1: \documentclass[12pt,letterpaper]{article}

2: 

3: \usepackage{amssymb}

4: \usepackage{amsmath}

5: \usepackage{amsthm}

6: %

7: %

8: \usepackage{graphicx}

9: \usepackage{graphics}

10: %

11: %

12: \usepackage{setspace}

13: \usepackage{subfigure}

14: %

15: \usepackage[margin=1in]{geometry}

16: 

17: 

18: \begin{document}

19: 

20: \newcommand{\pardfdx}{$\frac{\partial \mathbf{f}}{\partial \mathbf{x}} $}

21: \newcommand{\pardfdxnm}{\frac{\partial \mathbf{f}}{\partial \mathbf{x}}}

22: \newcommand{\pardfdynm}{\frac{\partial \mathbf{f}}{\partial \mathbf{y}}}

23: \newcommand{\boldf}{\mathbf{f}}

24: \newcommand{\boldg}{\mathbf{g}}

25: \newcommand{\boldx}{\mathbf{x}}

26: \newcommand{\boldz}{\mathbf{z}}

27: \newcommand{\boldT}{\mathbf{T}}

28: \newcommand{\boldM}{\mathbf{M}}

29: \newcommand{\boldF}{\mathbf{F}}

30: %

31: \newcommand{\boldR}{\mathbf{R}}

32: \newcommand{\boldy}{\mathbf{y}}

33: \newcommand{\boldu}{\mathbf{u}}

34: \newcommand{\pardfdxdel}{\frac{\partial \boldf_{\delta}}{\partial \boldx}}

35: \newcommand{\pardfdxgam}{\frac{\partial \boldf_{\gamma}}{\partial \boldx}}

36: 

37: 

38: \newcommand{\R}{{\mathbb R}}

39: 

40: 

41: \newtheorem{definition}{Definition}

42: \newtheorem{theorem}{Theorem}

43: \newtheorem{lemma}{Lemma}

44: 

45: \newcommand{\todo}[1]{\vspace{5 mm}\par \noindent \marginpar{\textsc{ToDo}}

46: \framebox{\begin{minipage}[c]{0.95 \columnwidth} \tt #1

47: \end{minipage}}\vspace{5 mm}\par}

48: 

49: \title{Stability and Robustness Analysis of Nonlinear Systems \\

50: via Contraction Metrics and SOS Programming }

51: 

52: \author{Erin M. Aylward$^1$ \qquad Pablo A. Parrilo$^1$ \qquad Jean-Jacques E. Slotine$^2$ \\

53: $^1$Laboratory for Information and Decision Systems\\

54: $^2$Nonlinear Systems Laboratory \\

55:  Massachusetts Institute of Technology \\

56: Cambridge, MA 02139, USA}

57: 

58: \date{LIDS Technical Report \#2691 \\ \vspace{.6cm} \today}

59: 

60: \maketitle

61: 

62: 

63:  \tableofcontents

64:  \pagebreak

65: \begin{abstract}

66: 

67: Contraction analysis is a stability theory for nonlinear systems where

68: stability is defined incrementally between two arbitrary

69: trajectories. It provides an alternative framework in which to study

70: uncertain interconnections or systems with external inputs, where it

71: offers several significant advantages when compared with traditional

72: Lyapunov analysis. Contraction-based methods are particularly useful

73: for analyzing systems with uncertain parameters and for proving

74: synchronization properties of nonlinear oscillators. Existence of a

75: contraction metric for a given system is a necessary and sufficient

76: condition for global exponential convergence of system

77: trajectories. For systems with polynomial or rational dynamics, the

78: search for contraction metrics can be made fully algorithmic through

79: the use of convex optimization and sum of squares (SOS)

80: programming. The search process is made computationally tractable by

81: relaxing matrix definiteness constraints, whose feasibility indicate

82: existence of a contraction metric, into SOS constraints on polynomial

83: matrices.  We illustrate the results through examples from the

84: literature, emphasizing the advantages and contrasting the differences

85: between the contraction approach and traditional Lyapunov techniques.

86: 

87: \end{abstract}

88: 

89: %

90: 

91: \section{Introduction}

92: Contraction analysis is a stability theory for nonlinear systems where

93: stability is defined incrementally between two arbitrary trajectories

94: \cite{OCA}. The existence of a contraction metric for a nonlinear

95: system ensures that a suitably defined distance between nearby

96: trajectories is always decreasing, and thus trajectories converge

97: exponentially and globally. One important application of contraction

98: theory is its use in studying the synchronization of nonlinear coupled

99: oscillators \cite{OPC-NO}.  These oscillators present themselves in a

100: variety of research fields such as mathematics, biology, neuroscience,

101: electronics, and robotics. The use of coupled oscillators in each of

102: these fields, as well as how contraction theory can be used to analyze

103: networks of coupled identical nonlinear oscillators can be can be

104: found in \cite{OPC-NO} and the references listed therein.

105: 

106: Contraction theory nicely complements Lyapunov theory, a standard

107: nonlinear stability analysis technique, as it provides an

108: alternative framework in which to study convergence and robustness

109: properties of nonlinear systems. For autonomous systems one can

110: interpret the search for a contraction metric as the search for a

111: Lyapunov function with a certain structure. This statement will be

112: explained further in Section \ref{UD}.  There are, however,

113: advantages to searching for a contraction metric instead of

114: searching explicitly for a Lyapunov function. In particular, as we

115: will show, contraction metrics are useful for analyzing uncertain

116: nonlinear systems. In general, nonlinear systems with uncertain

117: parameters can prove quite troublesome for standard Lyapunov

118: methods, since the uncertainty can change the equilibrium point of

119: the system in very complicated ways, thus forcing the use of

120: parameter-dependent Lyapunov functions in order to prove stability

121: for a range of the uncertain parameter values.

122: 

123: Much of the literature on parameter-dependent Lyapunov functions

124: focuses on linear systems with parametric uncertainty

125: \cite{PDL:Haddad,PDL:Gahinet,PDL:Feron, Rantzerpaper}. However, if

126: a linear model is being used to study a nonlinear system around an

127: equilibrium point, changing the equilibrium of the nonlinear

128: system, necessitates relinearization around the new equilibrium.

129: If the actual position of the equilibrium, in addition to the

130: stability properties of the equilibrium, of the nonlinear system

131: depends on the uncertainty, it may be impossible to obtain any

132: kind of closed form expression of the equilibrium in terms of the

133: uncertain parameters. Thus, parameterizing the linearization in

134: terms of the uncertainty may not be an option.

135: 

136: A well-studied method of dealing with specific forms of nonlinearities

137: is to model the nonlinear system as a linear system with bounded

138: uncertainty. In particular, in \cite{LMIBook} polytopic linear

139: differential inclusions (LDIs), norm-bound LDIs, and diagonal

140: norm-bound LDIs are considered. These techniques are computationally

141: tractable as they reduce to convex optimization problems. Though these

142: methods work for various kinds of uncertainty, it is also desirable to

143: find methods to study the stability of nonlinear systems that do not

144: easily admit linear approximations with the nonlinearities covered

145: with uncertainty bounds.

146: 

147: Contraction theory provides a framework in which to study the

148: stability behavior of more general uncertain nonlinear systems.

149: This framework eliminates many of the restrictions and problems

150: that may be encountered when trying to analyze uncertain nonlinear

151: systems with traditional linearization techniques or Lyapunov

152: methods. This results from the fact that if a nominal system is

153: contracting with respect to a certain contraction metric, it is

154: often the case that the uncertain system with additive or

155: multiplicative uncertainty within a certain range will still be

156: contracting with respect to the same metric, even if the

157: perturbation changes the position of the equilibrium of the

158: system. Thus, it is possible to determine stability of the system

159: for a range of values of the uncertain parameter without

160: explicitly tracking how the uncertainty changes the location of

161: the equilibrium. These ideas will be discussed further in

162: Section~\ref{UD}.

163: 

164: Another interesting feature of the contraction framework is its

165: relative flexibility in incorporating inputs and outputs. For

166: instance, to prove contraction of a class of systems with external

167: inputs, it is sufficient to show the existence of a contraction

168: metric with a certain structure.  This feature, which will be

169: discussed in Section~\ref{EI}, is central in using contraction

170: theory to prove synchronization of coupled nonlinear oscillators.

171: 

172: To translate the theoretical discussion above into effective

173: practical tools, it is desirable to have efficient computational

174: methods to numerically obtain contraction metrics. Sum of squares

175: (SOS) programming provides one such method. SOS programming is

176: based on techniques that combine elements of computational algebra

177: and convex optimization, and has been recently used to provide

178: efficient convex relaxations for several computationally hard

179: problems \cite{CAND}. In this paper we will show how SOS

180: programming enables the search for contraction metrics for the

181: class of nonlinear systems with polynomial dynamics. We discuss

182: how to use SOS methods to find bounds on the maximum amount of

183: uncertainty allowed in a system in order for the system to retain

184: the property of being contracting with respect to the contraction

185: metric of the unperturbed system. We also use SOS methods to

186: optimize the contraction matrix search to obtain a metric that

187: provides the largest symmetric uncertainty interval for which we

188: can prove the system is contracting.

189: 

190: This paper is organized as follows: in Section~\ref{CT} we give

191: background material on contraction theory. Section~\ref{SOSSect}

192: discusses sum of squares (SOS) polynomials and matrices.  We

193: present next an algorithm which uses SOS programming to

194: computationally search for contraction metrics for nonlinear

195: systems. We discuss why contraction theory is useful for studying

196: systems with uncertain dynamics in Section~\ref{UD} and external

197: inputs in Section~\ref{EI}. Finally, in Section~\ref{Conclusions}

198: we present our conclusions, and outline possible directions for

199: future work.

200: 

201: \section{Contraction Analysis}\label{CT}

202: 

203: Contraction analysis is a relatively recently developed stability

204: theory for nonlinear systems analysis \cite{OCA}. The theory attempts

205: to answer the question of whether the limiting behavior of a given

206: dynamical system is independent of its initial conditions.  More

207: specifically, contraction analysis is a theory in which stability is

208: defined incrementally between two arbitrary trajectories.  It is used

209: to determine whether nearby trajectories converge to one another. This

210: section summarizes the main elements of contraction analysis; a much

211: more detailed account can be found in \cite{OCA}.

212: 

213: We consider deterministic dynamical systems of the form

214: \begin{equation}

215: \dot{\boldx} = \boldf(\boldx(t),t), \label{sysdynamics}

216: \end{equation}

217: where $\boldf$ is a nonlinear vector field and $\boldx (t)$ is an

218: $n$-dimensional state vector. For this analysis it is assumed that

219: all quantities are real and smooth and thus that all required

220: derivatives or partial derivatives exist and are continuous. This

221: existence and continuity assumption clearly holds for polynomial

222: vector fields.

223: 

224: Under the assumption that all quantities are real and smooth, from

225: equation (\ref{sysdynamics}) we can obtain the differential

226: relation

227: \begin{equation}

228: \delta \dot{\boldx} (t) = \frac{\partial \boldf}{\partial \boldx

229: (t)}(\boldx (t),t)\delta \boldx (t), \label{diffdynamics}

230: \end{equation}

231: where $\delta \boldx (t)$ is an infinitesimal displacement at a

232: fixed time. For notational convenience from here on we will write

233: $\boldx$ for $\boldx(t)$, but in all calculations it should be

234: noted that $\boldx$ is a function of time.

