Proceedings of the 19th World Congress

The International Federation of Automatic Control

Cape Town, South Africa. August 24-29, 2014

On the use of algebraic geometry for the

design of high-gain observers for
continuous-time polynomial systems
Laura Menini ∗ Antonio Tornambè ∗
∗
Dip. Ingegneria Civile e Ingegneria Informatica
Università di Roma Tor Vergata, Roma, Italy
(e-mail: [menini,tornambe]@disp.uniroma2.it)

Abstract: The goal of this paper is to apply some concepts and techniques from algebraic
geometry to study the observability of nonlinear continuous-time polynomial systems. After
deriving some new results for observability and embeddings, it is shown how to use such concepts
to easily design high-gain observers. The proposed technique is illustrated by an application to
the well known Rossler oscillator.

Keywords: Algebraic geometry, observability, high-gain observers, Rössler oscillator.

1. INTRODUCTION 2. PRELIMINARIES AND NOTATION

The goal of this section is to briefly resume the basic

notions of algebraic geometry (Cox et al. (1998, 2007))
One of the classical problems in control theory is that of that will be used in the sequel.
designing state observers to obtain in real time estimates of
>
unmeasurable state variables, using the available measured Let x ∈ Rn , x = [ x1 ... xn ] , let R[x1 , ..., xn ] (or, shortly,
outputs. This has motivated the study of the structural R[x]) be the commutative ring of all scalar polynomials
property of observability and of various techniques for in x1 , ..., xn , and let R(x1 , ..., xn ) (or, shortly, R(x)) be
designing state observers, as in the works by Krener and the field of all scalar rational functions of x1 , ..., xn ; Rn [x]
Respondek (1985); Isidori (1995); Nijmeijer and van der denotes the set of vector functions with n entries in R[x].
Schaft (1990); Tsinias (1990); Kazantzis and Kravaris
Let p1 , ..., pm be polynomials in R[x]; the set
(1998); Menini and Tornambe (2002a,b, 2010a,b, 2011a).
Vn (p1 , ..., pm ) := {x ∈ Rn : pi (x) = 0 for all i = 1, ..., m}
Some of the works closer to the problems considered
here are by Sussmann (1978); Sontag (1979); Takens is called the affine variety 1 (briefly, variety) of Rn defined
(1981); Aeyels (1981a,b); Jouan and Gauthier (1996). In by p1 , ..., pm . As an example, V1 (x1 ) is a point of R1 ,
some of such works the problem of observability through V2 (x1 ) is a line of R2 , and so on. If Va = Vn (p1 , ..., pma )
sampled measurements is considered. In particular, an and Vb = Vn (q1 , ..., qmb ), for some polynomials pi and qj
application using a CAS (Computer Algebra System) of in R[x], then both Va ∩ Vb and Va ∪ Vb are varieties, and
the developments by Sontag (1979) was done by Nešić they are given by:
(1998).
Va ∩ Vb = Vn (p1 , ..., pma , q1 , ..., qmb ),
When dealing with polynomial systems, it is very useful
to use techniques from Algebraic Geometry, which provide Va ∪ Vb = Vn (p1 q1 , ..., pma q1 , ..., p1 qmb , ..., pma qmb ).
the natural tools to state clear and easily testable results.
This has been recognized by a variety of authors Diop A subset I of R[x] is a polynomial ideal (briefly, an ideal )
(1991); Helmke et al. (2003); Tibken (2004); Menini and if it satisfies:
Tornambe (2014); Kawano and Ohtsuka (2013); Menini
(1) if p, q ∈ I, then p + q ∈ I;
and Tornambe (2013a,b). The main novelty of this paper,
with respect to such works, is that here observability and (2) if p ∈ I and q ∈ R[x], then pq ∈ I.
embedding problems are studied with the focus of observer
design, whence results based on algebraic geometry are Let p1 , ..., pm be polynomials in R[x]; it is possible to show
stated that allow the easy writing of explicit forms for an that the set
inverse of the observability map and of an embedding of hp1 , ..., pm in := {q1 p1 + ... + qm pm : qi ∈ R[x], i = 1, ..., m}
the given system. Although we illustrate the application is an ideal of R[x] and, in particular, it is called the ideal
of our techniques with respect to high gain observers generated by p1 , ..., pm .
(see, e.g.,Tornambe (1989); Esfandiari and Khalil (1992);
Gauthier and Kupka (2001); Hammouri et al. (2010)), it 1 In some texts Vn (p1 , ..., pm ) is called algebraic set, whereas the
is stressed that, in principle, they can be applied to other term affine variety is reserved for the case when Vn (p1 , ..., pm ) is
ways of designing observers. irreducible.

