Least Squares Based Self-Tuning Control Systems: Supplementary Notes

Similar documents
Lyapunov Stability Theory

A sub-optimal second order sliding mode controller for systems with saturating actuators

Control, Stabilization and Numerics for Partial Differential Equations

Putzer s Algorithm. Norman Lebovitz. September 8, 2016

1. Find the solution of the following uncontrolled linear system. 2 α 1 1

Introduction to Nonlinear Control Lecture # 3 Time-Varying and Perturbed Systems

Iterative Methods for Solving A x = b

Nonlinear Control Lecture 5: Stability Analysis II

DS-GA 1002 Lecture notes 0 Fall Linear Algebra. These notes provide a review of basic concepts in linear algebra.

ECEEN 5448 Fall 2011 Homework #5 Solutions

Zeros and zero dynamics

Observer design for a general class of triangular systems

ALMOST SURE STABILITY OF CONTINUOUS-TIME MARKOV JUMP LINEAR SYSTEMS: A RANDOMIZED APPROACH. Paolo Bolzern Patrizio Colaneri Giuseppe De Nicolao

1 The Observability Canonical Form

Mildly degenerate Kirchhoff equations with weak dissipation: global existence and time decay

L2 gains and system approximation quality 1

Global Maxwellians over All Space and Their Relation to Conserved Quantites of Classical Kinetic Equations

Modeling and Analysis of Dynamic Systems

Positive Markov Jump Linear Systems (PMJLS) with applications

1 Lyapunov theory of stability

Average-Consensus of Multi-Agent Systems with Direct Topology Based on Event-Triggered Control

ẋ = f(x, y), ẏ = g(x, y), (x, y) D, can only have periodic solutions if (f,g) changes sign in D or if (f,g)=0in D.

Virtual Reference Feedback Tuning for non-linear systems

Analysis in weighted spaces : preliminary version

October 25, 2013 INNER PRODUCT SPACES

STABILITY ANALYSIS OF DYNAMIC SYSTEMS

Numerical Sequences and Series

Full-State Feedback Design for a Multi-Input System

Weighted Sums of Orthogonal Polynomials Related to Birth-Death Processes with Killing

Convergence Rate of Nonlinear Switched Systems

arxiv: v3 [math.ds] 22 Feb 2012

Qualitative behavior of global solutions to some nonlinear fourth order differential equations

Automatic Control Systems theory overview (discrete time systems)

RECURSIVE ESTIMATION AND KALMAN FILTERING

An Input-Output Approach to Structured Stochastic Uncertainty

Introduction to Nonlinear Control Lecture # 4 Passivity

Stability of an abstract wave equation with delay and a Kelvin Voigt damping

Math 331 Homework Assignment Chapter 7 Page 1 of 9

w T 1 w T 2. w T n 0 if i j 1 if i = j

Ahlswede Khachatrian Theorems: Weighted, Infinite, and Hamming

56 4 Integration against rough paths

Relative Controllability of Fractional Dynamical Systems with Multiple Delays in Control

Uniformly Uniformly-ergodic Markov chains and BSDEs

Calculus (Real Analysis I)

Lecture 2: Linear Algebra Review

ME 234, Lyapunov and Riccati Problems. 1. This problem is to recall some facts and formulae you already know. e Aτ BB e A τ dτ

Raktim Bhattacharya. . AERO 632: Design of Advance Flight Control System. Norms for Signals and Systems

(Refer Slide Time: 00:32)

Robust Stability. Robust stability against time-invariant and time-varying uncertainties. Parameter dependent Lyapunov functions

Nonlinear Control. Nonlinear Control Lecture # 6 Passivity and Input-Output Stability

A LaSalle version of Matrosov theorem

Math 350 Fall 2011 Notes about inner product spaces. In this notes we state and prove some important properties of inner product spaces.