235: 

236: The infinitesimal squared distance between two trajectories is

237: $\delta \boldx^T \delta \boldx$.  Using (\ref{diffdynamics}), the

238: following equation for the rate of change of the squared distance

239: between two trajectories is obtained:

240: \begin{equation} \frac{d}{dt}(\delta \boldx^T \delta \boldx) = 2\delta \boldx^T

241: \delta \dot{\boldx} =  2 \delta \boldx^T \frac{\partial

242: \boldf}{\partial \boldx} \delta \boldx. \label{eq:diffdynamics2}

243: \end{equation} If $\lambda_1(\boldx,t)$ is the largest eigenvalue of the

244: symmetric part of the Jacobian $\frac{\partial \boldf}{\partial

245: \boldx}$ (i.e. the largest eigenvalue of $\frac{1}{2}(\pardfdxnm +

246: \pardfdxnm^T)$),  then it follows from (\ref{eq:diffdynamics2}) that

247: \begin{equation}

248: \frac{d}{dt}(\delta \boldx^T \delta \boldx) \leq 2

249: \lambda_1(\boldx, t) \delta \boldx ^T \delta \boldx.

250: \end{equation} Integrating both sides gives

251: \begin{equation}

252: ||\delta \boldx|| \leq || \delta \boldx _o|| \, e^{\int_0^t

253: \lambda_1(\boldx,t)dt}.

254: \label{expdyn}

255: \end{equation} If $\lambda_1(\boldx,t)$ is uniformly strictly negative

256: (i.e. $ ( \frac{\partial \boldf}{\partial \boldx} + \frac{\partial

257: \boldf}{\partial \boldx}^T ) \prec 0 \; \forall \; \boldx, t)$, it

258: follows from (\ref{expdyn}) that any infinitesimal length

259: $||\delta \boldx||$ converges exponentially to zero.  By path

260: integration the distance of any finite path also converges

261: exponentially to zero.

262: 

263: A more general definition of length can be given by

264: \begin{equation}

265: \delta \boldz^T \delta \boldz =

266: \delta \boldx ^T \boldM (\bold x, t) \delta \boldx

267: \label{RiemannSp}

268: \end{equation}

269: where $\boldM(\boldx, t)$ is a symmetric, uniformly positive

270: definite and continuously differentiable metric (formally, this

271: defines a Riemannian manifold).  This notion of infinitesimal

272: distance defined with respect to a metric can be use to define a

273: finite distance measure between two trajectories with respect to

274: this metric. Specifically, the distance between two points $P_1$

275: and $P_2$ with respect to the metric $\mathbf{M}(\boldx,t)$ is

276: defined as the shortest path length, in other words the smallest

277: path integral $\int_{P_1}^{P_2} \sqrt{\delta \boldx^T

278: \boldM(\boldx,t) \delta \boldx}$, between these two points.

279: Accordingly a ball of center $\mathbf{c}$ with radius $R$ is

280: defined as the set of all points whose distance to $\mathbf{c}$

281: with respect to $\boldM (\boldx,t)$ is strictly less than $R$.

282: 

283: Under the definition of infinitesimal length given in

284: (\ref{RiemannSp}), the equation for its rate of change becomes

285: \begin{equation}\label{contractcond2} \frac{d}{dt}(\delta \boldx

286: ^T \boldM \delta \boldx) = \delta \boldx^T(\pardfdxnm ^T \boldM +

287: \boldM

288: \pardfdxnm + \dot{\boldM}) \delta \boldx \end{equation}

289: where $\boldM$ is shorthand notation for $\boldM(\boldx,t)$.

290: Convergence to a single trajectory occurs in regions where

291: $(\pardfdxnm ^T \boldM + \boldM

292: \pardfdxnm + \dot{\boldM})$ is

293: uniformly negative definite.   It should be noted that

294: $\dot{\boldM}  = \dot{\boldM}(\boldx, t)= \frac{\partial

295: \boldM(\boldx,t)}{\partial \boldx}\frac{d \boldx}{dt} +

296: \frac{\partial \boldM (\boldx, t)}{\partial t}$. The above

297: analysis leads to the following definition and theorem:

298: 

299: \begin{definition}[\cite{OCA}]

300: Given the system equations $\dot{\boldx} = \boldf(\boldx,t)$, a

301: region of the state space is called a \emph{contraction region}

302: with respect to a uniformly positive definite metric

303: $\boldM(\boldx, t)$  if $(\pardfdxnm ^T \boldM + \boldM

304: \pardfdxnm + \dot{\boldM})$ is uniformly negative definite in

305: that region.

306: \end{definition}

307: 

308: 

309: \begin{theorem}[\cite{OCA}]

310: Consider the system equations $\dot{\boldx} = \boldf(\boldx,t)$.

311: Assume a trajectory starts in a ball of constant radius that is

312: defined with respect to the metric $\boldM(\boldx, t)$, that is

313: centered at a another given trajectory, and that is contained at

314: all times in a contraction region with respect to the metric

315: $\boldM(\boldx,t)$. Then the first trajectory will remain in that

316: ball and converge exponentially to the center trajectory.

317: Furthermore, global exponential convergence to the center

318: trajectory is guaranteed if the whole state space is a contraction

319: region with respect to the metric $\boldM(\boldx,t)$.

320: \end{theorem}

321: 

322: Definition 1 provides sufficient conditions for a system to be

323: contracting. Namely, the following should be satisfied:

324: \begin{enumerate}

325: \item The matrix $\boldM(\boldx, t)$ must be a uniformly positive

326: definite matrix, i.e.,

327: \begin{equation}\label{Mpos}\boldM(\boldx, t)

328: \succeq \epsilon \mathbf{I} \succ 0 \qquad \forall \boldx, t.

329: \end{equation}

330: 

331: \item The metric variation $\pardfdxnm ^T \boldM + \boldM

332: \pardfdxnm + \dot{\boldM}$ must be a uniformly negative definite

333: matrix, i.e.,

334: \begin{equation}\label{Rneg}\boldR(\boldx,t) = \pardfdxnm ^T \boldM + \boldM

335: \pardfdxnm + \dot{\boldM} \preceq -\epsilon \mathbf{I} \prec 0 \; \; \; \forall \boldx,t.

336: \end{equation}

337: \end{enumerate}An explicit rate of convergence of trajectories $\beta$ can be

338: found by finding a $\boldM(\boldx, t)$ that satisfies (\ref{Mpos})

339: and

340: \begin{equation}

341: \pardfdxnm ^T \boldM + \boldM

342: \pardfdxnm + \dot{\boldM} \preceq -\beta \boldM .

343: \end{equation}

344: 

345: The notation above is standard; $\succ$, and $\succeq$ mean

346: positive definite and positive semidefinite respectively, while

347: $\prec$ and $\preceq$ mean negative definite and negative

348: semidefinite respectively. If the system dynamics are linear and

349: $\boldM(\boldx,t)$ is constant (i.e. $\boldM(\boldx,t) = M$), the

350: conditions above reduce to those in standard Lyapunov analysis

351: techniques. Lyapunov theory shows that the system $\dot{\boldx}(t)

352: = A\boldx(t)$ is stable (i.e., all trajectories converge to 0) if

353: and only if there exists a positive definite matrix $M$ (i.e., $M

354: \succ 0$) such that $A^TM + MA \prec 0$.

355: 

356: 

357: It should be noted that if a global contraction metric exists for

358: an autonomous system, all trajectories converge to a unique

359: equilibrium point, and we can always produce a Lyapunov function

360: for the system from the contraction metric \cite{OCA}. We assume,

361: without loss of generality, that the equilibrium is at the origin.

362: If the system dynamics are $\boldf(\boldx)$ and $\boldM(\boldx)$

363: is a time-invariant contraction metric for the system, then

364: $V(\boldx) = \boldf(\boldx)^T \boldM(\boldx)\boldf(\boldx)$ is a

365: Lyapunov function for the system since $V(\boldx) > 0$ and

366: $\dot{V} = \boldf(\boldx)^T(\frac{\partial \boldf}{\partial

367: \boldx}^T \boldM + \bold M \frac{\partial \boldf}{\partial \boldx}

368: + \dot{\boldM})\boldf(\boldx)\leq -\beta V$. This shows that

369: $\dot{\boldx} = \boldf(\boldx)$ tends to $\mathbf{0}$

370: exponentially, and thus that $\boldx$ tends towards a finite

371: equilibrium point.

372: 

373: %

374: %

375: %

376: %

377: %

378: %

379: %

380: 

381: 

382: For a constant metric $\boldM(\boldx,t) = M$, this reduces to

383: Krasovskii's Method \cite{Khalil}. We note that for systems with

384: uncertainty there are good reasons to search for a contraction

385: metric to create Lyapunov function of this structure instead of

386: searching for a Lyapunov function directly. These reasons will

387: become clear in Section \ref{UD}.

388: 

389: The problem of searching for a contraction metric thus reduces to

390: finding a matrix function $\boldM(\boldx,t)$ that satisfies the

391: conditions above. As we will see, SOS methods will provide a

392: computationally convenient approach to this problem.

393: 

394: \section{Sum of Squares (SOS) Polynomials and Programs}

395: \label{SOSSect}

396: 

397: The main computational difficulty of problems involving constraints

398: such as the ones in (\ref{Mpos}) and (\ref{Rneg}) is the lack of

399: efficient numerical methods that can effectively handle multivariate

400: nonnegativity conditions. A convenient approach for this, originally

401: introduced in \cite{Parrilo:Phd}, is the use of sum of squares (SOS)

402: relaxations as a suitable replacement for nonnegativity. We present

403: below the basic elements of these techniques.

404: 

405: A multivariate polynomial $p(x_1, x_2, ..., x_n) = p(\boldx) \in

406: \R[\boldx]$ is a sum of squares (SOS) if there exist polynomials

407: $f_1(\boldx), ..., f_m(\boldx) \in \R[\boldx]$ such that

408: \begin{equation}\label{SOSeqn}

409: p(\boldx) = \sum_{i=1}^m f_i^2(\boldx).

410: \end{equation}

411: The existence of a SOS representation for a given polynomial is a

412: sufficient condition for its global nonnegativity, i.e., equation

413: (\ref{SOSeqn}) implies that $p(\boldx) \geq 0 \; \forall \; \boldx

414: \in \R^n$. The SOS condition (\ref{SOSeqn}) can be shown to be

415: equivalent to the existence of a positive semidefinite matrix $Q$

416: such that

417: \begin{equation}\label{Q}

418: p(\boldx) = Z^T(\boldx)QZ(\boldx)

419: \end{equation} where $Z(\boldx)$ is a vector of monomials of

420: degree less than or equal to deg($p$)/2.  This equivalence of

421: descriptions between (\ref{SOSeqn}) and (\ref{Q}) makes finding an

422: SOS decomposition a computationally tractable procedure. Finding a

423: symmetric positive semidefinite $Q$ subject to the affine

424: constraint (\ref{Q}) is a semidefinite programming problem

425: \cite{Parrilo:Phd,SOSTOOLS}.

