AA!i' )'7 ASNHOOS DI R RIfCONIRO OF CONT INU PROC IJINO III AS I F N IN IVROS I NMA M fK I t A l INU A 'l II II 1 JJI IIAF SR TR83 0877 F4q620 82? 0080 1mohhmmhhhhhl mmmhhhmhhhl ImhomhohhiI ~I'll'.momo
AA!i' )'7 ASNHOOS DI R RIfCONIRO OF CONT INU PROCIJINO III AS I F N IN IVROS I NMA M fK I t A l
INU A 'l II II 1 JJI IIAF SR TR83 0877 F4q620 82? 0080
1mohhmmhhhhhlmmmhhhmhhhlImhomhohhiI~I'll'.momo
I m mi _ _ _ _ _----L- - W
• 4
+ ii
*La Uf02
-..
SI ijLI
,. ~ ICROCOP:Y RESOLUTION TEST CHAT
t ar ma Ma Or STmcaaos n-A
" - .. .. - ,--L.. --a "'i ':I
0
" 6 - it ! - . .S
,. .. . .. .. .. . ....4 -
kit-
SECURITY CLASSIFICATION OF TONS PA4E (tm m.. COW~S. ______________
REPOR DOCMENTATIO PAGE __________________
1jr'. REOR M5.f- OTACUS REMPICUT'S CATALOG NMBER
4. ITLE (d 8.MIej SL TPX OF REPORT & PCmo. CONVURED
Asynchronous Discrete Control of Annual Report7/l/82-6/30/8 3Continuous Processes ~PROSGOG EOTNME
,. umor6. CORACT" DO GRCT NSCR~.
Martin E. Kaliski _ _ _ _ _ _
Timothy L. Johnson F4 9620-82 -C-00 80
SPER? lUGN ORGANIZATION NAME AND ADDRESS go. Pr 6AELEMENZ PROJCT
Northeastern UniversityAmRKUITNDESAS360 Huntington Avenue0AFBoston, MA 02115 &3,___________
11. CONTROLLING OPPICZ "AMC AND AOONEM L 9PRTDT
Directorate of Math. & Info. Sciences 7L 30/83DTEAir Force Office of Scientific Research 98. NUMBER OF PAGES
ILS 1A MOIOIN NCY NAE ADDRESS(56 I..a ban .ua Om..ie IL. SECURITV CLASS. (of Oa. ,.p.,wj
Unclassified
Ickrts g~~FC ATI ON/ 00114 RDIN _
W. DSTRIOUTION STATEMNICT tot IN@ AasjpH) dfrpui.z'le
17. DISTRIOUTION STATEMENT to# 0v auWe. mgomE in..*te 20, It l aem be. RePWeO
IL. 1111PL6491TART NOTES
coding, automata theory, discrete control, switching theory,-feedback control
IF "ARAC V..u == u=aea. S6 mm'ie ~ebb rn N
This research concerns the analysis and synthesis of asynchronouidiscrete-state or hybrid-state feedback compensators forcontinuous or hybrid-state processes. New realisation theoriesfor asynchronous discrete sys team, based on automata indsemigroup theory, have been derived.- These theories suggest newarchitectures for asynchronous system
-D liWime. Mlos 9FI 10 as msr -O
____ ___ ___ __M I=l
ASYNCHRONOUS DISCRETE CONTROL OF CONTINUOUSPROCESSES
N.E. Kaliski and T.L. Johnson
July 1983 Accession ForNTIS iDTIC TABUnannoundQJustification
Distribution/Prepared by: Availability Codes nI
Northeastern University jAvail and/or It360 Huntington Avenue Dist SpecialBoston, MA 02115
Prepared fo r:
Mathematical and Information Sciences DirectorateAttfl: J. BurnsAir Force office of Scientific ResearchBolling Air Force BaseWashingt on D.C. 20332
AIR p~neumn3OwTI7I mSK (A7SO)
aW??v Id f Or rS1M,'.' 9 J AW AYR 19--12.
3,etiuo is nlmied
,un 3.
TABLE OF CONTENTS
Page
1 INTRODUCTION AND STATEMENT OF WORK I
2. RESEARCH PROGRESS 3
2.1 Analysis of Qualitative Properties of Asynchronous 3Hybrid Systems
2.2 Acceptance and Control for Asynchronous Hybrid 5Systems
2.3 Linguistic Approaches to Asynchronous System 8Modelling
2.4 Application to real-time Multitasking Systems; 8Stochastic Analysis
3. PUBLICATIONS 11
4. INTERACTIONS 13
7 7
: ' . .. . .
1. INTRODUCTION AND STATKM3NT OF WORK
This report summarizes the research undertaken during the
period from July 1, 1982 through June 39, 1983, the first year of
a three-year research program entitled *Asynchronous Discrete
Control of Continuous Processes*.
The research during this first contract year centered about
developing sound theoretical models for asynchronous systems,
models that extend and generalize previously developed research
of the co-investigators on synchronous discrete control. The
following sections of this report delineate this year's work.
The original proposal for this research described four
tasks:
(1.1) Analysis of Qualitative Properties ofAsynchronous Hybrid Systems
(1.2) Acceptance and Control for Asynchronous Hybrid
Systems
(1.3) Linguistic Approach to Asynchronous Systems
(1.4) Application to Real-time Multi-tasking Systems;Stochastic Analysis
These tasks are to be pursued in parallel over a 3-year period,
with the first two tasks receiving greatest emphasis during the
first year. The research involves a collaboration between 4Prof.
K.3. Kaliski of Northeastern University and Dr. T.L. Johnson of
Bolt Beranek and Newman Inc. Dr. Johnson has taken
responsibility for Tasks 1.1 and 1.4, while Prof. Kallski has
overseen Tasks 1.2 and 1.3.
..... .7
2. RESEARCH PROGRESS
2.1 Analysis of Qualitative Properties of Asynchronous HybridSystems
In order to focus on key issues in the realization of
asynchronous hybrid systems# the class of dynamic systems with
finite input, output and state sets -- termed Osimple
asynchronous machines* (SAR),-was treated during this year. The
time-variable was assumed to be real-valued. interesting results
were obtained under the mild assumption that the input is a
piecewise continuous (i.e., piecewise-constant) function. Uinder
these conditions, the semigroup properties can be used to show
that such system can be realized in terms of a finite set of
constant-input state-response functions that play a role
analogous to the impulse response functions of linear system
theory. Preliminary results reported in Johnson and Kaliski
(1983) show that in the timo-invariant case, such a rdalization
am be synthesised from ideal digital components similar to those
which ate, cmintclally available -- but in an architectural
cnfiguretion which i9 hoew.
2he key concept in deriving this new _relization is that
each input transition triggers a OcaspadeE ot subsequent state
transitions which Is pcedetermined by the machine structures this
cascade may be Overwitten by *ascades, due to subsequent input
tawemsints ot.I Sm~put tb*in beats a sequence of
mi..fte~ ~I~u tzIi~o tian.1arou~arohiteotura1
3
configurations have been described for generating the fundamental
state transition sequences; the applicable architecture depends
on the specializing assumptions which are introduced.