6.241 Dynamic Systems and Control

Solution. 1 Solution of Homework 7. Sangchul Lee. March 22, Problem 1.1

Statistics 612: L p spaces, metrics on spaces of probabilites, and connections to estimation

Nonlinear Control Systems

TOPOLOGICAL EQUIVALENCE OF LINEAR ORDINARY DIFFERENTIAL EQUATIONS

The Cayley-Hamilton Theorem and the Jordan Decomposition

ME Fall 2001, Fall 2002, Spring I/O Stability. Preliminaries: Vector and function norms

A linear algebra proof of the fundamental theorem of algebra

Spectral theory for compact operators on Banach spaces

Andrea Zanchettin Automatic Control AUTOMATIC CONTROL. Andrea M. Zanchettin, PhD Spring Semester, Linear systems (frequency domain)

Iowa State University. Instructor: Alex Roitershtein Summer Homework #5. Solutions

AUTOMATIC CONTROL. Andrea M. Zanchettin, PhD Spring Semester, Introduction to Automatic Control & Linear systems (time domain)

The Liapunov Method for Determining Stability (DRAFT)

Lecture 5: Lyapunov Functions and Storage Functions 1

Impulsive Stabilization and Application to a Population Growth Model*

Newtonian Mechanics. Chapter Classical space-time

1 Relative degree and local normal forms

Control of Mobile Robots

EE 380. Linear Control Systems. Lecture 10

GROUP THEORY PRIMER. New terms: so(2n), so(2n+1), symplectic algebra sp(2n)

Homework 4, 5, 6 Solutions. > 0, and so a n 0 = n + 1 n = ( n+1 n)( n+1+ n) 1 if n is odd 1/n if n is even diverges.

ON CALCULATING THE VALUE OF A DIFFERENTIAL GAME IN THE CLASS OF COUNTER STRATEGIES 1,2

Discrete and continuous dynamic systems

Bessel Functions Michael Taylor. Lecture Notes for Math 524

We denote the derivative at x by DF (x) = L. With respect to the standard bases of R n and R m, DF (x) is simply the matrix of partial derivatives,

BIBO STABILITY AND ASYMPTOTIC STABILITY

Linear ODE s with periodic coefficients

Dynamical Systems. August 13, 2013

arxiv: v1 [math.pr] 1 Jan 2013

Hybrid Control and Switched Systems. Lecture #7 Stability and convergence of ODEs

Exercises: Brunn, Minkowski and convex pie

Non-linear wave equations. Hans Ringström. Department of Mathematics, KTH, Stockholm, Sweden

Prove that this gives a bounded linear operator T : X l 1. (6p) Prove that T is a bounded linear operator T : l l and compute (5p)

A Systematic Approach to Extremum Seeking Based on Parameter Estimation

NOTES ON LINEAR ODES

Linear ODEs. Existence of solutions to linear IVPs. Resolvent matrix. Autonomous linear systems

Communication constraints and latency in Networked Control Systems

On the asymptotic dynamics of 2D positive systems

ECE504: Lecture 8. D. Richard Brown III. Worcester Polytechnic Institute. 28-Oct-2008

L 1 Adaptive Output Feedback Controller to Systems of Unknown

Autonomous navigation of unicycle robots using MPC

21 Linear State-Space Representations

u xx + u yy = 0. (5.1)

Topic # /31 Feedback Control Systems. Analysis of Nonlinear Systems Lyapunov Stability Analysis

Convex Optimization M2

Kernel Method: Data Analysis with Positive Definite Kernels

On the Existence of Bounded Solutions to a Class of Nonlinear Initial Value Problems with Delay

Lecture 4. Chapter 4: Lyapunov Stability. Eugenio Schuster. Mechanical Engineering and Mechanics Lehigh University.

Transcription:

Least Squares Based Self-Tuning Control Systems: Supplementary Notes S. Garatti Dip. di Elettronica ed Informazione Politecnico di Milano, piazza L. da Vinci 32, 2133, Milan, Italy. Email: simone.garatti@polimi.it July 17, 212 These notes provides some complimentary material to the paper Least Squares Based Self-Tuning Control Systems by S. Bittanti and M. Campi. In particular, they provide smoothed versions of the paper proofs by providing some reasonings left to the reader and simplifying the arguments when possible. The notes presume the reader got through the paper at least up to page 351 or that the reader attended the classes given by myself. I used here the same notation adopted in class, which is slightly different from that of the paper. The main differences with the paper are now briefly explained to ease the reader. 1