426: 

427: 

428: Using the notion of a SOS polynomial as a primitive, we can now

429: introduce a convenient class of optimization problems.  A

430: \emph{sum of squares program} is a convex optimization problem of

431: the form:

432: \begin{eqnarray}

433: & \min& \sum_{j = 1}^J w_j \, c_j \nonumber \\

434: & \text{subject to}&\; \; a_{i,0} + \sum_{j =1}^J a_{i,j}(\boldx)

435: \, c_j \quad \mbox{ is SOS for }\, i= 1, ..., I, \nonumber

436: \end{eqnarray}

437: where the $c_j$'s are the scalar real decision variables, the $w_j$'s

438: are given real numbers that define the objective function, and the

439: $a_{i,j}(\boldx)$ are given multivariate polynomials.  There has

440: recently been much interest in SOS programming and SOS optimization as

441: these techniques provide convex relaxations for various

442: computationally hard optimization and control problems; see e.g.

443: \cite{Parrilo:Phd,sdprelax,Lasserre,CAND} and the volume \cite{GarulliHenrion}.

444: 

445: 

446: A SOS decomposition provides an explicit certificate of the

447: nonnegativity of a scalar polynomial for all values of the

448: indeterminates. In order to design an algorithmic procedure to search

449: for contraction metrics, we need to introduce a similar idea to ensure

450: that a polynomial matrix is positive definite for every value of the

451: indeterminates. A natural definition is as follows:

452: 

453: \begin{definition}[\cite{ParriloGatermann}]

454: Consider a symmetric matrix with polynomial entries

455: $\mathbf{S}(\boldx) \in \R[\boldx]^{m \times m}$, and let $\boldy =

456: [y_1, \ldots, y_m]^T$ be a vector of new indeterminates. Then

457: $\mathbf{S}(\boldx) $ is a \emph{sum of squares matrix} if the scalar

458: polynomial $\boldy^T \mathbf{S}(\boldx) \boldy$ is a sum of squares in

459: $\R[\boldx, \boldy]$.

460: \end{definition} \noindent For notational convenience, we also define a stricter notion:

461: \begin{definition}

462: A matrix $\mathbf{S}(\boldx)$ is \emph{strictly SOS} if

463: $\mathbf{S}(\boldx) - \epsilon \mathbf{I}$ is a SOS matrix for some

464: $\epsilon > 0$.

465: \end{definition}

466: 

467: Thus, a strictly SOS matrix is a matrix with polynomial entries

468: that is positive definite for every value of the indeterminates.

469: An equivalent definition of an SOS matrix can be given in terms of

470: the existence of a polynomial factorization: $\mathbf{S}(\boldx)$

471: is a SOS matrix if and only if it can be decomposed as

472: $\mathbf{S}(\boldx) = \mathbf{T}(\boldx)^T \mathbf{T}(\boldx)$

473: where $\mathbf{T}(\boldx) \in \R[\boldx]^{p \times m}$.  For

474: example,

475: \begin{displaymath}

476: \boldM(\boldx) = \left[\begin{array}{cc}

477:   \omega^2 + \alpha^2(x^2 + k)^2    & \alpha(x^2 + k) \\

478:   \alpha(x^2 + k ) & 1 \\

479: \end{array}\right]\end{displaymath} is a SOS matrix for all values of $\alpha$ and $k$.

480: Indeed, this follows from the decomposition $\boldM(\boldx) =

481: \mathbf{T}(\boldx)^T \mathbf{T}(\boldx)$, where

482: \begin{displaymath} \mathbf{T}(\boldx) =

483: \left[\begin{array}{cc}

484:   \omega & 0 \\

485:   \alpha(x^2 + k) & 1 \\

486: \end{array} \right].\\

487: \end{displaymath}

488: SOS matrices have also been used recently by Hol and Scherer

489: \cite{HolScherer} and Kojima \cite{KojimaSOS} to produce

490: relaxations of polynomial optimization problems with matrix

491: positivity definiteness constraints.

492: 

493: \section{Computational Search for Contraction Metrics via SOS Programming}

494: 

495: As explained in Section~\ref{CT}, given a dynamical system, the

496: conditions for a contraction metric to exist in regions of the

497: state-space are given by a pair of matrix inequalities. In the

498: case of metrics $\boldM(\boldx)$ that do not depend explicitly on

499: time, relaxing the matrix definiteness conditions in (\ref{Mpos})

500: and (\ref{Rneg}) to SOS matrix based tests makes the search for

501: contracting metrics a computationally tractable procedure. More

502: specifically, the matrix definiteness constraints on

503: $\boldM(\boldx)$ (and $\boldR(\boldx)$) can be relaxed to SOS

504: matrix constraints by changing the inequality $\boldM(\boldx) -

505: \epsilon \mathbf{I} \succeq 0$ in~(\ref{Mpos}) (where $\epsilon$

506: is an arbitrarily small constant) to the weaker condition that

507: $\boldM(\boldx)$ be a strictly SOS matrix.  With these

508: manipulations we see that existence of SOS matrices

509: $\boldM(\boldx)$ and $\boldR(\boldx)$ is a sufficient condition

510: for contraction.

511: 

512: %

513: 

514: \begin{lemma}

515: Existence of a strictly SOS matrix $\boldM(\boldx)$ and a strictly SOS

516: matrix $-\boldR(\boldx) = -(\pardfdxnm ^T \boldM + \boldM\pardfdxnm +

517: \dot{\boldM})$ is a sufficient condition for global contraction of an

518: autonomous system $\dot{\boldx} = \boldf(\boldx)$ with polynomial

519: dynamics.

520: \end{lemma}

521: 

522: \begin{proof}

523: By Theorem 1, a sufficient condition for contraction of any nonlinear

524: system is the existence of uniformly positive definite

525: $\boldM(\boldx)$ and $-\boldR(\boldx)$. A sufficient condition for

526: uniform positive definiteness of $\boldM(\boldx)$ and

527: $-\boldR(\boldx)$ is the existence of strictly SOS matrices

528: $\boldM(\boldx)$ and $-\boldR(\boldx)$.

529: \end{proof}

530: %

531: %

532: 

533: This lemma can easily be extended to existence of certain SOS

534: matrices implying contraction with a convergence rate $\beta$ by

535: redefining $\boldR(\boldx)$ as $\boldR (\boldx) =

536: \pardfdxnm ^T \boldM + \boldM\pardfdxnm + \dot{\boldM} + \beta

537: \boldM$. At this point, we do not know if the full converse of

538: Lemma~1 holds. If a system is exponentially contracting, it is

539: known that a contraction matrix always exists \cite{OCA}.

540: Nevertheless, a system with polynomial dynamics may certainly be

541: contracting under non-polynomial metrics. Furthermore, even if a

542: positive definite contraction matrix with polynomial entries

543: $\bold M$ exists, it may not be the case that it is a SOS matrix.

544: We notice, however, that some of these issues, such as the gap

545: between ``true'' contracting metrics and SOS-based ones, can be

546: bridged by using the more advanced techniques explained in

547: \cite{sdprelax}.

548: 

549: %

550: 

551: \subsection{Search Algorithm}

552: \label{SearchAlgorithm}

553: 

554: One main contribution of this work is to show how sum of squares

555: (SOS) techniques can be used to algorithmically search for a

556: time-invariant contraction metric for nonlinear systems with

557: polynomial dynamics. Existence of a contraction metric for

558: nonlinear systems certifies contraction (or convergence) of system

559: trajectories. For systems with polynomial dynamics, we can obtain

560: a computationally tractable search procedure by restricting

561: ourselves to a large class of SOS-based metrics.

562: 

563: As suggested by Lemma~1, the main idea is to relax the search for

564: matrices that satisfy matrix definiteness constraints $\boldM(\boldx)

565: \succ 0$ and $-\boldR(\boldx) \succ 0$ into SOS-matrix sufficient

566: conditions. Equivalently, we want to find a polynomial matrix

567: $\boldM(\boldx)$ that satisfies SOS matrix constraints on

568: $\boldM(\boldx)$ and $\boldR(\boldx)$.

569: %

570: %

571: %

572: The SOS feasibility problem can then be formulated as finding

573: $\boldM(\boldx)$ and $\boldR(\boldx)$ such that

574: $\boldy^T\boldM(\boldx)\boldy$ is SOS and

575: $-\boldy^T\boldR(\boldx)\boldy$ is SOS.

576: 

577: More specifically, the detailed steps in the algorithmic search of

578: contraction metrics for systems with polynomial dynamics are as

579: follows: \\

580: 

581: \noindent 1. Choose the degree of the polynomials in the contraction

582: metric, and write an affine parametrization of the symmetric matrices

583: of that degree. For instance, if the degree is equal to two, the

584: general form of $\boldM(\boldx)$ is

585: 

586: 

587: %

588: 

589: {\small \[ \left[

590: \begin{array}{cc}

591:   a_1x_1^2 + a_2x_1x_2 + a_3x_2^2 + a_4x_1 + a_5x_2 + a_6 &  b_1x_1^2 + b_2x_1x_2 + b_3x_2^2 + b_4x_1 + b_5x_2 + b_6 \\

592:   b_1x_1^2 + b_2x_1x_2 + b_3x_2^2 + b_4x_1 + b_5x_2 + b_6 & c_1x_1^2 + c_2x_1x_2 + c_3x_2^2 + c_4x_1 + c_5x_2 + c_6 \\

593: \end{array}\right] \]}%

594: where $a_i$, $b_i$, and $c_i$ are

595: unknown coefficients.\\

596: 

597: \noindent 2. Calculate $\pardfdxnm$ and define $\boldR(\boldx) :=

598: \pardfdxnm ^T \boldM + \boldM \pardfdxnm + \dot{\boldM}$. Thus,

599: $\boldR(\boldx)$ will also be a symmetric matrix with entries that

600: depend affinely on the same unknown coefficients $a_i$, $b_i$, and

601: $c_i$. \\

602: 

603: \noindent 3. Change matrix constraints $\boldM(\boldx) \succ 0$

604: $\forall \boldx$, and $\boldR(\boldx) = \pardfdxnm ^T \boldM +

605: \boldM

606: \pardfdxnm + \dot{\boldM} \prec 0 \; \forall \: \boldx$ into scalar

607: constraints on quadratic functions $p(\boldx,\boldy)

608: =\boldy^T\boldM(\boldx)\boldy > 0 \; \forall \: \boldx , \:

609: \boldy$, and $r(\boldx,\boldy) = \boldy^T\boldR(\boldx)\boldy =

610: \boldy^T(\pardfdxnm ^T \boldM + \boldM

611: \pardfdxnm + \dot{\boldM})\boldy <0 \; \forall \: \boldx , \: \boldy$, where

612: $\boldy$ is an $n \times 1$ vector of new indeterminates. \\

613: 

614: \noindent 4. Impose SOS constraints on $p(\boldx,\boldy)$, and

615: $-r(\boldx,\boldy)$, and solve the associated SOS feasibility

616: problem. If a solution exists, the SOS solver will find values for

617: the unknown coefficients, such that the constraints are satisfied. \\

618: 

619: \noindent 5. Use the obtained coefficients $a_i, b_i, c_i$ to

620: construct the contraction metric $\boldM(\boldx)$ and the

621: corresponding $\boldR(\boldx)$. \\

622: 

623: \noindent 6. Optionally, for graphical presentation, independent

624: verification, or if the convex optimization procedure runs into

625: numerical error, further testing can be done to verify the

626: validity of the computed solution. To do this, we can check if the

627: matrix constraints $\boldM(\boldx) \succ 0$, and $\boldR(\boldx)