This realization theory overcomes a number of fundamental
problem associated with alternative representations. First, it
places the theory of asynchronous discrete systems in a clearIrelationship with the theory of semIgroups, which is almost
universally used in describing a wide variety of other classes of
dynamic systems: this. link has been-uissing from previous
results on asynchronous machines. Secondly# it successfully
resolves a problem which occurs if transition rates are bounded
by assumption -- i.e., that it is unreasonable to impose
constraints on input transition times that depend on internal
state dynamics. The ideal, mathematical reproestation is
well-defined even for unbounded or infinite transition rates; the
transition rate bounds arise ,only when the differences between
real and ideal 42inthesis components. are oonsidoco4. This is a
key requirement in the, corwsiderat ion of, genera; f eedback systems
composed of simple asynchronous mab.na sinoe a£q ack system
should itself be representable as a simple asynchronous machine I
For insth~ep unstable eedwak ste are well-detined and are
rakell"e so tuak x meahiihgl theory of stailty can be
#A Onticipstod. h tbeq "upps", tbeatuoor ajg slons
to.a boo CC p!beu"-&b ~ be p. v V04 1A ubsequent
774 -7j7~z--
reerh (1 th..- cae fhbi sae nu ndotu es
det mcdi (3) feedback system properties such as controllability,
observability and stability can be defined f or simple
asynchronous machines; (4) equivalent realizations which
partition the timing and data flow functions of asynchronous
machines can be sought.
From a practical Standpoint, the mwst significant
implication of the realization results is the possibility that a
general automatic synthesis procedure for aynchronous state
machines can be developed., Tbe theory swlqests a general
architecture which is significantly different from traditional
Von Neumann and Data Flow architectures, which require
considerable use of ad boo design methods by experienced
designers. A general astosistic Synthesis -procedure could have
major implications for 0eqmutor'aided ctscuit design and
achievable pert ormee at VX1SIC ebipe as well as Suggesting new
concepts f or the design of asynchronous feedback Systems. A
N~iatat* effort a£10.9 tbese limes Is anticipated.
2.2 AoPternee a Cstrol got £synchroaous sybtid System
Wfozta deaJa th1 Lust '0teract" Your:toolased on
dsvolop$ag *eraeWm ~e n to"*. for 4baacttatining
4P.
asynchronous coders. These coders, to interfaces between the
plant to be controlled, and the finite-state, controller itself,
are intrinsically hybrid devices which, as asynchronous devices,
react not only to input changes bust to the times at which these
changes occur. Thus models of coders explicitly incorporating
transition times were sought.
A mathematical abstraction of the asynchronous coding
process can be made by viewing such coders as mappings from a
subset of (RxR)+, the set of non-null, finite length sequences of
points in RxR (R the real numbers) in-to the binary alphabet
fo,1). This abstraction,, detailed in the attached informal
memoranduma Otwards a Theory of Finitary Asynchronous Codersm,
ad summarised ini the pap~er NTOlards a Theory of Asynchronous,
Real-Tim Codess and.Their Application* to IDiscrete-Control of
Continuous Pr9sesses', 1903 American Control Conferentce, allows
us to view asynchronous coders as special forms of their
synchronous counterparts -- forms defined over a limited subset
of allowable Input atuings. TVhis sabset cotisists of strings
whose second componsnts:(namely.tI-me), always increases.
Having made this abstraction, research turned to finding
meaningful ways to extend such *partially specified" maps to all
of RIM. so aS to permit our previously, developed theory of
synchronous coders to come into play. The memorandum cited above
dov4Upe am*, an. =esusiono , %eort' "or th* case at f initary
iS t*. )**deft: &no- jot a.( .' a ke )rCstY o1f such
codes - t~a~t stuct*A fip aoprty s ancbtrlflal onewhose isplications, as Of this vriting, are, still, beingresearched.
IY 11 tihl ' ji~ sequnces that form the domain ofasynchronousg codera may,# #e1r ~A-j.OCwge tsttmpig
onth tewonep,~d it 4t 00l1 aitural that -at tentionturned to cha~arss ~ al fivet Eittibutivou of theObreak points!, p~,a fthe.f ts Oftl research-- the breakpoints, after allp represent eve 'switcbtfig 'time".
We desired to develop nots general characterizations Ofbreakpoint 4tb ip t aple sodcl Le 19eE n hcited fnmranium above.
-suck a- ro4ia 4 ~ ero- fa to, it ad som e wasGownd to'-c r~thej by
no Stna ttWTIooai Atu**,46(Y Slwankamt) to develop
41 as *rbital b aio'and
5 ~ rnz~~rAJv Aflf
MWA6"In ipf womlitar tgem*a
A wit4
A .g4
MY
4414 ~ 51i4
l~ 1
Afi
AM=A
other side. The study of such timing functions is currently
being done.)
2.3 Linguistic Approaches to Asynchronous System Modelling
The above theoretical models will be used in subsequent
research to characterize wider classes of asynchronous coders.
We will attempt to extend the linguistic models that were so
successful in characterizing synchronous coders to the
asynchronous case.
2.4 Application to real-time Multitasking Systems; StochasticAnalysis
Only preliminary investigations of this topic have been
completed this year. The key issue of interest in multitasking
systems is the development of a fundamental model for process
synchronization; i.e. two concurrent processes may be represented
as parallel asynchronous machines connected by a communication
channel which may impose requirements for synchronization at
certain stages of a computation. A dual view of this problem is
to determine when a single process can be split into two
concurrent processes with minimal communications requirements.
The representation of the (asynchronous) Communications channel
is being viewed (in the deterministic case) from the perspective
of the realization theory of asynchronous machines.. As a
preliminary example, the partial realization of a UAM (univet"sal
asynchronous reeiver-transsitter) has been worked out this year.
Stochastic problems have not been considered this year.
However, it should be noted that the semigroup theory ofstochastic systems is well-developed, and this provides a point
of departure for developing a theory of stochastic asynchronous
machines via the results cited in Section 2.1.
I-
I -.
p!
3. PUBLICATIONS
The following works have been published and/or submitted for
publication this year.
Kaliski, N.E., *Towards a Theory of Finitary Asynchronous
Coders', Northeastern University Memo~randum, January 1983.
Kaliski, N.E., Kwankam, S.Y., and Halpern, P., 8A Theory of
Orbital Behavior in a Class of Nonlinear Systems: 'Chaos' and a
Signature-based Approachn, submitted to Intl-.3 .SysAt2M 5Sj±ja&,
February 1983.
Kaliski, N.E. and Wimpey, D.G., *Towards a Theory of
Asynchronous, Real-time Coders and Their Applications to Discrete
Control of Continuous Processess, Procg. Amerian CotlQ1
Cofrne San Francisco, CA, June 1.983, pp. 731-733.
Wimpey, D.G. and Johnson, T.L., "Finite-State Control of
Continuous-State Processes: The Discrete Time Case", £WL- 2.lLt
LBU nf.. 2a Dec±iin and Control, Orlando, FL, December 1982,
pp. 1230-1232.
Johnson, T.L. and Kaliski, N.E., *Realization Of
Asynchronous Finite-State N4achines*, S~ubmitted to I= Trans-
£Auto-. Contr. and 2=a Z=Z Cn. 2a R~a±Qn and Cnnkr.B, San
Antonio, TX, December 1983.
4. INTERACTIONS
Prof. Kaliski and Dr. Johnson met regularly every 1-2 weeks
to coordinate research progress during the year. Prof. D.G
wimpey of Northeastern University has continued to participate as
an unfunded collaborator in this line of research.
