CQF笔记M1L3泰勒级数和转移密度函数
Module 1 Building Blocks of Quant Finance
Lecture 2 Taylor Series and Transition Density Functions
泰勒级数
泰勒展开使用逐次逼近的思想,用多项式函数对函数
f
(
x
)
f(x)
f(x)在
x
0
x_0
x0的进行近似拟合,将难以研究的函数转换为简单的多项式形式
f
(
x
)
=
∑
n
=
0
∞
a
n
(
x
−
x
0
)
n
\begin{aligned} f(x) = \sum_{n=0}^{\infin} a_n (x-x_0)^n\end{aligned}
f(x)=n=0∑∞an(x−x0)n
其中 a n a_n an是常数,推导过程如下:
f
(
x
0
)
=
a
0
f
′
(
x
)
=
a
1
+
2
a
2
(
x
−
x
0
)
+
3
a
3
(
x
−
x
0
)
2
+
⋯
f
′
(
x
0
)
=
a
1
f
′
′
(
x
)
=
2
a
2
+
6
a
3
(
x
−
x
0
)
+
⋯
f
′
′
(
x
0
)
=
2
a
2
f
′
′
′
(
x
)
=
6
a
3
+
⋯
f
′
′
′
(
x
0
)
=
6
a
3
⋮
f
(
n
)
(
x
0
)
=
n
!
a
n
\begin{aligned} & f(x_0)=a_0 \\ & f'(x)= a_1 + 2a_2(x-x_0) + 3a_3(x-x_0)^2 + \cdots \\ & f'(x_0)=a_1 \\ & f''(x) = 2a_2 + 6a_3(x-x_0) + \cdots \\ & f''(x_0)=2a_2 \\ & f'''(x) = 6a_3 + \cdots \\ & f'''(x_0)=6a_3 \\ & \vdots \\ & f^{(n)}(x_0)=n!a_n \end{aligned}
f(x0)=a0f′(x)=a1+2a2(x−x0)+3a3(x−x0)2+⋯f′(x0)=a1f′′(x)=2a2+6a3(x−x0)+⋯f′′(x0)=2a2f′′′(x)=6a3+⋯f′′′(x0)=6a3⋮f(n)(x0)=n!an
得到
a
n
=
f
(
n
)
(
x
0
)
n
!
\begin{aligned} a_n=\frac{f^{(n)}(x_0)}{n!} \end{aligned}
an=n!f(n)(x0)
代入泰勒展开式得到
f
(
x
)
=
∑
n
=
0
∞
f
(
n
)
(
x
0
)
n
!
(
x
−
x
0
)
n
\begin{aligned} f(x) = \sum_{n=0}^{\infin} \frac{f^{(n)}(x_0)}{n!} (x-x_0)^n\end{aligned}
f(x)=n=0∑∞n!f(n)(x0)(x−x0)n
将
x
−
x
0
x-x_0
x−x0记为
h
h
h,则
x
=
x
0
+
h
x=x_0+h
x=x0+h,得到
f
(
x
0
+
h
)
=
∑
n
=
0
∞
f
(
n
)
(
x
0
)
n
!
h
n
\begin{aligned} f(x_0+h) = \sum_{n=0}^{\infin} \frac{f^{(n)}(x_0)}{n!} h^n\end{aligned}
f(x0+h)=n=0∑∞n!f(n)(x0)hn
再将
x
0
x_0
x0改为
x
x
x,得到泰勒级数的最终形式:
f
(
x
+
h
)
=
∑
n
=
0
∞
f
(
n
)
(
x
)
n
!
h
n
\begin{aligned} f(x+h) = \sum_{n=0}^{\infin} \frac{f^{(n)}(x)}{n!} h^n\end{aligned}
f(x+h)=n=0∑∞n!f(n)(x)hn
注意:
- 这是一个近似表达式
- 等式右边只使用函数在 x x x处的信息
常用的泰勒级数
e
x
=
∑
n
=
0
∞
x
n
n
!