628: \prec 0$ hold over a range of the state space by finding and

629: plotting the eigenvalues over this range.  If a true feasible

630: solution does not exist, the minimum eigenvalue of

631: $\boldM(\boldx)$ will be negative or the maximum eigenvalue of

632: $\boldR(\boldx)$ will be positive. Either one of these cases

633: violates the matrix constraints which certify contraction. In most

634: semidefinite programming solvers, the matrix $Q$ in (\ref{Q}) is

635: computed with floating point arithmetic.  If $Q$ is near the

636: boundary of the set of positive semidefinite matrices, it is

637: possible for the sign of eigenvalues that are zero or close to

638: zero to be computed incorrectly from numerical roundoff and for

639: the semidefinite program solver to encounter numerical

640: difficulties. Numerical issues are further discussed in Section

641: \ref{NumericalProblems}. \\

642: 

643: \noindent 7. An explicit lower bound on the rate of convergence

644: can be found by using bisection to compute the largest $\beta$ for

645: which there exist matrices $\boldM(\boldx) \succ 0$ and $

646: \boldR_{\beta}(\boldx) =

647: \pardfdxnm ^T \boldM + \boldM

648: \pardfdxnm + \dot{\boldM} + \beta \boldM \prec 0$.\\

649: 

650: For the specific examples presented later in the paper, we have

651: used SOSTOOLS, a SOS toolbox for MATLAB developed for the

652: specification and solution of sums of squares programs

653: \cite{SOSTOOLS}. The specific structure of SOS matrices, or

654: equivalently, the bipartite form of the polynomials $p(\boldx,

655: \boldy)$ and $r(\boldx, \boldy)$ is exploited through the option

656: \texttt{sparsemultipartite} of the command \texttt{sosineq} that

657: defines the SOS inequalities. Future versions of SOSTOOLS will

658: allow for the direct specification of matrix SOS constraints.

659: 

660: We present next are two examples of using this procedure to search

661: for contraction metrics for nonlinear systems with polynomial

662: dynamics. The systems studied are a model of a jet engine with

663: controller, and a Van der Pol oscillator.

664: 

665: \subsection{Example: Moore-Greitzer Jet Engine Model}\label{jetexamples}

666: 

667: The algorithm described was tested on the following dynamics,

668: corresponding to a Moore-Greitzer model of a jet engine, with

669: stabilizing feedback operating in the no-stall mode \cite{KKK}. In

670: this model, the origin is translated to a desired no-stall

671: equilibrium.  The state variables correspond to $\phi = \Phi - 1$,

672: $\psi = \Psi - \Psi_{co} - 2$, where $\Phi$ is the mass flow, $\Psi$

673: is the pressure rise and $\Psi_{co}$ is a constant \cite{KKK}.

674: The dynamic equations take the form:

675: \begin{equation}\label{JetDynamics}

676: \left[%

677: \begin{array}{c}

678:   \dot{\phi} \\

679:   \dot{\psi} \\

680: \end{array}%

681: \right] =\left[%

682: \begin{array}{c}

683:   -\psi - \frac{3}{2}\phi^2 - \frac{1}{2}\phi^3 \\

684:   3\phi - \psi \\

685: \end{array}%

686: \right]

687: \end{equation}

688: The only real-valued equilibrium of the system is $\phi = 0$, $ \psi=

689: 0$. This equilibrium is stable.

690: 

691: The results of the algorithmic search for SOS matrices

692: $\boldM(\boldx)$ and $-\boldR(\boldx)$ of various orders are given

693: in Table~\ref{JetTable}. Values in the table, except the final

694: row, are output values from SeDuMi \cite{SeDuMi}, the semidefinite

695: program solver used as the optimization engine in solving the SOS

696: program. \texttt{CPU time} is the number of seconds it took for

697: SeDuMi's interior point algorithm to find a solution. As expected,

698: the computation time increases with the degree of the polynomial

699: entries of $\boldM(\boldx)$. \texttt{Feasibility ratio} is the

700: final value of the feasibility indicator. This indicator converges

701: to $1$ for problems with a complementary solution, and to $-1$ for

702: strongly infeasible problems. If the feasibility ratio is

703: somewhere in between, this is usually an indication of numerical

704: problems.  The values \texttt{pinf} and \texttt{dinf} detect the

705: feasibility of the problem.  If $\texttt{pinf}=1$, then the primal

706: problem is infeasible. If $\texttt{dinf}=1$, the dual problem is

707: infeasible.  If \texttt{numerr} is positive, the optimization

708: algorithm (i.e., the semidefinite program solver) terminated

709: without achieving the desired accuracy. The value

710: $\texttt{numerr}=1$ gives a warning of numerical problems, while

711: $\texttt{numerr}=2$ indicates a complete failure due to numerical

712: problems.

713: 

714: As shown in Table~\ref{JetTable}, for this system no contraction

715: metric with polynomial entries of degree 0 or 2 could be found.

716: This can be certified from the solution of the dual optimization

717: problem. Since SeDuMi is a primal-dual solver, this infeasibility

718: certificates are computed as a byproduct of the search for

719: contraction metrics.

720: 

721: An explicit lower bound for the rate of convergence of the

722: trajectories of the jet engine model, i.e., the largest value

723: $\beta$ for which matrices $\boldM(\boldx) \succ 0$ and $

724: \boldR_{\beta}(\boldx) =

725: \pardfdxnm ^T \boldM + \boldM

726: \pardfdxnm + \dot{\boldM} + \beta \boldM \prec 0$ were found, was $\beta = 0.818$.

727: 

728: %

729: %

730: %

731: %

732: 

733: 

734: We remark that for this system, it is also possible to prove

735: stability using standard Lyapunov analysis techniques. However, we

736: illustrate stability of this example from a contraction viewpoint

737: because contraction theory offers a good approach to study this

738: system when there is parametric uncertainty in the plant dynamics

739: or feedback equations. For example, in the no-stall mode, the jet

740: dynamics equations are

741: \begin{equation}\label{JetDynamicscontrol}

742: \left[%

743: \begin{array}{c}

744:   \dot{\phi} \\

745:   \dot{\psi} \\

746: \end{array}%

747: \right] =\left[%

748: \begin{array}{c}

749:   -\psi - \frac{3}{2}\phi^2 - \frac{1}{2}\phi^3 \\

750:   -u\\

751: \end{array}%

752: \right]

753: \end{equation} where $u$ is a control variable.

754: If a nominal stabilizing feedback control $u$ can be found (e.g.,

755: using backstepping \cite{KKK} or some other design method), the SOS

756: techniques described in Section \ref{Uncert-alg-des} provide a way to

757: find other stabilizing feedback controls which are centered around the

758: nominal control.  For example, if a stabilizing linear feedback

759: control $u = k_{1} \phi + k_{2} \psi$ can be found, we can interpret

760: $k_{1}$ and $k_{2}$ as uncertain parameters and use the methods

761: described in Section \ref{Uncert-alg-des} to search for ranges of gain

762: values centered around the nominal values $k_{1}$ and $k_{2}$ that

763: will also stabilize the system.

764: 

765: \begin{table}

766: \begin{footnotesize}\begin{center}

767: \begin{tabular}{|l||c|c|c|c|}

768:   \hline

769:   %

770:   Degree of polynomials in $\boldM(\boldx)$& 0 &2 & 4  & 6  \\

771:   \hline

772:   CPU time (sec)  &0.140 & 0.230 & 0.481& 0.671  \\

773:   \hline

774:   Feasibility ratio& -1.000  &-0.979 & 1.003 & 0.990  \\

775:   \hline

776:   \texttt{pinf} &1 & 1& 0 & 0  \\

777:   \hline

778:   \texttt{dinf} & 0& 0 & 0 & 0  \\

779:   \hline

780:   \texttt{numerr} &0 & 1 & 0 & 0  \\

781:   \hline

782:   $M \succ 0$, $R \prec 0$ conditions met? & no & no & yes & yes \\

783:   \hline

784: \end{tabular} \end{center}\end{footnotesize}

785: \caption{Contraction matrix search results for closed-loop jet

786: engine dynamics.}\label{JetTable}

787: \end{table}

788: 

789: 

790: \subsection{Example: Van der Pol Oscillator}

791: 

792: A classic example that has played a central role in the

793: development of nonlinear dynamics is given by the Van der Pol

794: equation

795: \begin{equation}\label{vdposc}

796: \ddot{x} + \alpha(x^2 + k)\dot{x} + \omega^2x = 0,

797: \end{equation}

798: with $\alpha \geq 0$, $k$, and $\omega$ as parameters.  Historically

799: this equation arose from studying nonlinear electric circuits used in

800: the first radios \cite{Strogatz}. When $k<0$, the solutions of

801: (\ref{vdposc}) behave like a harmonic oscillator with a nonlinear

802: damping term $\alpha (x^2 + k)\dot{x}$.  The term provides positive

803: damping when $|x|>k$ and negative damping when $|x|<k$. Thus, large

804: amplitude oscillations will decay, but if they become too small they

805: will grow larger again \cite{Strogatz}. If $k >0$ all trajectories

806: converge to the origin.

807: 

808: In Table~\ref{OscTableu0} we present the results of running the

809: contraction matrix search algorithm for the system

810: \begin{displaymath}

811: \left[

812: \begin{array}{c}

813:   \dot{x}_1 \\

814:   \dot{x}_2 \\

815: \end{array}\right] = \left[

816: \begin{array}{c}

817:   x_2 \\

818:   -\alpha (x_1^2 + k)x_2 - \omega^2 x_1  \\

819: \end{array}

820: \right], \end{displaymath} with $\alpha = 1, \omega = 1$, which is the

821: state-space version of the Van der Pol oscillator (\ref{vdposc}). We

822: present solution for various values of $k$, with a contraction matrix

823: with entries that are quartic polynomials.

824: 

825: As a natural first step we searched for a constant contraction

826: metric.   None could be found algorithmically.  This was expected

827: as it is easily shown analytically that a constant contraction

828: matrix for this system does not exist. If $\boldM$ is constant,

829: then {\small

830: \begin{eqnarray}

831: \boldM  = \left[%

832: \begin{array}{cc}

833:   a & b \\

834:   b & c \\

835: \end{array}%

836: \right], \qquad \pardfdxnm = \left[%

837: \begin{array}{cc}

838:   0 & 1 \\

839:   -1 - 2x_1x_2 & -x_1^2 - k \\

840: \end{array}%

841: \right], \nonumber \\ \nonumber \\

842: \boldR = \left[%

843: \begin{array}{cc}

844:   -2b - 4bx_1x_2 & a - bx_1^2 - kb - c - 2cx_1x_2 \\

845:   a - bx_1^2 - kb - c - 2cx_1x_2 & 2b-2cx_1^2 - 2kc \\

846: \end{array}%

847: \right].

848: \end{eqnarray}} For $\boldR$ to be negative definite $\boldR_{11}$ must be

849: negative for all values of $x_1$, $x_2$. In other words $-2b -

850: 4bx_1x_2 \leq 0$ or $-1 \leq 2 x_1x_2$.  This clearly does not hold

851: for all values of $x_1$, and $x_2$. A more complicated analysis (or a

852: duality argument) also shows why there is no contraction matrix with

853: quadratic entries for this system.