A proposal by BBN, Inc. to develop control system design
algorithms based on our earlier results for the discrete-time
case was submitted to Wright-Patterson AFB but not funded this
year for lack of program funds. A similar effort has since been
funded by the Army R&D Command.
Dr. Johnson has been involved in related research on modular
hierarchical (hybrid-state) control systems and has given a
number of presentations and lectures during the year on this
topic.
'I
!1
FIZ-STATE CONTROL Of CON UW-#ZTR lSaCuuSES !EZ 0!SZ TME CAW1
David c. ViapeyDeprtent of Slectrical Engineering
TmtyL. Jolmiecs
Laboratory f or Infomaetian ad DecisionSytmRom. 35-2053
Visseachusetta basttte of TcnlgCamridge. Mass. 02139
An lgoith fo th deig offinte-tade omsign.
penstor fo nolinw dscrte inisytemdn or Csk'S~kg
e isdevlope. Te prble toformlatd asa rs-- with ehe breslsi eT9 oIwee9i mulster ~ ~ ~ ~ f problem andth oum.iiteis I- Piiv Intgr
catesecfcaition opf: al A? and quantizattheleel an to exac folte sne the p lant t- Th taeea fth lut inghoiorpe ilatorrned. Tfhresoluthen plantestae zIolve problem oilr excet tandw tolg reushrefioite-rsguated agregon of theplan tadte deigno 1 sard h
of acontolle forthe esuling andeermiist celoledacheT witn hoe pann otmW ha ~ ~ ~ ~ ~ Nek rarget sto tCR
8 fac that gihe timesste Initial0- 1. orte sa b
straint~~tat In tixee reuatr orul bepnatr ia sthan itulto for Ptm(umust ~ ~ ~ ~ ~ col be finte-tat realtraable. Thn genpenralo iap1 so (3
t 1u thestrod - u tion , Pre ,i) wlaithoanit It~ prh e 0 omlt sas otetretrobe eateiltyt prolatof fotos syot exit
st plan to be coanrle to: ftp * o'. ait sethe planing desrion, ofuther cotditieneaP(Sir Tsmares req]). I hspprecnie
& ictarget sys e £ an (In. qetzeio levels f.urtherh-X
steofiin-t sap te f: isx l n ods u ny eino a weak regulator forr amcws Ceiuu Infalinies arslte: ll and fruanth Cntl state as a vdSeeiiaiso h
ifguter isrobte hT I suchl that forlant stat x isore ifarolem fruisn ad iteisming that thesolvn t euled-ro sow Intialsaegtmet s sltaedlard Teprobem os defntion I ebeue
callte to a synchru proisis gie bys f0~lt- aroa agt ot TC~ i s g~iven tie facTh00- lwothe oo wher aP tropaidtriiti c finit
ths reeac was N permed at t th aI? fiiesoy The prbe omlto seless otetrefore Infeseatiln peroblem for seocwittissystemspwo-
sieb the RAir ermOf lo n: NZPStfi a eseara The d dlescto of afne-tte uertmaitrife
Utesetorate~~~ of shJtca an"ov~i. lses ias soe aftat fte atre~ copnstr Is-( a arge met T04C-0and as Intlawsaes i thi muig mall iD oband De fquazat leels. a1 erpaptere do nat atee bil8reachen the aor mo e lo 5.1fa.14lsisfud it tomnI ra
oover-~ belationite artitionole Is deine by hevd euwLe1Il-f% ~k (r))Teps t l sli n borsm
relatIon I Tenm the agregata model *tata-ttretutionU" 6p: P z In pis dfned as 4 .7 - Q a
SPPoU) u (P C P- f(7'i. U)fpj 0I 01 for With U.: Wa ) a UE4 Pr.U)t p C ;AG(w)].
u 4 tP Fommao.11 1: The solution of
Thus the nondterejaistic finitestate system (see E61) Xkl "(k k8e aC?. 0. 8p) silat~ae Pin the following somae. If.7 i
the Initial stae. of 6 ia the equivalence class 1-.12 2 k~l N(k, "k' q(S(xk)))
i.e. the block containing %0.and the input sequence w witX Zl resultsin Z k e 3' o Al W-.
is applied to both 8 and P , then the aet of possible Vt htteqetsto o eed nyastates reached by 8 contais the equivalence class.O htte4fitzttf a eed nyo(sodulo 1) of the state reached by P. 6 end P. The mop; chowe explicitly the decompoeition
of eekFlate-tat Coposteof a. *Aleo, In arder to comput and q we require11I. EIwicenc ofWa iieSaaCmeitows thP fI~*h
Clary a necessary cndtio for the existtee taf d I be wle ake foninuo-us. eglto
0of a weak finite-state regulator for P sta th ftl eiu JeI~ijMi 2.0 if twea euite o rea to ma
tis conttSion In berm cotollagrae t del 8 we Ks 2~ x W -0 0L with the property that solutios of themycompute the attainability seti 3:sse
T A i p 9 P: 3u CRm i.e. 8p(p~u)CT 1 1) zk s-f% '(k' k q(g(Wk))))
k a 1. 2, .. kl- ~';( k' (g( k ),q'( k )
where To a, A9( c P: pCT)Is athe aggregate target for any Za C0 X w6LPresult isnrk£T, Otty.
set. Then if X. is covered by Tk for isow k, i.e.- enX cVjlp: p g T AJ), there exsts a sata-fedback law H than ha U n
for 8 (and hae P) which solves the finite horims(hy (h. I~h, q(y)), q(y))regulator problem. Since the etteof P die not di-rectly miurable to general, the @electios of the con- n~h~y) - 19(h. q(y)).trol Inputs to he a*"lied at emy Uim stop mt hebased on an estimoa of the state of P. To obtain a I.slto fteIfnt oio rbefinite-itate state-eatimetor we will flee agoregateIV ouoafgetlfeRrinPobmieate-estimntea i.e. estimtes of the state of 8. In practice, construction of the successor teen
In principle the, the design of H my he carried for ; from the root down Is highly Inefficieet. Aout by the construction of a tree of aggregate state- "bottom-up" procedure to desirable, but this requiresaicinotci of P. The root 7f the tree is the Aggregate knowledge of a cot of etate-easntes in Tifor uhichinita state set In A (Ex3s a P: x e X0 and the all successive itate-sntent will remain innodes are elementa of 2 . If Pk~is a successor node Esentially. thia involves the comptation of a steady-to node Pk. With the connecting are labelled (u.7), crate reduced target tube E33 which we denotethen in' the -a aggregate itate etimate of P t dto2:Dfngiven that the Plant output vae 7 sad then input a Wa 4ZaU I DeieT to beth ae sbtapplied. Of comm.e the area would hae to he labelled of 2 with the proeet lo ':T are st.i~eby subsets, Instead of elements. of 0 x 279, sin1e aX)f' n I P , there exists ucRe et. ',luav)sthere are only finitely smy distinct state-eatimages. 4VThe design Is completed by selecting a feedbck lAw An aloih for the computation of 4: is gieK: 27 z a - I3 which results in all poesible #adoi of in [3. TAUn asoih actlly competes a minimalthe tree having scecessive ades ohich ea eveudoly ro hci ' hmal&Iweys subes at oe of It-Pa, smlyfi thu I fPe ha a
The deeip of en (open-loop) fitoz-esate state- Set.- subsets of 1' are not Included iseatluster for T psoeed s Mumsa . tofiva em sqwiV- Ow phac beow composed. the tree for ';my bealmso relationSe & a e II& YS 2 46 (EXIS cps I(X U1) constructed an follow.,
(C1 p S(a) - 1) - oa y1 *72ifft eat oeeaggregate states that P may he in gives output yl ios+ Pidentical to the a"c gie 7 . am, 41WIM ai Wls lP, for OeN& waW at. GMu6i1' #*,49 IV - gP/W (A I) end a, V + 2: 1a Ctxjl-S iaSw 0161""exsts Wei gse. ;Mq's)e xThen the statsaesimsen to a (9.0I X3 20,49, 11'as 2P a If: it R 2P 7ndetedi29Ma f p A wea fiaift-stau regeLator H suatg
for 1, T, aud ifthr e SiAM a fliie partitionP
of such that XaC I' for some I' i ' and some non- (51 Wimpey, D.G.. "'Tite-State Cntrol of Discrete-
n i Tim Continuous Processe: An Automata Motivatednelat~e L~teler J.Approach-, XIt..T., Ph.D. Thmeis, Dept. of Else.
