浅谈有限时间分布式一致跟踪算法

写在前面

分布式一致算法(Distributed Consensus Algorithm)可以用来设计分布式估计器/观测器,在控制中应用广泛。本文主要关注的是有限时间收敛且参考信号时变的情况,即有限时间(finite-time)分布式一致跟踪(Consensus Tracking)算法。

无向图的情况

第一个算法是最常见的有限时间速度估计器,如下所示:

ζ˙i=βsgn(jNiaij(ζiζj)+bi(ζiζd))\dot \zeta_i=-\beta\operatorname{sgn}\left(\sum_{j\in\mathcal N_i}a_{ij}(\zeta_i-\zeta_j)+b_i(\zeta_i-\zeta_d)\right)。

其中β>ζ˙d\beta>\|\dot \zeta_d\|_\infty。该算法直接将标准分布式一致跟踪算法的速度项去掉,然后加上了符号函数。

eζ=ζ1nζde_\zeta=\zeta-1_n\otimes \zeta_d。误差动力学为

e˙ζ=βsgn(Leζ+Beζ)1nζ˙d\dot e_\zeta=-\beta\operatorname{sgn}(Le_\zeta+Be_\zeta)-1_n\otimes \dot \zeta_d

H=L+BH=L+B,其中B=diag(b1,,bn)B=\operatorname{diag}(b_1,\cdots,b_n)。如果bi=1b_i=1即为leader,可以获得参考信号;否则为follower,无法获得参考信号。

假设图是一个连通无向图,容易得到HH是一个对称正定矩阵,证明可以参考定理1。

定理1(Ren 2005[1]):如果有向图存在一个有向生成树,则HH所有特征根拥有正实部。即HH是一个非奇异的M-矩阵,且主对角占优。

参考Mehdifar 2019[2],设计李雅普诺夫函数

V=12eζTHeζV=\frac{1}{2}e_\zeta^THe_\zeta

求导得

V˙=eζTH(βsgn(Heζ)1nζ˙d)βHeζ1+ζ˙dHeζ1=(βζ˙d)Heζ1(βζ˙d)λ2eζ(βζ˙d)2λ22λnV12\begin{aligned} \dot V&=e_\zeta^TH(-\beta\operatorname{sgn}(He_\zeta)-1_n\otimes \dot \zeta_d)\\ &\leq -\beta\|He_\zeta\|_1+\|\dot \zeta_d\|_\infty\|He_\zeta\|_1\\ &=-(\beta-\|\dot \zeta_d\|_\infty)\|He_\zeta\|_1\\ &\leq -(\beta-\|\dot \zeta_d\|_\infty)\lambda_2\|e_\zeta\|\\ &\leq -(\beta-\|\dot \zeta_d\|_\infty)\sqrt{\frac{2\lambda_2^2}{\lambda_n}}V^{\frac{1}{2}} \end{aligned}

V˙Vc\frac{\dot V}{\sqrt{V}}\leq -c积分,可以得到2V(t)2V(0)ct2\sqrt{V(t)}\leq 2\sqrt{V(0)}-ct,即有限时间收敛,其中c=(βζ˙d)2λ22λn>0c=(\beta-\|\dot \zeta_d\|_\infty)\sqrt{\frac{2\lambda_2^2}{\lambda_n}}>0

定理 2(Long 2010[3], Lemma 3):假设V(t)V(t)可微且

dV(t)dtcV(t)α\frac{dV(t)}{dt}\leq -cV(t)^\alpha

其中c>0c>00<α<10<\alpha<1。则V(t)V(t)有限时间收敛于0,且收敛时间为TV(0)1αc(1α)T\leq \frac{V(0)^{1-\alpha}}{c(1-\alpha)}

由定理2可知,收敛时间为tT=2cV(0)1/2t\leq T=\frac{2}{c}V(0)^{1/2}

有向图的情况

假设图是一个有向图,有以leader为根的有向生成树。

定理3(Li 2015[4],Lemma 4):存在正对角矩阵GG使得GH+HTG>0GH+H^TG>0成立。矩阵GG可以由diag(q1,,qn)\operatorname{diag}(q_1,\cdots,q_n)给出,其中q=[q1,,qn]T=(HT)11nq=[q_1,\cdots,q_n]^T=(H^T)^{-1}1_n