854: 

855: %

856: %

857: 

858: \begin{table*}

859: \begin{footnotesize} \begin{center}

860: 

861: \begin{tabular}{|c||c|c|c|c|c||c|c|c|c|c|c|c|}

862: \hline

863:   %

864:   Degree of polynomials in $\boldM(\boldx)$ &4 &4 &4 &4 &4 &4 &4 &4

865:   &4 &4\\

866:   \hline

867:   k   & -10 & -1 & -0.1 & -0.01 & -0.001 & 0.001 & 0.01 & 0.1 & 1 & 10 \\

868:   \hline

869:   pinf  & 1 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\

870:   \hline

871:   dinf  & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0& 0\\

872:   \hline

873:   numerr  & 1 & 1 & 1 & 1 & 1 & 0 & 0 & 0 & 0 & 0 \\

874:   \hline

875:  $M \succ 0$, $R \prec 0$ conditions met? & no & no & no & no & no & yes & yes & yes & yes & yes \\

876:   \hline

877: 

878: \end{tabular}

879: \end{center} \end{footnotesize}

880: \caption{Contraction matrix search results for oscillator

881: dynamics.}\label{OscTableu0}

882: \end{table*}

883: 

884: The algorithm finds a contraction function for the system

885: $\ddot{x} + (x^2 + k)\dot{x} + x = 0$ when $k >0$ but not when

886: $k<0$. As shown in Figure \ref{pp_vdpo} the trajectories of the

887: oscillator converge to zero when $k > 0$, and converge to a

888: limit-cycle when $k <0$. Thus, the results of the contraction

889: metric search is as expected.  Since all trajectories converge to

890: the origin when $k >0$ we expect that a contraction metric exists

891: for the system.  In the case where $k <0$ the origin is an

892: unstable fixed point and thus the system is not contracting.

893: 

894: Since for $k<0$ the system is not contracting, we should not be able

895: to find a contraction function. It should be noted that the converse

896: does not hold. The fact that we cannot find a contraction function

897: does not necessarily mean that the system is not contracting. This is

898: because finding an SOS representation of the constrained quadratic

899: functions is a sufficient condition for their positivity, not a

900: necessary one.

901: 

902: 

903: \begin{figure}

904:   \begin{center}

905:     \subfigure[$ k  = 0.5$]{\label{pp_osc_kpos}\includegraphics[scale=0.4]{pp_osc_kpt5.eps}}

906:    \hspace{1.5 cm}

907:     \subfigure[$k = -0.5$]{\label{pp_osc_kneg}\includegraphics[scale=0.4]{pp_osc_kmpt5.eps}}

908:   \end{center}

909:   \caption{Phase plots of Van der pol Oscillator.}

910:   \label{pp_vdpo}

911: \end{figure}

912: 

913: %

914: %

915: 

916: It should be noted that for the example above, we can prove

917: stability through Lyapunov analysis, and SOS programming can also

918: be used to find Lyapunov functions \cite{SOSlyap}.  However, we

919: illustrate this example here as contraction theory applied to a

920: slightly modified version of this system provides a nice way to

921: prove synchronization of coupled Van der Pol oscillators. This

922: will be discussed in Section \ref{EI}. This synchronization

923: property is much more difficult to prove with standard Lyapunov

924: methods.

925: 

926: \section{Contraction Metrics and Systems with Uncertain Dynamics}

927: \label{UD}

928: \subsection{Uncertainty Analysis with Contraction Metrics and

929: SOS Programming}\label{Uncert-alg-des}

930: 

931: From the robust control perspective, one of the most appealing

932: features of contraction theory is the fact that it provides a natural

933: framework in which to study uncertain nonlinear systems where the

934: parametric uncertainty changes the location of the equilibrium

935: points. In general, standard Lyapunov analysis does not handle this

936: situation particularly well, since the Lyapunov function must track

937: the changes in the location of the steady-state solutions, thus

938: forcing the use of parameter-dependent Lyapunov functions.  However,

939: in general it may be impossible to obtain any kind of closed form

940: expression of the equilibria in terms of the parameters, thus

941: complicating the direct parametrization of possible Lyapunov

942: functions.

943: 

944: Much attention has been given to robust stability analysis of linear

945: systems (e.g.,

946: \cite{PDL:Haddad,PDL:Gahinet,PDL:Feron,LMIBook,Zhou}). Less attention,

947: however, has been paid to nonlinear systems with moving

948: equilibria. Two papers addressing this issue are

949: \cite{Wang-Michel,Rantzerpaper}. The approach in \cite{Wang-Michel} is

950: to consider systems described by the equations

951: \begin{equation}

952: \dot{\mathbf{x}} = \mathbf{f}(\mathbf{x}) + \mathbf{h}(\mathbf{x}),

953: \label{wangeq}

954: \end{equation}

955: where $\mathbf{x}$ is a real $n$-vector, $\mathbf{f}$ and $\mathbf{h}$

956: are continuously differentiable functions, and

957: $\mathbf{h}(\mathbf{x})$ represents the uncertainties or perturbation

958: terms. Given an exponentially stable equilibrium $\mathbf{x}_e$,

959: \cite{Wang-Michel} establishes sufficient conditions by using the

960: linearization of the system to produce Lyapunov functions which prove

961: existence and local exponential stability of an equilibrium

962: $\tilde{\mathbf{x}}_e$ for (\ref{wangeq}) with the property

963: $|\mathbf{x}_e - \tilde{\mathbf{x}}_e| < \varepsilon$ where

964: $\varepsilon$ is sufficiently small.

965: 

966: Since the approach in \cite{Wang-Michel} is essentially based on a

967: fixed Lyapunov function, it is more limited than our approach using

968: contraction theory and SOS programming, and can prove stability only

969: under quite conservative ranges of allowable uncertainty.  By

970: ``allowable'' we mean that if the uncertainty is in this range, the

971: equilibrium remains exponentially stable under the uncertainty. A

972: quantitative measure of this conservativeness will be given in Section

973: \ref{Scalar Additive Uncertainty} where we discuss the results of the

974: method applied to an uncertain model of a Moore-Greitzer jet engine

975: and compare them to an approach via contraction theory and SOS

976: programming.

977: 

978: The approach in \cite{Rantzerpaper} is to linearize the dynamics

979: around an equilibrium which is a function of the uncertain

980: parameter ($\boldx_0 = \boldg(\delta), \; \delta \in \Omega$) and

981: then use structured singular values to determine the eigenvalues

982: of the linearized system $\frac{d\boldz}{dt} = A(\delta)\boldz$ if

983: $A(\delta)$ is rational in $\delta$.

984: %

985: %

986: If $A(\delta)$ is marginally stable, no conclusions can be made

987: about the stability of the nonlinear system.

988: 

989: The contraction theory framework eliminates the need for

990: linearization, and even the need to know the exact position of the

991: equilibrium, in order to analyze stability robustness in uncertain

992: nonlinear systems. In contrast to the Lyapunov situation, when

993: certain classes of parametric uncertainty are added to the system,

994: a contraction metric for the nominal system will often remain a

995: contraction metric for the system with uncertainty, even if the

996: perturbation has changed the equilibrium of the nonlinear system.

997: 

998: As noted in Section \ref{CT}, if a global time-invariant

999: contraction metric exists for an autonomous system, and an

1000: equilibrium point exists for the system, all trajectories converge

1001: to a unique equilibrium point, and we can always produce a

1002: Lyapunov function of the form $V(\boldx) =

1003: \boldf(\boldx)^T\boldM(\boldx)\boldf(\boldx)$. When a system

1004: contains parametric uncertainty, this formula yields the

1005: parameter-dependent Lyapunov function $V(\boldx, \delta) =

1006: \boldf(\boldx, \delta)^T\boldM(\boldx)\boldf(\boldx, \delta)$ for

1007: ranges of the parametric uncertainty $\delta$ where the

1008: contraction metric for the nominal system is still a contraction

1009: metric for the system with perturbed dynamics. Thus, if a

1010: contraction metric can be found for the system, we can easily

1011: construct a Lyapunov function which tracks the uncertainty for a

1012: certain range.

1013: 

1014: \subsubsection{Case 1: Bounds on the uncertainty range

1015: for which the system remains contractive with respect to the nominal

1016: metric.}

1017: 

1018: We can estimate the range of uncertainty under which the

1019: contraction metric for the nominal system is still a contraction

1020: metric for the perturbed system. To calculate this range, a SOS

1021: program can be written to minimize or maximize the amount of

1022: uncertainty allowed subject to the constraint

1023: $\boldR_{\delta}(\boldx) =

1024: \pardfdxdel ^T \boldM + \boldM

1025: \pardfdxdel + \dot{\boldM}(\boldf_{\delta}(\boldx)) \prec 0$, where $\boldf_{\delta}(\boldx)$ are the dynamics

1026: for the system with parametric uncertainty.  The uncertainty bound

1027: is a decision variable in the SOS program and enters the

1028: constraint above in the $\pardfdxdel$ and

1029: $\boldf_{\delta}(\boldx)$ terms.

1030: 

1031: If we have more than one uncertain parameter in the system, we can

1032: find a polytopic inner approximation of the set of allowable

1033: uncertainties with SOS Programming. For example, if we have two

1034: uncertain parameters, we can algorithmically find a polytope in

1035: parameter space for which the original metric is still a contraction

1036: metric. The convex hull of four points, each which can be found by

1037: entering one of the four combinations, $(\delta _1, \delta_2) =

1038: (\gamma, \gamma)$, $(\delta _1, \delta_2) = (\gamma, -\gamma)$,

1039: $(\delta _1, \delta_2) = (-\gamma, \gamma)$, or $(\delta _1, \delta_2)

1040: = (-\gamma, -\gamma)$, into the uncertainty values in

1041: $\boldf_{\mathbf{\delta} = [\delta_1, \delta_2]^T}(\boldx)$ and then

1042: maximizing $\gamma$ subject to the constraint $\boldR_{\gamma}(\boldx)

1043: = \pardfdxgam ^T \boldM + \boldM \pardfdxgam +

1044: \dot{\boldM}(\boldf_{\gamma}(\boldx)) \prec 0$, defines a polytope

1045: over which stability is guaranteed.

1046: 

1047: \subsubsection{Case 2: Search for a contraction metric that guarantees the largest

1048: symmetric uncertainty interval for which the system is contractive.}

1049: 

1050: Alternatively, we can instead optimize the search for a metric

1051: $\boldM(\boldx)$ that provides the largest symmetric uncertainty

1052: interval for which we can prove the system is contracting. If the

1053: scalar uncertainty $\delta$ enters the system dynamics affinely, in

1054: other words if $\boldf(\boldx) = \mathbf{f_1}(\boldx) + \delta

1055: \, \mathbf{f_2}(\boldx)$, we can perform this optimization as

1056: follows. First write $\boldR(\boldx, \delta) = \mathbf{R_0}(\boldx) +

1057: \delta \mathbf{R_1}(\boldx)$.  To find the largest interval $(-\gamma,

1058: \gamma)$ such that for all $\delta$ that satisfy $-\gamma < \delta <

1059: \gamma$ the system is contracting, introduce the following constraints

1060: into an SOS program: \begin{displaymath} \boldM(\boldx) \succ 0, \quad

1061: \mathbf{R_0}(\boldx) + \gamma \mathbf{R_1}(\boldx) \prec 0, \quad

1062: \mathbf{R_0}(\boldx) -\gamma \mathbf{R_1}(\boldx) \prec 0.

