Chapter 4 An introduction to multivariable Systems In this chapter a number of tools for the analysis of multivariable systems will be given. In particular, this will include poles and zeros of multivariable sy stems the generalized Ny quist Theorem for st ability analysis of multivariable sy stems, and frequency ses for multi- variable sy stems using singular values. 4.1 Poles and zeros of Multivariable s ystems For a single input single output(SISO) system with the transfer function G(s (4.1) the poles are defined as the values of the complex values s for which G(s)=oo, and the zero as the values of s for which G(s)=0 For a multiple input multiple output(MIMO) system with the transfer matria Gn,1(s Gn,2(s) the poles can be defined in analogy with the definition for SISo system. I.e., as the poles of every scalar transfer function Gll(s). Gn,n,s) in G(s). This definition is reasonable, as at ole. Hence, at least one of the entries in the transfer matrix equals Zeros for multivariable systems, however, can not in a reasonable way be defined simply as the values for which one of the entries of the transfer matrix equals zero, so the multivariable zeros are not directly related to the zeros of the indiviual transfer functions G11(s).Nun,(s) Instead, the zeros of a multivariable sy stem are defined as the values of the complex variable s for which G(s) loses rank. The rank of a matrix A is defined as the number of linearly independent columns of A and denoted by
! "#! $ % && ! ' "(! $ $ $ ) % * $ % $ % + + % $ + ",! (-
Thus, the definition of zeros for multivariable systems becomes Definition 4.1(Zeros for Multivariable Syst ems) The zeros of a transfer matriz G(s are defined as those values of the compler variable s for which rank G(s) is less than its morima valule Zeros defined in this way are called transmission zeros. The reason is the following transfer function G(s)loses rank for 8= %o, it can be shown, see Section 4.3.3, that the exist s an input vector uo#0, such that =0 Hence, the transmission of cert ain input signals is blocked for 8= Zo 4.1.1 Smith-McMillan form of a Transfer matrix Poles and transmission zeros of a transfer matrix G(s can be found, e.g. by transforming G(s)to its Smith-McMillan form. It can be shown that any propertransfer matrix G (s)can be written in its Smith-MCMillan form G(8)=U1(s)M(s)U2(s (4.5) U1(8)(1(8) 0}U2(s) (4.6) where U1(s)and U2(s)are unimodular matrices U(s)is said to be a unimodular matrix,if and only if its determinant det U (s is in dependent of s, i. e. if det U (s) is const ant. M(s) is a pseudo-diagonal matrix, and is called the Smith-McMillan form of G(s). G(s)and M(s) said to be similar, denoted as G(s) M(s). The polyn Ixi(s),i(s) have to be common factors or-equivalently -no common root s). Finally, xi(s),i(s) have to possess the following divisibility pr x(8)ki+1(s) +1(8)(8) The not ati s that the polynomial xi+1(s) is a factor of the poly nomial xi(s)(with no remainder ). Next, the following pole and zero polynomials are defined p(8)=1(s)φ2(s)…q(s x(8)=X1(8)X2(s)…Xr(s) (4.9) Now, it can be shown, that the poles and transmission zeros of the transfer function G(s)can e founds as the roots of p(s) and z(s), respectively. The degree of the pole polynomial p(s) called the McMillan degree of G (s) It can be shown that the Smith-McMillan form M(s of a transfer matrix G(s) can be deter mined by a series of elementary row and column operations on G(). let d(s) be the smallest I A transf (s)is said to be proper if all its entries satisfy Gi;(s)1-Ci< oo for rictly proper if Gil(s)|→0fors→∞
$ ' * $ + ",, % ""! ) + . % % && ' "/! "-! % % && 0 ! ! ' "1! ! % $ ' "2! "3! && % 4
