New zero-input overflow stability proofs based on Lyapunov
theory
Citation for published version (APA):
Werter, M. J., & Ritzerfeld, J. H. F. (1989). New zero-input overflow stability proofs based on Lyapunov theory. In
Proc. ICASSP 89, Int. Conf. Acoustics, Speech, and Signal Processing, Glasgow, Scotland, 23-26 May 1989
(pp. 876-879)
Document status and date:
Published: 01/01/1989
Document Version:
Publisher’s PDF, also known as Version of Record (includes final page, issue and volume numbers)
Please check the document version of this publication:
• A submitted manuscript is the version of the article upon submission and before peer-review. There can be
important differences between the submitted version and the official published version of record. People
interested in the research are advised to contact the author for the final version of the publication, or visit the
DOI to the publisher's website.
• The final author version and the galley proof are versions of the publication after peer review.
• The final published version features the final layout of the paper including the volume, issue and page
numbers.
Link to publication
General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. • Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain
• You may freely distribute the URL identifying the publication in the public portal.
If the publication is distributed under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license above, please follow below link for the End User Agreement:
www.tue.nl/taverne Take down policy
If you believe that this document breaches copyright please contact us at: openaccess@tue.nl
providing details and we will investigate your claim.
D2.7
Mew zero-input overflow stability proof8 based cm Lyapunov theory
nichiel J . Werter and John H.F. Ritzerfeld
Department of Electrical Engineering Eindhoven University of Technology
Eindhoven, The Netherlands
ABSTRACT
In this paper we demonstrate some new proofs of sup-
pressing zero-input overflow oscillations in re- sive digital filters. These proofs are based on the second method of Lyapunov
For second-order digital filters with complex conju- gated poles the state describes a trajectory in the
phase plane, spiralling towards the origin, as long
as no overflow correction is applied. Following this
state signal an energy function can be defined,which is a natural candidate for a Lyapunov function. For the second-order direct form digital filter with a saturation characteristic this energy function is a Lyapunov function.
However, this function is not the only possible Lya- punov function of this filter. All energy functions with anenergy matrix that is diagonally dominant, guarantee zero-input stability, if a saturation characteristic is used for overflow correction. In this paper we determine the condition a general
second-order digital filters has to fulfil so that
there exists at least one energy function with a matrix, which is diagonally dominant.
1. Introduction
In this paper we demonstrate sope new proofs of sup-
pressing zero-input overflow oscillations in recur- sive digital filters, which have also been published
in [I]. These proofs are based on the second rethal
of L y e p m v , which starts with a properly chosen
energy function B(n), preferably in quadratic form:
B(n) = ,rT(n) .P.x(n). (1.1)
Without loss of generality, w e choose P syametrical:
PT = P, (1.2)
The Lyapunov theory d-ds that
E = aT.p.X
>
o
for allr
f0,
(1-3)so that B ( n ) = 0 implies ,r(n) =
0.
From (1.3) itfollows that P has to be positive definite.
Further, the system dynamics muat be such that if no overflow correction is applied, according to
~ ' ( n ) = A * ~ ( p l ) , (1.4)
the energy strictly decreases with increasing time:
S'(n)
<
B(n-1) for all n, (1.5)where B'(n) is the energy pertaining to the state
~ ' ( n ) . This inequality is satisfied if
P
-
AT*P.A is positive definite. (1.6)Every matrix P satisfying condition (1.21, (1.3) and
(1.6) defines an energy function E(n), which is a
candidate for a Lyapunov function. If moreover for one such a matrix a subsequent overflow correction
x ( n ) = F{Z'(JJ)} (1.7)
lcuers the energy for all possible states,
8 ( n ) 5 B'(n) for all n (1.8)
the function E(n) is called a L y e p m v function of
the nonlinear system under consideration.
The existence of a Lyapunov function in a digital filter guarantees freedom from zero-input overflow oscillations. The idealized linear filter is assumed
to be stable, so the state in this filter amymptoti-
cally approaches the zero-state, for which also the energy is zero. In the actual digital filter o v e r flow correction lowers the energy B(n), implying that also here the energy asymptotically reaches zero, implying zero-state, which completes the prove of zero-input stability.
Care must be taken if
"<"
is replaced by "I" in(1.5) so that the energy can remain constant. Such a
situation occurs for a marginal choice of the matrix
P, for which the energy function is called a semi-
L y a p m v function. If, moreover, the equality sign
in (1.8) applies it can occur that the energy re-
mains constant, associated with the risk of zero- input oscillations.