The ar e esentially attainability sets can- ZnJ. end Com. 3d., January 1962.
structed from T,. A detaLled algorithm for the coupu- [61 mesna", F.C.. Fine-tate Models or Loeical
tation of the slnay be found in [51. The tree speci- ! , .
fying ;. wb represeted as a grah. to the state- t71 Tou, J.T., Darr .ip of PI Control
transition graph for a system with Input set Re x V £sZ!_ = Academic Fress. N York, 1963.
i.e. each we is labelled as (V, w) for son VCI, [81 Wgs, P.K.C., "A Method of Approximating DynamicalProcesses by Finite-State System". 1st. 3. Control
w a V. Thus. corresponding to each P' a 1 + and each Voc.8,eo by pinit8-2t , 19a8.
wsuch that a(*)flP, # s. we hem a traitiosat V ol. 8, No. 3, pp. 285-29%, 1948.
Ithen if utV. (F*', u, v)uI . since there are only [91 Sorkar, V. and Varaiya, P., "Fnite Chain Approx.finitely wany sets V emy select one representative for a Continuous Stochastic Cnerl Problem",
ftm each and define I as K(]P'. v) , I.e. K has . AC-24, No. 2, Arl L11I.
finite range YC . The structure of H t shoam in [101 Sorneoshenmk, 1.&., "'Finite Automaton Approim-Figure 1. tice of Behaviour of Linear Stationary Continuous
Plants", Lut. and aem. Control. Io. 7, pp. 1092-Clearly, the existence of the regulator H (if it fl02, 1977.
exists at &UtI) depends eD the choice of P. Tatialchoices for P should be coarse to simplify the con-
troller. and refinements should be mde if H cannot befound. The design process is therefore iterative. Figure 1. Structure of Reaulaor H.Techniques for chooeing the Initial partition and forupdating or refining P are under investigation. Theseare complex problems which are problem dependent. Tosimplify the design, a heirarchical controller struc-ture has beon suggested [51.
V. Discussio. Conclusions q
We have described a n w approach to the design ofdynamic digital compensetors. These compensators arefinite-state and are therefore directly realizableusing digital logic circuitry. Our procedure speci-fies all quantization levels and takes all quantize-tion effects Into account.
Previous attempts to design finite-state control-lets for coatinuous-state systems Mro concerned withmemoryless switching controllers. These problems wore
often formuated as optimal control problems and thedesign often employed approximations which resulted inclosed-loop systems with performances that war diffi-cult (if not Impossible) to predict [4, 71.
The problem of quantitation (or finite-stateaggregation) of cotinuous-state dynamic systems isnot new (8, 91. Our approach follow mot closely theIdeas of Kornousbenko [101 and is related to the prob-le- of obtaining structure-preserving covers forfinite automata (6].
Reforences
[(1 Gatto, M., and Guardabasi, G., "The RegulatorTheory for Finite Automata", taf. and Control,Vol. 31, pp. 1-16, 1976.
[21 Sontag, S.D., "Nonlinear Regulation: The Plece-
wise Linear Approach", U Trm. AC-26, No. 2,April 1981.
(31 -emires, S., "Set Theoretic Control of LargeSeale Uncertain Systems", M.Z.T., Ph.D. Thesis,Dept. of glee. ISn. and Coup. Set., May 1977.
(41 tarden, J.S., "Optimal Feedback Characteristicsfrom Stochastic Automalon Models", Use.AC-lA, go. 1, February 1969.
A
Rea ation aL haini ka- a Mchinna
T.L. Johnson** N.H. KaliskiBBN, Inc. Northeastern University,1 Moulton Street 405 Dana HallCambridge# NA 12238 360 Huntington Avenue
Boston, MA 12115
Ph. 617/497-3413 Ph. 617/437-3034
A
Asynchronous finite-state machines accurately represent the jaction of computer control systems on continuous and discrete
processes. Despite their importance, the lack of a general
analytic representation for asynchronous machines has prevented
the development of adequate techniques for analysis and design of
finite-state asynchronous control systems# while ad hoc real-timecontrol software has proliferated. In this paper, a general
realization method is presented and is applied to a portion of a
universal asynchronous receiver/transmitter (UART).
*Research supported by APO Matbematical and Information
Sciences Directorate under Contract ?4ff20-82-C-008.
Preferred address for correspondence and return of proofs.
-..... . _ .. . . ... .-.-
,L,.
A simple real-time control program (e.g., one which
repeatedly polls a number of input lines, performs certain
logical tests and computations, and then deposits results in
several output registers) operates asynchronously. While such
programs undoubtedly characterize most computer control
applications in operation today, very little theory for therigorous analysis (let alone designi) of such systems is
currently available. In fact, technical standards for defining
the behavior of such systems, short of multichannel timing
diagrams, do not exist. Some consequences of thie situation areas followss (1) systems are designed by trial and errori (2)control software is not generally transferable from oneapplication to the nezt; (3) prior performance estimates of
proposed systems cannot be obtainedp (4) system functions cannot
be efficiently documented, short of providing complete
hardware/software descriptions. Clearly, these consequences are
undesirable.
A. number of strategies for avoiding these problems have
evolved, but none are very adequate. The practitioner, if he is
even aware of the preceding problems, is generally quite contentwith the status quo. Trial-and-error debugging of softvare is
fast and relatively inexpensive compared to analytical design
procedures. For simple systems, this approach indeed works well;
but for large systems (e.g., the U.S. Space Shuttle [1]), thecomplexity becomes overwhelming, and even exhaustive simulationcan fail to reveal subtle timing errors. The response of thetraditional deedeatc eastrol designer, by coatrast, is to forcethe system to operate synabronously (or with mltirate sampling),and to recuse logLcal operations (as distinct from real-nuaber
"- " "1
1.:. n"....i I -
operations) to an absolute minimum. While this approach is
usually 30re amenable to analysis and documentation, it forcesthe designer to overlook certain attractive control strategiesand forces the processor to operate inefficiently; thus,performance may be significantly compromised [2].
