【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 Graphs[1].

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

minf(x,t)=i=1nfi(xi,t)s.t.Lx=0,xΩ(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)

其中x=[x1T,,xnT]Tx=[x_1^T,\cdots,x_n^T]^TΩ=Ω1××Ωn\Omega=\Omega_1\times\cdots\times \Omega_n

分布式算法

改成罚函数形式

minF(x,t)=i=1nfi(xi,t)+12βxTLx,xΩ(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)

罚函数投影算法

x˙i=kPΩi(xiαfi(xi,t)xiαβj=1naij(xixj))kxi(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)

其中α\alphakk是待定参数。

结论与证明

Lemma 3[2]: For all x,yRnx,y\in\mathbb R^n and a nonempy compact set KRnK\subset \mathbb R^n, PK(x)PK(y)xy\|P_K(x)-P_K(y)\|\leq \|x-y\|.

Lemma 6: x(t)x^*(t) is an optimal solution of (2) if and only if for some fixed α>0\alpha>0, PΩ(αF(x(t),t)x+x(t))=x(t)P_\Omega(-\alpha\frac{\partial F(x^*(t),t)}{\partial x}+x^*(t))=x^*(t), xΩ\forall x\in\Omega, t>t0t>t_0.

注意到Lemma 6成立,是因为0F(x)+NΩ(x)0\in\partial F(x^*)+N_\Omega(x^*)(Lemma 2.1[3])。

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)\| uniformly ultimately bounded as tt\to \infty with ultimate bound x(t)x(t)12k(1εL)1ωˉL\|x(t)-x^*(t)\|\leq \sqrt{\frac{1}{2k(1-\varepsilon_L)-1}}\bar \omega_L, where 0<εL<10<\varepsilon_L<1 is a positive constant, ωˉL\bar \omega_L is the upper bound of x˙(t)\|\dot x^*(t)\|, provided that there exist positive constants q1q_1 and q2q1q_2\geq q_1 such that q1IR(t)q2Iq_1 I\leq R(t)\leq q_2I, x˙(t)\dot x^*(t) exists, the elements in R(t)R(t) and x(t)x^*(t), x˙(t)\dot x^*(t) are bounded and the control parameters εL,α,β\varepsilon_L, \alpha,\beta and kk are chosen such that

q2+βλmax(L)q1q2+βλmax(L)+q1<εL<1,k>12(1εL),1εLq1<α<1+εLq2+βλmax(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}

证明:注意到原论文考虑二次代价函数,即

F(x,t)=12xT(R(t)+βL)x+sT(t)x+h(t)F(x,t) = \frac{1}{2}x^T(R(t)+\beta L)x+s^T(t)x+h(t)

则有

F(x,t)x=R(t)x+βLx+s(t)\frac{\partial F(x,t)}{\partial x} = R(t)x+\beta Lx+s(t)

式(3)重写为

x˙=kPΩ(xαR(t)xβLxαs(t))x\dot x=kP_\Omega(x-\alpha R(t)x-\beta Lx-\alpha s(t))-x

定义李雅普诺夫函数V=12eTeV=\frac{1}{2}e^Te,其中e=xx(t)e=x-x^*(t)x(t)x^*(t)满足PΩ(x(t)αR(t)x(t)αβLx(t)αs(t))=x(t)P_\Omega(x^*(t)-\alpha R(t)x^*(t)-\alpha\beta L x^*(t)-\alpha s(t))=x^*(t)

那么,可以得到

PΩ(xαR(t)xαs(t))x=PΩ(e+x(t)αR(t)eαR(t)x(t)αβLeαβLx(t)αs(t))PΩ(x(t)αR(t)x(t)αβLx(t)αs(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}

李雅普诺夫函数对时间导数为

V˙=eT(kPΩ(xαR(t)xαβLxαs(t))kxx˙(t))keTe+keζ+ex˙(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}

其中

ζ=PΩ(e+x(t)αR(t)eαR(t)x(t)αβLeαβLx(t)αs(t))PΩ(x(t)αR(t)x(t)αβLx(t)αs(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}

利用Lemma 3,可得ζeαR(t)eαβLe\|\zeta\|\leq \|e-\alpha R(t)e-\alpha \beta L e\|,故

V˙keTe+keeαR(t)eαβLe+ex˙(t)k(112kIαR(t)αβL)e2+12x(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}

由于1εLq1<α<1+εLq2+βλmax(L)\frac{1-\varepsilon_L}{q_1}<\alpha<\frac{1+\varepsilon_L}{q_2+\beta\lambda_\text{max}(L)}(注意λmin(L)=0\lambda_\text{min}(L)=0),可得IαR(t)αβLεL\|I-\alpha R(t)-\alpha \beta L\|\leq \varepsilon_L,即

V˙k(112kεL)e2+12ωˉL2\dot V\leq -k(1-\frac{1}{2k}-\varepsilon_L)\|e\|^2+\frac{1}{2}\bar \omega_L^2

因此V˙<0\dot V<0e2>12k(1εL)1ωˉL2\forall \|e\|^2>\frac{1}{2k(1-\varepsilon_L)-1}\bar \omega_L^2。即误差一致最终有界,最终界为e12k(1εL)1ωˉL\|e\|\leq \sqrt{\frac{1}{2k(1-\varepsilon_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 ↩︎