2. Overfla, stability in secolld-order filtera
In this section we investigate the overflow stabi-
lity of second-on3er d l g l t a l fllters. For sake of
conciseness, we restrict the following discussion to sections with complex poles. Compared with real poles, they generally favour all forms of parasitic oscillations (particularly for high Q-values) and thus deserve special consideration.
The 2 x 1 state vector x(n) = (x,,~,)~ in an auto-
nomous second-order system satisfies the fundmental difference equation
x ( n ) = FL4.x(wl)l, (2.1)
In this paper it is understood that
[F~A.III~ = F([A.all}, (2.3)
i.e. the individual components of A - x undergo the
same memoryless and local overflow correction.
876
The quastion to be analyzed is: under which circum-
stances (choice of A, F and ~ ( 0 ) ) does (or does not)
(2.1) admit periodic solutions?
Due to the overflow bound, which is henceforth n o r malized to unity, the state variables conform to
Ixt(n)( I 1 , (2.4)
resulting in a state vector confined to the interior
of the unit square (cf. Fig.2.1). Without overflow
(as long as (2.4) holds) the solution of (2.1) is
where q l , q l = eriJO denotes the ccmplex eigenvalues
of A and trij.ti denotes the pertinent eigenvectors.
It is tacitly assumed that
r <
0, express,ing linearstability. Further, the constants of integration
( X , P) are determined by the initial state ~ ( 0 ) .
If, for the time being, time n is viewed as a conti-
nuous variable, ~ ( n ) describes a trajectory in the
phase plane. For the case
r
= 0 this would be an el-lipse with main axes in the direction of tr and ti.
For
r <
0 (corresponding to poles inside the unitcircle), we obtain a non-closed, ellipselike curve
spiralling towards the origin, cf. Fig. 2.1.
Fig. 2.1. Trajectory of the state vector r(n)
in the state plane.
Of course, these results only apply to the digital
filter as long as overflow does not occur. If this
is not the case, the linearly determined x(n) leaves
the unit square, and overflow correction has to be applied. This correction introduces one of two basic
atate modifications: (a) x ( n ) is moved towards the
origin, (b) ~ ( n ) is roved away from the origin.
Case (a) is wanted because it supports the natural linear motion; no oscillation occurs if all over- flows are corrected this way.
Case (b) is dangerous, because it colpensates or even overcompensates the linear behaviour and, hence can (but need not) lead to oscillations. Of course. these statements ask for an unambiguous definition of "distance from the origin". Instead of the widely used Euclidean norm our definition is guided by the linear state motion, according to (2.5). Following
~ ( n ) = x ( n ) . [~.c~s(r(n))-~~.sin(r(n))] (2.6)
two variables I(n),f(n) can be associated with each
state ~ ( n ) . Particularly, the variable X(n) is de-
termined from ~ ( n ) as
Comparing (2.6) with (2.5) one recognizes
w(n) = X.ern, (2.8)
i.e. a monotonically decreasing function. Therefore, the function
R(n) = X2(n) (2.9)
is a natural candidate for a Lyapunov function. We
choose x ( n ) as the "distance from the origin".
Overflow correction is now visualized in Fig. 2.2.
An uncorrected state point B is mapped into B', B",
or B"' after applying saturation. zeroing and two's
complement, respectively. For this example all types
lead to an increase of E ( n ) and, hence, to a move-
ment away from the origin. On the other hand, for
point C this is only true for zeroing and two's com- plement overflow correction.
For some ellipse gecmetries it is possible to use appropriate overflow characteristics such that the
state always moves t-ds the origin and wcilla-
tions are suppressed.
Obviously this is not the case for the arbitrarily oriented ellipse of Fig. 2.2. However, it is easily
recognized that for an ellipse whose ax- coincide
with the xI-x.-axes, each of the three overflow cot-
rections satisfies the stability condition, while
for an ellipse with a 450 inclination stabilization
can be obtained at least with a saturation charar teristic.
*2
't
-'
I
Fig. 2.2. Ellipse X = constant in the state plane.
3. Ovema &ability in direct forr filters
In this section we investigate the zero-input
stability of the second-order dfrect form dlgltel
fllter with overflow correction. This filter is described by the syster matrix
It turns out to be free from zero-input overflow oscillations, if saturation is used. This is true for all pairs of filter coefficients a and b in the “stability triangle”, described by
l - l a ( - b > O
l + b
>
0. (3.2)The analytic proof of this statement has been re-
ported by Ebert e.a. [Z]. In this section we present
two novel proofs using Lyapunov theory, one proof only for complex conjugated poles, the other for all pairs a and b in the stability triangle.
