变分法与固定端点动态优化
目录
高级宏观经济学涉及到动态优化问题,解决动态优化问题的方法包括变分法(欧拉方程),最优控制理论(变分法推广)以及动态规划(贝尔曼方程)。本文简单介绍了经济学中动态优化的事实、动态优化的类型以及固定端点类型的动态优化问题。预备知识:高等数学与最优化理论
1经济学中动态优化事实
1.1 Ramsey(1928)最优储蓄
问题1:设个人效用函数为
U
=
u
(
c
)
−
v
(
L
)
U = u(c)-v(L)
U=u(c)−v(L),
c
c
c为消费,
L
L
L为劳动供给。
k
k
k为个人财富水平,
f
(
k
,
L
)
f(k,L)
f(k,L)表示
k
k
k水平财富同时拥有
L
L
L劳动供给所得到的收入水平,它决定了财富水平的增长率
k
˙
\dot k
k˙。如果人的生命无限长,目标函数为加总一生的效用水平最大化,应如何选择消费与劳动供给路径使得
max
c
(
t
)
,
k
(
t
)
:
∫
0
+
∞
[
u
(
c
(
t
)
)
−
v
(
L
(
t
)
)
]
d
t
,
s
.
t
.
:
k
˙
=
f
(
k
,
L
)
−
c
,
k
(
0
)
=
k
0
\begin{aligned} \max _{c(t), k(t)}:\quad & \int_{0}^{+\infty}[u(c(t))-v(L(t))] \mathrm{d} t, \\ \\ \quad \mathrm{s.t.}: \quad &\dot{k}=f(k, L)-c, \quad k(0)=k_0 \end{aligned}
c(t),k(t)max:s.t.:∫0+∞[u(c(t))−v(L(t))]dt,k˙=f(k,L)−c,k(0)=k0
1.2 Eisner and Strotz(1963)最优规模扩张
问题2:企业在
t
t
t时期利润函数为
π
(
k
)
\pi(k)
π(k),
k
k
k资本存量,投资成本函数
C
(
k
˙
)
C(\dot k)
C(k˙),资本利率水平
r
r
r。企业如何选择最优投资路径使得所有期的利润贴现之和最大?即
max
k
(
t
)
:
∫
0
+
∞
[
π
(
k
(
t
)
)
−
C
(
k
˙
(
t
)
)
]
d
t
,
s.t.
:
k
(
0
)
=
k
0
\begin{aligned} \max _{k(t)}: &\int_{0}^{+\infty}[\pi(k(t))-C(\dot{k}(t))] \mathrm{d} t,\\ \\ \quad \text { s.t. }:& k(0)=k_0 \end{aligned}
k(t)max: s.t. :∫0+∞[π(k(t))−C(k˙(t))]dt,k(0)=k0
1.3 Hotelling(1931)资源最优抽取
设某种资源总量为
S
0
S_0
S0,资源社会总需求
P
(
Q
)
P(Q)
P(Q),
Q
Q
Q资源供给量,
P
P
P价格。社会供给为
Q
Q
Q时,社会价值函数
V
(
Q
)
=
∫
0
Q
P
(
x
)
d
x
V(Q) =\int_0^QP(x)dx
V(Q)=∫0QP(x)dx。资源采掘成本
C
(
Q
)
C(Q)
C(Q),
t
t
t时刻采掘
Q
Q
Q产量资源社会净价值
N
(
Q
)
=
V
(
Q
)
−
C
(
Q
)
=
∫
0
Q
P
(
x
)
d
x
−
C
(
Q
)
N(Q)=V(Q)-C(Q)=\int_{0}^{Q} P(x) \mathrm{d} x-C(Q)
N(Q)=V(Q)−C(Q)=∫0QP(x)dx−C(Q)
设利率水平为
r
r
r,该资源如何使社会总价值最大化
max
Q
(
t
)
∫
0
t
N
(
Q
(
t
)
)
e
−
r
t
d
t
,
s. t.
:
N
(
Q
)
=
∫
0
Q
P
(
x
)
d
x
−
C
(
Q
)
,
∫
0
t
Q
(
x
)
d
x
=
S
0
\begin{aligned} \max _{Q(t)} \quad &\int_{0}^{t} N(Q(t)) \mathrm{e}^{-r t} \mathrm{~d} t,\\ \text { s. t. }:\quad &N(Q)=\int_{0}^{Q} P(x) \mathrm{d} x-C(Q), \quad \int_{0}^{t} Q(x) \mathrm{d} x=S_{0} \end{aligned}
Q(t)max s. t. :∫0tN(Q(t))e−rt dt,N(Q)=∫0QP(x)dx−C(Q),∫0tQ(x)dx=S0
2 动态优化一般形式
2.1一般形式
形如
max
y
(
x
)
(
min
y
(
x
)
)
:
∫
x
1
x
2
F
(
x
,
y
,
y
′
)
d
x
,
s.t.