- The delay operator z 1 is used in place of q 1 - The true system is indicated with S - The RLS estimate is indicated by ϑ t and the RLS algorithm is given by the following equations: ϑ t+1 = ϑ ] t + S(t + 1) 1 ϕ(t) [y(t + 1) ϕ(t) T ϑt S(t + 1) = S(t) + ϕ(t)ϕ(t) T ϑ = ϑ S() = S (here S(t) is equal to P (t) 1 in the paper). - The control law is as follows: C(ϑ) : A c (ϑ, z 1 )u(t) = B c (ϑ, z 1 )y(t) + C c (ϑ, z 1 )r(t), where r(t) is the reference signal and A c (ϑ, z 1 ) = 1 α 1 (ϑ)z 1 α nα (ϑ)z nα B c (ϑ, z 1 ) = β (ϑ) + β 1 (ϑ)z 1 + + β nβ (ϑ)z n β C c (ϑ, z 1 ) = γ (ϑ) + γ 1 (ϑ)z 1 + + γ nγ (ϑ)z nγ 1 RLS estimate convergence In this section we want to prove that the RLS estimate ϑ t converges for t to a finite value for any possible data record {..., y(1), u(1),..., y(t), u(t),...} 2

generated according to the true system, i.e. such that y(t + 1) = ϕ(t) T ϑ o. From the RLS equations: ϑ t+1 = ϑ ] t + S(t + 1) 1 ϕ(t) [y(t + 1) ϕ(t) T ϑt [recall that y(t + 1) = ϕ(t) T ϑ o ] ϑ t+1 ϑ o = ϑ [ t ϑ o + S(t + 1) 1 ϕ(t)ϕ(t) T ϑ o ϑ ] t. Calling ϑ t = ϑ t ϑ o we have that ϑ t+1 = ϑ t S(t + 1) 1 ϕ(t)ϕ(t) T ϑt S(t + 1) ϑ t+1 = S(t + 1) ϑ t ϕ(t)ϕ(t) T ϑt [note that S(t + 1) ϕ(t)ϕ(t) T = S(t)] S(t + 1) ϑ t+1 = S(t) ϑ t, (1) i.e. the quantity S(t) ϑ t is an invariant, and, recursively applying (1), S(t) ϑ t = S() ϑ = S (ϑ ϑ o ), t. We need to recall here the definition of positive (semi-)definite matrix. Definition 1 M, N are symmetric quadratic matrices. M means that M is positive definite, i.e. x T Mx >, x. M means that M is positive semi-definite, i.e. x T Mx, x. We will write M N and M N when M N and M N, respectively. Since S(t + 1) = S(t) + ϕ(t)ϕ(t) T we have that x T S(t + 1)x = x T S(t)x + x T ϕ(t)ϕ(t) T x = [note that x T ϕ(t) = ϕ(t) T x is a scalar] = x T S(t)x + ( x T ϕ(t) ) 2. 3

Hence, x T (S(t + 1) S(t))x = ( x T ϕ(t) ) 2, for all x, i.e. S(t + 1) S(t). Recursively reasoning, the following relationship is obtained: S(t + 1) S(t) S(1) S, where the last relation is because the user-chosen initialization of the RLS algorithm is such that S is positive definite. Since S(t), t, the inverse of S(t) exists for all t and moreover S(t) 1, t (this is a basic property of positive definite matrices). Hence, we can write ϑ t = S(t) 1 S ϑ and it is clear that if S(t) 1 converges then ϑ t does as well. Another well known property of positive definite matrices is that M N N 1 M 1, leading to the following relationship which is equivalent to S 1 S(t) 1 S(t + 1) 1, x T S 1 x x T S(t) 1 x x T S(t + 1) 1 x >, x. (2) Indicate with a i,j (t) the generic element of S(t) 1, that is a 1,1 (t) a 1,n (t) S(t) 1 =...... a n,1 (t) a n,n (t) 4

By taking x = e 1 = 1. in (2) we have (x T S(t) 1 x indeed selects the element a 1,1 (t) of S(t) 1 ): a 1,1 () a 1,1 (t) a 1,1 (t + 1) >, i.e., as t increases, the a 1,1 (t) s form a decreasing sequence which is bounded from below by. Therefore, a 1,1 (t) must converge i.e. a 1,1 (t) t ā 1,1. Similarly, letting x equal to e 2 = 1.,, e n 1 =. 1, e n =. 1, it can be shown that all the elements on the diagonal a i,i (t) must converge. Finally, letting x = e i + e j, i j we have a i,i (t) + a j,j (t) + 2a i,j (t) a i,i (t + 1) + a j,j (t + 1) + 2a i,j (t + 1) >, that is, a i,i (t)+a j,j (t)+2a i,j (t) must converge. Yet, because a i,i (t) and a j,j (t) converge, we deduce that a i,j (t) converges too. 5