For complex Conjugated pole pairs q , , q a = er*J’ we
define an energy function E(n), according to (2.10): (3.3) 2
E ( n ) = x,(n)-a.x,(n) .x,(n)-b-x;(n),
(3.4)
This energy function is characterized by the matrix (3.5) which is symmetrical. It is positive definite, due to
Det[P] = -b
-
a2/4 = ez.sinz(0)>
0 (3.6)and Tr[P] = 1
-
b = 1+
ez>
0. (3.7)The system dynamics is such that if no overflow cor- rection is applied, the energy decreases:
E’(n) = ez.E(n-l)
<
E(n-1). (3.8)In the nonlinear system, saturation causes an addit- ional energy reduction. If before correction we have xl’(n)>;, then after saturat’ion x,(n)=l. The compo-
nent xt ( n ) can never overflow since x,’(n)=x,(n-l).
So E ( n ) = 1
-
a.x,’(n)-
b.xZs2(n) (3.9) andE(n)-E’ ( n ) = -[xI’ (01-11 * [l+x,’ (n)-a.x,’(n)]
<
0 for ell n. (3.10)The latter inequality is due to the stability re-
quirement la
<
2. The same conclusion can be drawnE(n) is a Lyapunov function, which guarantees zero- input stability in the second-order direct form fil- ter with saturation and complex conjugated poles.
for values o
1
x,’(n)<
-1. Hence the energy functionThe zero-input stability of this filter can also be
proved with another Lyapunov function E(n), which is
characterized by the symnetrical matrix 1-b -a
= [-a 1-b]. (3.11)
Matrix P is positive definite for all pairs of coef-
ficients a and b in the stability triangle, due to
Oet[P] = (l+a-b).(l-a-b)
>
0and
Tr[Pl = 2.(1-b)
>
0.(3.12) (3.13) Without overflow correction the energy cannot in- crease, since
E’(n)-E(n-l) = -(l+b). [a*x,(n-l)+(b-l) .~,(n-l)]~
5 0 for all n. (3.14)
In the nonlinear system saturation causes a reduc-
tion of the energy. If xI’(n)
>
1 then after satura-tion a,(n) = 1 and with x,(n) = xa’(n):we have
E(n)-E’ (n)= -[xI’(n)-ll- [(l-b).(l+x,’(n))-2a.x,’(n)]
<
0 for all n. (3.15)The last inequality is a consequence of the
stability condition 1- a1 b
>
0. The same conclusioncan be drawn for x,’(n] <--1.
In a strict sense function E(n) is not a Lyapunov function since the energy can remain constant. This
situation can only appear if no overflow correction
is applied. But then the filter responds linearly and the state will asymptotically reach zero; no zero-input overflow oscillation is possible in the second-order direct form filter with saturation. This result is now proved for all pairs of filter coefficients a and b in the stability triangle.
4. Overflas stability in filtera with saturation It is a c o m o n property of normal, wave digital and lattice filters of second and higher order that the
”energy, matrix” P is diagonal and that E(n)=constant
are ellipses oriented parallel to the coordinate
axes. Only this ellipse georetry all- for all
overflow characteristics applied to the individual state variables, without risk of zero-input overflow oscillations.
The question arises: Which A matrices a d d t a diago-
nal ”energy matrix”? This question has been solved in [3] for the second-order filter, where it is shown that this is only possible if the system
matrix A fulfils the condition:
laII
-
a,,l<
1-
Det[A]. (4.1)If a saturation characteristic is used in a sewnd-
order digital filter the condition (4.1) for zero- input stablility (valid for all overflow character- istics) can be relaxed. We shall prove the following theorem (see also [41):
If E(n) = xT(n) .P.g(n), (4.2)
satisfies all conditions for an energy function (see Section 1) and if matrix P is diagonally dominant:
IPlZl I PI1 and lP1rl I Pra. (4.3)
then the function E ( n ) is a Lyapunov function of the
filter with saturation. Proof:
The difference in energy between the uncorrected
signal ~ ’ ( n ) and the actual state ~ ( n ) is
E’(n)-E(n) =
= pII * (XI *2-x7:)+pa,. (xz* 2-x,)+2p1,. (X1’Xt’-xIxr) 2
,2 2 2 2
1 P I , (XI -XI )+P*r(X,’ -Xr)-21P1,
I
(IXI’XZ’ l-lxlx.1)
+ l P . t l ~ ~ l ~ I ’ ~ - ~ x t ’ l ~ - ~ I ~ 1 ~ - ~ x z ~ ~
1 .
2 2 2 2 = ( P L 1- If I t1 ) .