1063: \end{displaymath}  We note that $\gamma$ multiplies the scalar decision coefficients

1064: $a_i$, $b_i$, and $c_i$ in $\boldR_1(\boldx)$ and thus we must use

1065: a bisection procedure to find the maximum value of $\gamma$ for

1066: which there exists SOS matrices $\boldM(\boldx)$,

1067: $\mathbf{R_0}(\boldx)$ and $\mathbf{R_1}(\boldx)$ that satisfy the

1068: constraints above.

1069: 

1070: If there are two uncertain parameters that enter the system

1071: dynamics affinely, we can extend the procedure above as follows:

1072: To find the largest uncertainty square with width and height

1073: $\gamma$ such that for all $\delta_1$ and $\delta_2$ that satisfy

1074: $-\gamma < \delta_1 < \gamma$ and $-\gamma < \delta_2 < \gamma$

1075: the system is contracting, first write $\boldR(\boldx, \delta_1,

1076: \delta_2) = \mathbf{R_0}(\boldx) + \delta_1 \mathbf{R_1}(\boldx) +

1077: \delta_2 \mathbf{R_2}(\boldx)$,  Then introduce the following

1078: constraints into and SOS program:

1079: \begin{eqnarray}

1080: \boldM(\boldx) \succ 0, \quad \mathbf{R_0}(\boldx) + \gamma

1081: \mathbf{R_1}(\boldx) + \gamma \mathbf{R_2}(\boldx) \prec 0, \quad

1082: \mathbf{R_0}(\boldx) + \gamma \mathbf{R_1}(\boldx) - \gamma \mathbf{R_2}(\boldx) \prec 0 \nonumber \\

1083: \mathbf{R_0}(\boldx) - \gamma \mathbf{R_1}(\boldx) + \gamma

1084: \mathbf{R_2}(\boldx) \prec 0, \quad \mathbf{R_0}(\boldx) - \gamma

1085: \mathbf{R_1}(\boldx) - \gamma \mathbf{R_2}(\boldx) \prec 0.

1086: \end{eqnarray}

1087: Next, as in the scalar uncertainty case, use a bisection procedure to

1088: find the maximum value of $\gamma$ for which there exists SOS matrices

1089: $\boldM(\boldx)$, $\mathbf{R_0}(\boldx) $, $\mathbf{R_1}(\boldx)$ and

1090: $\mathbf{R_2}(\boldx)$ that satisfy the constraints above.  In the

1091: case of a large number of uncertain parameters, standard relaxation

1092: and robust control techniques can be used to avoid an exponential

1093: number of constraints.

1094: 

1095: \subsection{Example: Moore-Greitzer Jet Engine Model with Uncertainty}

1096: \subsubsection{Scalar Additive Uncertainty}\label{Scalar Additive

1097: Uncertainty}

1098: 

1099: As described above, SOS programming can be used to find ranges of

1100: uncertainty under which a system with uncertain perturbations is

1101: still contracting with the original contraction metric. The

1102: contraction metric found for the deterministic system continues to be

1103: a metric for the perturbed system over a range of uncertainty even if

1104: the uncertainty shifts the equilibrium point and trajectories of the

1105: system. For the Moore-Greitzer jet engine model, the dynamics in

1106: (\ref{JetDynamics}) were perturbed by adding a constant term $\delta$

1107: to the first equation.

1108: \begin{equation}\label{JetDynamics_eps}

1109: \left[%

1110: \begin{array}{c}

1111:   \dot{\phi} \\

1112:   \dot{\psi} \\

1113: \end{array}%

1114: \right] =\left[%

1115: \begin{array}{c}

1116:   -\psi - \frac{3}{2}\phi^2 - \frac{1}{2}\phi^3 + \delta \\

1117:   3\phi - \psi \\

1118: \end{array}%

1119: \right]

1120: \end{equation}

1121: 

1122: 

1123: \begin{table}

1124: \begin{center}

1125: \footnotesize

1126: \begin{tabular}[h]{|l||c|c|c|c|}

1127:   \hline

1128:   %

1129:   Degree of polynomials in $\boldM(\boldx)$ & 4  & 6  \\

1130:   \hline

1131:   $ \delta$ range &(-0.126,0.630)  &( -0.070, 0.635) \\

1132:     \hline

1133: \end{tabular} \normalsize \end{center}

1134: \caption{Range of perturbation where closed-loop uncertain jet

1135: engine dynamics given in (\ref{JetDynamics_eps}) are contracting

1136: with respect to the nominal metric.} \label{epsJettable}

1137: \end{table}

1138: 

1139: In Table \ref{epsJettable} we display the ranges of $\delta$ where

1140: the system was still contracting with the original contraction

1141: metric for $4^{th}$ and $6^{th}$ degree contraction metrics. Note

1142: the range of allowable uncertainty is not symmetric.

1143: 

1144: When instead we optimized the contraction metric search to get the

1145: largest symmetric $\delta$ interval we obtained the results listed

1146: in Table~\ref{tbl:optdelta}.  A $6^{th}$ degree contraction

1147: function finds the uncertainty range $|\delta| \leq 1.023$.

1148: Because a Hopf bifurcation occurs in this system at $\delta

1149: \approx 1.023$, making the system unstable for $\delta > 1.023$,

1150: we can conclude that the 6th degree contraction metric is the

1151: highest degree necessary to find the maximum range of uncertainty

1152: for which the system is contracting. The Hopf bifurcation is shown

1153: in Figure~\ref{fig:Hopf}.

1154: 

1155: Using the techniques in \cite{Wang-Michel} we computed the

1156: allowable uncertainty range for the system given in

1157: (\ref{JetDynamics_eps}) as $|\delta| \leq 5.1 \times 10^{-3}$. In

1158: the notation of \cite{Wang-Michel}, we calculated the other

1159: parameters in Assumption 1  of \cite{Wang-Michel} as: $\mathbf{h}

1160: = [\delta, \; \; 0]^T$, $|A^{-1}|_\infty = 1$, $|D

1161: \mathbf{h}(\boldx_e)|_\infty = 0$, $a = \frac{1}{30}$, and

1162: $|\mathbf{h}(\boldx_e)|_\infty = |\delta|$, where $\delta$ is the

1163: perturbation term in (\ref{JetDynamics_eps}).  The allowable range

1164: $|\delta| \leq 1.023$ computed via contraction theory and SOS

1165: programming is much larger than the allowable uncertainty range

1166: $|\delta| \leq 5.1 \times 10^{-3}$ computed with the techniques in

1167: \cite{Wang-Michel}.

1168: 

1169: \begin{table*}

1170: \begin{footnotesize}\begin{center}

1171: \begin{tabular}{|l||c|c|c|}

1172:   \hline

1173:   %

1174:   Degree of polynomials in $\boldM(\boldx)$ & 4  & 6 & 8  \\

1175:   \hline

1176:    $ \delta$ range  &$|\delta| \leq 0.938$  & $|\delta| \leq 1.023$ & $|\delta| \leq 1.023$ \\

1177:   \hline

1178: \end{tabular} \end{center}\end{footnotesize}

1179: \caption{Symmetric range of perturbation for the jet engine

1180: model.} \label{tbl:optdelta}

1181: \end{table*}

1182: 

1183: \begin{figure}[]

1184:   \begin{center}

1185:     \subfigure[$\delta =

1186:     -0.5$]{\label{JetDelMpt5}\includegraphics[scale=0.5]{JetPhasePlotd_Minus0pt5.eps}}

1187:     \hspace{1cm}

1188:     \subfigure[$\delta = -1.01$]{\label{JetDelM1pt01}\includegraphics[scale=0.5]{JetPhasePlotd_Minus1pt01.eps}}

1189:     \hspace{1cm}

1190:     \subfigure[$\delta = -1.1$]{\label{JetDelM1pt1}\includegraphics[scale =0.5]{JetPhasePlotd_Minus1pt1.eps}}

1191:   \end{center}

1192:   \caption{Hopf bifurcation in uncertain jet dynamics.}

1193:   \label{fig:Hopf}

1194: \end{figure}

1195: 

1196: 

1197: \subsubsection{Scalar Multiplicative Uncertainty}

1198: 

1199: The approaches in Section \ref{Uncert-alg-des} also apply to

1200: multiplicative uncertainty, since the multiplicative coefficients

1201: enter affinely in the constraints in the SOS program. Tables

1202: \ref{epsJettable-mult} and \ref{tbl:optdelta:mult} present the results

1203: of the described uncertainty analysis on the following system, which

1204: is equation (\ref{JetDynamics}) with multiplicative uncertainty.

1205: \begin{equation}\label{JetDynamics:multuncert}

1206: \left[%

1207: \begin{array}{c}

1208:   \dot{\phi} \\

1209:   \dot{\psi} \\

1210: \end{array}%

1211: \right] =\left[%

1212: \begin{array}{c}

1213:   -\psi - \frac{3}{2}\phi^2 - \frac{1}{2} \delta \phi^3 \\

1214:   3\phi - \psi \\

1215: \end{array}%

1216: \right].

1217: \end{equation}

1218: 

1219: \begin{table}

1220: \begin{center}

1221: \footnotesize

1222: \begin{tabular}{|l||c|c|c|c|}

1223:   \hline

1224:   %

1225:   Degree of polynomials in $\boldM(\boldx)$ & 4  & 6  \\

1226:   \hline

1227:   $ \delta$ range & (0.9767, 5.8686) & (0.9796, 3.9738)  \\

1228:     \hline

1229: \end{tabular} \normalsize \end{center}

1230: \caption{Range of perturbation for which the uncertain system given in

1231: (\ref{JetDynamics:multuncert}) is contracting with respect to the

1232: nominal metric.}

1233: \label{epsJettable-mult}

1234: \end{table}

1235: 

1236: \begin{table*}

1237: \begin{footnotesize}\begin{center}

1238: \begin{tabular}{|l||c|c|c|}

1239:   \hline

1240:   %

1241:   Degree of polynomials in $\boldM(\boldx)$ & 4  & 6 & 8  \\

1242:   \hline

1243:    $ \delta$ range  &$(1- 0.247,1+ 0.247)$  & $(1 - 0.356  ,1+ 0.356 )$ & $(1 - 0.364  ,1+0.364)$ \\

1244:   \hline

1245: \end{tabular} \end{center}\end{footnotesize}

1246: \caption{Symmetric range of perturbation where uncertain

1247: closed-loop jet engine dynamics given in

1248: (\ref{JetDynamics:multuncert}) are contracting.}

1249: \label{tbl:optdelta:mult}

1250: \end{table*}

1251: 

1252: \subsubsection{Multiple Uncertainties}

1253: 

1254: We consider next the system that results from introducing two additive

1255: uncertainties to the jet dynamics in equation (\ref{JetDynamics}). We

1256: computed an uncertainty polytope (shown in Figure \ref{poly-uncert})

1257: for which the system

1258: \begin{equation}\label{JetDynamics:2Duncert}

1259: \left[%

1260: \begin{array}{c}

1261:   \dot{\phi} \\

1262:   \dot{\psi} \\

1263: \end{array}%

1264: \right] =\left[%

1265: \begin{array}{c}

1266:   -\psi - \frac{3}{2}\phi^2 - \frac{1}{2}\phi^3 + \delta_1 \\

1267:   3\phi - \psi + \delta_2 \\

1268: \end{array}%

1269: \right]

1270: \end{equation}

1271: is guaranteed to be contracting with respect to the original metric.