Thus, altogether, we have proved that S(t) 1 S 1, t so that ϑ t = S(t) 1 S ϑ S 1 S ϑ, t, i.e. ϑ t converges to the finite value S 1 S ϑ which will be indicated by ϑ. Since ϑ t = ϑ t ϑ o, the convergence of ϑ t to a finite value is easily obtained from the convergence of ϑ t : ϑ t ϑ := ϑ o + ϑ, t. Clearly, if S(t) 1, then ϑ t and ϑ t ϑ o. Yet, we cannot rely on S(t) 1 in the context of adaptive control. 2 Characterization of ϑ For a given realization of {..., y(1), u(1),..., y(t), u(t),...} we can define the unexcitation subspace as E = { } ϑ R n : lim ϑ T S(t)ϑ < +, t i.e. E is the space of directions where the information does not diverge (ϑ T S(t)ϑ can be interpreted as the projection of S(t) on the direction of ϑ). Note that since S(t + 1) S(t) for all t, we have that ϑ T S(t + 1)ϑ ϑ T S(t)ϑ, i.e. ϑ T S(t)ϑ is an increasing sequence of scalars. Hence, lim t ϑ T S(t)ϑ exists and it can be either a finite value or +. 6

The excitation subspace E is the orthogonal complement of E, i.e. E = E = { λ R n : λ T ϑ =, ϑ E }. Note that every ϑ R n can be decomposed as ϑ = ϑ u + ϑ e, where ϑ u is the orthogonal projection of ϑ on the unexcitation subspace E, while ϑ e is the projection on E. Clearly, ϑ u ϑ e, i.e. (ϑ u ) T ϑ e =. First, we want to prove that ϑ always belongs to the unexcitation subspace. Let v(t) = ϑ T t S(t) ϑ t. v(t) is of course a scalar, and, since S(t), v(t) >, for all t. We want to understand what kind of sequence is formed by v(t) as t is let increase. We have that v(t + 1) = ϑ T t+1s(t + 1) ϑ t+1 = ϑ T t+1s(t + 1)S(t + 1) 1 S(t + 1) ϑ t+1 = [recall that S(t + 1) ϑ t+1 = S(t) ϑ t ] = ϑ T t S(t)S(t + 1) 1 S(t) ϑ t. From the matrix inversion lemma we have that S(t + 1) 1 = S(t) 1 S(t) 1 ϕ(t)ϕ(t) T S(t) 1 1 + ϕ(t) T S(t) 1 ϕ(t), 7

and substituting we obtain that v(t + 1) = ϑ T t S(t) ϑ t ϑ T t ϕ(t)ϕ(t) T ϑt 1 + ϕ(t) T S(t) 1 ϕ(t) = [note that ϑ T t ϕ(t) = ϕ(t) T ϑt is a scalar] = v(t) ( ϑ T t ϕ(t)) 2 1 + ϕ(t) T S(t) 1 ϕ(t). Hence, v(t + 1) is equal to v(t) minus a term which is positive (the numerator is a square, while the denominator is positive in view of the positive definiteness of S(t) 1 ). This means that v(t) is decreasing with t: This proves that v() v(t) v(t + 1) >. lim v(t) = lim ϑ T t S(t) ϑ t v() < +. t t Unfortunately, this latter statement does not permit us to draw the final conclusion from the convergence of ϑ t to ϑ only, see Appendix A. Yet, it is not required to exploit the particular structure of ϑ t, since it suffices to consider the monotonicity of S(t) to finalize the proof. Since S(τ) S(t) for every t and τ t, we have that ϑ τ S(t) ϑ τ ϑ τ S(τ) ϑ τ, t, τ t. The right-hand-side of this inequality is equal to v(τ) which is smaller than or equal to v(), so that ϑ τ S(t) ϑ τ v(), t, τ t. Since this inequality holds for all τ t, it applies also to the limit for τ, that is (remember that ϑ τ converges to ϑ ): ϑ S(t) ϑ = lim τ ϑτ S(t) ϑ τ v(), t. 8