\begin{aligned} e^x = \sum_{n=0}^{\infin} \frac{x^n}{n!} \end{aligned}
ex=n=0∑∞n!xn
l
n
(
1
+
x
)
=
∑
n
=
1
∞
(
−
1
)
n
+
1
x
n
n
\begin{aligned} ln(1+x) = \sum_{n=1}^{\infin} \frac{(-1)^{n+1}x^n}{n} \end{aligned}
ln(1+x)=n=1∑∞n(−1)n+1xn
期权价格的泰勒级数
偏微分
对期权价格函数 V ( S , t ) V(S,t) V(S,t),定义
∂ V ∂ S = lim δ S → 0 V ( S + δ S , t ) − V ( S , t ) δ S \begin{aligned} \frac{\partial V}{\partial S} = \lim_{\delta S \to 0} \frac{V(S+\delta S, t) - V(S,t)}{\delta S} \end{aligned} ∂S∂V=δS→0limδSV(S+δS,t)−V(S,t)
∂ V ∂ t = lim δ t → 0 V ( S , t + δ t ) − V ( S , t ) δ t \begin{aligned} \frac{\partial V}{\partial t} = \lim_{\delta t \to 0} \frac{V(S, t+\delta t) - V(S,t)}{\delta t} \end{aligned} ∂t∂V=δt→0limδtV(S,t+δt)−V(S,t)
高阶偏微分是低阶偏微分的偏微分
∂
2
V
∂
S
2
=
∂
∂
V
∂
S
∂
S
\begin{aligned} \frac{\partial ^ 2V}{\partial S^2} = \frac{\partial \frac{\partial V}{\partial S}}{\partial S} \end{aligned}
∂S2∂2V=∂S∂∂S∂V
二维泰勒展开
V ( S + δ S , t + δ t ) = V ( S , t ) + δ t ∂ V ∂ t + δ S ∂ V ∂ S + 1 2 ( δ S ) 2 ∂ 2 V ∂ S 2 + 1 2 ( δ t ) 2 ∂ 2 V ∂ t 2 + δ t δ S ∂ 2 V ∂ t ∂ S + ⋯ ≈ V ( S , t ) + δ t ∂ V ∂ t + δ S ∂ V ∂ S + 1 2 ( δ S ) 2 ∂ 2 V ∂ S 2 + ⋯ \begin{aligned} &V(S+\delta S, t+\delta t) \\ = & V(S,t) + \delta t \frac{\partial V}{\partial t} + \delta S \frac{\partial V}{\partial S} + \frac{1}{2} {(\delta S)}^2 \frac{\partial ^ 2V}{\partial S^2} + \frac{1}{2} {(\delta t)}^2 \frac{\partial ^ 2V}{\partial t^2} + \delta t \delta S \frac{\partial ^ 2V}{\partial t\partial S} + \cdots\\ \approx & V(S,t)+\delta t \frac{\partial V}{\partial t} + \delta S \frac{\partial V}{\partial S} + \frac{1}{2} {(\delta S)}^2 \frac{\partial ^ 2V}{\partial S^2} + \cdots \end{aligned} =≈V(S+δS,t+δt)V(S,t)+δt∂t∂V+δS∂S∂V+21(δS)2∂S2∂2V+21(δt)2∂t2∂2V+δtδS∂t∂S∂2V+⋯V(S,t)+δt∂t∂V+δS∂S∂V+21(δS)2∂S2∂2V+⋯
备注:由于资产价格被建模为维纳过程,展开式中 ( δ S ) 2 ∼ δ t {(\delta S)}^2 \sim \delta t (δS)2∼δt, 不能忽略,但是 δ t \delta t δt的二次及以上的高阶项, δ t δ S \delta t \delta S δtδS及高阶项, δ S \delta S δS三次及以上的高阶项被忽略。后面前向和反向方程中的 y y y也有相同的处理。