1272: \begin{figure}[]

1273:   \begin{center}

1274:    \includegraphics[scale=0.4]{2Duncertainty.eps}

1275:    \end{center}

1276:   \caption{Polytopic region of uncertainty where closed-loop jet engine

1277: dynamics given in (\ref{JetDynamics:2Duncert}) are contracting with respect to nominal metric.}

1278:   \label{poly-uncert}

1279: \end{figure} Table \ref{tbl:optdelta:2D} shows the results of optimizing the

1280: contraction metric to find the largest uncertainty square with

1281: width and height $\gamma$ such that for all $\delta_1$ and

1282: $\delta_2$ that satisfy $-\gamma < \delta_1 < \gamma$ and $-\gamma

1283: < \delta_2 < \gamma$ the system is contracting.

1284: \begin{table*}

1285: \begin{footnotesize}\begin{center}

1286: \begin{tabular}{|l||c|c|c|}

1287:   \hline

1288:   %

1289:   Degree of polynomials in $\boldM(\boldx)$ & 4  & 6 & 8  \\

1290:   \hline

1291:     Height and width of allowed uncertainty box  & 0.7093  & 0.7321  & 0.7346  \\

1292:   \hline

1293: \end{tabular} \end{center}\end{footnotesize}

1294: \caption{Symmetric range of perturbation where uncertain

1295: closed-loop dynamics given in (\ref{JetDynamics:2Duncert}) are

1296: contracting.} \label{tbl:optdelta:2D}

1297: \end{table*}

1298: 

1299: 

1300: \section{Contraction Metrics for Systems with External

1301: Inputs}\label{EI}

1302: \subsection{Stability Analysis of Systems with External Inputs}

1303: 

1304: Another interesting feature of the contraction framework is the

1305: relative flexibility in incorporating inputs and outputs. For

1306: instance, to prove contraction of a class of systems with external

1307: inputs, it is sufficient to show the existence of a polynomial

1308: contraction metric with a certain structure. This is described in

1309: the following theorem. \\

1310: 

1311: \begin{theorem}\label{Thm:EI}

1312: Let

1313: \begin{eqnarray}\label{Thm2:eqn}

1314: %

1315:  \nonumber \dot{x}_1 &=& f_1(x_1, x_2, \ldots, x_n) \\

1316:   %

1317: \nonumber  \vdots &=& \vdots \\

1318:  \nonumber \dot{x}_k &=& f_k(x_1, x_2, \ldots, x_n) \\

1319:  \nonumber \dot{x}_{k+1} &=& f_{k+1}(x_1, x_2, \ldots, x_n) + v_{k+1}(u) \\

1320:  \nonumber \vdots &=& \vdots\\

1321:    \dot{x}_n &=& f_n(x_1, x_2, \ldots, x_n)+ v_n(u)

1322: \end{eqnarray}

1323: be a set of nonlinear coupled differential equations where only the

1324: last $n-k$ depend explicitly on $u(t)$.  If there exists a $n \times

1325: n$ matrix $\boldM(x_1, \ldots , x_k)$ such that $\boldM \succ 0$ and

1326: $\dot{\boldM} + \frac{\partial f}{\partial x}^T\boldM + \boldM

1327: \frac{\partial f}{\partial x} \prec 0$ then the system is contracting

1328: for all possible choices of $u(t)$.

1329: \label{thm:inputs}

1330: \end{theorem}

1331: 

1332: \begin{proof}

1333: For notational convenience, let $\dot{\mathbf{x}}_1 = [

1334: \begin{array}{ccc}

1335:   \dot{x}_1 & ... & \dot{x}_k \\

1336: \end{array}]^T = \mathbf{f_1}(\mathbf{x_1}, \mathbf{x_2})$ and

1337: $\dot{\mathbf{x}}_2 = [ \begin{array}{ccc}

1338:   \dot{x}_{k+1} & ... & \dot{x}_n \\

1339: \end{array}]^T = \mathbf{f_2}(\mathbf{x_1}, \mathbf{x_2}, u)$. The metric

1340: $\boldM(x_1, x_2 ..., x_k) = \boldM(\mathbf{x_1})$ is independent

1341: of $\mathbf{x_2}$, and thus $\frac{\partial \boldM_{ij}}{\partial

1342: \mathbf{x_2}} = \mathbf{0} \; \; \forall i,j.$ Since

1343: $\frac{\partial \boldM_{ij}}{\partial t}$ also vanishes, it

1344: follows that $\forall i, j$, $\dot{\boldM}_{ij} = \frac{\partial

1345: \boldM_{ij}}{\partial \mathbf{x_1}}\frac{d \mathbf{x_1}}{dt} +

1346: \frac{\partial \boldM_{ij}}{\partial \mathbf{x_2}}\frac{d

1347: \mathbf{x_2}}{dt} + \frac{\partial \boldM}{\partial t} =

1348: \frac{\partial \boldM_{ij}}{\partial \mathbf{x_1}}\frac{d

1349: \mathbf{x_1}}{dt}$. Thus $\dot{\boldM}(\mathbf{x_1})$ is not a

1350: function of $u(t)$. In addition, $\frac{\partial \boldf}{\partial

1351: \boldx}$ has no dependence on $u(t)$ because $\boldf(\boldx, u) =

1352: \mathbf{h}(\boldx) + \mathbf{v}(u)$. Thus, if there exists a $n

1353: \times n$ matrix $\boldM(x_1, ..., x_k)$ such that $\boldM \succ

1354: 0$ and $\dot{\boldM} + \pardfdxnm^T\boldM + \boldM

1355: \pardfdxnm \prec 0$, then the system in (\ref{Thm2:eqn}) is contracting for any value of

1356: $u(t)$.

1357: \end{proof}

1358: 

1359: Through the example considered in the following section, we will

1360: illustrate how Theorem~\ref{Thm:EI} is particularly useful in proving

1361: synchronization of nonlinear oscillators, an issue explored in more

1362: detail in \cite{OPC-NO}. Theorem~\ref{thm:inputs} can be easily

1363: extended to the case where $u(t)$ is a vector (i.e.  $\boldu(t) =

1364: [u_1(t), ... , u_m(t)]^T$).

1365: 

1366: \subsection{Coupled Oscillators}

1367: Contraction Theory is a useful tool to study synchronization behaviors

1368: of various configurations of coupled oscillators.  For simplicity, we

1369: only consider here the case a pair of unidirectionally coupled

1370: oscillators; more complicated and general couplings are discussed in

1371: \cite{OPC-NO}.

1372: 

1373: %

1374: A state-space model of two unidirectionally coupled oscillators (only

1375: one oscillator influences the other) is

1376: \begin{equation}

1377: \begin{aligned}

1378:     \dot{\boldx} &= \boldf(\boldx,t) \\

1379:     \hfill \dot{\boldy} &= \boldf(\boldy,t) + \boldu(\boldx) -

1380:     \boldu(\boldy),

1381: \end{aligned}

1382: \label{OPC-thm2-eqn}

1383: \end{equation}

1384: where $\boldx, \; \boldy \in \mathbb{R}^m$, are the state vectors,

1385: $\boldf(\boldx,t)$ and $\boldf(\boldy,t)$ are the dynamics of the

1386: uncoupled oscillators, and $\boldu(\boldx) - \boldu(\boldy)$ is

1387: the coupling force\footnote{An example of coupled oscillators

1388: whose state-space representation is in this form is

1389: $\begin{cases}

1390:   \ddot{x} + \alpha(x^2 + k)\dot{x} + \omega^2x  &= 0 \\

1391:  \hfill \ddot{y} + \alpha(y^2 + k)\dot{y} + \omega^2y  &= \alpha \eta (\dot{x} - \dot{y}) \\

1392: \end{cases}$

1393: where $\alpha >0$, $\omega > 0$, $k$  are arbitrary constants.}.

1394: The following theorem is a slightly modified version of Theorem 2

1395: in \cite{OPC-NO}.

1396: 

1397: \begin{theorem}\label{thm:OPC2}

1398: If $\dot{\boldy} = \boldf(\boldy) + \boldu(\boldy) -

1399: \boldu(\boldx)$ in (\ref{OPC-thm2-eqn}) is contracting with

1400: respect to  $\boldy$ over the entire state space for arbitrary

1401: $\boldu(\boldx)$\footnote{By contracting with respect to $\boldy$

1402: for arbitrary $\boldu(\boldx)$ we mean that the system

1403: $\dot{\boldy} = \boldf(\boldy) - \boldu(\boldy) + \boldu(\boldx)$,

1404: where $\boldy$ is the state vector and $\boldu(\boldx)$ is an

1405: arbitrary driving function, is contracting for all inputs

1406: $\boldu(\boldx)$.}, the two systems will reach synchrony (i.e.

1407: $\boldy(t)$ and $\boldx(t)$ will tend toward the same trajectory)

1408: regardless of initial conditions.

1409: \end{theorem}

1410: 

1411: \begin{proof}

1412: The system $\dot{\boldy} = \boldf(\boldy) - \boldu(\boldy) +

1413: \boldu(\boldx)$ with input $\boldu(\boldx)$ is contracting with

1414: respect to $\boldy$ over the entire state space and $\boldy (t) =

1415: \boldx (t)$ is a particular solution.  Thus, by the properties of

1416: contraction, all solutions converge exponentially to $\boldy (t) =

1417: \boldx (t)$.

1418: \end{proof}

1419: 

1420: Theorem \ref{thm:inputs} becomes especially powerful when the

1421: vector field appearing in the second subsystem of

1422: (\ref{OPC-thm2-eqn}) has the structure described in

1423: equation (\ref{Thm2:eqn})\footnote{If it does not have such a structure

1424: and $\boldu(\boldx)$ drives each component of $\boldy$, we lose

1425: degrees of freedom in the possible forms of our contraction

1426: metric.  If $\boldu(\boldx)$ drives each component of $\boldy$ the

1427: only possible contraction metric is a constant.}. We illustrate

1428: this in the next example.

1429: 

1430: \subsection{Example: Coupled Van der Pol Oscillators}\label{coupled_vdpo}

1431: 

1432: Consider two identical Van der Pol oscillators coupled as

1433: \begin{equation}\label{coupled-osc-unid}

1434: \begin{cases}

1435:   \ddot{x} + \alpha(x^2 + k)\dot{x} + \omega^2x  &= 0 \\

1436:  \hfill \ddot{y} + \alpha(y^2 + k)\dot{y} + \omega^2y  &= \alpha \eta (\dot{x} - \dot{y}) \\

1437: \end{cases}\end{equation}

1438: where $\alpha >0$, $\omega > 0$, $k$  are arbitrary constants. We

1439: note that if $k<0$, trajectories of the individual oscillator

1440: dynamics converge to a limit cycle. See Figure \ref{pp_osc_kneg}.