Since this other inequality holds for all t we eventually have lim ϑ S(t) ϑ v() < +, t showing that ϑ belongs to unexcitation subspace E according to its very definition. In virtue of this result, it holds that ϑ = ϑ u since, belonging ϑ to E, ϑ e =. Thanks to the convergence theorem, we then have that ϑ e t. Recalling that ϑ = ϑ o + ϑ, we also have ϑ e = (ϑ o ) e + ϑ e = (ϑ o ) e and ϑ e t (ϑ o ) e, i.e. the RLS estimate converges to the true parameter value along the direction of the excitation subspace only. Clearly if the excitation subspace were the whole R n, then we would have ϑ t ϑ o. Yet, since the input sequence is determined by the adaptive control scheme it is not possible to a-priori know what is the excitation subspace in the current system operation condition. As for ϑ u and ϑ u nothing can be said, they depend on the u, y signals realization. Yet, if we consider the regression vector ϕ(t) and its projections on E and E, i.e. ϕ(t) = ϕ u (t) + ϕ e (t), we can prove that ϕ u (t) gets smaller and smaller till to as t increases. In other words, the information about the real system carried by the regression vector ϕ(t) vanishes along the directions of the unexcitation subspace. We now prove that ϕ u (t) as t. Consider lim t ϑ T [ t i=1 ϕu (i 1)ϕ u (i 1) T ] ϑ. The following chain of 9

equalities/inequalities holds true. [ ] ϕ u (i 1)ϕ u (i 1) T ϑ lim t ϑt i=1 = lim t (ϑ e + ϑ u ) T [ ] ϕ u (i 1)ϕ u (i 1) T (ϑ e + ϑ u ) i=1 = [ϕ u (t) ϑ e since they belong to E and E, respectively] [ = lim t (ϑ u ) T i=1 ϕ u (i 1)ϕ u (i 1) T ] = [ϕ(t) = ϕ u (t) + ϕ e (t) and ϑ u ϕ e (t)] [ = lim t (ϑ u ) T < lim t (ϑ u ) T [ i=1 S + = lim t (ϑ u ) T S(t)ϑ u < +. ϕ(i 1)ϕ(i 1) T ] ϑ u ϑ u ] ϕ(i 1)ϕ(i 1) T i=1 ϑ u Let ϕ u 1(t) ϕ ϕ u (t) = u 2(t). ϕ u n(t) 1

Taking ϑ = e 1 we have [ lim t et 1 i=1 ϕ u (i 1)ϕ u (i 1) T ] e 1 = lim t This implies that ϕ u 1(t). = lim t = lim t < +. e T 1 ϕ u (i 1)ϕ u (i 1) T e 1 i=1 (e 1 ϕ u (i 1)) 2 i=1 (ϕ u 1(i 1)) 2 Letting ϑ = e 2, e 3,..., e n we also obtain ϕ u i (t), i = 2, 3,..., n, showing that ϕ u (t). i=1 3 BIBO stability and self-optimality of the self-tuning adaptive scheme Throughout, ϑ t has to be intended as the RLS estimate obtained based on the measurements of u and y (input and output of the true system) up to time t. When not explicitly required, the equations for ϑ t will be omitted. Let s start from the equations of the self-tuning control scheme: Σ(ϑ o, ϑ y(t) = [1 A(ϑ o, z 1 )]y(t) + B(ϑ o, z 1 )u(t d) t ) : u(t) = [1 A c ( ϑ t, z 1 )]u(t) + B c ( ϑ t, z 1 )y(t) + r (t) where to ease the notation we have put r (t) = C c ( ϑ t, z 1 )r(t) = γ ( ϑ t )r(t) + γ 1 ( ϑ t )r(t 1) + + γ nγ ( ϑ t )r(t n γ ). 11

Since ϑ t ϑ and by continuity, it holds that γ i ( ϑ t ) γ i ( ϑ ), so that if r(t) is bounded then r (t) is bounded too 1. Hence, BIBO stability and self-optimality can be proved with reference to r (t) instead of r(t). The first equation of Σ(ϑ o, ϑ t ) can be also written as y(t) = ϕ(t 1) T ϑ o = [ ϕ(t 1) T ϑ o ϑ t + ϑ ] t = ϕ(t 1) T ϑt ϕ(t 1) T ϑt = [1 A( ϑ t, z 1 )]y(t) + B( ϑ t, z 1 )u(t d) + e(t), where we have defined e(t) = ϕ(t 1) T ϑt. Hence, the self-tuning scheme Σ(ϑ o, ϑ t )can be re-written as y(t) = 1 A(ϑo, z 1 ) B(ϑ o, z 1 )z d y(t) + e(t) (3a) u(t) 1 A c ( ϑ t, z 1 ) B c ( ϑ t, z 1 ) u(t) r (t) e(t) = ϕ(t 1) T ϑt (3b) In (3), the first (2 inputs/2 outputs) equation is exactly the same equation of the so-called imaginary system Σ( ϑ t, ϑ t ), with an additional input e(t). The second equation then reveals that e(t), however, is not an exogenous input as it is calculated based on the values taken by the signals u and y. Hence, 1 Indeed, let g i = sup t γ i ( ϑ t ). Since ϑ t converges, g i exists and is finite. Let also R such that r(t) R t. Then r (t) (g + g 1 + + g nγ )R < +. 12