证明:(第一句)直接取LL的左特征向量即可。第二句也给出了范例。

(第二句)因为HH非奇异,所以(LT)T(L^T)^T存在且非负,没有0行,故q>0q>0。由于HH主对角占优,即H1n=0H1_n=0,左乘GG相当于行元素同时放大缩小,故有GH1n>0GH1_n>0。又注意到HTG1n=HTq=1nH^TG1_n=H^Tq=1_n,相加得到(GH+HTG)1n>0(GH+H^TG)1_n>0,所以GH+HTGGH+H^TG严格主对角占优,即所有特征根为正,再加上对称性,得到正定性。

对于xRdx\in\mathbb R^d,定义sigμ(x)=[x1μsgn(x1),,xdμsgn(xd)]T\operatorname{sig}^{\mu}(x)=[|x_1|^\mu\operatorname{sgn}(x_1),\cdots,|x_d|^\mu\operatorname{sgn}(x_d)]^T。设计第二个算法

ζ˙i=αsigμ(δi)αsig1μ(δi)βsgn(δi)δi=jNiaij(ζiζj)+bi(ζiζd)\begin{aligned} \dot \zeta_i&=-\alpha\operatorname{sig}^\mu(\delta_i)-\alpha\operatorname{sig}^{\frac{1}{\mu}}(\delta_i)-\beta\operatorname{sgn}(\delta_i)\\ \delta_i&=\sum_{j\in\mathcal N_i}a_{ij}(\zeta_i-\zeta_j)+b_i(\zeta_i-\zeta_d) \end{aligned}

其中0<μ<10<\mu<1β>ζ˙d\beta>\|\dot \zeta_d\|_\infty。可以看出该算法是在第一个算法基础上增加了两个sig函数。

eζ=ζ1nζde_\zeta=\zeta-1_n\otimes \zeta_d,则δ=Heζ\delta=He_\zeta。由定理1可知,在有向图中,HH不一定对称,但仍然是满秩的。因此只需证明δ\delta有限时间收敛于0。

δ˙=He˙ζ=αHsigμ(δ)αHsig1μ(δ)βHsgn(δ)H(1nζ˙d)\dot \delta=H\dot e_\zeta=-\alpha H\operatorname{sig}^\mu(\delta)-\alpha H\operatorname{sig}^{\frac{1}{\mu}}(\delta)-\beta H\operatorname{sgn}(\delta)-H (1_n\otimes \dot \zeta_d)

定义xγ=(i=1dxiγ)1/γ\|x\|_\gamma=(\sum_{i=1}^d |x_i|^\gamma)^{1/\gamma}。参考Ning 2020[5],设计李雅普诺夫函数

V=i=1nqi(αμ+1δiμ+1μ+1+αμμ+1δi1μ+11μ+1)V=\sum_{i=1}^{n}q_i(\frac{\alpha}{\mu+1}|\delta_i|_{\mu+1}^{\mu+1}+\frac{\alpha\mu}{\mu+1}|\delta_i|_{\frac{1}{\mu}+1}^{\frac{1}{\mu}+1})

求导得

V˙=i=1nαqi(sigμ(δi)+sig1μ(δi))δ˙i=α2(sigμ(δ)+sig1μ(δ))GH(sigμ(δ)+sig1μ(δ))α(sigμ(δ)+sig1μ(δ))GH(βsgn(δ)+1nζ˙d)\begin{aligned} \dot V&=\sum_{i=1}^n\alpha q_i(\operatorname{sig}^\mu(\delta_i)+\operatorname{sig}^{\frac{1}{\mu}}(\delta_i))\dot \delta_i\\ &=-\alpha^2(\operatorname{sig}^\mu(\delta)+\operatorname{sig}^{\frac{1}{\mu}}(\delta))GH(\operatorname{sig}^\mu(\delta)+\operatorname{sig}^{\frac{1}{\mu}}(\delta))\\ &\quad-\alpha(\operatorname{sig}^\mu(\delta)+\operatorname{sig}^{\frac{1}{\mu}}(\delta))GH(\beta\operatorname{sgn}(\delta)+1_n\otimes \dot \zeta_d)\\ \end{aligned}

注意到,第二个等式第二项中GH=GL+GBGH=GL+GB,对于GLGL这一部分,可以看出GLGL是有向平衡图的拉普拉斯矩阵,我们有GL(1nζd)=0GL(1_n\otimes \zeta_d)=0,且

sigμ(δ)GLsgn(δ)=i=1nsgn(δi)jNiδiμqiaij(sgn(δi)sgn(δj))\operatorname{sig}^\mu(\delta)GL\operatorname{sgn}(\delta)=\sum_{i=1}^n\operatorname{sgn}(\delta_i)\sum_{j\in\mathcal N_i}|\delta_i|^\mu q_ia_{ij}(\operatorname{sgn}(\delta_i)-\operatorname{sgn}(\delta_j))