符号
Δ = ∂ V ∂ S θ = ∂ V ∂ t Γ = ∂ 2 V ∂ S 2 \begin{aligned} \Delta = \frac{\partial V}{\partial S} \\ \theta = \frac{\partial V}{\partial t} \\ \Gamma = \frac{\partial ^ 2V}{\partial S^2} \\ \end{aligned} Δ=∂S∂Vθ=∂t∂VΓ=∂S2∂2V
trinomial random walk和转移密度函数
trinomial random walk
多期trinomial树的模拟结果
最终价格分布非常接近正态分布
转移概率密度函数transition probability density function
p
(
y
,
t
;
y
′
,
t
′
)
p(y,t;y',t')
p(y,t;y′,t′)定义为
P
r
o
b
(
a
<
y
′
<
b
a
t
t
i
m
e
t
′
∣
y
a
t
t
i
m
e
t
)
=
∫
a
b
p
(
y
,
t
;
y
′
,
t
′
)
d
y
′
\begin{aligned} Prob(a<y'<b \ at \ time \ t' | y\ at\ time\ t) = \int_a^b p(y,t;y',t')dy' \end{aligned}
Prob(a<y′<b at time t′∣y at time t)=∫abp(y,t;y′,t′)dy′
t
t
t表示当前时刻
y
y
y表示当前价格
t
′
t'
t′表示未来时刻
y
′
y'
y′表示未来价格
转移概率密度函数表示,以前当前 t t t的价格为 y y y,在未来 t ′ t' t′时刻,价格 y ′ y' y′位于给定区间 [ a , b ] [a,b] [a,b]的概率
forward equation
已知:当前时刻
t
t
t的状态
y
y
y,
求解:在未来时刻
t
′
t'
t′到达状态
y
′
y'
y′的概率
p ( y , t ; y ′ , t ′ ) = α A + ( 1 − 2 α ) B + α C \begin{aligned} p(y,t;y',t') = \alpha A + (1-2\alpha) B + \alpha C \end{aligned} p(y,t;y′,t′)=αA+(1−2α)B+αC
其中
A
=
p
(
y
,
t
;
y
′
+
δ
y
,
t
′
−
δ
t
)
B
=
p
(
y
,
t
;
y
′
,
t
′
−
δ
t
)
C
=
p
(
y
,
t
;
y
′
−
δ
y
,
t
′
−
δ
t
)
A
≈
p
(
y
,
t
;
y
′
,
t
′
)
−
δ
t
∂
p
∂
t
′
+
δ
y
∂
p
∂
y
′
+
1
2
δ
y
2
∂
2
p
∂
y
′
2
+
⋯
C
≈
p
(
y
,
t
;
y
′
,
t
′
)
−
δ
t
∂
p
∂
t
′
−
δ
y
∂
p
∂
y
′
+
1
2
δ
y
2
∂
2
p
∂
y
′
2
+
⋯
B
≈
p
(
y
,
t
;
y
′
,
t
′
)
−
δ
t
∂
p
∂
t
′
+
⋯
\begin{aligned} & A = p(y,t;y'+\delta y, t'-\delta t) \\ & B = p(y,t;y', t'-\delta t) \\ & C = p(y,t;y'-\delta y, t'-\delta t) \\ & A \approx p(y,t;y',t') - \delta t \frac{\partial p}{\partial t'} + \delta y \frac{\partial p}{\partial y'} + \frac{1}{2} \delta y^2 \frac{\partial ^2p}{\partial y'^2} + \cdots \\ & C \approx p(y,t;y',t') - \delta t \frac{\partial p}{\partial t'} - \delta y \frac{\partial p}{\partial y'} + \frac{1}{2} \delta y^2 \frac{\partial ^2p}{\partial y'^2} + \cdots \\ & B \approx p(y,t;y',t') - \delta t \frac{\partial p}{\partial t'} + \cdots \\ \end{aligned}