:
f
(
x
,
y
,
y
′
)
=
0
;
y
(
x
1
)
=
y
1
,
y
(
x
2
)
=
y
2
\begin{aligned} &\max _{y(x)}\left(\min _{y(x)}\right): \quad\int_{x_{1}}^{x_{2}} F\left(x, y, y^{\prime}\right) \mathrm{d} x,\\ \\ &\text { s.t. }: \quad f\left(x, y, y^{\prime}\right)=0 ; \quad y\left(x_{1}\right)=y_{1}, \quad y\left(x_{2}\right)=y_{2} \end{aligned}
y(x)max(y(x)min):∫x1x2F(x,y,y′)dx, s.t. :f(x,y,y′)=0;y(x1)=y1,y(x2)=y2
- 自变量 x x x:在经济学中一般为 t t t;
- 可选择变量或路径 y ( x ) y(x) y(x):在经济学中称为时间路径 y ( t ) y(t) y(t)。
- 对于每一时刻 t t t, y y y为状态变量 y ′ y' y′为控制变量
- 单期目标函数:积分里的函数 F F F,它是单期目标值同 x x x, y , y ′ y,y' y,y′之间的函数关系。
- 约束条件:即 f ( x , y , y ′ ) = 0 f(x,y,y')=0 f(x,y,y′)=0对选择路径进行约束;
- 边界条件:函数 y y y的端点值。
2.2 特点
- 目标函数是函数的函数,即泛函。目标函数值取决于函数形式
- 被选择的是一个函数,不是向量。这区别于静态优化
2.3 目标泛函形式
- 一般形式或标准形式
{ V ( y ) = ∫ t 1 t 2 F ( t , y ( t ) , y ˙ ( t ) ) d t V ( y ) = ∫ t 1 t 2 F ( t , y ( t ) , y ˙ ( t ) , z ( t ) , z ˙ ( t ) ) d t \left\{\begin{array}{l} V(y)=\int_{t_{1}}^{t_{2}} F(t, y(t), \dot{y}(t)) \mathrm{d} t\\ \\ V(y)=\int_{t_{1}}^{t_{2}} F(t, y(t), \dot{y}(t), z(t), \dot{z}(t)) \mathrm{d} t \end{array}\right. ⎩ ⎨ ⎧V(y)=∫t1t2F(t,y(t),y˙(t))dtV(y)=∫t1t2F(t,y(t),y˙(t),z(t),z˙(t))dt
- 迈耶问题
{ V ( y ) = G ( T , y ( T ) ) V ( y ) = G ( T , y ( T ) , z ( T ) ) \left\{\begin{array}{l} V(y)=G(T, y(T)) \\ \\ V(y)=G(T, y(T), z(T)) \end{array}\right. ⎩ ⎨ ⎧V(y)=G(T,y(T))V(y)=G(T,y(T),z(T))
- 博尔扎问题
{ V ( y ) = ∫ 0 T F ( t , y , y ˙ ) d t + G ( T , y ( T ) ) V ( y ) = ∫ 0 T F ( t , y , y ˙ , z , z ˙ ) d t + G ( T , y ( T ) , z ( T ) ) \left\{\begin{array}{l} V(y)=\int_{0}^{T} F(t, y, \dot{y}) \mathrm{d} t+G(T, y(T)) \\ \\ V(y)=\int_{0}^{T} F(t, y, \dot{y}, z, \dot{z}) \mathrm{d} t+G(T, y(T), z(T)) \end{array}\right. ⎩ ⎨ ⎧V(y)=∫0TF(t,y,y˙)dt+G(T,y(T))V(y)=∫0TF(t,y,y˙,z,z˙)dt+G(T,y(T),z(T))
- 双重积分问题
V ( y , z ) = ∫ s 1 s 1 ∫ t 1 t 2 F ( t , s , , y ( t , s ) y t ( t , s ) , y s ( t , s ) ) d t d s V(y, z)=\int_{s_{1}}^{s_{1}} \int_{t_{1}}^{t_{2}} F\left(t, s,, y(t, s) y_{t}(t, s), y_{s}(t, s)\right) \mathrm{d} t \mathrm{~d} s V(y,z)=∫s1s1∫t1t2F(t,s,,y(t,s)yt(t,s),ys(t,s))dt ds
