【论文笔记】无向图下的分布式时变二次优化(Quadratic Optimization)

写在前面

原论文：Distributed Time-Varying Quadratic Optimization for Multiple Agents under Undirected Graphs¹.

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
$$\begin{array}{rl}\min & f(x,t)=\sum _{i=1}^n{f}_i(x_i,t)\\ \text{s.t.}& Lx=0,x\in \Omega \end{array}(1)$$
其中$x=[x_1^T,\cdots ,x_n^T{]}^T$，$\Omega ={\Omega }_1\times \cdots \times {\Omega }_n$。

分布式算法

改成罚函数形式
$$\min F(x,t)=\sum _{i=1}^n{f}_i(x_i,t)+\frac{1}{2}\beta x^{T}Lx,x\in \Omega (2)$$
罚函数投影算法
$${\dot{x}}_i=k{P}_{{\Omega }_i}\left(x_i-\alpha \frac{\partial f_i(x_i,t)}{\partial x_i}-\alpha \beta \sum _{j=1}^n{a}_{ij}(x_i-x_j)\right)-k{x}_i(3)$$
其中$\alpha$、$k$是待定参数。

结论与证明

Lemma 3²: For all $x,y\in {\mathbb{R}}^n$ and a nonempy compact set $K\subset {\mathbb{R}}^n$, $\Vert P_K(x)-P_K(y)\Vert \le \Vert x-y\Vert$.

Lemma 6: $x^{\ast }(t)$ is an optimal solution of (2) if and only if for some fixed $\alpha >0$, $P_{\Omega }(-\alpha \frac{\partial F(x^{\ast }(t),t)}{\partial x}+x^{\ast }(t))=x^{\ast }(t)$, $\forall x\in \Omega$, $t>t_0$.

注意到Lemma 6成立，是因为$0\in \partial F(x^{\ast })+N_{\Omega }(x^{\ast })$（Lemma 2.1³）。

最优条件得到了，下一步就是收敛性。

Theorem 5: The updating law in (3) guarantees $\Vert x(t)-x^{\ast }(t)\Vert$ uniformly ultimately bounded as $t\to \infty$ with ultimate bound $\Vert x(t)-x^{\ast }(t)\Vert \le \sqrt{\frac{1}{2k(1-{\epsilon }_L)-1}}{\bar{\omega }}_L$, where $0<{\epsilon }_L<1$ is a positive constant, ${\bar{\omega }}_L$ is the upper bound of $\Vert {\dot{x}}^{\ast }(t)\Vert$, provided that there exist positive constants $q_1$ and $q_2\ge q_1$ such that $q_{1}I\le R(t)\le q_{2}I$, ${\dot{x}}^{\ast }(t)$ exists, the elements in $R(t)$ and $x^{\ast }(t)$, ${\dot{x}}^{\ast }(t)$ are bounded and the control parameters ${\epsilon }_L,\alpha ,\beta$ and $k$ are chosen such that
$$\begin{array}{rl}\frac{q_2+\beta {\lambda }_{\text{max}}(L)-q_1}{q_2+\beta {\lambda }_{\text{max}}(L)+q_1}& <{\epsilon }_L<1,k>\frac{1}{2(1-{\epsilon }_L)},\\ \frac{1-{\epsilon }_L}{q_1}& <\alpha <\frac{1+{\epsilon }_L}{q_2+\beta {\lambda }_{\text{max}}(L)}.\end{array}$$

证明：注意到原论文考虑二次代价函数，即
$$F(x,t)=\frac{1}{2}{x}^T(R(t)+\beta L)x+s^T(t)x+h(t)$$
则有
$$\frac{\partial F(x,t)}{\partial x}=R(t)x+\beta Lx+s(t)$$
式(3)重写为
$$\dot{x}=k{P}_{\Omega }(x-\alpha R(t)x-\beta Lx-\alpha s(t))-x$$
定义李雅普诺夫函数$V=\frac{1}{2}{e}^{T}e$，其中$e=x-x^{\ast }(t)$，$x^{\ast }(t)$满足$P_{\Omega }(x^{\ast }(t)-\alpha R(t)x^{\ast }(t)-\alpha \beta L{x}^{\ast }(t)-\alpha s(t))=x^{\ast }(t)$。