978-3-902823-62-5/2014 © IFAC 43
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

An ideal I is said to be finitely generated if there exist ∀x ∈ Rn (algebraically independent, otherwise). Necessary
polynomials p1 , ..., pm in R[x] such that I = hp1 , ..., pm in ; and sufficient condition for q1 ,..., qm to be algebraically
the set of the ideal generators {p1 , ..., pm } is called a dependent is that rankR(x) ∂x ∂q
< m.
basis of I. By the Hilbert Basis Theorem (Cox et al.,
2007), every ideal of R[x] is finitely generated. Clearly, Fix a total ordering > of the monomials of R[x]. For any
if hp1 , ..., pma in = hq1 , ..., qmb in , for some polynomials pi p ∈ R[x], with p 6= 0, one can write
and qj in R[x], then Vn (p1 , ..., pma ) = Vn (q1 , ..., qmb ).
p(x) = a1 xα1 + a2 xα2 + ... + a` xα` ,
Let V = Vn (p1 , ..., pm ) be the variety defined by p1 , ..., pm where ai ∈ R, αi is a multi-index, i = 1, 2, ..., `, and
in R[x]. By definition, polynomials p1 , ..., pm vanish on V, xα1 > xα2 > ... > xα` ; this allows one to define the
whence all polynomials belonging to hp1 , ..., pm in vanish on leading monomial LM(p) = xα1 and the leading coefficient
V. Note that there may be polynomials p ∈ / hp1 , ..., pm in LC(p) = a1 of p.
that vanish on V; this renders necessary the following
definitions. Let {p1 , ..., pm } be a basis of an ideal I of R[x]. A
polynomial r ∈ R[x] is said to be reduced with respect
Let V be a variety of Rn ; it is possible to show that the to {p1 , ..., pm } if either r = 0 or no monomial that appears
set in r is divisible by LM(pi ), i = 1, ..., m. A polynomial
In (V) := {p ∈ R[x1 , ..., xn ] : p(x) = 0 for all x ∈ V} (1) r ∈ R[x], which is reduced with respect to {p1 , ..., pm },
is an ideal and, in particular, it is called the ideal of V; note is called a remainder of the division of p ∈ R[x] by
that the empty set ∅ is a variety and that In (∅) = h1in = {p1 , ..., pm } if p − r ∈ hp1 , ..., pm in .
R[x]. Clearly, hp1 , ..., pm in ⊆ In (Vn (p1 , ..., pm )), and the Example 1. Fix the deglex (degree lexicographic) mono-
equality hp1 , ..., pm in = In (Vn (p1 , ..., pm )) need not occur, mial ordering on R[x1 , x2 ] with x2 > x1 . Let p(x) = x1 x22 −
e.g., if n = 1, all the polynomials x21 q(x1 ) of x21 1 , with x1 and hp1 , p2 i2 = x1 x2 − x2 , x22 − x1 2 . It is easy to see
q ∈ R[x1 ], vanish on the variety V = V1 (x21 ) = {0}, but that r1 = 0 and r2 = x21 − x1 are both reduced with
I1 (V1 (x21 )) = hx1 i1 and x21 1 6= hx1 i1 since p(x1 ) = x1 respect to {p1 , p2 } and that p = x2 p1 + p2 + r1 and
p = 0p1 + x1 p2 + r2 , which show that the remainder of
vanishes on {0}, but does not belong to x21 1 .
the division by an arbitrary basis need not be unique.
Let Va and Vb be two varieties of Rn , then
Let {g1 , ..., gm } be a set of non-zero polynomials in R[x]. A
(1) Va ⊂ Vb if and only if In (Va ) ⊃ In (Vb ); basis {g1 , ..., gm } of an ideal I of R[x] is called a Gröbner