- 高阶导数问题
{ V ( y ) = ∫ 0 T F ( t , y ( t ) , y ˙ ( t ) , y ¨ ( t ) ) d t V ( y ) = ∫ 0 T F ( t , y ( t ) , y ˙ ( t ) , y ¨ ( t ) , ⋯ , y ( n ) ( t ) ) d t \left\{\begin{array}{l} V(y)=\int_{0}^{T} F(t, y(t), \dot{y}(t), \ddot{y}(t)) \mathrm{d} t \\ \\ V(y)=\int_{0}^{T} F\left(t, y(t), \dot{y}(t), \ddot{y}(t), \cdots, y^{(n)}(t)\right) \mathrm{d} t \end{array}\right. ⎩ ⎨ ⎧V(y)=∫0TF(t,y(t),y˙(t),y¨(t))dtV(y)=∫0TF(t,y(t),y˙(t),y¨(t),⋯,y(n)(t))dt
2.4 边界约束条件
-
固定端点
-
垂直终结性条件
-
水平终结性条件
-
终结性曲线
-
截断垂直性终结
-
阶段水平性终结
以上六种边界约束分别对应下图
2.5 非边界约束条件
- 无任何约束条件;
- 应变量代数等式约束 y = τ ( t , z ) y =\tau(t,z) y=τ(t,z);
- 微分等式约束非边界约束条件 g ( t , y , y ˙ ) = 0 g(t,y,\dot y)=0 g(t,y,y˙)=0
- 不等式约束方程非边界约束条件 g ( t , y , y ˙ ) ≥ ( ≤ ) 0 g(t,y,\dot y)\ge (\le)0 g(t,y,y˙)≥(≤)0;
- 等周问题的非边界约束条件 ∫ t 1 t 2 g ( t , y , y ˙ ) d t = L \int_{t_{1}}^{t_{2}} g(t, y, \dot{y}) \mathrm{d} t=L ∫t1t2g(t,y,y˙)dt=L
3 固定端点动态优化
形式:
max
x
(
t
)
∫
t
1
t
2
F
(
t
,
x
(
t
)
,
x
˙
(
t
)
)
d
t
,
s.t. :
x
(
t
1
)
=
x
1
,
x
(
t
2
)
=
x
2
\begin{aligned} \max _{x(t)} \int_{t_{1}}^{t_{2}} F(t, x(t), \dot{x}(t)) \mathrm{d} t,\\ \\ \quad \text { s.t. : } \quad x\left(t_{1}\right)=x_{1},x\left(t_{2}\right)=x_{2} \end{aligned}
x(t)max∫t1t2F(t,x(t),x˙(t))dt, s.t. : x(t1)=x1,x(t2)=x2
3.1 一阶必要条件:欧拉公式
设
x
∗
(
t
)
x^{*}(t)
x∗(t)为为最优解则
{
x
∗
(
t
1
)
=
x
(
t
1
)
=
x
1
;
x
∗
(
t
2
)
=
x
(
t
2
)
=
x
2
;
\left\{\begin{array}{l} x^{*}(t_1) =x(t_1)=x_1;\\ \\ x^{*}(t_2) =x(t_2)=x_2; \end{array}\right.
⎩
⎨
⎧x∗(t1)=x(t1)=x1;x∗(t2)=x(t2)=x2;
定义变差
h
(
t
)
h(t)
h(t)
{
h
(
t
)
=
x
(
t
)
−
x
∗
(
t
)
h
(
t
1
)
=
h
(
t
2
)
=
0
\left\{\begin{array}{l} h(t)=x(t)-x^{*}(t)\\ \\ h(t_1)=h(t_2)=0 \end{array}\right.
⎩
⎨
⎧h(t)=x(t)−x∗(t)h(t1)=h(t2)=0
如果
x
(
t
)
=
x
∗
(
t
)
+
h
(
t
)
x(t) =x^{*}(t)+h(t)
x(t)=x∗(t)+h(t)是可行的,则
h
(
t
)
h(t)
h(t)是一条可行变分;对
∀
R
\forall R
∀R,
{
y
(
t
)
=
x
∗
(
t
)
+
a
h
(
t
)
y
(
t
1
)
=
x
1
y
(
t
2
)
=
x
2
\left\{\begin{array}{l} y(t)=x^{*}(t)+a h(t)\\ y\left(t_{1}\right)=x_{1}\\ y\left(t_{2}\right)=x_{2} \end{array}\right.