The realization problem for asynchronous finite-statemachines is a first step in the development of analytical methods
for the design of asynchronous computer control systems. Thaprimary issues of in &rat are todefine a natural set of
primitives that can ba used Ito haracteri.. the hehaviior of
asynhronous discreto-State nyste. and to illustrate a aneral
synthesis graodutra whereby speCIfication of the primitives leadaI - to a realistic imlenaon in previous research of the
authors and a colleague [31, these objectives have beensubstantially achieved for discrete synhannum systems. Inanyncghn = systems, the actual timing of events may effect theevolution of states, whereas in synchronous systems,, only thechronological ordering of events is important (i.e.,*synchronous* does not necessarily imply a uniform samplingrate).
several simplifications are made here in order to isolatethe main Issues: (1) only the purely finite-state case isconsidered, although the technical framework admits an extensionto hybrid-state systems;g (2) feedback interconnection of systemsIs not considered# though with certaibn additional conditions, theissue of closodness under feedback could be addressed; (3)existence of invariantsp equivalent representationa, and
* miniality are not treated here. While the present results arethus Incomplete, they do foaus on pragmatic Issues of Interest.
2
II. Background and Notation
Conceptual models for asynchronous state machines have been
employed primarily in automata theory, switching and
pomnications theory, and digital circuit design. Kohavi [4
provides a reasonable overview of the available methods and
technical Ksues, while Unger [5] Miller [6] have given detailed
design exatples for such syste. Petri nets 171 have als been
used to represent the behavior of asynchronous machines, though
they are often impractical for large systems. Kalman, Falb and
Arbib [8] have developed a conceptual framework for mathematical
systems thery which includes both automata theory and dynamic
systems theory, but this is based primarily on formal analogiesbetween existing theories; they did not explore the specific
properties of hybrid systems, and in particular discrete-state,
continuous time systems. In 191, an axiomatic framework for therepresentation of hybrid-state systems was introduced, and the
special case of asynchronous machines was briefly discussed.
The key technical issue in developing a representation for
asynchronous machines, aside from the choice of state variables,
is essentially one of existence of solutions. Two aspects of
this problem are: (1) On multiple input lines, transitions can
occur at arbitrarily close tines, which may lead to indeterminacy
in the state transition function. (2) Even with constant Inputs,
the state may switch at a rate which increases so rapidly with
tims that its value in not continuable with respect to time.
Both of these phenomena suggest the need to limit the maxiaum
Iwitchitig rate, or to impose a minimum switching delay time oan
the system. The avoidance of hazards and races is also an
important practical design consideration for switching circuits
that is closely related to these matbematical difficulties. it
an asynchronous circuit is not'designed to avoid these problems,
its actual behavior may be indeterminate In the sense that
- _ _ _ _ _ _ l- --i i I I
successive trials with identical" inputs produce different
results, i.e., the behavior depends on details of design and
implementation which are impractical to model. Limit cycle
oscillations, one manifestation of the second problem above, may
also occur in actual circuits, and their detailed structure is
again implementation-dependent. These phenomena are almost
always undesirable in practice; however, they are inherent in the
discontinuous nature of the systems under study. In the next
section, a further discussion is given of the way in which these
phenomena are represented.
The general setting for these results is dynamic systems
theory [l]. Only a slight generalization of the usual
definition of a dynamic system is required for the current
problem, due to the fact that the input and output sets admit
only a discrete topology (8, pp. 163-1641. The general notation
is introduced here, and is specialized in the following section.
nafIn~ton 2.1: A continuous-time ftnamig myjELt, E , on an
open interval TCR, consists of
U an InLt x",u an LAht AVm of functions u:?-G,Y an mLti MIk,I an gUiak UmqW of functions ytT-Y,
X a AtAls t,a Ltati M-W js TzxTxUzX, and a xjadan" uprsT&lUxX-. The State transition map is assumed to satisfy the
following semigroup axioms
i) IdNAuUk. JCt,t,u(.),zo)=z o for all tT, u(.)eU, zoeX.
(ii) £suinQ A" For any elemente tegtel, toct, and any.zoCe,
if ul(T)ft2(T) , toii the & o
4
(III) !zanitxitx. For any t 2 ltllto, all elements of T,any U.)CU, and any xd X, Jd(t 21to1u(-),xo).
The dynamic system Em(U#U,YvV,X,r) is thought of as thepair of equations
x(t) - 0(t't0,'u(.),X(tofl
for t~t0, both in T.
in this section, the properties of asynchronous machines# asa subclass of dynamic systems, are examined, and correspondingideal circuit realization results are given. some preliminarynotions are the following: notation follows that of the previoussection.
D5LD±ifll3.,Lt: A gartIal fungtion of finite rgnga (PlPO?)Is a partial function zi T-, where T4kR andZum(zliz 2 ... ,zq),qk.l is a finite set. A fUnalao gkf finkto rangnMOMl is a PrOrR such that T.T (Overbar denotes closure withrespect to the set operations on It).
h~iJ~tM .41 A point te! is a point ag- Ai M duai ofa VOn xsa1Z# with qJ2 r if z takes, oa mJtle. values 4- s everyopen set in T containing t. (Notes Por qwsl, points ofdiscontinuity cant xst).ae
~LtDL M =tI A VOR $s142, is a mnraI& uEkma
pMA.SM f Wor *I'0Mr'~h'eli24i~ tV4 r
max.s9rA:j.i~,mt epc oth aoge*aaoo
3~ 4Las. A 3011 a :iw*Z with a'ooutitable Wamb~iv of Points Of&ts~nkpRJtyig evogy fluite ,*ubintrvol of I is mewrakbl*.
ibis fellows frcm prqorties -of lao u'eawSable #ts.
Io4R~n, 3.,A: A V(WARt' uitk, a fir4t&number of pintsOf 4iscoentinuaity in every finite subinterval of T in termed a
ammzev aon4inUMi -f nnetL-W' at& U&J"~ -XMM' PCVMU).
11r'k _=.: if 2. is a PCLFOFR~o It Is. measurable.. in suary,EUVW60f 'J CWV gff 4:;- A.P1t tebrait iilbluiiaon).
The special class of dynamic systems of tttt'her*eL*s tUeclass of simple asnchr6noum machinles.
M~iLLBR I: A ginL a~rau uaakW~ IBM) is a
u-{11 *.,~*,sW a finite integerus u1r*)t u isI Cight-60tipousj
to. atftr that tbeir z at* *"-*m got al1
toas %9.txA MYP~
a..i. *LAq1M~W
These formal definitions are of interest only to the extent
that the functions j6 and r can be simplified, which is next shown
to be the case. This is analogous to the expression of the
transition mapping in terms of the transition matrix, for linear
systems, and the expression of this in turn as a matrix of
impulse-response functions that can be written in terms of a
finite set of parameters for linear time-invariant systems.
A SAN can be thought of as a system
x(t) - O(t,toU('),Xo)EX ; t>t O (3.1)
y(t) - r(t,toU(t),x(t))EY; t,toeT (3.2)
Since uPCFOFR and [t,t O] is a finite interval, u has a finiteset of points of discontinuity, denoted tl,...,t3 , such that
to<tl<t 2 ...<tN<t andUoCU tol T < t (3.3)
u(T) du, U, tij T < ti+1 1 i{l,N}(by right continuity)
UN U, tN< T < t
and furthermore, ui+l#ui.* (The degenerate cases m-1 and/or N-1
are not written out explicitly here]. By the causality property,
0 in (3.1) depends only on u(.) in the interval [t,t o ] . By the
semigroup property of %,
x(t) - 0(ttNuM,¢(tNtluMl,0(.... ,¢(tiltooXo) ... )
y(t) - r(t, touN,x(t)) (3.4)
**he case of a single input line with multiple values isconsidered here. for multiple linea, it is necessary todistinguish which input line has changed.
7[
i
where (by slight abuse of notation), uo...uN represent constant
functions of values uo...u N , respectively, on the intervals in
question.
From (3.4), it is apparent that the fundamental solutionp(t, T,U,X): TxTxUX ; t>i~to0; t,Te T
is sufficient to define the transition mapping. This function isdefined for a finite t of possible input-state combinations,and hence admits a sum-of-products (mini sum) decomposition inthe form
0(t,T,u,x) = V [0.. (t,T) (uAui)A(xAxJ)] (3*5)i~j il, .. ,m;j-l,...n
where V is the *or" operator and A is the "andu operator.
Furthermore, the range of lj is finite for each i,j, and
hence this function may be written as k0ij(t,T) = V xk I(t,Ti (T)) (3.6)we k-1
where T j(T)-{CTaeTIjlj(a,:i-xk), and V is the characteristic
function defined as
1ttETJC0 { otherwise
Furthermore, by definition the sets Tij(T) have the completeness
and orthogonality properties that for each Trt o
nU Tk (T) ={aT,aeT1 - T-(tT)
Tk.(T)IT"j (T) - (the empty set)
The readout map may be similarly specifiedt
(3.7)r(t, ToUX0 = V [rij(t,T) (uAui)A(xAxJl) i-1...M,
- . .J -,... n.
8 l
P k k kkaTlrij(t,r) kr.(t,T) = V y *(tsi (T)) ; S.. =yT Y
k=l i.
(3.8)
Summarizing these results for time-varying SAM's, one obtains
Theorem 3.11: Any SAM is completely characterized bycomplete mutually orthogonal subsets TiilT) and Sijlr),
parameterized by T and defined on T-(toT) where to is theinfimal element of T. These sets have the followinginterpretation: T~j(T) is the set of times in T that the statetakes on value xk when the system starts in state x at timeT and the constant input u=ui is applied. Si (T) is the set oftimes in T that the output takes on value when the system
starts at time T and the final state and input have values xi and
u1.
In general, in the time-invariant case, it is well-inot*that 0(tT,u(.),x) depends only on the difference t-T, and thatr(t,T,u(t),x(t)) depends only on u(t) and x(t). The implicationsof the time-invariant assumption for the finite-state case are
clear from the preceding constructions.
oroll&u. ,i: Any time-invariant SAN is completelycharacterized by the complete mutually orthogonal subsets T j, of
T; and the indicator function 8 . These sets have the following
interpretation: Ttj is the set of times in T, relative to an
initial time tYCT that the state takes on value xk when the
system starts in state xJ at time to and the constant input usu1
Lis applied. SOj takes value 1 if the current output value
y occurs whenever the current state takes value 0i and the
current input takes, value u1, and Is zero otherwise.
At this juncture, it should be noted that the assumptions
___1---.
about the input space allowed 6 (and r) to be defined only on theset of all finite partitions of TxT (e.g., all possible Onetsu
that cover the plane); but this is all that is required.
Secondly, it should be noted that the definition of a dynamic
system does not specify any space of Ostate-functionsa. In fact,
the sets Tk in Corollary 3 may be almost arbitrarily ill-
behaved. This means, in turn, that the rate of switching between
different state values may actually be infinite. This situation
exists because no continuity conditions have been imposed on 0 or
r. However, the assumption that the output be piecewise
continuous does imply that certain compatibility conditions exist
between the state transition and readout mappings.
C 3.2: When a SAM is characterized as in Theorem
3.1 or Corollary 3.1, certain compatibility conditions among the
sets T~i and Skj must exist in order to assure that the input-
output mapping Wtoxo:UY defined by~r(trtorU(t) wx(t) ) a r(ttto'U(t) wA(t'ttoU(-)'Xo))
takes U into (piecewise continuous functions) y for all xoEX. A
necessary condition (in addition to completeness and
orthogonality) for satisfaction of these conditions is that the
sets Tj(T) and Stj(T) have a finite number of boundary points in
every finite subinterval of T-IteT] for each TET, in Theorem
3.1; or that the sets Tfj have a finite number of boundary points
in every finite subinterval of T. in Corollary 3.1.
The necessary conditions of the Corollary can be established
by noting that they imply that the state function x(t) isS pieeise continuous In time and that the readout sap preserves
piecewise-€ontinuity (in the tilme-varying case). These
conditions are far from sufficient, thought -the readout map may
Omamk* sequences of states which occur at Infinite rates in the
ifi
-9
general case, so that *irregularity" in the state-transition
function is compensated by *regularity* in the readout map.
For Corollary 3.1, it is apparent that the most interesting
realization results will be obtained in the time-invariant case,
and hence only time-invariant SANs will be considered in the
sequel. A useful insight in provided by (3.1)z each input
transition triggers a pattern of state transitions (depending on
what "state" the system is in when it occurs) which persists only
until the next input transition. The output thus may be viewed
as resulting from a sequence of truncated cascades, each
truncated cascade being drawn from a finite (but possibly large)
set of waveforms. The definition of the output space for a SAN
in such that the output can only make a transition whenever
either the *state* or input makes a transition.
In considering circuit realization and synthesis issues, it
is desirable to consider the subset of SANs which have piecewise-
continuous state functions.
DmtLLnII±LI:3. A SAN is termed xsgulu if for all torT,
u(.)CU and z0CX, x(t) as defined by (3.1) is a PCFOFR when viewed
as a function from T to X.
Necessary conditions for a SAM to be regular have been
stated in Corollary 3.2. The ideal circuit realization for
regular SANs requires that certain idealized elementary building-
blocks be defined.
naLaLot£laa L; A SWigiJa ,,,lpilaw-r (DNUX) is a
memoryless element witb addeses lines, al...aq , where input aican assume n1 discrete values (aij,"1,.e..,ni), nlqI . Ini digitalinput lines labelled va..a each taking no values, and one
output line, w, taking no valuls, defined such that
, 11 .
bq
--- -- --- -- ----- --
w (t) =Wa~~2t.. qt t1 a W.. qW(3.9)
Renark 3J.: An (ideal) digital multiplexor can be synthesizedfrom (ideal) digital switches.
D.8inition : A digital function generator (DFG) for thefunction f:T Z, where TC-R is an open interval and Z-{zl...z q} isa finite set, has a binary-valued input line, b, and a q-valuedoutput line, z, defined such that
z(t) = f(t-T)
where T is the time of the most recent *- l transition on b. }Remark 3.J: Formally, the set of all discontinuity points ofb(t) is partitioned into times {T9 I } and {TlMJ, and the supremumof { 3il} (--,t) is defined to be T.
Dalnition 3.,: An ideal trigger element (IT)it fox thetransition ij9. on the input line v taking values vl...v q has abinary output b(t) such that
'1 for all t such that v(t-) = V(T) - V and
b(t) v v(t + r v(T) = vZ
0 otherwise
An ideal transition &lement (IT) with input line v detects alldiscontinuity points of v (i.e., its output is the union of allideal trigger elements for that input line).
3mak j: An ideal transition element is defined for allveFOFRI, while an ideal trigger element may be only defined forvcPCOI'R. A one-shot circuit approximates an ideal 1*1transition element with a binary Input line.
.12
3/nJiitin : A logical transformation (LT) for the
memoryless transformation G: V1Z where V-(v 1 ...vs ) andZ-zl... zq} are finite sets with input line v and output line z,
is defined such that
z(t) = Gv(t)
for (almost) all tET.
Symbols for these elements are shown in Figure 1. From thediscussion preceding Theorem 3.1 and the remarks prior to
Corollary 3.2, it is apparent that a SAM may be realized as shown
in Figure 2. Every input transition initializes the function
generators for Oij(.,*) as defined in (3.5), and the multiplexor
performs the logical operations (uAui)A(xAxJ) to select the
appropriate function. The transformation 8 in Figure 2 is
defined as in Corollary 3.1, based on the more general expression
y V V V y Si(uAu)A(xAx)i-1 Jul 1=1 (3.16)
The interconnection of elements in Figure 2 has the effect of
computing the transition mapping using the semigroup property
according to (3.1). These results are restated as follows.
Theorm 2: A time-invariant SAM satisfying the conditions
of Corollary 3.1 may be realized by one IT, an DFG's and one LT
as shown in Figure 2.*
The nature of the definitions of the ideal elements is such
that regularity is not required in Theorem 3.2. However, in
practice, actual circuit building blocks corresponding to these
*One inPut line is assumed) if there are maltIple input lines,one IT element for each line is required (see example).
13
1qDMUX FM 4z
ni
a, q
a) digital multiplexor b) digital functiongenerator
vT1 b v LF (G) z
c) ideal trigger element d) logical transformation
FIG. 1. IDEAL CIRCUIT ELEKENTS
U- LF
• .. U** _.-.----yOPO to,,)
FiG. 2. UALIZATI01 OF llU ABNCYEROUMO NMUAII (SA)
t -- -. i Ji4
ideal elements have built-in switching delays which limit their
performance. The nature of these limitations becomes more
apparent when the problem of synthesizing the DFG's in Figure 2
is considered. In addressing this problem, the relationship of
the present theory with traditional state-transition based
approaches will also become apparentl in fact, these approaches
constitute a rather special case of the foregoing theory. This
process requires a few more definitions (see Figure 3)1
De. nitln 3ii: An ideal resettable intearator (RI) has a
binary-valued input b and a real-valued output, h defined by
h(t) - t-tb
where tb is the time of the most recent 061 transition on b.
f inition 3.12: An ideal compurator (C) is a memoryless
element with two real inputs h1 and h2 and a binary output, b,
defined such that
b(t) - 1 whenever h 1 (t) > h 2 (t)
0 otherwise
Dh inition 2.: An ideal hybrid Stack (RS) with data
sequences {hk} and {Zk], k-,l,... has two binary inputs, b1 and
b2 , one real output h, and one discrete output zeZ, a finite set.
These are related as follows
h(t) = hkERz(t) - zkCZ
where k is the number of 0-4 transitions of b2 since the most
recent 9+l transition of bl, i.e., b1 resets the stack pointer
and b2 increments the stack pointer.
mMK 3A s ft6 Ideal hybrid stack can be apOrOzimated by atandem Interconnection of very long analog and digital shift
registers.
.. ..
b f h ,'
RI
a) resattable integrator b) comparatorb2
ih*KS0
c) hybrid stack element
FIG. 3. MORE IDEAL CIRCUIT ELEMENTS
(!13-I~hRAN CSI)
Corollary 3.3: If the conditions of Theorem 3.2 hold and
the SAM is regular, it may be represented as in Figure 2, with
the DFG's synthesized as in Figure 4.
Since the SAN is regular, the number of transitions of any
Oij(.,O} in Figure 2 is finite for any finite interval and
countable on [,0). Let fhk} denote the sequence of
intertransition times and let [zk) denote the sequence of output
values (elements of X) of Alj(.uI). These are stored in the HS
of Figure 4. The values zk are delivered at the appropriate
times by the clocking circuit consisting of the RI, C and IT,
while b1 (the input shown on Figure 2) correctly resets the
circuit whenever a new input value occurs.
The obvious limitation of Corollary 3.3 is that the stacks
(in general) have infinite memory requirements. However, there
is great variety of interesting special cases in which the stack
elements can be finitely generated. For instance, if the number
of transitions for each j is finite, each corresponding hybrid
stack can be terminated (formally) with an element hxmD, fKtas.
Upon reaching this element, the comparator output of Figure 4
will never change state, and hence the output z. will remain
unless and until b1 is tetriggered. In this case, the hybrid
stacks can be replaced with hybrid shift registers. A more
interesting, and quite general, special case is where the
elements of the stack may be recursively generated. This case
corresponds to the one considered in [9J and (111. Two final
definitions art required for this case.
n3.Utt An ideal real recursive function (Mli)
generator for the function ftR*R with kernel hdR has two binary
input lines bl and b2 and one real-valued output line h. These
are related as follows
L -.
h(t) .fk(ho)
where k is the number of 0+11 transitions of b2 since the most
recent On1 transition of b1 , and fk(.) denotes the k-th iterate
of f. Thus, bi serves as a reset line and b2 causes successive
iterates to be generated.
maxk 3.1: The properties of iterates of certain functions have
been studied by Klein and Kaliski [12]; they point out relations
to ergodic theory and coding theory.
/±flJ IL'±2/ jal: An ideal digital recursive function
generator (DRF) for the function g:X'*Z with kernelzoz,:m{zl...zgj, has two binary input lines b1 and b2 , and one
discrete output ztZ. These are related as followsz(t)-gk(zo)
where k is the number of I-l transitions on b2 since the most
recent I-l transition of b1 , and gk(.) denotes the k-th iterate
of g.
Bauazah IJI: In Definition 3.15, since Z is a finite set, gK.g,
for some finite K which is bounded by q2 . Thus, in the absence
of resets, the sequence of values assumed by z in a DRF is
necessarily periodic.
SCaioaa LA. There exist subsets of the regular SANs
which admit a finite parameterization. Onesuch subset is termed
the inaMnden y raeurmivlv Senaramted SAN's (IRQ8AN). These
are characterized by the existence of independent functional
recursions for the stack elements in Corollary 3.3 (Figure 4).
The realisation for these systms is sbown in Figure 6.
|1
RRF(Ifho) h DRFMg, zo) z
a) r4i -r§ursive function b) digital rencursive functiongenerator g0nerator
FIG. 5. COMPUTATIONAL ELEMNTSba/OR
C I
rZG. 6 . RMZXAt~c U3PI RUU IV3L GRIERATED
As an example, one mode of operation of the Signetics 2651
Programmable Communications Interface tm (PCI) is considered [13].
This is a universal synchronous/asynchronous data communications
controller chip which accepts programmed instructions from a
microprocessor and supports several serial communications
disciplines, both synchronous and asynchronous. It serializes
parallel data characters output from the microprocessor for
transmission and can simultaneously receive serial data and
convert it to parallel data for input to the microprocessor. It
is packaged in a 28-pin DIP. A block diagram of the PCI is shown
in Figure 7.