(2) Va = Vb if and only if In (Va ) = In (Vb ); basis of I if, for any p ∈ R[x], the remainder of the division
of p by {g1 , ..., gm } is unique. Every non-zero ideal I of R[x]
this, in particular, shows that the map In (V) is one-to-one. has a Gröbner basis, which need not be unique 2 .
For instance if n = 1, Va = {1} and Vb = {1, 2}, then A monomial ordering on R[xa , xb ], with xa ∈ Rna and
I1 (Va ) = hx1 − 1i1 and I1 (Vb ) = h(x1 − 1)(x1 − 2)i1 , with xb ∈ Rnb , n = na + nb , eliminates xa if
I1 (Va ) ⊃ I1 (Vb ). xα β α γ β δ
a > xa ⇒ xa xb > xa xb ,
Let I be an ideal of R[x]; it is possible to show that the for all multi-indices α, β for which xα β
a > xa and for all
set multi-indices γ, δ. For instance, the lexicographic ordering
Vn (I) := {x ∈ Rn : p(x) = 0 for all p ∈ I} (2) with x1 > x2 > ... > xn eliminates x1 , ..., x` , for all 1 ≤
is a variety and that Vn (I) = Vn (p1 , ..., pm ) for any basis ` < n. Let I be an ideal of R[xa , xb ]. The elimination ideal
{p1 , ..., pm } of I. of I that eliminates xa is I ∩ R[xb ]. Let {g1 , ..., gm } be a
Gröbner basis for a monomial ordering that eliminates xa .
A variety V can be studied through the corresponding Then, the set obtained from {g1 , ..., gm } by retaining only
ideal In (V) given by (1) and, conversely, an ideal I can the elements that do not depend on xa (i.e., {g1 , ..., gm } ∩
be studied through the corresponding variety Vn (I) given R[xb ]) is a Gröbner basis of the elimination ideal I ∩ R[xb ]
by (2), but In (Vn (I)) need not coincide with I (e.g., for the monomial ordering on R[xb ] induced by >.
I1 (V1 ( x21 1 )) = hx1 i1 ), whence the map Vn (I) is not
Example 2. Let I = x1 + x22 , x1 − x2 2 . Fix the lexico-
one-to-one (e.g., V1 (h1i1 ) = V1 ( 1 + x21 ) and h1i1 6= graphic ordering with x1 > x2 , which eliminates x1 ; a
1 + x21 1 ). Gröbner basis of I for such an ordering is {x22 +x2 , x1 −x2 },
Given any subset S of Rn , it is easy to check that whence the elimination ideal of I that eliminates x1 is
I ∩ R[x2 ] = x22 + x2 1 ; in particular, {x22 + x2 } is a
In (S) := {p ∈ R[x1 , ..., xn ] : p(x) = 0 for all x ∈ S}
Gröbner basis of I ∩ R[x2 ] for the induced ordering (in
is an ideal even if S is not a variety; hence, Vn (In (S)) is a this case trivial).
variety that contains S. In particular, it is the smallest
variety containing S; Vn (In (S)) is called the Zariski 3. OBSERVABILITY AND EMBEDDINGS OF
closure of S. As an example that will be useful in the POLYNOMIAL SYSTEMS
following, if S ⊆ Rn is open, S = 6 ∅, then In (S) =
{0} and its Zariski closure coincides with the whole Rn Consider the following polynomial system
(Vn (In (S)) = Rn ).
Some polynomials q1 , ..., qm ∈ R[x] are algebraically de- ẋ = f (x), (3a)
pendent if there exists a non-zero polynomial p ∈ R[q], 2 By using the stronger notion of reduced Gröbner basis unicity is
>
where q = [ q1 . . . qm ] , such that p(q1 (x), ..., qm (x)) = 0, achieved, but this is not needed here.