⎩
⎨
⎧y(t)=x∗(t)+ah(t)y(t1)=x1y(t2)=x2
仍然可行。若
x
∗
(
t
)
x^{*}(t)
x∗(t)与
h
(
t
)
h(t)
h(t)固定,则目标泛函为
a
a
a的函数
g
(
a
)
=
∫
t
1
t
2
F
(
t
,
y
(
t
)
,
y
˙
(
t
)
)
d
t
=
∫
t
1
t
2
F
(
t
,
x
∗
(
t
)
+
a
h
(
t
)
,
x
˙
∗
(
t
)
+
a
h
˙
(
t
)
)
d
t
\begin{aligned} g(a) &=\int_{t_{1}}^{t_{2}} F(t, y(t), \dot{y}(t)) \mathrm{d} t \\ \\ &=\int_{t_{1}}^{t_{2}} F\left(t, x^{*}(t)+a h(t), \dot{x}^{*}(t)+a \dot{h}(t)\right) \mathrm{d} t \end{aligned}
g(a)=∫t1t2F(t,y(t),y˙(t))dt=∫t1t2F(t,x∗(t)+ah(t),x˙∗(t)+ah˙(t))dt
因为
x
∗
(
t
)
x^{*}(t)
x∗(t)为最优值,则对
∀
h
(
t
)
\forall h(t)
∀h(t),
g
′
(
a
)
∣
a
=
0
=
0
g'(a)|_{a=0}=0
g′(a)∣a=0=0。由于
g
′
(
a
)
=
∫
t
1
t
2
[
d
F
(
t
,
y
(
t
)
,
y
˙
(
t
)
)
/
d
a
]
d
t
=
∫
t
1
t
2
[
h
(
t
)
F
x
(
t
,
y
(
t
)
,
y
˙
(
t
)
)
+
h
˙
(
t
)
F
x
˙
(
t
,
y
(
t
)
,
y
˙
(
t
)
)
]
d
t
\begin{aligned} g^{\prime}(a) &=\int_{t_{1}}^{t_{2}}[\mathrm{~d} F(t, y(t), \dot{y}(t)) / \mathrm{d} a] \mathrm{d} t \\ \\ &=\int_{t_{1}}^{t_{2}}\left[h(t) F_{x}(t, y(t), \dot{y}(t))+\dot{h}(t) F_{\dot{x}}(t, y(t), \dot{y}(t))\right] \mathrm{d} t \end{aligned}
g′(a)=∫t1t2[ dF(t,y(t),y˙(t))/da]dt=∫t1t2[h(t)Fx(t,y(t),y˙(t))+h˙(t)Fx˙(t,y(t),y˙(t))]dt
故
g
′
(
0
)
=
∫
t
1
t
2
[
h
(
t
)
F
x
(
t
,
x
∗
,
x
˙
∗
)
+
h
˙
(
t
)
F
x
˙
(
t
,
x
∗
,
x
˙
∗
)
]
d
t
=
∫
t
1
t
2
[
h
(
t
)
F
x
(
t
,
x
∗
,
x
˙
∗
)
]
d
t
+
∫
t
1
t
2
[
F
x
˙
(
t
,
x
∗
,
x
˙
∗
)
]
d
h
(
t
)
=
∫
t
1
t
2
[
h
(
t
)
F
x
(
t
,
x
∗
,
x
˙
∗
)
]
d
t
+
h
(
t
)
F
x
˙
(
t
,
x
∗
,
x
˙
∗
)
∣
t
1
t
2
−
∫
t
1
t
2
h
(
t
)
d
(
F
x
˙
(
t
,
x
∗
,
x
˙
∗
)
)
=
∫
t
1
t
2
[
h
(
t
)
(
F
x
(
t
,
x
∗
,
x
˙
∗
)
−
d
F
x
˙
(
t
,
x
∗
,
x
˙
∗
)
/
d
t
)
]
d
t
\begin{aligned} g^{\prime}(0) &=\int_{t_{1}}^{t_{2}}\left[h(t) F_{x}\left(t, x^{*}, \dot{x}^{*}\right)+\dot{h}(t) F_{\dot{x}}\left(t, x^{*}, \dot{x}^{*}\right)\right] \mathrm{d} t \\ \\ &=\int_{t_{1}}^{t_{2}}\left[h(t) F_{x}\left(t, x^{*}, \dot{x}^{*}\right)\right] \mathrm{d} t+\int_{t_{1}}^{t_{2}}\left[F_{\dot{x}}\left(t, x^{*}, \dot{x}^{*}\right)\right] \mathrm{d} h(t) \\ \\ &=\int_{t_{1}}^{t_{2}}\left[h(t) F_{x}\left(t, x^{*}, \dot{x}^{*}\right)\right] \mathrm{d} t+\left.h(t) F_{\dot{x}}\left(t, x^{*}, \dot{x}^{*}\right)\right|_{t_{1}} ^{t_{2}}-\int_{t_{1}}^{t_{2}} h(t) \mathrm{d}\left(F_{\dot{x}}\left(t, x^{*}, \dot{x}^{*}\right)\right) \\ \\ &=\int_{t_{1}}^{t_{2}}\left[h(t)\left(F_{x}\left(t, x^{*}, \dot{x}^{*}\right)-\mathrm{d} F_{\dot{x}}\left(t, x^{*}, \dot{x}^{*}\right) / \mathrm{d} t\right)\right] \mathrm{d} t \end{aligned}