common denominator for the entries Gij (s)of G(s). Then, G(s) can be written as N1(s) 1n(s) G(s)=dsN(s)=ds where N(s)is a poly nomial matrix (not a transfer matrix). There are three element ary types of row or column operations to perform on N(s) Interchange of two rows or columns Multiplication of a row or a column with a const ant Addition of one row or column multiplied with a polynomial to another A common property of these operations is, that they do not change the rank of the matrix ( s). Each of these elementary operations can be represented as a pre- or post multiplication of N(s by a suit able matrix L(s) called an elementary matric. It can be shown, that all element ary matrices are unimodular. Now, N(s can be rewritten as a sequen ce of row and column operations (8)=L1(s)S(s)L2(8) Ln(s)diag{∈1(s),E2(8),……,6r(s),0,0,…,0}L2(s) (4.12) Here, S(s is a pseudo-diagonal polynomial matrix. S(s) is called the Smith form of N(s The polynomials Ei(s) are monic and have the following divisibility properties 1 (4.13) Hence, the Smith form of a polynomial matrix is equivalent to the Smith-McMillan form of a transfer matrix. Now, the point is that the polynomials Ei(s) can be determined from the determinant divisors (4.14) Di()=greatest common divisor for all i x i sub determinants of N(s) (4.15) where every of the greatest common divisors are normalized to a monic polynomial. It can be shown, see [Mac89, Pages 40-43 that the polynomials Ei() are given by (4.16) Hence, the Smith-McMillan form of G(s) is given by M()=(s) 17) In summary, this est ablishes the following procedure for determining the Smith-McMillan form of a transfer matrix G(s)
' "#5! % %! ' & 6 6 + % 7 % ' "##! "#(! ) % ' "#,! ) % && % ' "#"! "#/! 8&23 . "59",: ' "#-! ) && ' "#1! && % '
Procedure 4.1(Smith-McMillan form of a transfer mat rix) 1. Let G(s)be a proper transfer matrit. Find the smallest common denominator d(s) for all entries in G(s)and rewrite G(s)as G(s N(8) (4.18) 2. Determine the Smith form ofN(s) N(8)~S(8)=dag{∈1(8),E2(s),……,6r(s),0,0,…,0} (4.19) where Ei(s)is determined from the determinant divisors D(s) Di(s) ;(s)=D1 8. Then, the Smith-McMillan form of G(s)is given by (8)~M(8) (4.21) The following example has been taken from [ Mac89, Ex. 2.2 Example 4.1(The Smith-McMillan form of a transfer mat rix) Let G(s be given by 2= 22 2 2s-4 The smalles common denominator for Gi(s)is d(s)=82+38+2 and G(s)can be written as 1 G(8)=2+38+282+8-4282-8-8 423) (8-2)(s+2)(2s-4)(s+2) (4.24) The Smith form of the polynomial matrit N(s is given b (8)~S(s)=diag{∈1(8),E2(s)} where Ei( s)is determined from the determinant divisors D(s) e()=D-1(s) i=1,2
"#2! ! " # "#3! "(5! $ % # "(#! % + 8&23 7% ((: ! # "((! "(,! "("! # "(/! "(-!
(4.27) Dgtar a ged i, -i, a2n a-roa2-a-h,a2-Toa2-hg h D2tar a ged a2 n a-T h ' a2-T h h fata2-Tt;zta2-T aata2-fry a a2-Ta ta n orta-or (4.30) nanc ged ochr Ici Inc t ransaf s emi i nmr o T 'mrz Svnlncnx C (4.31) Datar ∈tara a ta n orta -or (4.32) Dstar Inc Na pIncT'wTrt h P tarp grl ch 1 ta n i rta a-0 433) a n ncnc Inc w a arh st wlrn Th hv a cit Irn tho ochr a pht Irn ntIc. cch wt hwcyyaoz chwcx Ic gryc tho bcnr gryhra pt yi srn p tar t gr ch ptar a ta n irta n or (4.34) 7t unix P tar i Inc gryci a a -1, -i, -o tho Inc ht hi a piprh bair a a oz (=2 +tate +p82e Des2riptions a dyce decer (bed by =he dt =e dlt ee relrece)=t =o) cttr a 2 rttrn Bott r (4.36) ttr a catton Do tt r et)be delle=ed tdo Fgure 4.1. SO) t)g=egrt =o)0=ine domt g eorredlo)dd =o as0 he a doto, =he =rt)fer mt =re P tar eorredlo)dog =0 =he dt=e dlt ee decer al=o) tbove et)be wrC=e)td where I Cdt)Cde)=y mt =re of =he ct me dame)do)dt d2. The 0)verce of t dqutre ) o)d gultr ce x Cdgeve) by djx where t djx Cd =he t djog=of X. He)ee, f aI-2 ()o)dgultr, p tar et) be wrC-e)td fental-2r dj taI-2
& % "(1! "(2! "(3! ",5! % ",#! ",(! % # ",,! '% # ","! ",/! % 6 ",-! ",1! "# % ",2! % % ; ",3! ; ; ) ; ""5!
