写在前面
原论文:Distributed Time-Varying Quadratic Optimization for Multiple Agents under Undirected Graphs.
由于个人喜好问题,本文只做原论文中的Problem 2,Section 3-C,Section 3-D相关笔记。
问题描述
优化问题
min f ( x , t ) = ∑ i = 1 n f i ( x i , t ) s.t. L x = 0 , x ∈ Ω ( 1 ) \begin{aligned}
\min&\quad f(x,t)=\sum_{i=1}^n f_i(x_i,t)\\
\text{s.t.}&\quad Lx=0,\quad x\in\Omega
\end{aligned} \qquad (1)
min s.t. f ( x , t ) = i = 1 ∑ n f i ( x i , t ) L x = 0 , x ∈ Ω ( 1 )
其中x = [ x 1 T , ⋯ , x n T ] T x=[x_1^T,\cdots,x_n^T]^T x = [ x 1 T , ⋯ , x n T ] T ,Ω = Ω 1 × ⋯ × Ω n \Omega=\Omega_1\times\cdots\times \Omega_n Ω = Ω 1 × ⋯ × Ω n 。
分布式算法
改成罚函数形式
min F ( x , t ) = ∑ i = 1 n f i ( x i , t ) + 1 2 β x T L x , x ∈ Ω ( 2 ) \min F(x,t)=\sum_{i=1}^n f_i(x_i,t)+\frac{1}{2}\beta x^TLx,\quad x\in\Omega\qquad (2)
min F ( x , t ) = i = 1 ∑ n f i ( x i , t ) + 2 1 β x T L x , x ∈ Ω ( 2 )
罚函数投影算法
x ˙ i = k P Ω i ( x i − α ∂ f i ( x i , t ) ∂ x i − α β ∑ j = 1 n a i j ( x i − x j ) ) − k x i ( 3 ) \dot x_i=kP_{\Omega_i}\left(x_i-\alpha \frac{\partial f_i(x_i,t)}{\partial x_i}-\alpha\beta\sum_{j=1}^n a_{ij}(x_i-x_j) \right)-kx_i\qquad (3)
x ˙ i = k P Ω i ( x i − α ∂ x i ∂ f i ( x i , t ) − α β j = 1 ∑ n a i j ( x i − x j ) ) − k x i ( 3 )
其中α \alpha α 、k k k 是待定参数。
结论与证明
Lemma 3: For all x , y ∈ R n x,y\in\mathbb R^n x , y ∈ R n and a nonempy compact set K ⊂ R n K\subset \mathbb R^n K ⊂ R n , ∥ P K ( x ) − P K ( y ) ∥ ≤ ∥ x − y ∥ \|P_K(x)-P_K(y)\|\leq \|x-y\| ∥ P K ( x ) − P K ( y ) ∥ ≤ ∥ x − y ∥ .
Lemma 6: x ∗ ( t ) x^*(t) x ∗ ( t ) is an optimal solution of (2) if and only if for some fixed α > 0 \alpha>0 α > 0 , P Ω ( − α ∂ F ( x ∗ ( t ) , t ) ∂ x + x ∗ ( t ) ) = x ∗ ( t ) P_\Omega(-\alpha\frac{\partial F(x^*(t),t)}{\partial x}+x^*(t))=x^*(t) P Ω ( − α ∂ x ∂ F ( x ∗ ( t ) , t ) + x ∗ ( t ) ) = x ∗ ( t ) , ∀ x ∈ Ω \forall x\in\Omega ∀ x ∈ Ω , t > t 0 t>t_0 t > t 0 .
注意到Lemma 6成立,是因为0 ∈ ∂ F ( x ∗ ) + N Ω ( x ∗ ) 0\in\partial F(x^*)+N_\Omega(x^*) 0 ∈ ∂ F ( x ∗ ) + N Ω ( x ∗ ) (Lemma 2.1)。
最优条件得到了,下一步就是收敛性。
Theorem 5: The updating law in (3) guarantees ∥ x ( t ) − x ∗ ( t ) ∥ \|x(t)-x^*(t)\| ∥ x ( t ) − x ∗ ( t ) ∥ uniformly ultimately bounded as t → ∞ t\to \infty t → ∞ with ultimate bound ∥ x ( t ) − x ∗ ( t ) ∥ ≤ 1 2 k ( 1 − ε L ) − 1 ω ˉ L \|x(t)-x^*(t)\|\leq \sqrt{\frac{1}{2k(1-\varepsilon_L)-1}}\bar \omega_L ∥ x ( t ) − x ∗ ( t ) ∥ ≤ 2 k ( 1 − ε L ) − 1 1 ω ˉ L , where 0 < ε L < 1 0<\varepsilon_L<1 0 < ε L < 1 is a positive constant, ω ˉ L \bar \omega_L ω ˉ L is the upper bound of ∥ x ˙ ∗ ( t ) ∥ \|\dot x^*(t)\| ∥ x ˙ ∗ ( t ) ∥ , provided that there exist positive constants q 1 q_1 q 1 and q 2 ≥ q 1 q_2\geq q_1 q 2 ≥ q 1 such that q 1 I ≤ R ( t ) ≤ q 2 I q_1 I\leq R(t)\leq q_2I q 1 I ≤ R ( t ) ≤ q 2 I , x ˙ ∗ ( t ) \dot x^*(t) x ˙ ∗ ( t ) exists, the elements in R ( t ) R(t) R ( t ) and x ∗ ( t ) x^*(t) x ∗ ( t ) , x ˙ ∗ ( t ) \dot x^*(t) x ˙ ∗ ( t ) are bounded and the control parameters ε L , α , β \varepsilon_L, \alpha,\beta ε L , α , β and k k k are chosen such that
q 2 + β λ max ( L ) − q 1 q 2 + β λ max ( L ) + q 1 < ε L < 1 , k > 1 2 ( 1 − ε L ) , 1 − ε L q 1 < α < 1 + ε L q 2 + β λ max ( L ) . \begin{aligned}
\frac{q_2+\beta \lambda_\text{max} (L)-q_1}{q_2+\beta\lambda_\text{max}(L)+q_1}&<\varepsilon_L<1,\quad k>\frac{1}{2(1-\varepsilon_L)},\\
\frac{1-\varepsilon_L}{q_1}&<\alpha<\frac{1+\varepsilon_L}{q_2+\beta\lambda_\text{max}(L)}.