g′(0)=∫t1t2[h(t)Fx(t,x∗,x˙∗)+h˙(t)Fx˙(t,x∗,x˙∗)]dt=∫t1t2[h(t)Fx(t,x∗,x˙∗)]dt+∫t1t2[Fx˙(t,x∗,x˙∗)]dh(t)=∫t1t2[h(t)Fx(t,x∗,x˙∗)]dt+h(t)Fx˙(t,x∗,x˙∗)∣t1t2−∫t1t2h(t)d(Fx˙(t,x∗,x˙∗))=∫t1t2[h(t)(Fx(t,x∗,x˙∗)−dFx˙(t,x∗,x˙∗)/dt)]dt
因此,若
x
∗
(
t
)
x^{*}(t)
x∗(t)为最优解,则对于
∀
h
(
t
)
\forall h(t)
∀h(t)都有
∫
t
1
t
2
[
h
(
t
)
(
F
x
(
t
,
x
∗
,
x
˙
∗
)
−
d
F
x
˙
(
t
,
x
∗
,
x
˙
∗
)
/
d
t
)
]
d
t
=
0
\int_{t_{1}}^{t_{2}}\left[h(t)\left(F_{x}\left(t, x^{*}, \dot{x}^{*}\right)-\mathrm{d} F_{\dot{x}}\left(t, x^{*}, \dot{x}^{*}\right) / \mathrm{d} t\right)\right] \mathrm{d} t=0
∫t1t2[h(t)(Fx(t,x∗,x˙∗)−dFx˙(t,x∗,x˙∗)/dt)]dt=0
借助如下面引理,则可推出一阶条件欧拉公式:
A1:设 φ ( t ) \varphi(t) φ(t)在 [ t 0 , t 1 ] [t_0,t_1] [t0,t1]上连续,对于任意在 [ t 0 , t 1 ] [t_0,t_1] [t0,t1]上连续的函数 h ( t ) h(t) h(t),且 h ( t 0 ) = h ( t 1 ) = 0 h(t_0)=h(t_1)=0 h(t0)=h(t1)=0,都有
∫ t 0 t 1 [ h ( t ) φ ( t ) ] d t = 0 \int_{t_{0}}^{t_{1}}[h(t) \varphi(t)] \mathrm{d} t=0 ∫t0t1[h(t)φ(t)]dt=0
根据引理A1,得到著名的欧拉公式:
F
x
(
t
,
x
∗
,
x
˙
∗
)
=
d
F
x
˙
(
t
,
x
∗
,
x
˙
∗
)
/
d
t
F_{x}\left(t, x^{*}, \dot{x}^{*}\right)=\mathrm{d} F_{\dot{x}}\left(t, x^{*}, \dot{x}^{*}\right) / \mathrm{d} t
Fx(t,x∗,x˙∗)=dFx˙(t,x∗,x˙∗)/dt
3.2 二阶必要条件:勒让德条件
x
∗
(
t
)
x^{*}(t)
x∗(t)取最值,则意味着对于任意可行的
h
(
t
)
h(t)
h(t)都有
g
′
(
a
)
∣
a
=
0
=
0
g'(a)|_{a=0}=0
g′(a)∣a=0=0。当
g
′
′
(
0
)
≤
0
g''(0)\le 0
g′′(0)≤0为最大值;当
g
′
′
(
0
)
≥
0
g''(0)\ge 0
g′′(0)≥0为最小值。
g
′
′
(
a
)
=
∫
t
1
t
2
[
d
F
x
(
t
,
y
,
y
˙
)
d
a
h
(
t
)
+
d
F
x
˙
(
t
,
y
,
y
˙
)
d
a
h
˙
(
t
)
]
d
t
=
∫
t
1
t
2
[
F
x
x
(
t
,
y
,
y
˙
)
(
h
(
t
)
)
2
+
2
F
x
˙
x
(
t
,
y
,
y
˙
)
h
(
t
)
h
˙
(
t
)
+
F
x
˙
x
˙
(
t
,
y
,
y
˙
)
(
h
˙
(
t
)
)
2
]
d
t
\begin{aligned} g^{\prime \prime}(a)=& \int_{t_{1}}^{t_{2}}\left[\frac{\mathrm{d} F_{x}(t, y, \dot{y})}{\mathrm{d} a} h(t)+\frac{\mathrm{d} F_{\dot{x}}(t, y, \dot{y})}{\mathrm{d} a} \dot{h}(t)\right] \mathrm{d} t \\ \\ =& \int_{t_{1}}^{t_{2}}\left[F_{x x}(t, y, \dot{y})(h(t))^{2}+2 F_{\dot{x} x}(t, y, \dot{y}) h(t) \dot{h}(t)+\right. \left.F_{\dot{x} \dot{x}}(t, y, \dot{y})(\dot{h}(t))^{2}\right] \mathrm{d} t \end{aligned}