分类讨论,若δi,δj\delta_i,\delta_j同号(或δi=0\delta_i=0),则sgn(δi)sgn(δj)=0\operatorname{sgn}(\delta_i)-\operatorname{sgn}(\delta_j)=0(或sgn(δi)=0\operatorname{sgn}(\delta_i)=0);

δi,δj\delta_i,\delta_j异号(或δj=0\delta_j=0),则sgn(δi)sgn(δj)=2sgn(δi)\operatorname{sgn}(\delta_i)-\operatorname{sgn}(\delta_j)=2\operatorname{sgn}(\delta_i)(或sgn(δi)\operatorname{sgn}(\delta_i)),

sgn(δi)δiμqiaij(sgn(δi)sgn(δj))=2δiμqiaij0\operatorname{sgn}(\delta_i)|\delta_i|^\mu q_ia_{ij}(\operatorname{sgn}(\delta_i)-\operatorname{sgn}(\delta_j))=2|\delta_i|^\mu q_ia_{ij}\geq 0(或δiμqiaij0|\delta_i|^\mu q_ia_{ij}\geq 0)。

综上所述,对于GLGL的这部分,恒大于等于0。

g=α(sigμ(δ)+sig1μ(δ))g=\alpha(\operatorname{sig}^\mu(\delta)+\operatorname{sig}^{\frac{1}{\mu}}(\delta))。对于GBGB这一部分,有

gGB(βsgn(g)+1nζ˙d)(βζd)iVlqibigi0gGB(\beta\operatorname{sgn}(g)+1_n\otimes \dot \zeta_d)\geq (\beta-\|\zeta_d\|_\infty)\sum_{i\in\mathcal V_l}q_ib_i|g_i|\geq 0

因此,当β>ζ˙d\beta>\|\dot \zeta_d\|_\infty时,有

V˙α2(sigμ(δ)+sig1μ(δ))GH(sigμ(δ)+sig1μ(δ))λmin(M)α2sigμ(δ)+sig1μ(δ)2\begin{aligned} \dot V&\leq -\alpha^2(\operatorname{sig}^\mu(\delta)+\operatorname{sig}^{\frac{1}{\mu}}(\delta))GH(\operatorname{sig}^\mu(\delta)+\operatorname{sig}^{\frac{1}{\mu}}(\delta))\\ &\leq -\lambda_{\min}(M)\alpha^2\|\operatorname{sig}^\mu(\delta)+\operatorname{sig}^{\frac{1}{\mu}}(\delta)\|^2 \end{aligned}

其中M=12(GH+HTG)M=\frac{1}{2}(GH+H^TG)

定理4(Wang 2019[6], Theorem 1):对任意向量xRdx\in\mathbb R^d,以及0<μ1<1<μ20<\mu_1<1<\mu_2,给定X1=xμ1+1μ1+1+xμ2+1μ2+1X_1=\|x\|_{\mu_1+1}^{\mu_1+1}+\|x\|_{\mu_2+1}^{\mu_2+1}X2=sigμ1(x)+sigμ2(x)X_2=\|\operatorname{sig}^{\mu_1}(x)+\operatorname{sig}^{\mu_2}(x)\|。则以下不等式成立:

X1μ1+μ2μ2+1X22,(2d)1μ21+μ2X12μ2μ2+1X22X_1^{\frac{\mu_1+\mu_2}{\mu_2+1}}\leq X_2^2,\quad (2d)^{\frac{1-\mu_2}{1+\mu_2}}X_1^{\frac{2\mu_2}{\mu_2+1}}\leq X_2^2。

由李雅普诺夫函数定义得到,

Vαq1+μ(δμ+1μ+1+δ1μ+11μ+1)V˙λmin(M)α2(δ2μ2μ+2δμ+1μμ+1μ+δ2μ2μ)\begin{aligned} V&\leq \frac{\alpha\|q\|_\infty}{1+\mu}(\|\delta\|_{\mu+1}^{\mu+1}+\|\delta\|_{\frac{1}{\mu}+1}^{\frac{1}{\mu}+1})\\ \dot V&\leq -\lambda_{\min}(M)\alpha^2(\|\delta\|_{2\mu}^{2\mu}+2\|\delta\|_{\mu+\frac{1}{\mu}}^{\mu+\frac{1}{\mu}}+\|\delta\|_{\frac{2}{\mu}}^{\frac{2}{\mu}}) \end{aligned}