\end{aligned}
q 2 + β λ max ( L ) + q 1 q 2 + β λ max ( L ) − q 1 q 1 1 − ε L < ε L < 1 , k > 2 ( 1 − ε L ) 1 , < α < q 2 + β λ max ( L ) 1 + ε L .
证明:注意到原论文考虑二次代价函数,即
F ( x , t ) = 1 2 x T ( R ( t ) + β L ) x + s T ( t ) x + h ( t ) F(x,t) = \frac{1}{2}x^T(R(t)+\beta L)x+s^T(t)x+h(t)
F ( x , t ) = 2 1 x T ( R ( t ) + β L ) x + s T ( t ) x + h ( t )
则有
∂ F ( x , t ) ∂ x = R ( t ) x + β L x + s ( t ) \frac{\partial F(x,t)}{\partial x} = R(t)x+\beta Lx+s(t)
∂ x ∂ F ( x , t ) = R ( t ) x + β L x + s ( t )
式(3)重写为
x ˙ = k P Ω ( x − α R ( t ) x − β L x − α s ( t ) ) − x \dot x=kP_\Omega(x-\alpha R(t)x-\beta Lx-\alpha s(t))-x
x ˙ = k P Ω ( x − α R ( t ) x − β L x − α s ( t ) ) − x
定义李雅普诺夫函数V = 1 2 e T e V=\frac{1}{2}e^Te V = 2 1 e T e ,其中e = x − x ∗ ( t ) e=x-x^*(t) e = x − x ∗ ( t ) ,x ∗ ( t ) x^*(t) x ∗ ( t ) 满足P Ω ( x ∗ ( t ) − α R ( t ) x ∗ ( t ) − α β L x ∗ ( t ) − α s ( t ) ) = x ∗ ( t ) P_\Omega(x^*(t)-\alpha R(t)x^*(t)-\alpha\beta L x^*(t)-\alpha s(t))=x^*(t) P Ω ( x ∗ ( t ) − α R ( t ) x ∗ ( t ) − α β L x ∗ ( t ) − α s ( t ) ) = x ∗ ( t ) 。
那么,可以得到
P Ω ( x − α R ( t ) x − α s ( t ) ) − x = P Ω ( e + x ∗ ( t ) − α R ( t ) e − α R ( t ) x ∗ ( t ) − α β L e − α β L x ∗ ( t ) − α s ( t ) ) − P Ω ( x ∗ ( t ) − α R ( t ) x ∗ ( t ) − α β L x ∗ ( t ) − α s ( t ) ) − e \begin{aligned}
P_\Omega&(x-\alpha R(t)x-\alpha s(t))-x\\
&=P_\Omega(e+x^*(t)-\alpha R(t)e-\alpha R(t)x^*(t)-\alpha\beta Le -\alpha\beta L x^*(t) -\alpha s(t))\\
&\quad -P_\Omega(x^*(t)-\alpha R(t)x^*(t)-\alpha \beta L x^*(t) -\alpha s(t))-e
\end{aligned}
P Ω ( x − α R ( t ) x − α s ( t ) ) − x = P Ω ( e + x ∗ ( t ) − α R ( t ) e − α R ( t ) x ∗ ( t ) − α β L e − α β L x ∗ ( t ) − α s ( t ) ) − P Ω ( x ∗ ( t ) − α R ( t ) x ∗ ( t ) − α β L x ∗ ( t ) − α s ( t ) ) − e
李雅普诺夫函数对时间导数为
V ˙ = e T ( k P Ω ( x − α R ( t ) x − α β L x − α s ( t ) ) − k x − x ˙ ∗ ( t ) ) ≤ − k e T e + k ∥ e ∥ ∥ ζ ∥ + ∥ e ∥ ∥ x ˙ ∗ ( t ) ∥ \begin{aligned}
\dot V&=e^T(kP_\Omega(x-\alpha R(t)x-\alpha \beta L x-\alpha s(t))-kx-\dot x^*(t))\\
&\leq -ke^Te+k\|e\|\|\zeta\|+\|e\|\|\dot x^*(t)\|
\end{aligned}
V ˙ = e T ( k P Ω ( x − α R ( t ) x − α β L x − α s ( t ) ) − k x − x ˙ ∗ ( t ) ) ≤ − k e T e + k ∥ e ∥ ∥ ζ ∥ + ∥ e ∥ ∥ x ˙ ∗ ( t ) ∥
其中
ζ = P Ω ( e + x ∗ ( t ) − α R ( t ) e − α R ( t ) x ∗ ( t ) − α β L e − α β L x ∗ ( t ) − α s ( t ) ) − P Ω ( x ∗ ( t ) − α R ( t ) x ∗ ( t ) − α β L x ∗ ( t ) − α s ( t ) ) \begin{aligned}
\zeta &= P_\Omega(e+x^*(t)-\alpha R(t)e-\alpha R(t)x^*(t)-\alpha \beta L e-\alpha \beta L x^*(t)-\alpha s(t))\\
&\quad -P_\Omega(x^*(t)-\alpha R(t)x^*(t)-\alpha \beta L x^*(t)-\alpha s(t))
\end{aligned}
ζ = P Ω ( e + x ∗ ( t ) − α R ( t ) e − α R ( t ) x ∗ ( t ) − α β L e − α β L x ∗ ( t ) − α s ( t ) ) − P Ω ( x ∗ ( t ) − α R ( t ) x ∗ ( t ) − α β L x ∗ ( t ) − α s ( t ) )
利用Lemma 3,可得∥ ζ ∥ ≤ ∥ e − α R ( t ) e − α β L e ∥ \|\zeta\|\leq \|e-\alpha R(t)e-\alpha \beta L e\| ∥ ζ ∥ ≤ ∥ e − α R ( t ) e − α β L e ∥ ,故
V ˙ ≤ − k e T e + k ∥ e ∥ ∥ e − α R ( t ) e − α β L e ∥ + ∥ e ∥ ∥ x ˙ ∗ ( t ) ∥ ≤ − k ( 1 − 1 2 k − ∥ I − α R ( t ) − α β L ∥ ) ∥ e ∥ 2 + 1 2 ∥ x ∗ ( t ) ∥ 2 \begin{aligned}
\dot V&\leq -ke^Te+k\|e\|\|e-\alpha R(t)e-\alpha \beta L e\|+\|e\|\|\dot x^*(t)\|\\
&\leq -k\left(1-\frac{1}{2k}-\|I-\alpha R(t)-\alpha\beta L\|\right)\|e\|^2+\frac{1}{2}\|x^*(t)\|^2
\end{aligned}
V ˙ ≤ − k e T e + k ∥ e ∥ ∥ e − α R ( t ) e − α β L e ∥ + ∥ e ∥ ∥ x ˙ ∗ ( t ) ∥ ≤ − k ( 1 − 2 k 1 − ∥ I − α R ( t ) − α β L ∥ ) ∥ e ∥ 2 + 2 1 ∥ x ∗ ( t ) ∥ 2
由于1 − ε L q 1 < α < 1 + ε L q 2 + β λ max ( L ) \frac{1-\varepsilon_L}{q_1}<\alpha<\frac{1+\varepsilon_L}{q_2+\beta\lambda_\text{max}(L)} q 1 1 − ε L < α < q 2 + β λ max ( L ) 1 + ε L (注意λ min ( L ) = 0 \lambda_\text{min}(L)=0 λ min ( L ) = 0 ),可得∥ I − α R ( t ) − α β L ∥ ≤ ε L \|I-\alpha R(t)-\alpha \beta L\|\leq \varepsilon_L ∥ I − α R ( t ) − α β L ∥ ≤ ε L ,即
V ˙ ≤ − k ( 1 − 1 2 k − ε L ) ∥ e ∥ 2 + 1 2 ω ˉ L 2 \dot V\leq -k(1-\frac{1}{2k}-\varepsilon_L)\|e\|^2+\frac{1}{2}\bar \omega_L^2
V ˙ ≤ − k ( 1 − 2 k 1 − ε L ) ∥ e ∥ 2 + 2 1 ω ˉ L 2
因此V ˙ < 0 \dot V<0 V ˙ < 0 ,∀ ∥ e ∥ 2 > 1 2 k ( 1 − ε L ) − 1 ω ˉ L 2 \forall \|e\|^2>\frac{1}{2k(1-\varepsilon_L)-1}\bar \omega_L^2 ∀ ∥ e ∥ 2 > 2 k ( 1 − ε L ) − 1 1 ω ˉ L 2 。即误差一致最终有界,最终界为∥ e ∥ ≤ 1 2 k ( 1 − ε L ) − 1 ω ˉ L \|e\|\leq \sqrt{\frac{1}{2k(1-\varepsilon_L)-1}}\bar \omega_L ∥ e ∥ ≤ 2 k ( 1 − ε L ) − 1 1 ω ˉ L 。