g′′(a)==∫t1t2[dadFx(t,y,y˙)h(t)+dadFx˙(t,y,y˙)h˙(t)]dt∫t1t2[Fxx(t,y,y˙)(h(t))2+2Fx˙x(t,y,y˙)h(t)h˙(t)+Fx˙x˙(t,y,y˙)(h˙(t))2]dt
故
g
′
′
(
0
)
=
∫
t
1
t
2
[
F
x
x
(
t
,
x
∗
,
x
˙
∗
)
(
h
(
t
)
)
2
+
2
F
x
˙
x
(
t
,
x
∗
,
x
˙
∗
)
h
(
t
)
h
˙
(
t
)
+
F
x
˙
x
˙
(
t
,
x
∗
,
x
˙
∗
)
(
h
˙
(
t
)
)
2
]
d
t
\begin{aligned} g^{\prime \prime}(0)=& \int_{t_{1}}^{t_{2}}\left[F_{x x}\left(t, x^{*}, \dot{x}^{*}\right)(h(t))^{2}+2 F_{\dot{x} x}\left(t, x^{*}, \dot{x}^{*}\right) h(t) \dot{h}(t)+\right.\left.F_{\dot{x} \dot{x}}\left(t, x^{*}, \dot{x}^{*}\right)(\dot{h}(t))^{2}\right] \mathrm{d} t \end{aligned}
g′′(0)=∫t1t2[Fxx(t,x∗,x˙∗)(h(t))2+2Fx˙x(t,x∗,x˙∗)h(t)h˙(t)+Fx˙x˙(t,x∗,x˙∗)(h˙(t))2]dt
又
∫
t
1
t
2
[
2
F
x
˙
x
(
t
,
x
∗
,
x
˙
∗
)
h
(
t
)
h
˙
(
t
)
]
d
t
=
∫
t
1
t
2
F
x
˙
x
(
t
,
x
∗
,
x
˙
∗
)
d
(
h
(
t
)
)
2
=
(
h
(
t
)
)
2
F
x
˙
x
(
t
,
x
∗
,
x
˙
∗
)
∣
t
1
t
2
−
∫
t
1
t
2
(
h
(
t
)
)
2
d
F
x
˙
x
(
t
,
x
∗
,
x
˙
∗
)
d
t
d
t
=
−
∫
t
1
t
2
(
h
(
t
)
)
2
d
F
x
˙
x
(
t
,
x
∗
,
x
˙
∗
)
d
t
d
t
\begin{aligned} & \int_{t_{1}}^{t_{2}}\left[2 F_{\dot{x} x}\left(t, x^{*}, \dot{x}^{*}\right) h(t) \dot{h}(t)\right] \mathrm{d} t \\ \\ =& \int_{t_{1}}^{t_{2}} F_{\dot{x} x}\left(t, x^{*}, \dot{x}^{*}\right) \mathrm{d}(h(t))^{2} \\ \\ =&\left.(h(t))^{2} F_{\dot{x} x}\left(t, x^{*}, \dot{x}^{*}\right)\right|_{t_{1}} ^{t_{2}}-\int_{t_{1}}^{t_{2}}(h(t))^{2} \frac{\mathrm{d} F_{\dot{x} x}\left(t, x^{*}, \dot{x}^{*}\right)}{\mathrm{d} t} \mathrm{~d} t \\ \\ =&-\int_{t_{1}}^{t_{2}}(h(t))^{2} \frac{\mathrm{d} F_{\dot{x} x}\left(t, x^{*}, \dot{x}^{*}\right)}{\mathrm{d} t} \mathrm{~d} t \end{aligned}
===∫t1t2[2Fx˙x(t,x∗,x˙∗)h(t)h˙(t)]dt∫t1t2Fx˙x(t,x∗,x˙∗)d(h(t))2(h(t))2Fx˙x(t,x∗,x˙∗)
t1t2−∫t1t2(h(t))2dtdFx˙x(t,x∗,x˙∗) dt−∫t1t2(h(t))2dtdFx˙x(t,x∗,x˙∗) dt
故
g
′
′
(
0
)
=
∫
t
1
t
2
{
[
F
x
x
(
t
,
x
∗
,
x
˙
∗
)
−
d
F
x
˙
x
(
t
,
x
∗
,
x
˙
∗
)
d
t
]
(
h
(
t
)
)
2
+
F
x
˙
x
˙
(
t
,
x
∗
,
x
˙
∗
)
(
h
˙
(
t
)
)
2
}
d
t