A=p(y,t;y′+δy,t′−δt)B=p(y,t;y′,t′−δt)C=p(y,t;y′−δy,t′−δt)A≈p(y,t;y′,t′)−δt∂t′∂p+δy∂y′∂p+21δy2∂y′2∂2p+⋯C≈p(y,t;y′,t′)−δt∂t′∂p−δy∂y′∂p+21δy2∂y′2∂2p+⋯B≈p(y,t;y′,t′)−δt∂t′∂p+⋯
代入化简得到
∂
p
∂
t
′
=
α
δ
y
2
δ
t
∂
2
p
∂
y
′
2
\begin{aligned} \frac{\partial p}{\partial t'} = \frac{\alpha \delta y^2}{\delta t} \frac{\partial ^2p}{\partial y'^2} \\ \end{aligned}
∂t′∂p=δtαδy2∂y′2∂2p
当 δ t \delta t δt趋近于0时, δ y \delta y δy也趋近于0。 当 α δ y 2 δ t \begin{aligned} \frac{\alpha \delta y^2}{\delta t}\end{aligned} δtαδy2的极限趋近于有限值时,上述方程才有意义
δ y 2 δ t \begin{aligned} \frac{\delta y^2}{\delta t}\end{aligned} δtδy2 的三个场景:
- 分子比分母更快的变为0:随机过程collapse为0
- 分子比分母更慢的变为0:随机过程变为无穷
- 分子和分母的阶数相同,即 δ y 2 δ t ∼ O ( 1 ) , δ y ∼ O ( δ t ) \begin{aligned} \frac{\delta y^2}{\delta t} \sim O(1), \delta y \sim O(\sqrt{\delta t}) \end{aligned} δtδy2∼O(1),δy∼O(δt),
备注:这里似乎是个循环论证,因为能推导到这一步,已经假设过
δ
y
2
∼
δ
t
{\delta y}^2 \sim \delta t
δy2∼δt
定义
α
δ
y
2
δ
t
=
c
2
\begin{aligned} \frac{\alpha \delta y^2}{\delta t} = c^2 \\ \end{aligned}
δtαδy2=c2
c c c用于表达波动性: c 2 = 1 2 σ 2 c^2 = \frac{1}{2} \sigma ^2 c2=21σ2
最终的方程变为
∂
p
∂
t
′
=
c
2
∂
2
p
∂
y
′
2
\begin{aligned} \frac{\partial p}{\partial t'} =c^2 \frac{\partial ^2p}{\partial y'^2} \\ \end{aligned}
∂t′∂p=c2∂y′2∂2p
这个方程称为Fokker–Planck或者forward Kolmogorov方程,是一个线性抛物线偏微分方程forward parabolic partial differential equation(方程的解的线性组合仍然是方程的解),用于计算未来时刻 t ′ t' t′的 y y y的概率分布
backward equation
已知:假定未来时刻
t
′
t'
t′的状态为
y
′
y'
y′,
求解:当前时刻
t
t
t的状态为
y
y
y的概率
反向方程与隐马尔可夫链存在关联
p ( y , t ; y ′ , t ′ ) = α A + ( 1 − 2 α ) B + α C \begin{aligned} p(y,t;y',t') = \alpha A + (1-2\alpha) B + \alpha C \end{aligned} p(y,t;y′,t′)=αA+(1−2α)B+αC
其中
A
=
p
(
y
+
δ
y
,
t
+
δ
t
;
y
′
,
t
′
)
B
=
p
(
y
,
t
+
δ
t
;
y
′
,
t
′
)
C
=
p
(
y
−
δ
y
,
t
+
δ
t
;
y
′
,
t
′
)
A
≈
p
(
y
,
t
;
y
′
,
t
′
)
+
δ
t
∂
p
∂
t
′
+
δ
y
∂
p
∂
y
′
+
1
2
δ
y
2
∂
2
p