D A Figure 4.1: State space rep res entation of a system. Since C adj(sI-A)B is a poly nomial matrix, obviously all poles of G(s)(i.e. the values of s for which G(s)=oo) have to be roots of the polynomial det(sI-A). The roots of det(sI-A) coincide with the eigenvalues of A. Hence, all poles of G(s) have to be eigenvalues of A. Th opposite needs not be the case al ways, since roots of det(sI - A)might be cancelled in(4.38) and consequently they will not appear as poles of G(s). This is the case, when the realiz ation (A, B, C, D)is uncontrollable, unob servable, or both. On the other hand, if the realiz (A, B, C, D)is both controllable and observable, the roots of det(sI- A)equals the poles of G(s and the pole polynomial p(s) will be given by ps=det(sI-A) 441) This means, that the dimension of A can not be smaller than the McMillan degree of G(s) Hence, a st ate sp ace realization which is both controllable and observable is called a minimal realization. These results can be summarized as the following theorem ThId li maldlaaaw K aaamranlavpacta Let G(s)be a tra fer matric with a minimal realization(A, B, C, D)and let p(s be the Smith-McMillan pole polynomial of G(s. Then dim a= degp(s (4.42) Hence, the McMillan degree of G(s equals the dimension cf a minimal realization. Moreot the eigenvalues f A equal the poles af G(s) Note, that if(A, B, C, D) is a non-minimal realization, then the poles of G(s) constitute a proper sub set of the eigenvalues of A aea Fdh rrtahrvelce dbaspranAauact The transformation of a state space description to a transfer function description is unique given by(4.38 ). In contrast, there are several ways in which a transfer function can be transformed into a st ate space description. A straightforward approach would be to derive separate st ate space descriptions for each column in G(s), i.e. for each input, and then collect these separate state space descriptions into an overall state space model. Let Gp) be the ith ch that G(s)=(Gn(s), Gt(s).,Ghs) O=tt C)ntr)
"#' ; % ! ) ",2! ""#! && ) ! " ""(! '% # ( % # ( # $ ! ",2! 6 4 "",!
32 of Write each column as T(8)+ 444) where di(s)is the common denominator polynomial of G:(s) d(s)=s4+a4-1+…+4 Note, that di(s) is monic. ni (s is a vector of polynomials, each having a degree strictly less than ki. The jth entry of n; (s) can be written as the polynom n1()=n,一1+-2+…+m is a vector consisting entirely of constants a state sp ace description in controllable canonical form of the column Gi(s) is then given by the realization(Ai, Bi, Ci,&)where n B 448) Finally, a realiz ation(A, B, C, D)of G(s) can be found as B=diag D=[51,2,…,na This realization is controllable, but not necessarily observable. If a minimal is required, there exist algorithms to remove the unob servable modes, see e.g. [Mac89, Section 8.3.5 The MATLAB function tfm2ss m from MATLABs Robust Control Toolbox pro duces a similar st ate sp ace realization and minreal m from MATLABTM's Control Toolbox can extract a minimal realiz ation from a non-minimal one Remark In computer aided design, and especially in MATLABTM, it is easier in general to work with st ate space descriptions, since it is difficult to represent transfer matrices, as this requires three dimensional structures. Robust control. howe sponse analysis of a number of transfer matrices, such as the sensitivity function S(s)and the complementary sensitivity function T(s). It is no problem, however, to compute the resp onse based on a st ate sp ace description of the system. Hence, a multivariable often be represented in st ate space representation, although the analy sis is performed in th frequency domain. This mixture of time and frequency domain is actually quite typical for
? % = ? % % % & + 0 > ) %