g^{\prime \prime}(0)=\int_{t_{1}}^{t_{2}}\left\{\left[F_{x x}\left(t, x^{*}, \dot{x}^{*}\right)-\frac{\mathrm{d} F_{\dot{x} x}\left(t, x^{*}, \dot{x}^{*}\right)}{\mathrm{d} t}\right](h(t))^{2}+F_{\dot{x} \dot{x}}\left(t, x^{*}, \dot{x}^{*}\right)(\dot{h}(t))^{2}\right\} \mathrm{d} t
g′′(0)=∫t1t2{[Fxx(t,x∗,x˙∗)−dtdFx˙x(t,x∗,x˙∗)](h(t))2+Fx˙x˙(t,x∗,x˙∗)(h˙(t))2}dt
对于动态优化问题,
max
x
(
t
)
∫
t
1
t
2
F
(
t
,
x
(
t
)
,
x
˙
(
t
)
)
d
t
,
s.t. :
x
(
t
1
)
=
x
1
,
x
(
t
2
)
=
x
2
\begin{aligned} \max _{x(t)} \int_{t_{1}}^{t_{2}} F(t, x(t), \dot{x}(t)) \mathrm{d} t,\\ \\ \quad \text { s.t. : } \quad x\left(t_{1}\right)=x_{1},x\left(t_{2}\right)=x_{2} \end{aligned}
x(t)max∫t1t2F(t,x(t),x˙(t))dt, s.t. : x(t1)=x1,x(t2)=x2
必要条件为
- 一阶条件欧拉方程
F x ( t , x ∗ , x ˙ ∗ ) = d F x ˙ ( t , x ∗ , x ˙ ∗ ) / d t F_{x}\left(t, x^{*}, \dot{x}^{*}\right)=\mathrm{d} F_{\dot{x}}\left(t, x^{*}, \dot{x}^{*}\right) / \mathrm{d} t Fx(t,x∗,x˙∗)=dFx˙(t,x∗,x˙∗)/dt
- 二阶条件勒让德条件
F x ˙ x ˙ ( t , x ∗ ( t ) , x ˙ ∗ ( t ) ) ⩽ 0 , ∀ t ∈ [ t 1 , t 2 ] , 如果 x ∗ ( t ) 是最大值解 ; F x ˙ x ˙ ( t , x ∗ ( t ) , x ˙ ∗ ( t ) ) ⩾ 0 , ∀ t ∈ [ t 1 , t 2 ] , 如果 x ∗ ( t ) 是最小值解。 \begin{aligned} &F_{\dot{x} \dot{x}}\left(t, x^{*}(t), \dot{x}^{*}(t)\right) \leqslant 0, \forall t \in\left[t_{1}, t_{2}\right], \text { 如果 } x^{*}(t) \text { 是最大值解 } ;\\ &F_{\dot{x} \dot{x}}\left(t, x^{*}(t), \dot{x}^{*}(t)\right) \geqslant 0, \forall t \in\left[t_{1}, t_{2}\right], \text { 如果 } x^{*}(t) \text { 是最小值解。 } \end{aligned} Fx˙x˙(t,x∗(t),x˙∗(t))⩽0,∀t∈[t1,t2], 如果 x∗(t) 是最大值解 ;Fx˙x˙(t,x∗(t),x˙∗(t))⩾0,∀t∈[t1,t2], 如果 x∗(t) 是最小值解。
- 边界条件
x ( t 1 ) = x 1 , x ( t 2 ) = x 2 x\left(t_{1}\right)=x_{1}, \quad x\left(t_{2}\right)=x_{2} x(t1)=x1,x(t2)=x2
3.3 最优路径充分条件
对于动态优化问题
max
x
(
t
)
∫
t
1
t
2
F
(
t
,
x
(
t
)
,
x
˙
(
t
)
)
d
t
,
s.t. :
x
(
t
1
)
=
x
1
,
x
(
t
2
)
=
x
2
\begin{aligned} \max _{x(t)} \int_{t_{1}}^{t_{2}} F(t, x(t), \dot{x}(t)) \mathrm{d} t,\\ \\ \quad \text { s.t. : } \quad x\left(t_{1}\right)=x_{1},x\left(t_{2}\right)=x_{2} \end{aligned}
x(t)max∫t1t2F(t,x(t),x˙(t))dt, s.t. : x(t1)=x1,x(t2)=x2
若函数
F
(
t
,
x
(
t
)
,
x
˙
(
t
)
)
F(t, x(t), \dot{x}(t))
F(t,x(t),x˙(t))关于
x
(
t
)
x(t)
x(t)与
x
˙
(
t
)
\dot x(t)
x˙(t)是凹函数,且
