p(xt,t,xs,s)]+21i,j,k∑∂xti∂xtj∂2[p(xt,t,xs,s)gik(xt,t)gjk(xt,t)]
对应地，我们可以在等式两侧同除
p
(
x
s
,
s
)
p(x_s,s)
p(xs,s) 或进一步对
d
x
s
dx_s
dxs 积分，分别能够得到条件转移概率和边缘密度函数（对应 Fokker-Planck 方程形式）的两个等式：
−
∂
p
(
x
t
,
t
∣
x
s
,
s
)
∂
t
=
∑
i
∂
∂
x
t
i
[
f
ˉ
i
(
x
t
,
t
)
p
(
x
t
,
t
∣
x
s
,
s
)
]
+
1
2
∑
i
,
j
,
k
∂
2
[
p
(
x
t
,
t
∣
x
s
,
s
)
g
i
k
(
x
t
,
t
)
g
j
k
(
x
t
,
t
)
]
∂
x
t
i
∂
x
t
j
-\dfrac{\partial p(x_t,t|x_s,s)}{\partial t}=\sum_i\dfrac{\partial}{\partial x_t^i}\left[\bar f^i(x_t,t)p(x_t,t|x_s,s)\right]+\dfrac{1}{2}\sum_{i,j,k}\dfrac{\partial^2\left[p(x_t,t|x_s,s)g^{ik}(x_t,t)g^{jk}(x_t,t)\right]}{\partial x_t^i\partial x_t^j}
−∂t∂p(xt,t∣xs,s)=i∑∂xti∂[fˉi(xt,t)p(xt,t∣xs,s)]+21i,j,k∑∂xti∂xtj∂2[p(xt,t∣xs,s)gik(xt,t)gjk(xt,t)]
以及
−
∂
p
(
x
t
,
t
)
∂
t
=
∑
i
∂
∂
x
t
i
[
f
ˉ
i
(
x
t
,
t
)
p
(
x
t
,
t
)
]
+
1
2
∑
i
,
j
,
k
∂
2
[
p
(
x
t
,
t
)
g
i
k
(
x
t
,
t
)
g
j
k
(
x
t
,
t
)
]
∂
x
t
i
∂
x
t
j
-\dfrac{\partial p(x_t,t)}{\partial t}=\sum_i\dfrac{\partial}{\partial x_t^i}\left[\bar f^i(x_t,t)p(x_t,t)\right]+\dfrac{1}{2}\sum_{i,j,k}\dfrac{\partial^2\left[p(x_t,t)g^{ik}(x_t,t)g^{jk}(x_t,t)\right]}{\partial x_t^i\partial x_t^j}
−∂t∂p(xt,t)=i∑∂xti∂[fˉi(xt,t)p(xt,t)]+21i,j,k∑∂xti∂xtj∂2[p(xt,t)gik(xt,t)gjk(xt,t)]
至此为止，我们做第二次总结：（1）我们仍然只关注了过程（1）的各种条件转移概率，边缘概率密度函数，所涉及的倒向概念，也只是针对该过程的时间先后，目前的随机过程的时间方向只有一个；（2）类比 Kolmogorov 方程，我们对倒向的转移概率
p
(
x
t
,
t
∣
x
s
,
s
)
,
t
⩽
s
p(x_t,t|x_s,s), t\leqslant s
p(xt,t∣xs,s),t⩽s 给出了一个偏导方程，但注意，我们只描述了倒向的转移概率的关于
t
t
t 的变化，关于
s
s
s 的变化暂不讨论。
本篇先整理到这里，在下一篇中就要引入 reverse time，时间的方向会出现两个，希望能够尽量以一种不使人困扰的方式把 diffusion model 所用的部分介绍清楚。
扫一扫
1834
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