44
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

y = h(x), (3b) A solution of Problem 2 is called an explicit embedding of

n
where h ∈ R[x] and f ∈ R [x]. Consider the observability system (3).
map of order h + 1, h ∈ Z, h ≥ 0, given by ye,h = Oh (x), Let p(ye,N ) = yN − p̄(ye,N −1 ) be a solution to Problem 2;
> i
where ye,h = [ y0 ... yh ] , yi = ddtyi , for i = 0, 1, ..., h, and the differential equation corresponding to yN = p̄(ye,N −1 )
 0  can be rewritten in the first-order normal form (or state-
Lf h(x) space form) as follows:
 Lf h(x) 
Oh (x) :=  .. , (4)
 
 .  ẏ0 = y1 ,
Lhf h(x) ..
.
∂Lif h(x)
with Li+1
f h(x) = ∂x f (x) and L0f h(x) = h(x). ẏN −1 = yN ,
It is classical to define system (3) to be (N + 1)- ẏN = p̄(ye,N −1 ),
differentially observable if the observability map ye,N =
y = y0 .
ON (x) is injective, i.e., if it is left invertible with a (not
−1
necessarily unique) left inverse x = ON (ye,N ) such that Now, let the polynomial λN +1 + k1 λN + ... + kN +1 have all
−1 n
ON ◦ ON (x) = x, ∀x ∈ R (see Jouan and Gauthier roots with negative real part and let ε > 0 be a sufficiently
(1996)). Such a notion can be extended as follows. small parameter. Under the assumptions and conditions
Definition 1. System (3) is (N + 1)-polynomially (respec- of Theorems 1 and 2 of Tornambe (1992) (essentially,
tively, rationally) observable if it is (N + 1)-differentially boundedness of p̄(ye,N −1 (t)) as a function of t), a high-
−1 gain “practical” observer for such a system is given by
observable and ON (ye,N ) is a polynomial (respectively,
rational) function.
k1
ŷ˙ 0 = ŷ1 + (y0 − ŷ0 ), (7a)
Clearly, if system (3) is (N + 1)-differentially (respectively, ε
polynomially or rationally) observable for a certain N , ..
then it is (N + 1)-differentially (respectively, polynomially . (7b)
or rationally) observable for all N ≥ N . k
ŷ˙ N −1 = ŷN + N (y0 − ŷ0 ),
N
(7c)
Now we consider the problem of finding an embedding for ε
kN +1
system (3). ŷ˙ N = N +1 (y0 − ŷ0 ). (7d)
ε
Problem 1. Given an integer N ≥ 0, find, if any, a
polynomial p(ye,N ) ∈ R[ye,N ] such that The observer (7) guarantees that the estimation error
ye,N (t) − ŷe,N (t) can be rendered arbitrarily small, by
p ◦ ON (x) = 0, ∀x ∈ Rn . (5)
decreasing ε, and has the advantage that the obtained
Note that, if the entries of ON (x) are algebraically depen- practical stability is global (i.e., the initial errors can be
dent and have degree less than or equal to d, then there arbitrary).
exists a polynomial p(ye,N ), having degree less than or Once the output and its derivatives have been estimated
equal to dN +1 , that solves Problem 1. for a sufficiently high N , under the assumption that
Any polynomial p(ye,N ) ∈ R[ye,N ] can be thought of as an the system is (N + 1)-differentially observable (hopefully,
ordinary differential equation, polynomially or rationally observable), an estimate of
−1
the state variables can be obtained by x̂ = ON (ŷe,N ),
p(ye,N ) = 0; (6) especially when ye,N (t) is bounded as a function of t.
in particular, a polynomial p(ye,N ) solution of Problem 1
is called an embedding of (3). Any embedding has the The following lemma follows easily from the definition
following property: if x(t) is a solution of (3a), then of algebraic dependence (see e.g., Menini and Tornambe
y(t) = h(x(t)) is a solution of (6), i.e., the corresponding (2011b, 2009, 2010c,d); Menini and Tornambè (2012)).
ye,N (t) satisfies (6) identically. Lemma 1. There exists a solution p(ye,N ) ∈ R[ye,N ] to
Problem 1 with p 6= 0 if and only if the entries of
About any regular point of ye,N = ON (x), by the Implicit the observability map ye,N = ON (x) are algebraically
Function Theorem (see, e.g., Fleming (1987)), there ex- dependent.
ists locally a function ϕ(ye,N −1 ) such that the following
relation holds locally about such a regular point: As well known, the entries of the observability map
p(ye,N ) = 0 ⇔ yN = ϕ(ye,N −1 ); ye,N = ON (x) are
 algebraically dependent if and only if
the differential equation corresponding to yN = ϕ(ye,N −1 ) ∂ON (x)
rankR(x) ∂x ≤ N.
is often called in normal form. Unfortunately, ϕ(ye,N −1 )
need not exist about a singular point of ye,N = ON (x) The following theorem gives, in terms of algebraic geome-
and, even in the case it exists, it need not be polynomial try, a characterization for the entries of the observability
nor rational. Thus, the following particularization of Prob- map to be algebraically independent, whence for Prob-
lem 1 will be considered in this paper. lem 1 to have no solution.
Problem 2. Given an integer N ≥ 0, find, if any, Theorem 1. Let YN = ON (Rn ) be the image of Rn
p(ye,N −1 ) ∈ R[ye,N −1 ] (respectively, p(ye,N −1 ) ∈ R(ye,N −1 )) through the observability map; let VN +1 (IN +1 (YN )) be
such that p(ye,N ) = yN − p(ye,N −1 ) satisfies (5). the Zariski closure of YN . The entries of ON (x) are