x
∗
(
t
)
x^{*}(t)
x∗(t)满足
-
欧拉方程: F x ( t , x ∗ , x ˙ ∗ ) = d F x ˙ ( t , x ∗ , x ˙ ∗ ) / d t F_{x}\left(t, x^{*}, \dot{x}^{*}\right)=\mathrm{d} F_{\dot{x}}\left(t, x^{*}, \dot{x}^{*}\right) / \mathrm{d} t Fx(t,x∗,x˙∗)=dFx˙(t,x∗,x˙∗)/dt
-
边界条件: x ( t 1 ) = x 1 , x ( t 2 ) = x 2 x\left(t_{1}\right)=x_{1}, \quad x\left(t_{2}\right)=x_{2} x(t1)=x1,x(t2)=x2
则 x ∗ ( t ) x^{*}(t) x∗(t)为该优化问题的最大值;
若函数 F ( t , x ( t ) , x ˙ ( t ) ) F(t, x(t), \dot{x}(t)) F(t,x(t),x˙(t))关于 x ( t ) x(t) x(t)与 x ˙ ( t ) \dot x(t) x˙(t)是凸函数,且 x ∗ ( t ) x^{*}(t) x∗(t)满足
-
欧拉方程: F x ( t , x ∗ , x ˙ ∗ ) = d F x ˙ ( t , x ∗ , x ˙ ∗ ) / d t F_{x}\left(t, x^{*}, \dot{x}^{*}\right)=\mathrm{d} F_{\dot{x}}\left(t, x^{*}, \dot{x}^{*}\right) / \mathrm{d} t Fx(t,x∗,x˙∗)=dFx˙(t,x∗,x˙∗)/dt
-
边界条件: x ( t 1 ) = x 1 , x ( t 2 ) = x 2 x\left(t_{1}\right)=x_{1}, \quad x\left(t_{2}\right)=x_{2} x(t1)=x1,x(t2)=x2
则 x ∗ ( t ) x^{*}(t) x∗(t)为该优化问题的最小值;
4 例子
求解如下最优化
min
x
(
t
)
:
∫
0
T
(
x
˙
(
t
)
)
2
d
t
,
s.t.
:
x
(
0
)
=
0
,
x
(
T
)
=
B
\min _{x(t)}: \int_{0}^{\mathrm{T}}(\dot{x}(t))^{2} \mathrm{~d} t, \quad \text { s.t. }: x(0)=0, \quad x(T)=\mathrm{B}
x(t)min:∫0T(x˙(t))2 dt, s.t. :x(0)=0,x(T)=B
其中
T
,
B
T,B
T,B已知。
解:根据欧拉公式有
d
F
x
˙
(
t
,
x
∗
,
x
˙
∗
)
/
d
t
=
F
x
(
t
,
x
∗
,
x
˙
∗
)
=
0
⇔
F
x
˙
(
t
,
x
∗
,
x
˙
∗
)
=
2
x
˙
(
t
)
=
C
1
⇔
x
(
t
)
=
C
1
t
+
C
2
\begin{aligned} &\mathrm{d} F_{\dot{x}}\left(t, x^{*}, \dot{x}^{*}\right) / \mathrm{d} t=F_{x}\left(t, x^{*}, \dot{x}^{*}\right)=0\\ \\ &\Leftrightarrow F_{\dot{x}}\left(t, x^{*}, \dot{x}^{*}\right)=2 \dot{x}(t)=C_{1}\\ \\ &\Leftrightarrow x(t)=C_{1} t+C_{2} \end{aligned}
dFx˙(t,x∗,x˙∗)/dt=Fx(t,x∗,x˙∗)=0⇔Fx˙(t,x∗,x˙∗)=2x˙(t)=C1⇔x(t)=C1t+C2
根据边界条件得到
x
∗
(
t
)
=
B
t
/
T
,
C
1
=
B
t
/
T
,
C
2
=
0
x^{*}(t)=B t / T, \quad C_{1}=B t / T, \quad C_{2}=0
x∗(t)=Bt/T,C1=Bt/T,C2=0
二阶条件
F
x
˙
x
˙
=
2
>
0
F_{\dot{x} \dot{x}}=2>0
Fx˙x˙=2>0
故
x
∗
(
t
)
=
B
t
/
T
x^{*}(t)=B t / T
x∗(t)=Bt/T为最小值解
根据充分条件
F
x
x
=
F
x
˙
x
=
0
,
F
x
˙
x
˙
=
2
>
0
F_{x x}=F_{\dot{x} x}=0, F_{\dot{x} \dot{x}}=2>0
Fxx=Fx˙x=0,Fx˙x˙=2>0
故
F
F
F是关于
x
(
t
)
x(t)
x(t)与
x
˙
(
t
)
\dot x(t)
x˙(t)的凸函数,所求解为最小值解。
参考文献
王弟海 . 经济学中的优化方法 [M]. 清华大学出版社,2012