(x,’
-x, )+(P.t-IP I t1).
(x*’ -xz) (4.4) 2 2878
The inequality is valid due to s@(~,')=sgn(X,) and
sgn(r,')=sgn(x,). If only one of the c y n e n t s
overflows, f.e. x,'
>
1 then x, = 1, x, = X, andE' (n)-E(n) 2
2
(PII-(PI.
1)
-(=1~2-1)+~p1, I.(~x,~-x,l -[1-x,I2) L 0 -(4.5)
If both components overflow then l x , l =
I X , ~
= 1 andE' (n)-B(n) 1 (PI ,-1plt
1)
(II *2-l)+(p,,-~p,I1)
(x2*'-1)+ IPI.I(IxI'l-lx*'
I)
1 0. (4.6)So in a system with an energy E(n) satisfying (4.3)
saturation causes an additional energy reduction.
Now we solve the question: Which second-order A
matrices posse88 at least one energy function E ( f l )
whose energy matrix P is diagonally dominant? Therefore the energy in the idealized system must strictly decrease, which will be satisfied if
P
-
A~.P.A is positive definite. (4.7)For a second-order system this condition is equiva-
lent to the inequalities
I Det[P
-
A''*P.A]>
0 and Tr[P-
AT.P.A]>
0. 2 where a = e 1 2 b = a,,' + a,,' -1 - Det2[A] nSince e1,-a,,
<
0, it is possible to choose para-R t e r p , , in such a way that the centre point of the
ellipse ( a - p , ,
,
p . p , , ) lies in the first quadrantof the p,,-p,,-plane.
For
lal.l
<
p?,!,
there exists some point ( p , & , p , = )within the e lipse (4.11) satisfying condition (4.3)
if the ellipse curves the line p . , = (pit(. Such a
point of intersection is found if
epl ,2+[b.sgn(p,,)+dlpI ,p.,+[c+e.san(~,,)+flp,,~
<
0has real solutions. So
(4.13)
[b.sgn(p,,)+d12 2 4a[c+e.sgn(pl,)+f1, (4.14)
or
[ ( 1+Det [AI 1 2- (a ,+a
,
,
)'1
*[ (l-Det [AI )2-(a, I - a , , - ~ , , s B n ( p , ,) 1
1
1 0 . (4.15)The first factor of (4.15) is positive due to stabi- lity conditions. The second factor shows the reein- ing condition which can be written in the form
2
1al1-aZ11
<
2 . l a , , l + 1-
Det[A]. (4.16) All matrices P that satisfy (4.15) and therefore (4.8) does also satisfy (4.9).The case
4
a?.I
Ia,,l is equivalent to the previouscase resu ting in the condition
~ a , , - a a z ~
<
Z . l a , , I + 1 - Det[Al. (4.17)The conditions (4.16) and (4.17) form together the
result of this section: All second-order digital
filters with a system matrix A satisfying:
Ia,,-a..I i? 2~min(~eI,~,~a,,J)+1-Det[A], (4.18)
possess at least one energy function E(n), with a matrix P satisfying (4.3), and thus are free from zero-input overflow oscillations for saturation. As it should be, this condition is less restrictive than (4.1). Contrary to the former condition, all stable direct form filter satisfy (4.18).
REFERENCES
[l] M.J. Werter, "Suppression of parasitic oscflla-
tions due t o overflow and quantization In recur-
sfve dfgital filters," Ph.D.-Thesis, Eindhoven
University of Technology, The Netherlands, 1989. [2] P.M. Ebert, J.B. Mazo and M.G. Taylor, "Overflow oscillations in recursive digital filters," Bell
Syst. Tech. J . , vol. 48, pp. 2999-3020, 1969. [31 W.L. Mills, C.T. Mullis and R.A. Roberts, "Di-
gital filter realizations without overflow os-
cillations," IEEE Trans. Acoust., Speech, Signal
Processing, vol. ASSP-26, pp. 334-338, 1978. [ 4 ] J.H.F. Ritzerfeld, "A condition for the overflow
stability of second-order digital filters that is satisfied by all scaled state-space structu- res using saturation," submitted to IEEE Trans.
C f r c u i t s S y s t .
For p , . = 0 the energy matrix P is diagonal which
has a solution for a system matrix A satisfying con-
dition (4.1). If condition (4.1) is & satisfied,
which can only be true for filters with a l , . a l l
<
0,we have according to (4.12) a discriminant which is
negative, with as result that (4.11) describes the
inner part of an elliptical curve in the p,,-pIz-
plane, which is non-empty, since R