45
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

algebraically independent (whence, there is no solution to the set of all polynomials being solution of Problem 1, and
Problem 1) if and only if VN +1 (IN +1 (YN )) = RN +1 . define
Corollary 1. (1.1) If N < n, the entries of the observability JN := y0 − L0f h(x), ..., yN − LNf h(x) n+N +1 .
map ye,N = ON (x) are algebraically dependent if and only Theorem 3. (3.1) IN is a polynomial ideal of R[ye,N ];
if system (3) is not (N + 1)-differentially observable;
(3.2) IN = JN ∩ R[ye,N ];
(1.2) If N ≥ n, the entries of the observability map
ye,N = ON (x) are always algebraically dependent. (3.3) if YN = ON (Rn ) is the image of Rn through the
observability map, then VN +1 (IN ) is the Zariski closure
The following theorem gives two sufficient conditions un- of YN , i.e., the smallest variety containing YN ; in addition,
der which there exists a solution to Problem 2. if system (3) is (N + 1)-differentially observable, then
Theorem 2. (2.1) If system (3) is N -polynomially (respec- VN +1 (IN ) = YN .
tively, rationally) observable, then there exists p(ye,N −1 ) ∈
R[ye,N −1 ] (respectively, p(ye,N −1 ) ∈ R(ye,N −1 )) such that According to the proof of Theorem 3, if one fixes the
p(ye,N ) = yN − p(ye,N −1 ) satisfies (5). lexicographic monomial ordering with x1 > x2 > . . . >
xn > yN > yN −1 > . . . > y0 , then a Gröbner basis of
(2.2) If N ≥ n+1, then there exists p(ye,N −1 ) ∈ R(ye,N −1 ) IN can be obtained from a Gröbner basis G of JN as
such that p(ye,N ) = yN − p(ye,N −1 ) satisfies (5). G ∩ R[ye,N ], i.e., simply by taking all the elements of G
Remark 1. Statement (2.1) of Theorem(2) guarantees that that do not depend on the entries of x.
the N -polynomial (rational) observability, for some fixed By the second part of Statement (3.3), if system (3) is (N +
N , implies the existence of a polynomial (rational) em- 1)-differentially observable, then x(t) is a solution of (3a)
bedding, and Statement (2.2) guarantees that there al- if and only if y(t) = h(x(t)) is a solution of p(ye,N ) = 0,
ways exists a rational embedding for N ≥ n + 1. As for all p ∈ IN .
will be illustrated in Section 4, in practice, the rational
embeddings can be more useful if the poles of the rational 4. APPLICATION TO THE ROSSLER OSCILLATOR
function p(ye,N −1 ) are guaranteed not to be attainable by
the system trajectories. For some specific system, such an The Rössler oscillator is described by the equations:
additional requirement may be possible only for values of
N greater than n + 1.
ẋ1 = x1 x2 + b − cx1 , (8a)
Example 3. If pair f, h is in the observability form,
ẋ2 = −x1 + x3 , (8b)
x2 ẋ3 = −x2 + ax3 , (8c)
 
..

f (x) =  .

, y = x2 , (8d)
 xn 
where a, b and c are three scalar positive parameters,
ϕ(x1 , . . . , xn ) and the choice of the measured output is just an exam-
h(x) = x1 , ple, the studies reported here below can be analogously
developed for, e.g., y = x1 and y = x2 . Despite the
with ϕ ∈ R[x], then Problem 2 is solvable by apparent simplicity of such dynamic equations, for many
yN = ϕ(y0 , . . . , yN −1 ). values of the parameters the Rössler oscillator exhibits a
Example 4. Let n = 1, f (x) = 12 x and h(x) = x2 . It is chaotic behaviour, and the presence of an attractor renders
easy to see that yi = x2 , for all i = 0, 1, . . ., whence it a perfect benchmark for observer design, since it is
the considered system is not N -differentially observable easy to generate bounded chaotic trajectories not conver-
for any N ≥ 0. Nevertheless, for such a system one has gent to the origin. All the computations in this section,
the linear explicit embedding y1 − y0 = 0. and in particular the computation of all the mentioned
Gröbner bases, have been performed by using the freeware
Problems 1 and 2 have been extensively studied in the ana- Macaulay 2, a Computer Algebra System (CAS) special-
lytic case by Jouan and Gauthier (1996) and Gauthier and ized in algebraic geometry computations (see Grayson and
Kupka (2001) (earlier similar studies were given in Takens Stillman (2013)).
(1981)), showing that if N ≥ 2n + 1, then Problems 1 By computing Lf h(x), L2f h(x) and L3f h(x), it can be
and 2 are generically solvable, thus implying the existence
of an analytic function ϕ, locally about any regular point of seen that the observability map O3 (x) is not injective.
p(ye,N ) (a point ye,N = ON (xr ) is regular for p(ye,N ) if the By computing L4f h(x) and a Gröbner basis G4 of the
Jacobian matrix of the observability mapping has full col- ideal J4 = hy0 − h(x), y1 − Lf h(x), y2 − L2f h(x), y3 −
umn rank about xr ), such that yN = ϕ(y0 , ..., yN −1 ). The L3f h(x), y4 −L4f h(x)i of R[x1 , x2 , x3 , y4 , y3 , y2 , y1 , y0 ] w.r.t.
above mentioned results do not give a constructive pro- the lexicographic monomial order (with x1 > x2 > x3 >
cedure for the computation of such a polynomial p(ye,N ), y4 > . . . > y0 ), it can be seen that the observability
which can be determined by eliminating the x variables map ye,4 = O4 (x) is globally invertible. In fact, G4 is
from the equations y0 = L0f h(x), ..., yN = LN f h(x) consti- constituted by 8 polynomials, of which the relevant ones
tuting the observability map ye,N = ON (x). are
Clearly, the elimination theory for polynomial ideals seems
to be the correct instrument for such computations in the g4,4 = x3 y0 + (−a − c)x3 + γ4,4 (ye,4 ),
polynomial case. To pursue such an approach, let IN be g4,5 = x3 y1 + γ4,5 (ye,4 ),

46
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

g4,6 = x3 y2 + γ4,6 (ye,4 ),

g4,8 = x1 − x3 + y1 .
The complete expressions of the above polynomials are
omitted for brevity. To see that ye,4 = O4 (x) is injective,
note that h(x) − a − c, Lf h(x) and L2f h(x) do not vanish 10

jointly at any x ∈ R3 , being a + b + c 6= 0 by hypoth- 5

esis. This can be checked immediately by computing the

x3
0
Gröbner basis of the ideal hh(x) − a − c, Lf h(x), L2f h(x)i, −5
that is given by {a + b + c, x2 + b, x1 − x3 }, and noting −10
that one of its elements never vanishes. Whence, by equat-
ing to zero (y0 − a − c)g4,4 , y1 g4,5 and y2 g4,6 , and summing
10
up the three resulting equations, one obtains the following
20
rational expression 0 15
x2
10
(y0 −a−c)γ4,4 (ye,4 )+y1 γ4,5 (ye,4 )+y2 γ4,6 (ye,4 ) −10 5 x1
x3 = − , (9) 0
(y0 − a − c)2 + y12 + y22
that is globally valid, along the trajectories of the system.
Furthermore, by equating to zero g4,8 , one obtains: Fig. 1. The chosen trajectory of the Rössler system for
t ∈ [0, 40].
x1 = x3 − y1 ; (10)
expressions (9), (10), with x3 replaced by its expression
given by (9), and the trivial x2 = y0 , constitute an explicit 20
expression of a left inverse x = O4−1 (ye,4 ), that is needed
x1 , x̂1

to design the fifth order “practical” observer (7). 10

0
To perform some simulations, the values of the system 0 5 10 15 t 20 25 30 35 40
parameters have been chosen as a = 0.1, b = 0.1 and
c = 10. The initial condition for the system has been 10
x2 , x̂2

chosen as 0
>
x(0) = [ 6 16 −0.6 ] , −10

which, in the time interval [0, 40], gives the trajectory 0 5 10 15 t 20 25 30 35 40

reported in Fig. 1. The initial condition for the observer
has been chosen as ŷi = 0, i = 0, . . . , 4. The polynomial 10
x3 , x̂3

λN +1 + k1 λN + ... + kN +1 has been chosen as 0

(λ + 1)(λ2 + 2λ + 2)(λ2 + 3λ + 4), −10

and the parameter ε has been set as ε = 0.25 · 10−3 . 0 5 10 15 t 20 25 30 35 40

The estimates of the state variables have been obtained

by x̂(t) = O4−1 (ŷe,4 (t)), being ŷe,4 (t) the observer state.
Fig. 2. The time behaviour of the three state variables and
In Fig. 2 the time behaviour of the three state variables of their estimates, for the chosen trajectory.
x1 (t), x2 (t) and x3 (t) is reported together with the time
behaviour of the respective estimates x̂1 (t), x̂2 (t) and
x̂3 (t); they are practically undistinguishable. But, since
the observer is just a “practical” one, a non vanishing 5
estimation error is expected: such an error is shown in
x1 − x̂1

Fig. 3. The obtained error is acceptable, it can be reduced 0

arbitrarily by decreasing ε, at the expense of a worst initial
transient and more expensive computations. −5

0 5 10 15 20 25 30 35 40
5. CONCLUSIONS
t
After characterizing observability and high-gain observer 5
design for polynomial systems in terms of algebraic ge-
x3 − x̂3

ometry notions, the main theoretical result of this paper 0

has been derived, that is Theorem 3, which allows to
find all possible implicit embeddings for a polynomial −5
system by using Gröbner bases and elimination theory.
Then, through the application to the Rössler system, it 0 5 10 15 20 25 30 35 40

has been shown that the same techniques, that can be t

used by means of powerful modern Computer Algebra
Systems, effectively allow the design of high-gain observers Fig. 3. The time behaviour of the estimation errors for the
for polynomial systems. chosen test.

47
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014

REFERENCES Systems & Control Letters, 59(1), 33–41.