1441: We first write these coupled systems in state-space form to get

1442: the equations in the form of (\ref{OPC-thm2-eqn}). Their

1443: state-space form is

1444: \begin{equation}\label{cou-ssform}

1445: \begin{cases}

1446: \left[%

1447: \begin{array}{c}

1448:   \dot{x}_1 \\

1449:   \dot{x}_2 \\

1450: \end{array}%

1451: \right] &= \left[

1452: \begin{array}{c}

1453:   x_2 \\

1454:   -\alpha (x_1^2 + k)x_2 - \omega^2 x_1  \\

1455: \end{array}%

1456: \right] \\

1457: \hfill \\

1458: \hfill \left[%

1459: \begin{array}{c}

1460:   \dot{y}_1 \\

1461:   \dot{y}_2 \\

1462: \end{array}%

1463: \right] &= \left[

1464: \begin{array}{c}

1465:   y_2 \\

1466:   -\alpha (y_1^2 + k + \eta)y_2 - \omega^2 y_1  + \alpha \eta x_2 \\

1467: \end{array}%

1468: \right]. \\

1469: \end{cases}

1470: \end{equation} By Theorem \ref{Thm:EI}, this pair of unidirectional oscillators

1471: will reach synchrony regardless of initial conditions if

1472: \begin{equation}\label{co:eqn4cont}

1473: \dot{\boldy} = \boldf(\boldy) - \boldu(\boldy)+ \boldu(\boldx) = \left[%

1474: \begin{array}{c}

1475:   y_2 \\

1476:   -\alpha(y_1^2 +k + \eta)y_2 - w^2y_1 + \alpha \eta x_2 \\

1477: \end{array}%

1478: \right] \end{equation} is contracting with respect to $\boldy$ for

1479: arbitrary values of $\boldu(\boldx) = x_2$. We see by Theorem

1480: \ref{thm:inputs} that for this to occur, we must find a

1481: contraction metric $\boldM(\boldy)$ that is only  a function of

1482: $y_1$ (i.e. $\boldM(\boldy) = \boldM(y_1)$).

1483: 

1484: When the search algorithm described in Section~\ref{SearchAlgorithm},

1485: was applied to find a metric that satisfied $\boldM(\boldy) =

1486: \boldM(y_1)$ as well as $\boldM(\boldy)\succ 0$ and $\boldR(\boldy)

1487: \prec 0$, none were found.  However, it is shown in the appendix,

1488: which is a modified version of the appendix of \cite{OPC-NO}, that a

1489: metric that satisfies $\boldM(\boldy)\succ 0$ and $\boldR(\boldy)

1490: \preceq 0$ implies asymptotic convergence of trajectories of system

1491: (\ref{co:eqn4cont}). A system with this metric that satisfies

1492: $\boldM(\boldy)\succ 0$ and $\boldR(\boldy) \preceq 0$ is called

1493: \emph{semi-contracting} \cite{OCA, OPC-NO}.

1494: 

1495: \noindent The metric

1496: \begin{equation}\label{MAnalyOsc}\boldM(\boldy) =

1497: \left[\begin{array}{cc}

1498:   \omega^2 + \alpha^2(y_1^2 + k + \eta)^2    & \alpha(y_1^2 +k + \eta) \\

1499:   \alpha(y_1^2 + k +\eta ) & 1 \\

1500: \end{array}\right]\end{equation} that appears in \cite{OPC-NO} is only a function of $y_1$

1501: and satisfies $\boldM(\boldy)\succ 0$ and $\boldR(\boldy) \preceq

1502: 0$ for the system dynamics (\ref{co:eqn4cont}) if $\alpha >0$ and

1503: $(k + \eta) \geq 0$.  For this $\boldM$ and the system equation

1504: (\ref{co:eqn4cont}), we have

1505: %

1506: \begin{equation}\label{Rnou} \boldR = \dot{\boldM}  + \pardfdynm^T \boldM + \boldM \pardfdynm = \left[%

1507: \begin{array}{cc}

1508:   -2\alpha \omega^2 y_1^2 - 2\alpha \omega^2(k + \eta ) & 0 \\

1509:   0 & 0 \\

1510: \end{array}%

1511: \right]. \end{equation} For $\alpha >0$, $(k + \eta)>0$,

1512: $\boldM(\boldy) \succ 0$ and $\boldR(\boldy) \preceq 0$. Since

1513: (\ref{MAnalyOsc}) and (\ref{Rnou}) show analytically that the

1514: system (\ref{co:eqn4cont}) is semi-contracting we used our search

1515: algorithm to search for a metric with $\boldM(\boldy)\succ 0$ and

1516: $\boldR(\boldy) \preceq 0$.

1517: 

1518: \subsubsection{Search for a Semidefinite $\boldR$ Matrix: Numerical

1519: Problems and Solutions}\label{NumericalProblems}

1520: 

1521: A minor problem that one may encounter when searching for contraction

1522: metrics, depending on the structure of polynomial constraints, is that

1523: the resulting optimization problem may be feasible, but not strictly

1524: feasible. This can cause numerical difficulties in the algorithms used

1525: in the solution procedure. In many cases, however, this can be

1526: remedied by a introducing a presolving stage in which redundant

1527: variables are eliminated. When we ran the search algorithm based on

1528: Theorem~\ref{thm:inputs} and only searched for $\boldM$ as a function

1529: of $y_1$, no valid solution was found even if we only constrained

1530: $\boldR$ to be negative semidefinite and not strictly negative

1531: definite. Since the analytic solution (\ref{Rnou}) was feasible but

1532: not strictly feasible, we hypothesized there was numerical error in

1533: the algorithm. Based on knowledge of the analytic solution

1534: (\ref{Rnou}), we thus constrained $R_{22} = 0$ and $R_{12} = 0$,

1535: eliminated redundant variables, and then searched for a solution in

1536: the resulting lower dimensional space\footnote{Setting $R_{22} = 0$,

1537: and $R_{12} = 0$ leads to redundant decision coefficients in the

1538: polynomial entries of $\boldM$ and $\boldR$.  If these redundant

1539: variables are eliminated through a presolving stage, the search

1540: algorithm finds $\boldM \succ 0$ and $\boldR \preceq 0$.}. With these

1541: constraints in place, a solution was found with the search algorithm.

1542: 

1543: 

1544: \section{Conclusions}\label{Conclusions}

1545: 

1546: In this paper we have described how SOS programming enables an

1547: algorithmic search for contraction metrics for the class of

1548: nonlinear systems with polynomial dynamics.  We also have

1549: illustrated the results through several examples.

1550: 

1551: These examples illustrate how contraction analysis offers several

1552: significant advantages when compared with traditional Lyapunov

1553: analysis. Contraction analysis provides relative flexibility in

1554: incorporating inputs and outputs. It is also particularly useful

1555: in the analysis of nonlinear systems with uncertain parameters

1556: where the uncertainty changes the equilibrium points of the

1557: system. It is often the case that if the nominal system is

1558: contracting with respect to a metric, the uncertain system with

1559: additive or multiplicative uncertainty will still be contracting

1560: with respect to the original metric, even if the perturbation

1561: changes the equilibrium of the system.  In addition, a slightly

1562: modified version of the standard algorithmic search allows us to

1563: optimize the search to obtain a contraction metric that provides

1564: the largest uncertainty interval for which we can prove the system

1565: is contracting.

1566: 

1567: Subjects of future research include a careful evaluation of how the

1568: computational resources needed by the algorithm scale with system

1569: size, as well as the benefits and limitations of this approach in the

1570: context of other nonlinear system analysis techniques.

1571: 

1572: \appendix

1573: 

1574: \section{Proving Asymptotic Convergence of Coupled Van der Pol Oscillators With a Negative

1575: Semidefinite $\boldR$ Matrix.}

1576: 

1577: This appendix is a modified version of the appendix in

1578: \cite{OPC-NO}. Consider the system given in (\ref{co:eqn4cont}).

1579: Consider a $2 \times 2$ matrix $\boldM(\boldy)$ that is uniformly

1580: positive definite, and a corresponding $\boldR(\boldy)$ matrix that is

1581: uniformly negative semidefinite, but not uniformly negative

1582: definite. Since (\ref{co:eqn4cont}) is a two-dimensional system, we

1583: can assume without loss of generality that $\boldR(\boldy)$ is of the

1584: form \[\boldR(\boldy) = \left(%

1585: \begin{array}{cc}

1586:   -K(\boldy) & 0 \\

1587:   0 & 0 \\

1588: \end{array}

1589: \right) \]where $K(\boldy)>0 \;\forall \; \boldy$. Let $\delta

1590: \boldy

1591: = \left(%

1592: \begin{array}{cc}

1593:   \delta y_1 & \delta y_2 \\

1594: \end{array}%

1595: \right)^T = \left ( \begin{array}{cc}

1596:   \delta y & \delta \dot{y} \\

1597: \end{array} \right)^T$.  where $y_1$ and $y_2$ are the variables in equation (\ref{co:eqn4cont}). With this $\boldR(\boldy)$ and $\boldM(\boldy)$ matrices, the

1598: general definition of differential length given in

1599: (\ref{RiemannSp}) and associated equation for rate of change of

1600: length (\ref{contractcond2}) are

1601: \[ \delta \boldz^T \delta \boldz = \delta \boldy^T \boldM (\boldy)

1602: \delta \boldy \] and

1603: \begin{eqnarray} \frac{d}{dt} \delta \boldz^T

1604: \delta \boldz &=& \frac{d}{dt}(\delta \boldy^T \boldM (\boldy)

1605: \delta \boldy) \nonumber \\

1606: & = & \delta \boldy ^T \boldR(\boldy) \delta \boldy \nonumber \\

1607: & = & -K(\boldy) \delta y_1^2. \label{negsdR}

1608: \end{eqnarray} Since $\frac{d}{dt}(\delta \boldy^T \boldM (\boldy)

1609: \delta \boldy) \leq 0$ and $\delta \boldz ^T \delta \boldz \geq 0$,

1610: $\delta \boldz ^T \delta \boldz$ has a limit at $t$ goes to

1611: infinity. We will prove through a Taylor series argument that all

1612: trajectories of this system converge asymptotically. If $\delta y =

1613: \delta y_1 \neq 0$, then

1614: 

1615: \begin{displaymath} \delta \boldz ^T \delta \boldz (t + dt) - \delta \boldz ^T

1616: \delta \boldz (t) = -K(\boldy) (\delta y_1)^2 dt + O((dt)^2)

1617: \end{displaymath} while

1618: if $\delta y_1 = 0$, \[\delta \boldz ^T \delta \boldz (t + dt) -

1619: \delta \boldz ^T \delta \boldz (t) = -2K(\boldy) (\delta

1620: y_2)^2\frac{dt^3}{3!} + O((dt)^4). \] Since $\delta \boldz ^T

1621: \delta \boldz$ converges, $\delta \boldz ^T \delta

1622: \boldz (t + dt) - \delta \boldz ^T \delta \boldz (t)$ approaches

1623: zero asymptotically and hence $\delta y_1$ and $\delta y_2$ or

1624: equivalently $\delta y$ and $\delta \dot{y}$ both tend to zero.

1625: Thus, for any input $\boldu(\boldx)$ all solutions of system

1626: (\ref{co:eqn4cont}) converge asymptotically to a single trajectory

1627: independent of initial conditions, and the unidirectional

1628: oscillators given in (\ref{cou-ssform}) will reach synchrony

1629: asymptotically regardless of initial conditions.

1630: 

1631: \pagebreak

1632: %

1633: \nocite{SOS_LMI_Pablo}

1634: \bibliographystyle{plain}

1635: \bibliography{AylwardParriloSlotineJournal}

1636: 

1637: \end{document}

1638: