Part 1 Stability Analysis of Linear Switched Systems: An Optimal Control Approach Michael Margaliot School of Elec. Eng. Tel Aviv University, Israel Joint work with: Gideon Langholz (TAU), Daniel Liberzon (UIUC), Michael S. Branicky (CWRU), Joao Hespanha (UCSB). 1 Overview Switched systems Global asymptotic stability The edge of stability Stability analysis: An optimal control approach A geometric approach An integrated approach Conclusions 2 Switched Systems Systems that can switch between several possible modes of operation. Mode 1 Mode 2 3 Example 1 x1 a1 (t ) C x2 a2 (t ) a1 (t ) a2 ( t ) x1 x2 C server x1 a1 (t ) x2 a2 (t ) C 4 Example 2 Switched power converter 100v linear filter 50v 5 Example 3 A multi-controller scheme + plant controller1 switching logic controller2 Switched controllers are stronger than “regular” controllers. 6 More Examples Air traffic control Biological switches Turbo-decoding …… 7 Synthesis of Switched Systems Driving: use mode 1 (wheels) Braking: use mode 2 (legs) The advantage: no compromise 8 Linear Systems x Ax. Solution: x(t ) exp( At ) x(0). Definition: The system is globally asymptotically x(0). stable if lim x(t ) 0, t Theorem: stability Re( λ) 0, λ eig( A). A is called a Hurwitz matrix. 9 Linear Switched Systems Two (or more) linear systems: x(t ) A1 x(t ), x(t ) A2 x(t ). A system that can switch between them: x(t ) Aσ (t ) x(t ), σ (t ) 2 1 σ : R {1,2}. ... t1 t1 t2 t x(T ) ...exp(t4 A1 )exp(t3 A2 )exp(t2 A1 )exp(t1 A2 ) x0 . 10 10 Stability Linear switched system: x(t ) Aσ (t ) x(t ), σ : R {1,2}. Definition: Globally uniformly asymptotically x(t ) 0, x(0), σ . stable (GUAS): In other words, lim x(t ) lim(...exp(t3 A2 ) exp(t2 A1 ) exp(t1 A2 ) x0 ) 0 t t for any t1 , t2 ,.... 0 AKA, “stability under arbitrary switching”. 11 11 A Necessary Condition for GUAS The switching law σ (t ) 1 yields x(t ) A1 x(t ). Thus, a necessary condition for GUAS is that both A1 , A2 are Hurwitz. Then instability can only arise due to repeated switching. 12 12 Why is the GUAS problem difficult? Answer 1: The number of possible switching laws is huge. 13 13 Why is the GUAS problem difficult? Answer 2: Even if each linear subsystem is stable, the switched system may not be GUAS. 0 1 x x 2 1 14 0 1 x x 12 1 14 Why is the GUAS problem difficult? Answer 2: Even if each linear subsystem is stable, the switched system may not be GUAS. 15 15 Stability of Each Subsystem is Not Enough A multi-controller scheme + plant controller1 switching logic controller2 Even when each closed-loop is stable, the switched system may not be GUAS. 16 Easy Case #1 A trajectory of the switched system: x(T ) ...exp(t4 A1 )exp(t3 A2 )exp(t2 A1 )exp(t1 A2 ) x0 . Suppose that the matrices commute: A1 A2 A2 A1. Then x(T ) exp((T s) A1 )exp(sA2 ) x0 , and since both matrices are Hurwitz, the switched system is GUAS. 17 17 Easy Case #2 Suppose that both matrices are upper triangular: 1 1 x x, 0 2 3 7 x x. 0 2.5 Then x 2 x , x 2.5x , so 2 Now 2 2 2 | x (t ) | exp(2t ) | x (0) |. 2 2 x x x , x 3x 7 x so x (t ) 0. 1 1 1 2 1 1 2 This proves GUAS. 18 18 Optimal Control Approach Pioneered by E. S. Pyatnitsky (1970s). Basic idea: (1) A relaxation: linear switched system → bilinear control system (2) characterize the “most destabilizing” control u * (3) the switched system is GUAS iff x * (t ) 0 19 19 Optimal Control Approach Relaxation: the switched system: x Aσ x, σ : R {1,2}, → a bilinear control system: x ( A1 ( A2 A1 )u ) x, u U , where U is the set of measurable functions taking values in [0,1]. u0 x A1 x u 1/ 2 u 1 x (1/ 2)( A1 A2 ) x x A2 x 20 Optimal Control Approach The bilinear control system (BCS) x ( A1 ( A2 A1 )u ) x, u U , is globally asymptotically stable (GAS) if: x(t ) 0, x(0), u U . Theorem The BCS is GAS if and only if the linear switched system is GUAS. 21 Optimal Control Approach The most destabilizing control: x ( A1 ( A2 A1 )u ) x, u U , x(0) x0 . Fix a final time t f 0.. Let J (u) | x(t f ; u) |. Optimal control problem: find a control that maximizes J (u). u*U Intuition: maximize the distance to the origin. 22 Optimal Control Approach and Stability Theorem The BCS is GAS iff x *(t f ) 0. 23 Edge of Stability The BCS: x ( A1 Bu) x, B A2 A1. Consider x ( A1 Bk u ) x, Bk ( A2 A1 )k . k 0 x ( A1 0u ) x GAS k 1 k ε0 x ( A1 Bεu ) x, GAS x ( A1 B1u ) x, original BCS 24 Edge of Stability The BCS: x ( A1 Bu) x, B A2 A1. Consider x ( A1 Bk u ) x, Bk ( A2 A1 )k . k 0 x ( A1 0u ) x GAS k 1 k ε0 x ( A1 Bεu ) x, GAS x ( A1 B1u ) x, original BCS Definition: k* is the minimal value of k>0 such that GAS is lost. 25 Edge of Stability The BCS: x ( A1 Bu) x, B A2 A1. Consider x ( A1 Bk u ) x, Bk ( A2 A1 )k . Definition: k* is the minimal value of k>0 such that GAS is lost. The system x ( A1 Bk*u) x is said to be on the edge of stability. 26 Edge of Stability The BCS: x ( A1 Bu) x, B A2 A1. Consider x ( A1 Bk u ) x, Bk ( A2 A1 )k . Definition: k* is the minimal value of k>0 such that GAS is lost. 0 k* 1 k 0 1 k* k Proposition: our original BCS is GAS iff k*>1. 27 Edge of Stability The BCS: x ( A1 Bu) x, B A2 A1. Consider x ( A1 Bk u ) x, Bk ( A2 A1 )k . Proposition: our original BCS is GAS iff k*>1. → we can always reduce the problem of analyzing GUAS to the problem of determining the edge of stability. 28 Edge of Stability When n=2 Consider x ( A1 Bk u ) x. The trajectory x* corresponding to u*: k k* k k* k k* x0 x0 A closed periodic trajectory 29 Solving Optimal Control Problems 2 | x(t f ; u ) | is a functional: x(t f ; u) F (u(t), t [0, t f ]) Two approaches: 1. The Hamilton-Jacobi-Bellman (HJB) equation. 2. The Maximum Principle. 30 Solving Optimal Control Problems 1. The HJB equation. Intuition: there exists a function V (, ) : R R n R V (t , x *(t )) const, and V can only decrease on any other trajectory of the system. 31 The HJB Equation Find V (, ) : R R n R such that V (t f , y) || y ||2 / 2, d MAX V (t , x(t )) 0. u [0,1] dt (HJB) Integrating: V (t f , x(t f )) V (0, x(0)) 0 2 | x(t f ) | / 2 V (0, x(0)). or 2 An upper bound for | x(t f ) | / 2, obtained for the u * maximizing (HJB). 32 The HJB for a BCS: 0 max{Vt Vx x} u max{Vt Vx (uAx (1 u ) Bx )} u max{Vt Vx Bx uVx ( A B ) x} u Hence, 1, Vx ( A B) x 0, u* 0, Vx ( A B) x 0, ?, V ( A B) x 0. x In general, finding V is difficult. Note: u* depends on Vx only. 33 The Maximum Principle 2 Let (t ) : Vx (t , x *(t )). Then, (t f ) x (t f ) / 2 x(t f ). x Differentiating 0 Vt Vx x, we get 0 Vtx Vxx x Vx (uA (1- u ) B) d dt Vx Vx (uA (1- u ) B) (uA (1- u ) B) A differential equation for (t ), with a boundary condition at t f . 34 Summarizing, T (uA (1- u) B) , (t f ) x(t f ) x (uA (1- u) B) x, x(0) x0 The WCSL is the u * maximizing Vt Vx x Vt T (uA (1 u) B) x that is, 1, T (t )( A B) x(t ) 0, u *(t ) T 0, (t )( A B) x(t ) 0. We can simulate the optimal solution backwards in time. 35 Result #1 (Margaliot & Langholz, 2003) An explicit solution for the HJB equation, when n=2, and {A,B} is on the “edge of stability”. This yields an easily verifiable necessary and sufficient condition for stability of second-order switched linear systems. 36 Basic Idea The HJB eq. is: Thus, 0= max{Vx Bx uVx ( A B ) x}. u u 0 0=Vx Bx u 1 0=Vx Ax Let H A : R2 R x(t ) Ax(t ), be a first integral of that is, d A 0 H ( x(t )) H xA Ax. dt Then V is a concatenation of two first integrals H A ( x) and H B ( x). 37 1 1 0 0 Example: A 2 1 B 2 k 1 x Ax 1 7 x1 2 A T H ( x) x P0 x exp( arctan( )) x1 2 x2 7 1 x Bx H B ( x) xT Pk x exp( 2 k 1/ 2 where Pk 1 1/ 2 7 4k x1 2 arctan( )) x1 2 x2 7 4k and k * 6.985... 38 Nonlinear Switched Systems x { f ( x), f ( x)} 1 with 2 (NLDI) x f ( x) GAS. i Problem: Find a sufficient condition guaranteeing GAS of (NLDI). 39 Lie-Algebraic Approach For the sake of simplicity, we present the approach for LDIs, that is, x { Ax , Bx} and x(t ) ...exp( Bt2 ) exp( At1 ) x(0). 40 Commutation and GAS Suppose that A and B commute, AB=BA, then x(t ) ...exp( At3 )exp( Bt2 )exp( At1 ) x(0) exp( A(... t3 t1 ))exp( B(... t4 t2 )) x(0) Definition: The Lie bracket of Ax and Bx is [Ax,Bx]:=ABx-BAx. Hence, [Ax,Bx]=0 implies GAS. 41 Lie Brackets and Geometry x { Ax, Ax, Bx, Bx}. Consider x Ax x ( 0) x Bx x ( 4 ) x Bx x Ax A calculation yields: x(4ε ) x(0) ε 2[ A, B]( x(0)). 42 Geometry of Car Parking This is why we can park our car. The 2 term is the reason it takes so long. Bx Ax Ax Bx [ A, B]( x) 43 Nilpotency We saw that [A,B]=0 implies GAS. What if [A,[A,B]]=[B,[A,B]]=0? Definition: k’th order nilpotency all Lie brackets involving k terms vanish. [A,B]=0 → 1st order nil. [A,[A,B]]=[B,[A,B]]=0 → 2nd order nil. 44 Nilpotency and Stability We saw that 1st order nilpotency Implies GAS. A natural question: Does k’th order nilpotency imply GAS? 45 Some Known Results Switched linear systems: k=2 implies GAS (Gurvits,1995). k order nilpotency implies GAS (Liberzon, Hespanha, and Morse, 1999). (The proof is based on Lie’s Theorem) Switched nonlinear systems: k=1 implies GAS. An open problem: higher orders of k? (Liberzon, 2003) 46 A Partial Answer Result #2 (Margaliot & Liberzon, 2004) 3rd order nilpotency implies GAS. Proof: Consider the WCSL 1, T (t )( A B) x(t ) 0 u *(t ) T 0, (t )( A B) x(t ) 0 Define the switching function m(t ) : (t )Cx(t ), C A B T 47 Differentiating m(t) yields m(t ) (t )Cx(t ) (t )Cx(t ) T T (t )[C, A]x(t ). T 2nd order nilpotency m 0 m(t ) const no switching in the WCSL! Differentiating again, we get T T m [ C , A ] x [C , A] x T [[C , A], A] x uT [[C , A], B ] x 3rd order nilpotency m 0 m(t ) at b up to a single switching in the WCSL. 48 Singular Arcs If m(t)0, then the Maximum Principle provides no direct information. Singularity can be ruled out using the auxiliary system. 49 Summary Parking cars is an underpaid job. Switched systems and differential inclusions are important in various scientific fields, and pose interesting theoretical questions. Stability analysis is difficult. A natural and useful idea is to consider the worst-case trajectory. 50 Summary: Optimal Control Approach Advantages: reduction to a single control u * leads to necessary and sufficient conditions for GUAS allows the application of powerful tools (high-order MPs, HJB equation, Liealgebraic ideas,….) applicable to nonlinear switched systems Disadvantages: requires characterizing u * explicit results for particular cases only 51 More Information 1. Margaliot. “Stability analysis of switched systems using variational principles: an introduction”, Automatica, 42: 2059-2077, 2006. 2. Sharon & Margaliot. “Third-order nilpotency, nice reachability and asymptotic stability”, J. Diff. Eqns., 233: 136-150, 2007. 3. Margaliot & Branicky. “Nice reachability for planar bilinear control systems with applications to planar linear switched systems”, IEEE Trans. Automatic Control, 54: 1430-1435, 2009. Available online: 52 www.eng.tau.ac.il/~michaelm 52