Menini, L. and Tornambe, A. (2010c). Semi-invariants
Aeyels, D. (1981a). Generic observability of differentiable and their use for stability analysis of planar systems.
systems. SIAM Journal on Control and Optimization, International Journal of Control, 83(1), 154–181.
19(5), 595–603. Menini, L. and Tornambe, A. (2010d). Stability analysis
Aeyels, D. (1981b). On the number of samples necessary to of planar systems with nilpotent (non-zero) linear part.
achieve observability. Systems & Control Letters, 1(2), Automatica, 46(3), 537–542.
92–94. Menini, L. and Tornambe, A. (2011a). Design of state
Cox, D., Little, J., and O’Shea, D. (1998). Using algebraic detectors for nonlinear systems using symmetries and
geometry, volume 185. Springer Verlag. semi-invariants. Systems & Control Letters, 60(2), 128–
Cox, D., Little, J., and O’Shea, D. (2007). Ideals, vari- 137.
eties, and algorithms: an introduction to computational Menini, L. and Tornambe, A. (2011b). Symmetries and
algebraic geometry and commutative algebra, volume 10. Semi-invariants in the Analysis of Nonlinear Systems.
Springer Verlag. Springer, London.
Diop, S. (1991). Elimination in control theory. Mathemat- Menini, L. and Tornambè, A. (2012). Exact and approx-
ics of control, signals and systems, 4(1), 17–32. imate feedback linearization without the linear control-
Esfandiari, F. and Khalil, H. (1992). Output feedback lability assumption. Automatica, 48(9), 2221–2228.
stabilization of fully linearizable systems. International Menini, L. and Tornambe, A. (2013a). Immersion and
Journal of Control, 56(5), 1007–1037. Darboux polynomials of Boolean networks with appli-
Fleming, W.H. (1987). Functions of several variables. cation to the Pseudomonas syringae HRP regulon. In
Springer-Verlag, New York Heidelberg Berlin, third edi- 52th IEEE Conference on Decision and Control, 4092–
tion. 4097. Firenze, Italy.
Gauthier, J. and Kupka, I. (2001). Deterministic observa- Menini, L. and Tornambe, A. (2013b). Observability and
tion theory and applications. Cambridge Univ Pr. dead-beat observers for Boolean networks modeled as
Grayson, D.R. and Stillman, M.E. (2013). Macaulay2, polynomial discrete-time systems. In 52th IEEE Con-
a software system for research in algebraic geometry. ference on Decision and Control, 4428–4433. Firenze,
Available at http://www.math.uiuc.edu/Macaulay2/. Italy.
Hammouri, H., Bornard, G., and Busawon, K. (2010). Menini, L. and Tornambe, A. (2014). On a Lyapunov
High gain observer for structured multi-output nonlin- equation for polynomial continuous-time systems. In-
ear systems. Automatic Control, IEEE Transactions on, ternational Journal of Control, 87(2), 393–403.
55(4), 987–992. Nešić, D. (1998). A note on observability tests for general
Helmke, U., Martin, C.F., and Zhou, Y. (2003). Gener- polynomial and simple wiener–hammerstein systems.
icity of algebraically observable polynomial systems. Systems & control letters, 35(4), 219–227.
preprint. Nijmeijer, H. and van der Schaft, A.J. (1990). Nonlinear
Isidori, A. (1995). Nonlinear control systems. Communi- dynamical control systems. Springer-Verlag, New York
cations and Control Engineering. Springer, Berlin Hei- Heidelberg Berlin.
delberg New York, third edition. Sontag, E.D. (1979). On the observability of polynomial
Jouan, P. and Gauthier, J. (1996). Finite singularities systems, i: Finite-time problems. SIAM Journal on
of nonlinear systems. output stabilization, observabil- Control and Optimization, 17(1), 139–151.
ity, and observers. Journal of Dynamical and Control Sussmann, H.J. (1978). Generic single-input observability
systems, 2(2), 255–288. of continuous-time polynomial systems. In Decision
Kawano, Y. and Ohtsuka, T. (2013). Observability at an and Control including the 17th Symposium on Adaptive
initial state for polynomial systems. Automatica. Processes, 1978 IEEE Conference on, volume 17, 566–
Kazantzis, N. and Kravaris, C. (1998). Nonlinear observer 571. IEEE.
design using Lyapunov’s auxiliary theorem. Systems and Takens, F. (1981). Detecting strange attractors in turbu-
Control Lett., 34, 241–247. lence. In Dynamical systems and turbulence, Warwick
Krener, A. and Respondek, W. (1985). Nonlinear observers 1980, 366–381. Springer.
with linearizable error dynamics. SIAM J. Control and Tibken, B. (2004). Observability of nonlinear systems-
Optimization, 23, 197–216. an algebraic approach. In Decision and Control, 2004.
Menini, L. and Tornambe, A. (2002a). State estimation CDC. 43rd IEEE Conference on, volume 5, 4824–4825.
of (otherwise unobservable) linear mechanical systems IEEE.
through the use of non-smooth impacts: the case of two Tornambe, A. (1989). Use of asymptotic observers having-
mating gears. Automatica, 38(10), 1823–1826. high-gains in the state and parameter estimation. In
Menini, L. and Tornambe, A. (2002b). Velocity ob- Decision and Control, 1989., Proceedings of the 28th
servers for non-linear mechanical systems subject to IEEE Conference on, 1791–1794. IEEE.
non-smooth impacts. Automatica, 38(12), 2169–2175. Tornambe, A. (1992). High-gain observers for non-linear
Menini, L. and Tornambe, A. (2009). Linearization systems. International Journal of Systems Science,
through state immersion of nonlinear systems admitting 23(9), 1475–1489.
Lie symmetries. Automatica, 45(8), 1873–1878. Tsinias, J. (1990). Further results on the observer design
Menini, L. and Tornambe, A. (2010a). Analysis and problem. Syst. & Control Lett., 14, 411–418.
observer design for the Bullard and Rikitake dynamos.
IET Control Theory and Applications, 4(8), 1353–1365.
Menini, L. and Tornambe, A. (2010b). Computation of
the real logarithm for a discrete-time nonlinear system.

48