那么，可以得到
$$\begin{array}{rl}{P}_{\Omega }& (x-\alpha R(t)x-\alpha s(t))-x\\ & =P_{\Omega }(e+x^{\ast }(t)-\alpha R(t)e-\alpha R(t)x^{\ast }(t)-\alpha \beta Le-\alpha \beta L{x}^{\ast }(t)-\alpha s(t))\\ & -P_{\Omega }(x^{\ast }(t)-\alpha R(t)x^{\ast }(t)-\alpha \beta L{x}^{\ast }(t)-\alpha s(t))-e\end{array}$$
李雅普诺夫函数对时间导数为
$$\begin{array}{rl}\dot{V}& =e^T(k{P}_{\Omega }(x-\alpha R(t)x-\alpha \beta Lx-\alpha s(t))-kx-{\dot{x}}^{\ast }(t))\\ & \le -k{e}^{T}e+k\Vert e\Vert \Vert \zeta \Vert +\Vert e\Vert \Vert {\dot{x}}^{\ast }(t)\Vert \end{array}$$
其中
$$\begin{array}{rl}\zeta & =P_{\Omega }(e+x^{\ast }(t)-\alpha R(t)e-\alpha R(t)x^{\ast }(t)-\alpha \beta Le-\alpha \beta L{x}^{\ast }(t)-\alpha s(t))\\ & -P_{\Omega }(x^{\ast }(t)-\alpha R(t)x^{\ast }(t)-\alpha \beta L{x}^{\ast }(t)-\alpha s(t))\end{array}$$
利用Lemma 3，可得$\Vert \zeta \Vert \le \Vert e-\alpha R(t)e-\alpha \beta Le\Vert$，故
$$\begin{array}{rl}\dot{V}& \le -k{e}^{T}e+k\Vert e\Vert \Vert e-\alpha R(t)e-\alpha \beta Le\Vert +\Vert e\Vert \Vert {\dot{x}}^{\ast }(t)\Vert \\ & \le -k\left(1-\frac{1}{2k}-\Vert I-\alpha R(t)-\alpha \beta L\Vert \right)\Vert e{\Vert }^2+\frac{1}{2}\Vert x^{\ast }(t){\Vert }^2\end{array}$$
由于$\frac{1-{\epsilon }_L}{q_1}<\alpha <\frac{1+{\epsilon }_L}{q_2+\beta {\lambda }_{\text{max}}(L)}$(注意${\lambda }_{\text{min}}(L)=0$)，可得$\Vert I-\alpha R(t)-\alpha \beta L\Vert \le {\epsilon }_L$，即
$$\dot{V}\le -k(1-\frac{1}{2k}-{\epsilon }_L)\Vert e{\Vert }^2+\frac{1}{2}{\bar{\omega }}_L^2$$
因此$\dot{V}<0$，$\forall \Vert e{\Vert }^2>\frac{1}{2k(1-{\epsilon }_L)-1}{\bar{\omega }}_L^2$。即误差一致最终有界，最终界为$\Vert e\Vert \le \sqrt{\frac{1}{2k(1-{\epsilon }_L)-1}}{\bar{\omega }}_L$。

---

1. Sun, C., Ye, M., & Hu, G. (2017). Distributed Time-Varying Quadratic Optimization for Multiple Agents under Undirected Graphs. IEEE Transactions on Automatic Control, 62(7), 3687–3694. https://doi.org/10.1109/TAC.2017.2673240 ↩︎

2. Kinderlehrer, D. and Stampacchia, G. (1980). An Introduction to Variational Inequalities and Their Applications. Philadelphia, PA, USA: SIAM. ↩︎

3. Liang, S., Zeng, X., & Hong, Y. (2018). Distributed Nonsmooth Optimization with Coupled Inequality Constraints via Modified Lagrangian Function. IEEE Transactions on Automatic Control, 63(6), 1753–1759. https://doi.org/10.1109/TAC.2017.2752001 ↩︎