altogether these equations reveal the fundamental fact that the self-tuning scheme is nothing but the imaginary system in a feedback configuration as pictorially represented in Figure 1. This suggest the following steps to com- Figure 1: The self-tuning controller scheme is the imaginary system feedback connected with the perturbation system. plete the proof about the BIBO-stability and self-optimality of Σ(ϑ o, ϑ t ). 1. We first consider e(t) = and show that the imaginary system Σ( ϑ t, ϑ t ) (which is a linear time-variant system) is asymptotically stable for ϑ t generated by the RLS algorithm. 2. Thanks to the convergence property of the RLS estimate and the asymptotic stability of Σ( ϑ t, ϑ t ) we will show that u and y in (3) must be bounded whenever r is bounded (BIBO stability of Σ(ϑ o, ϑ t )). 3. Thanks to the boundedness of u and y and the convergence property of 13

the RLS estimate we can easily show that e(t) always. But then, by (3), we have that the output of Σ(ϑ o, ϑ t ) behaves as the output of Σ( ϑ t, ϑ t ) additionally fed by a signal which goes to zero. Since Σ( ϑ t, ϑ t ) is linear and asymptotically stable, this output will tend to the output of Σ( ϑ t, ϑ t ) when e(t) =, i.e. Σ(ϑ o, ϑ t ) asymptotically behaves like Σ( ϑ t, ϑ t ). Since ϑ t ϑ and by a continuity property, we have that the output of Σ( ϑ t, ϑ t ) will tend to the output of Σ( ϑ, ϑ ), so that we eventually have that Σ(ϑ o, ϑ t ) asymptotically behaves like Σ( ϑ, ϑ ). To ease the notation is better to work with state-space representations rather than I/O representation. Let then ξ(t + 1) = F ( ϑ t )ξ(t) + G( ϑ t )w(t) z(t) = H( ϑ t )ξ(t) + M( ϑ t )w(t) be a state space representation of (3a) where z(t) denotes the 2-dimensional signal y(t) and w(t) the signal e(t). Moreover, for simplicity we will u(t) r (t) write F (t), G(t), H(t), and M(t) in place of F ( ϑ t ), G( ϑ t ), H( ϑ t ), and M( ϑ t ). As is clear, ξ(t + 1) = F (t)ξ(t) is the autonomous part of the imaginary system Σ( ϑ t, ϑ t ), while (4) ξ(t + 1) = F ( )ξ(t) is the autonomous part of the asymptotic imaginary system. 14

3.1 Stability issues Being Σ( ϑ t, ϑ t ) and Σ( ϑ, ϑ ) linear, the stability depends on their autonomous parts only. Let consider ξ(t + 1) = F ( )ξ(t) first. As is clear since ϑ Ξ and, by definition of Ξ, we have that the matrix F ( ) is Hurwitz 2 so that ξ(t + 1) = F ( )ξ(t) (and hence Σ( ϑ, ϑ )) is asymptotically stable. This in particular means that c 1 and ρ (, 1) such that F ( ) t c 1 ρ t and ξ(t) c 1 ρ t ξ(). Turn now to the imaginary system ξ(t + 1) = F (t)ξ(t). In this case, since ϑ t Ξ, t, we still have that each matrix F (t) at various time instants is Hurwitz, but unfortunately in this time-varying case, this not not enough to guarantee that the system ξ(t + 1) = F (t)ξ(t) is asymptotically stable as shown by the counterexample given in Appendix B. On the other hand, we also have that ϑ t ϑ so that, thanks to continuity, we have that F (t) F ( ). This property, together with the fact that F (t) is Hurwitz t and that F ( ) is Hurwitz, guarantee indeed the sought result, i.e. that ξ(t + 1) = F (t)ξ(t) is asymptotically stable. This result is taken here for granted, while the interested reader may find a proof in the books on time-varying linear systems. Before moving to the next step, we need to remark that in view of the asymptotic stability c 2 and ν (, 1) such that t i=1 F (i 1) c 2ν t so that ξ(t) = t i=1 F (i 1)ξ() t i=1 F (i 1) ξ() c 2ν t ξ(). 2 That is, all the eigenvalues are strictly inside the unit circle in the complex domain 15