∂
y
′
2
+
⋯
C
≈
p
(
y
,
t
;
y
′
,
t
′
)
+
δ
t
∂
p
∂
t
′
−
δ
y
∂
p
∂
y
′
+
1
2
δ
y
2
∂
2
p
∂
y
′
2
+
⋯
B
≈
p
(
y
,
t
;
y
′
,
t
′
)
+
δ
t
∂
p
∂
t
′
+
⋯
\begin{aligned} & A = p(y+\delta y,t+\delta t;y', t') \\ & B = p(y,t+\delta t;y', t') \\ & C = p(y-\delta y,t+\delta t;y', t') \\ & A \approx p(y,t;y',t') + \delta t \frac{\partial p}{\partial t'} + \delta y \frac{\partial p}{\partial y'} + \frac{1}{2} \delta y^2 \frac{\partial ^2p}{\partial y'^2} + \cdots \\ & C \approx p(y,t;y',t') + \delta t \frac{\partial p}{\partial t'} - \delta y \frac{\partial p}{\partial y'} + \frac{1}{2} \delta y^2 \frac{\partial ^2p}{\partial y'^2} + \cdots \\ & B \approx p(y,t;y',t') + \delta t \frac{\partial p}{\partial t'} + \cdots \\ \end{aligned}
A=p(y+δy,t+δt;y′,t′)B=p(y,t+δt;y′,t′)C=p(y−δy,t+δt;y′,t′)A≈p(y,t;y′,t′)+δt∂t′∂p+δy∂y′∂p+21δy2∂y′2∂2p+⋯C≈p(y,t;y′,t′)+δt∂t′∂p−δy∂y′∂p+21δy2∂y′2∂2p+⋯B≈p(y,t;y′,t′)+δt∂t′∂p+⋯
代入化简得到
∂
p
∂
t
′
+
α
δ
y
2
δ
t
∂
2
p
∂
y
′
2
=
0
\begin{aligned} \frac{\partial p}{\partial t'} + \frac{\alpha \delta y^2}{\delta t} \frac{\partial ^2p}{\partial y'^2} = 0 \\ \end{aligned}
∂t′∂p+δtαδy2∂y′2∂2p=0
定义
α
δ
y
2
δ
t
=
c
2
\begin{aligned} \frac{\alpha \delta y^2}{\delta t} = c^2 \end{aligned}
δtαδy2=c2
最终的方程变为
∂
p
∂
t
′
+
c
2
∂
2
p
∂
y
′
2
=
0
\begin{aligned} \frac{\partial p}{\partial t'} + c^2 \frac{\partial ^2p}{\partial y'^2} = 0 \\ \end{aligned}
∂t′∂p+c2∂y′2∂2p=0
与前向方程仅仅只有符号上的差别
BS方程是反向方程
注意:
- 这里的y可以为负值,但是金融资产的价格,例如股价,不可能是负值
- 三叉树模型只允许 y y y往三个特定值变化,是极其简单的特例
- 不同的金融量,需要不同的模型:股价、利率
Similarity solutions 求解过程
也叫Similarity Reduction Method,通过引入新的变量并换元减少维度
大部分微分方程没有显式解(无法用初等函数表示,可以类比为大部分问题无法用编程语言的标准库解决),通常用数值解法求解,能求解的是一些形式比较特殊的微分方程。
考虑前向方程 ∂ p ∂ t ′ = c 2 ∂ 2 p ∂ y ′ 2 \begin{aligned} \frac{\partial p}{\partial t'} = c^2 \frac{\partial ^2p}{\partial y'^2} \\ \end{aligned} ∂t′∂p=c2∂y′2∂2p
微分方程的初始条件和边界条件
- 初始条件 initial conditions:how solution starts off
- 边界条件 boundary conditions:函数在给定y’时的行为
(1) 考察一个如下形式的解: p = t ′ a f ( y ′ t ′ b ) \begin{aligned} p = t'^a f(\frac{y'}{t'^b}) \end{aligned} p=t′af(t′by′), 目标是解出常数项 a , b a, b a,b和函数 f ( x ) f(x) f(x)的解析式
令
ξ
=
y
′
t
−
b
\xi = y' t^{-b}
ξ=y′t−b
用链式法则和乘法法则求导
∂
p
∂
y
′
=
t
′
a
d
f
d
ξ
d
ξ
d
y
′
=
t
′
a
−
b
d
f
d
ξ
∂
2
p
∂
y
′
2
=
t
′
a
−
2
b
d
2
f
d
ξ
2
d
p
d
t
′
=
a
t
′
a
−
1
f
(
ξ
)
+
t
′
a
d
f
d
ξ
∂
ξ
∂
t
′
=
a
t
′
a
−
1
f
(
ξ
)
−
b
y
′
t
′
a
−
b
−
1
d
f
d
ξ
\begin{aligned} \frac{\partial p}{\partial y'} = & {t'}^a \frac{df}{d \xi} \frac{d \xi}{dy'} = {t'}^{a-b} \frac{df}{d \xi} \\ \frac{\partial ^2 p}{\partial {y'}^2 } = & {t'}^{a-2b} \frac{d^2f}{d \xi^2} \\ \frac{dp}{dt'} = & a{t'}^{a-1}f(\xi) + {t'}^a \frac{df}{d\xi} \frac{\partial \xi}{\partial t'} \\ = & a{t'}^{a-1}f(\xi) -by'{t'}^{a-b-1} \frac{df}{d\xi} \\ \end{aligned}
∂y′∂p=∂y′2∂2p=dt′dp==t′adξdfdy′dξ=t′a−bdξdft′a−2bdξ2d2fat′a−1f(ξ)+t′adξdf∂t′∂ξat′a−1f(ξ)−by′t′a−b−1dξdf
(2)带入原方程,得到
a
t
′
a
−
1
f
(
ξ
)
−
b
y
′
t
′
a
−
b
−
1
d
f
d
ξ
=
c
2
t
′
a
−
2
b
d
2
f
d
ξ
2
\begin{aligned} a{t'}^{a-1}f(\xi) -by'{t'}^{a-b-1} \frac{df}{d\xi} = c^2 {t'}^{a-2b} \frac{d^2f}{d \xi^2} \end{aligned}
at′a−1f(ξ)−by′t′a−b−1dξdf=c2t′a−2bdξ2d2f
消去y’和部分t’得到
a
f
(
ξ
)
−
b
ξ
d
f
d
ξ
=
c
2
t
′
−
2
b
+
1
d
2
f
d
ξ
2
\begin{aligned} af(\xi) - b \xi \frac{df}{d\xi}=c^2{t'}^{-2b+1} \frac{d^2f}{d \xi ^2} \end{aligned}
af(ξ)−bξdξdf=c2t′−2b+1dξ2d2f
左边与 t ′ t' t′无关,说明右边应该也没有 t ′ t' t′项,所以 b = 1 2 b=\frac{1}{2} b=21,
(3) 带入b得到
a
f
(
ξ
)
−
1
2
ξ
d
f
d
ξ
=
c
2
d
2
f
d
ξ
2
\begin{aligned} af(\xi) - \frac{1}{2} \xi \frac{df}{d\xi}=c^2 \frac{d^2f}{d \xi ^2} \end{aligned}
af(ξ)−21ξdξdf=c2dξ2d2f
备注:这里没有给出推导过程
解微分方程得到原方程的一个解形式为:
p
=
t
′
a
f
(
y
′
t
′
)
\begin{aligned} p={t'}^a f(\frac{y'}{\sqrt{t'}}) \end{aligned}
p=t′af(t′y′)
因为可以选择不同的常数
a
a
a,这是一个解集
(4)考虑(问题的出发点:转移概率密度函数)
∫
−
∞
∞
p
(
y
′
,
t
′
)
d
y
′
=
1
=
∫
−
∞
∞
t
′
a
f
(
y
′
t
′
)
d
y
′
\begin{aligned} \int_{-\infin}^{\infin} p(y',t')dy' = 1 = \int_{-\infin}^{\infin} {t'}^a f(\frac{y'}{\sqrt{t'}})dy' \end{aligned}
∫−∞∞p(y′,t′)dy′=1=∫−∞∞t′af(t′y′)dy′
令
y
′
=
t
′
1
2
u
y' = {t'}^{\frac{1}{2}} u
y′=t′21u
带入得到
t
′
a
+
1
2
∫
−
∞
∞
f
(
u
)
d
u
=
1
\begin{aligned} {t'}^{a + \frac{1}{2}} \int_{-\infin}^{\infin} f(u)du=1 \end{aligned}
t′a+21∫−∞∞f(u)du=1
同样因为右边与
t
′
t'
t′无关,所以
a
=
−
1
2
a=-\frac{1}{2}
a=−21
同时有: ∫ − ∞ ∞ f ( u ) d u = 1 \begin{aligned} \int_{-\infin}^{\infin} f(u)du=1 \end{aligned} ∫−∞∞f(u)du=1
(5)带入
a
a
a得到:
−
1
2
f
(
ξ
)
−
1
2
ξ
d
f
d
ξ
=
c
2
d
2
f
d
ξ
2
\begin{aligned} -\frac{1}{2}f(\xi) - \frac{1}{2} \xi \frac{df}{d\xi}=c^2 \frac{d^2f}{d \xi ^2} \end{aligned}
−21f(ξ)−21ξdξdf=c2dξ2d2f
再次求解微分方程
−
1
2
d
(
ξ
f
)
d
ξ
=
c
2
f
′
′
→
−
1
2
ξ
f
=
c
2
f
′
+
C
o
n
s
t
\begin{aligned} -\frac{1}{2} \frac {d(\xi f)}{d\xi}=c^2 f'' \rightarrow -\frac{1}{2} \xi f = c^2 f' + Const \end{aligned}
−21dξd(ξf)=c2f′′→−21ξf=c2f′+Const
当
ξ
\xi
ξ趋向于
∞
\infin
∞时,
f
(
ξ
)
f(\xi)
f(ξ)和
f
′
(
ξ
)
f'(\xi)
f′(ξ)都趋向于0, 因此
C
o
n
s
t
Const
Const应该为0
c 2 d ( l n f ) d ξ = − 1 2 ξ → f ( ξ ) = A e − ξ 2 4 c 2 \begin{aligned} c^2 \frac{d(lnf)}{d\xi} = -\frac{1}{2} \xi \rightarrow f(\xi) = Ae^{-\frac{\xi^2}{4c^2}} \end{aligned} c2dξd(lnf)=−21ξ→f(ξ)=Ae−4c2ξ2
(6)由于 ∫ − ∞ ∞ f ( u ) d u = 1 \begin{aligned} \int_{-\infin}^{\infin} f(u)du=1 \end{aligned} ∫−∞∞f(u)du=1
最终得到
p
p
p的解析式为
p
=
1
2
c
π
t
′
e
x
p
(
−
y
′
2
4
c
2
t
′
)
\begin{aligned} p = \frac{1}{2c \sqrt{\pi t'}} exp(- \frac{{y'}^2}{4c^2t'}) \end{aligned}
p=2cπt′1exp(−4c2t′y′2)
表示 y ′ y' y′服从正态分布 y ′ ∼ N ( 0 , 2 c 2 t ′ ) y' \sim N(0, 2c^2t') y′∼N(0,2c2t′)
(7)最终解
p
(
y
,
t
;
y
′
,
t
′
)
=
1
2
c
π
(
t
′
−
t
)
e
x
p
(
−
(
y
′
−
y
)
2
4
c
2
(
t
′
−
t
)
)
\begin{aligned} p(y, t; y', t') = \frac{1}{2c \sqrt{\pi (t' - t)}} exp(- \frac{{(y' - y)}^2}{4c^2 (t'-t)}) \end{aligned}
p(y,t;y′,t′)=2cπ(t′−t)1exp(−4c2(t′−t)(y′−y)2)
也就是,如果在t时刻看t’时刻, 转移PDF的均值为 y y y, 方差为 4 c 2 ( t ′ − t ) 4c^2 (t'-t) 4c2(t′−t)
这就是trinomial random walk前向方程的转移概率密度函数
预测时间越长,方差越大(表示预测准确的概率越低)