由定理4,容易证明

V˙c22((Vc1)μ2+1μ+1+(2nd)μ1μ+1(Vc1)2μ+1)\dot V\leq -\frac{c_2}{2}((\frac{V}{c_1})^{\frac{\mu^2+1}{\mu+1}}+(2nd)^{\frac{\mu-1}{\mu+1}}(\frac{V}{c_1})^{\frac{2}{\mu+1}})

其中c1=αq1+μc_1=\frac{\alpha\|q\|_\infty}{1+\mu}c2=λmin(M)α2c_2=\lambda_{\min}(M)\alpha^2

定理5(Polyakov 2012[7]):考虑如下ODE

x˙(t)=g(x(t),t),x(0)=x0,\dot x(t)=g(x(t),t),\quad x(0)=x_0,

其中xRdx\in\mathbb R^dg:RdRdg:R^d\to R^d是非线性函数。假设存在连续正定函数V(x):RdRV(x):R^d\to R和一些标量a,b>0a,b>0μ>0\mu>0以及0<ν<10<\nu<1使得V˙(x)+aVμ(x)+bVν(x)0\dot V(x)+aV^\mu(x)+bV^\nu(x)\leq 0xRd\{0}x\in\mathbb R^d\backslash\{0\}。那么,该ODE的原点是全局固定时间稳定(globally fixed-time stable),且稳定时间TT满足

T1/a(1μ)+1/b(1ν)T\leq 1/a(1-\mu)+1/b(1-\nu)。

由于μ2+1μ+1<1<2μ+1\frac{\mu^2+1}{\mu+1}<1<\frac{2}{\mu+1},由定理5,可以得到VV有限时间收敛到0,收敛时间为

tT=1c22(1μ2+1μ+1)+1(2nd)μ1μ+1(2μ+11)=(2+c2μ(2nd)1μ1+μ)(1+μ)c2μ(1μ)t\leq T=\frac{1}{\frac{c_2}{2}(1-\frac{\mu^2+1}{\mu+1})}+\frac{1}{(2nd)^{\frac{\mu-1}{\mu+1}}(\frac{2}{\mu+1}-1)}=\frac{(2+c_2\mu(2nd)^\frac{1-\mu}{1+\mu})(1+\mu)}{c_2\mu(1-\mu)}


  1. W. Ren and R. Beard, “Consensus seeking in multiagent systems under dynamically changing interaction topologies,” IEEE Transactions on Automatic Control, vol. 50, no. 5, pp. 655–661, 2005 ↩︎

  2. F. Mehdifar, F. Hashemzadeh, M. Baradarannia, and M. de Queiroz, “Finite-Time Rigidity-Based Formation Maneuvering of Multiagent Systems Using Distributed Finite-Time Velocity Estimators,” IEEE Trans. Cybern., vol. 49, no. 12, pp. 4473–4484, Dec. 2019. ↩︎

  3. W. Long and X. Feng, “Finite-time consensus problems for networks of dynamic agents,” IEEE Trans. Autom. Control, vol. 55, no. 4, pp. 950–955, 2010. ↩︎

  4. Z. Li, G. Wen, Z. Duan, and W. Ren, “Designing Fully Distributed Consensus Protocols for Linear Multi-Agent Systems with Directed Graphs,” IEEE Trans. Automat. Contr., vol. 60, no. 4, pp. 1152–1157, Apr. 2015. ↩︎

  5. B. Ning, Q.-L. Han, and Q. Lu, “Fixed-Time Leader-Following Consensus for Multiple Wheeled Mobile Robots,” IEEE Trans. Cybern., vol. 50, no. 10, pp. 4381–4392, Oct. 2020. ↩︎

  6. H. Wang, W. Yu, G. Wen, and G. Chen, “Fixed-Time Consensus of Nonlinear Multi-Agent Systems With General Directed Topologies,” IEEE Trans. Circuits Syst. II Express Briefs, vol. 66, no. 9, pp. 1587–1591, Sep. 2019. ↩︎

  7. A. Polyakov, “Nonlinear feedback design for fixed-time stabilization of linear control systems,” IEEE Trans. Automat. Contr., vol. 57, no. 8, pp. 2106–2110, 2012. ↩︎