3.2 BIBO stability of Σ(ϑ o, ϑ t ) In this section, we want to prove that if r (t) M, t (i.e. r (t) is a bounded input), then the ouput of the self-tuning scheme z(t) keeps bounded too. Based on (4), we have that (Lagrange formula): [ ] t 1 t 1 t 1 z(t) = H(t) F (i)ξ() + H(t) F (i)g(τ) w(τ) + M(t)w(t). i= τ= i=τ+1 We have then t 1 t 1 z(t) = H(t) F (i)ξ() + H(t) i= + M(t) w(t) [ bounding k 1 = max t k 2 = max t 1 i=τ+1 τ= t 1 i=τ+1 F (i) with νt τ 1 and letting t 1 H(t) F (i)ξ(), { max t,τ i= k 1 + k 2 ν t τ w(τ). τ= H(t) G(τ) ν } ], max M(t) t F (i) G(τ) w(τ) + Moreover, recalling that w(τ) = e(τ) e(τ) and since r (τ) r (τ) = e(τ) 2 + r (τ) 2 e(τ) + r (τ), we have that: z(t) k 1 + k 2 ν t τ e(τ) + k 2 τ= ν t τ r (τ), τ= 16

which in turn, by noting that e(τ) = ϕ(τ 1) T ϑτ = ϕ u (τ 1) T ϑu τ + ϕ e (τ 1) T ϑe τ ϕ u (τ 1) T ϑu τ + ϕ e (τ 1) T ϑe τ, can be split as follows z(t) k 1 + k 2 ν t τ r (τ) + k 2 ν t τ ϕ u (τ 1) T ϑu τ + τ= τ= +k 2 ν t τ ϕ e (τ 1) T ϑe τ. (5) τ= Since r (t) M, t, we have that ν t τ r (τ) M τ= ν t τ M τ= τ= ν t τ = M 1 ν, and hence the second term in the sum keeps bounded. Similarly, since, as we proved previously, ϕ u (τ 1) T ϑu τ so that ϕ u (τ 1) T ϑu τ M, t, the third term keeps bounded too. Bounding the first three terms in (5) with a unique constant, we have that z(t) k 3 + k 2 ν t τ ϕ e (τ 1) T ϑe τ τ= [by means of the Cauchy-Schwarz inequality] k 3 + k 2 ν t τ ϕ e (τ 1) ϑ e τ. (6) τ= 17

Now, let consider the regression vector ϕ(t). We have that y(t) y(t 1) z(t). z(t 1) ϕ(t) = y(t n) z(t) + z(t 1) + + z(t q),. u(t + 1 d) z(t q). u(t + 1 d m) where q = max{n, d+m 1}. Using repeatedly, for each term in the previous sum, the bound in (6), we obtain ϕ(t) k 4 + k 5 ν t τ ϕ e (τ 1) ϑ e τ τ= k 4 + max i t 1 ϕe (i) k 5 ν t τ ϑ e τ τ= [ ϕ e (i) ϕ e (i) + ϕ u (i) = ϕ(i) ] k 4 + max 5 ν t τ ϑ e τ. i t 1 (7) τ= Suppose now that ϕ(t) does not keep bounded. Since, as we proved in a previous section, ϑ e τ, it turns out ν t τ ϑ e τ, τ= when t. Indeed, t τ= νt τ ϑ e τ is nothing but the motion of an asymptotically stable system (x(t + 1) = νx(t) + ϑ e τ ) with vanishing input. 18

Hence, a t can be founded such that, simultaneously: k 5 max ϕ(i) 2k 4 i t ν t τ ϑ e τ 1 2, t t τ= (since ϕ(t) is not bounded) (since the convergence to zero of this term). Plugging these two properties in (7), we have that (note that max i t ϕ(i) 2k 4 k 4 1 2 max i t): Moreover, ϕ( t + 1) k 4 + max ϕ(i) k 5 i t 1 2 max i t τ= ν t τ ϑ e τ ϕ(i) + 1 2 max ϕ(i) i t max ϕ(i). (8) i t ϕ( t + 2) k 4 + max ϕ(i) k 5 i t+1 t+1 ν t+1 τ ϑ e τ τ= 1 2 max ϕ(i) + 1 i t 2 max ϕ(i) i t+1 [ thanks to (8) ] 1 2 max i t max ϕ(i). i t ϕ(i) + 1 2 max ϕ(i) i t Proceeding recursively for t + 3, t + 4,... it can be shown that ϕ(t) max ϕ(i), t > t, i t contradicting the assumption that ϕ(t) was unbounded. Hence, eventually, we proved that ϕ(t) must keep bounded, and this implies that z(t) is bounded, i.e. Σ(ϑ o, ϑ t ) is BIBO stable. 19

3.3 Final derivations We have that e(t) = ϕ(t 1) T ϑt ϕ e (t 1) T ϑe t + ϕ u (t 1) T ϑu t. Since and ϕ u (t 1) T ϑu t ϕ u (t 1) T ϑu t leq ϕ e (t 1) ϑ e t (because ϕ e (t 1) keeps bounded and ϑ e t ), we obtain that e(t) when t. In conclusion, getting back to Figure 1, since Σ(ϑ o, ϑ t ) is equal to Σ( ϑ t, ϑ t ) additionally fed by e(t) which goes to zero and since Σ( ϑ t, ϑ t ) is linear and asymptotically stable, the output of Σ(ϑ o, ϑ t ) will tend to the output of Σ( ϑ t, ϑ t ), and by a continuity property, this latter will tend to the output of Σ( ϑ, ϑ ). This means that i.e. Σ(ϑ o, ϑ t ) is self-optimal. lim t y(t) y i (t) =, A Appendix: a counterexample Consider S(t) = 1 + t2 t 3, ϑ = 1, and ϑ t = S(t) 1 S() ϑ. t 3 1 + t 4 2

We have that ϑ t = 1+t 4 t 3 1+t 2 +t 4 1+t 2 +t 4 t 3 1+t 2 1+t 2 +t 4 1+t 2 +t 4 Simple calculations show that ϑ T t S(t) ϑ t = 1 = 1+t 4 1+t 2 +t 4 t 3 1+t 2 +t 4 t 1 + t4 1 + t 2 + t 4 t 1, 1 = ϑ. while ϑ T S(t) ϑ = 1 + t 2 which clearly tends to + as t increases. Note however that S(t) is not monotonically increasing, i.e. it is not true that S(t + 1) S(t) (hence this S(t) cannot be generated according to the rule S(t + 1) = S(t) + ϕ(t)ϕ(t) T ). This is clear even considering S(2) S(1) = 5 8 2 1 = 3 7 8 17 1 2 7 15 whose determinant is equal to 3 15 7 7 = 4 (a positive semi-definite matrix has determinant always greater than or equal to ). B Appendix: another counterexample Consider the 2-dimensional system ξ(t + 1) = F (t)ξ(t), where the matrix F (t) obeys the following law:.9989.1998 F 1 = if ξ 1 (t)ξ 2 (t).999.9989 F (t) =.9989.999 F 2 = else.1998.9989 21

The systems ξ(t + 1) = F 1 ξ(t) and ξ(t + 1) = F 2 ξ(t) are both asymptotically stable and their free motions is depicted in Figure 2 (first two sub-figures). As it appears they correspond to convergent spirals elongated on the x-axis Figure 2: Motion associated to F 1 (first sub-figure), to F 2 (second sub-figure), and to F (t) (third sub-figure). and on the y-axis respectively. When F 1 and F 2 are combined together in F (t), however, it happens that we switch from one dynamics to the other exactly when we reach the point of maximum elongation according to the current dynamics. In other words, switching is such that it is not given enough time to the dynamics of F 1 or F 2 to bring the state close the origin. Rather, the state is driven further and further away, and an unstable behavior is obtained. This can be appreciate again in Figure 2 (third sub-figure) where the motion of ξ(t + 1) = F (t)ξ(t) is depicted. Perhaps, it is worth remarking that the unstable behavior is obtained proper because of switching. F (t) continues to oscillate between F 1 and F 2 and F (t) does not converge to any matrix when t. 22