1 ode3a 3obusg ooag3ol gee 3y. Tee 1 osg 9o9ulf 3 solughba go gee H ooagol 93oblel 2 fo3 ulf ged 2a ge31 s of sgf ge s9f oe 1 fg32 que oy dol f 2a. To ff o2 2gf gege 2 dufl 3e 3eseagf gaba, gee follow lag aogf gaba fo3 f g f a sfe 3 fuaogaa of f sysgel bf sed oa 2gs sgf ge s9f oe 1 f g320es efs beea 2ag3o duoed (I)nt CI)I-An DC AlB 451) 4.2 Nominal Stability for Multivariable Systems Fa gee sequel, gee ooad 2Zbas fo3 ao1 aaf l sgf gab 2gy fo3 1 ulgaf 3f ble sysgel s wall be 9 3eseagec Teese ooad2g2ba s f3e eagaely f aflogous go geose fo3 scf If 3 (S So) sy sge l s. Jusg fs fo3 SeSO sysgel s, f l ugf 3/f ble sysgel 2 sgf ble, if fll 4s(S1 4e-MoMAlIf a )goles f 3e locf ged 2a gee fg ef If 9lf 4.2.1 Internal Stability Coa sade 3 gee ooag3ol sysgel 2 F2gu3e 4.2. Pa ooag f sg go 93ev 2ously, 2g 2 f ssul ed gef g fll gaf ls f 3e veco 3, f ad gef g fll gf a sfe fua ogpoas'f3e 1 fg32es of oo1 9fgzble d2l eas2oas noelke)nl). 2gu3e 4.2: Multivariable control system Agf a, 2age3afl sgf b 22gy wall be uade3sgood fs gee 9309e3gy, gef g f boua ded sagafl 2agoduoec f aywee3e 2a gee loo9 does aog sgal ulf ge f a uab ouaded s2gaf l f aywee 3e else 2a gee loo9. Jusg f s fo3 SRSO sysgel s, gee ooa agoa fo3 age afl sgf b21 2gy 3eduoes go sgf b 2gy of gee ool osage gf asfe3 1 fg f3o1 I)n(o3)n fad e I)ngo yI)n fad eI)n Hea oe, 2ageafl sgf b 212gy fo3 gee olosed loo9 sy gel 2 equ2vf leag go gee sysgel yI)n IIh(I)nkI)m (I)nkI)n IIh(I)nkI)m(I)n KI)nII h(I)nKI)m. -KI)nIIh(InKI)m(I)nEi)n efv2gfⅡ29oles2gee019 lex leng ef If 9lf ae.N08,ea8efo31的eM3eoe3名 gef a deal aaf go 3s 2(4.52), fs deaol af go3s do aog l fk 2(ve3y)21 903gf ag go 3el el be3 gef g 1 f g3% 1 ulg2912f g?oa 2 non-commutative, 2e. gef g gee ode 3 of gee ff ogo3s f 3e saga 2fiof ag fs i)nkint/ Kin( I)n 2 geae f I Fo3 gee sysge 1 2(4.52) go be sgf ble, fll fou3 g f a sfe3 1 f goes 1 usg be sgf ble Example 4.2(Internal st ability) If the controller ki)n is stable, internal stability can be checked by checking stability of the transfer function IIh( I)nKI)m (I)n Indeed, assume that KI)n and ITh( I)nKI)m (I)n
' "/#! ! @ &&! ? "( = = ! "(' 6 @ % ! ) "/(! % % "/(! + & ! % $ "/(! # ' ) % # )%
1eltAvliiy CAlyie 41Fde s 14aEo, 1P14 cede LAdP(sdt d Aecean1iy le s 14akeonaii yec4lyie EMK 点1 nai ye c4lyie s4lyaaay dl cAii feo dye ec4ljiaareMlyv4fao IPyeceeP Lds Atm aaacl-Atm ana aa an 1FMLas Are c4lyaiAy Atm aa an anr dh过ASHP小akA程A的中:H(我a 414d4e trd cs Siy yy 1cceccaPg 4re hicceMdis (2)2 Tn e Generalized NyqutsC T'n eore0 Tae Geseral ded Nyqu ab Taeorem plays a cesbral role s robust cos brol baeory. Jusb lke facllbabes as assesme lyp cally 4s berms of Ahab y of a cos bol sysem by as aly. dg dePid S proper点 cX=LF=ar如p a,m知 (4.5) Waes cos ao be loon eal aa esbablsaed. as d aes ce De At Darlan 457) Dar waere calatD ce asd da latp b. moreover At I tA at A ( at bell 62 w aere At i Now, bae dSPidSbrLI企4mA4Me4 BatAc(I AtDi 2A AlCAN (4.67) O=uCt C)nt ]
) * + % % "/,! ) % % % , # -! ) % # ! A @ + A ? "( 4 % ! "/"! ! "//! < ! "/-! "/1! "/2! "/3! & ' "-5! "-#! "-(! "-,! "-"! "-/! "--! $ ' "-1!
aldar and the liccevdshrllte acabsiyIds ali dci+ eCL n=de-nI a A4 (4.68) As described in Section 4.1.2, the st ability of the closed loop system is determined by the of the characteristic poly nomial ocL+). To reach a Ny quist-similar'expression, cl-n) be expressed in terms of the open loop transfer matrix m n)K-n) To that end, the eAv PNfeeFte matrix F-n)is introduced as F (4.69) Moreover, the following lemma, which is presented without a proof, will be instrument al LImma 4. 1(Schur's formula for block partitioned ditirminants) Ie4l Cvle s1- 4anP ye SL 44dReMlo n PP zreP, 4e Nees aFIf4d P HPyeeSecceMlo de-P=de-Pll de-P22 a P21 Pi P12), SctalaMie-Pl1 b0 (4.71) d de-p=de-P22 de-Pu a Pi2 Pr p2), Sctalevie-P22b0 472) Now, write the determinant of the return difference matrix Fn)as de-F+)=de+crIa lBr+ DI With the following variable substitutions P Ar P1 =a cr P21= bl P22=I+Dr (475) the right hand side of(4.73) corresponds to the last term of the right hand side of (4.71) Thus, multiplying by de-Pil= de nI a al gives r a de nI a al de+ Dr+cEnl a Albr=de acr I+ Dr 476) r a nia ar Br AD cr I+Dr Once again, applying Schur's formula, it can be shown that r a de- a a BrDr--de-ae-h-oraBrI +Dr
' "-2! 6 "#( = = % % % . % " " "-3! & ( ) & ( # # # # # # "15! % # # # # ## # # "1#! # # # ## # # "1(! B % " " "1,! < # # "1"! # # "1/! "1,! "1#! # "1-! " "11! = "12! "13!