Since the PCI has 11 8-bit binary user-accessible registers,
a rough bound on the number of possible states is 288. In the
design of such devices, it is common to employ non-minimal(l)
realizations to simplify the internal logic. As a special case,
consider the lCZ 9onfigursd for as an asynchronous transmitter
for 5-bit haractets with no parity and two stop bits. In this
mod, the input lines can be taken as
?TC ezternal transmitter clock (pin 9)T = 8transmit enable (command register, bit 0)al chip enable (pin 11)
and the relevant outputs are
TxR0 1 transmitter ready (pin 15, and status registerbit I)
Tx=T transmitter empty (pin 18, and status register,bit 2)
TD transmitted data (pin 19)
In this mode, CR causes parallel data to be clocked into the
2L
a.w
oumo.
now7 WE I3KO IGUIB25 PORKALcOKWIAIOS I!3FCE(CI.KDIID RM 11
'moIt.wim31
"F-
transmit data holding register (THR). When TxEN is asserted by
programming the command register (and CTS is low), transmission
will commence and TIREN will be asserted (low). Normally, the
microprocessor will test this status bit, deposit new parallel
data and pulse CE, which resets TIRM (high). However, if new
data has not been received from the microprocessor by the timethe last data bit is being set, T;ERT is asserted (low) andmarking bits are transmitted at the output. If data is
subsequently received from the microprocessor, the end of the CE
pulse resets TjERT and TjO5; as soon as the data is clocked into
the THR, then TRDY is reset (low) and the new data istransmitted, etc.
For convenience, the state set may be chosen from the timingdiagram (Figure 8), to consist of states A, SlS 2 ,S3,S4, S5 ,BC,and D, where
A - start bit transmissionS1 - Do transmission--- DI transmissionS3 = D2 transmissionS4~ D3 transmission,S D, transmissionB5 sop bit 1 transmissionC stop bit 2 transmissionD = mark transmission
Thus there are a total of (2)3 3 8 input values and 9 statevalues, for a possible maximum of 8 x 9 = 72 state transition
functions to be generated. However, most of these are trivial.
Two examples will be considered: (1) the type of transition
occurring between A and 81, and (2) a transition between D and A
after !IR! has been recently cleared by a data transmission.
CaMa IJI: This case is defined by an input transition to
22
-- - - --- -- tmmm m
Im
TxRDY, TxEMT
ITeD° --°-' 1 -""" ' a' -3 4""0 A "-2 a'
BY S-IGETICS 2651 PCI. (5-BI CHRCES NO PRT,2
o p ts FI, TD (9,) h c anbe a
PIG. S. TIMING DIAGRAM FOR SYNCRONOUS CHARACTER TRANSMISSIONBY SIGNETICS 2651 PCI (5-BIT CHARACTERS, No PARITY, 2STOP BITS). MODIFIED PROM 113]
(TjC,TxEN, E) (0,0,) when the state is A. In this case there
is an almost immedpate (650 nsec) transition to state S1 , and theoutputs are (TxfT, TrT, TXD) - (,ilDo) .This can be realizedby the one-state timing circuit of Figure 6 which switches frob Ato h at 95o nsec after the transition.
Cain tims This cae is defined by an Input transition from
to (0,1,1) in state D, as hown at the end of the timingdiagram. In this case, there is a significant delay while the
holding register accepts the data and transfers it to the .
transmitter shift register for serialization. Thus the state
transition function shifts from D to A after & time which may be
note than one clock cycle. The (current) output is (,,)
This time delay may be realized as in Case (1), however one
22
should also proceed to consider the situation after the next
clock cycle, as shown in the timing diagram, which illustrates
the difference between a single input line having 8 values and 3
input lines having 2 values. For a single input line having 8
values, every new clock transition would reset this function
generator, which is incorrect in the present case. with 3 input
lines, only a transition on the CK line will reset this function
generator, while transitions on the other input lines (in
particular, T; C) will leave it unaffected. (See footnote follow-ing (3.3).)
Most of the other state transitions are well-approximated as
instantaneous, as in Case (1), and admit simplified realizations
when this approximation is acceptable. The state transition
functions are simple in this mode of operation because there is
an external clock, i.e., generally, state transitions cannot
occur without a clock transition, which is by definition an input
transition. The state transition functions are defined for
nstagt inputs, which is equivalent to predicting what happens
if the clock sta following the input transitionl The PCI
admits another mode of operation, where the transmitter clock is
taken from an internal baud rate generator. In this case, the
state transition functions for constant inputs would switch
spontaneously as dictated by the internal clock, and the
realization would contain an internal clocking mechanism that was
still practical to implement as described in Corollary 3.4.
_ _ _ _ 24
The realization theory presented in Section 3 is based on
semigroup theory and thus, unlike many common asynchronous
machine representations, allows one to draw on a wealth ot
intuition about dynamic system theory. In fact, only the
simplest case has been presented here: generalizations to
hybrid-state and pulse modulated systems are clearly possible,
for example. The results have some favorable properties wnich
one would not necessarily expect: it is relatively easy to
identify candidates for state variables for representative
applications, and furthermore, the functions (ij) required for
the realization may often be readily determined from experiments
on model systems. Finally, there is a significant class of
systems which can be accurately realized by a finite number of
elements and which correspond to practical digital devices. The
theory provides a rigorous basis for the delineation of Ocontrol"
(sequencing, and/or timing elements), and computational or "data
flow" elements.
References
1. AILUsLon N*k A EAAR TachnoLOUs Nov. 91, 1981,"2Shttla Launch Delay Attributed to Oil. Boftwareftablemis, pp. 21-21.
2. young# R.D., and Ewatny, B.G., rormmlation and Dynamicbehavior of a Variable Btructure Bevomaecanisd"n, Pnac.L22 nnt.m A&IL Cantzl, Cwd., Paper 3P-3,. (June,
Srlotteaville, VA.)
3. Wiapey, D.G., Johnson, ?.L., and Kaliski, N.S.,*Real nation of A/D and D/A Coders", Izac.. =11 ainL°Aan" CALn. Coa rn Paper WA-sA. (June 1981,Charlottesville, VA.)
4. Kohavi, 3., klc±B a4hl~a£~a a ~ rzNcGraw-BEll Publishing Co., Ltd., New Delhi, 1978.
< 25.
-- -,-v,&,--~
5. Unger, S.., Aynrnous g aicngCircuits, J. Wiley & Sons, Inc., New York, 1969.
6. Miller, R.E., Switching Thgory (Vol. II), J. Wiley &Sons, Inc., New York, 1966.
7. Peterson, J.L., 'Petri Nets", = mputing M LY-= ,Vol. 9, No. 3, pp. 223-252, 1977.
8. Kalman, R.E., Falb, P.L. and Arbib, K.A., Xg ir& inHathaaica1System Theory, McGraw-Hill, New York,1969.
9. Johnson, T.L., 'Finite-State Control of ContinuousP, Prc ,th I= Cngress, Helsinki, Finland,
1978.
10. Willems, J.C., and Hitter, S.K., "Controllability,Observability, Pole Allocatoin and StateReconstruction', I=i raD. Auto , Conftrol, Vol AC-16,No. 6, pp. 582-596 (1971).
11. Johnson, T.L., 'Analytic Models of Multitask,Processes Proc. 2th I= Conf. an DAinn And
Control, pp. 738-740, December 1981, San Diego, CA.
12. Klein, Q., and Kaliski, M., OFunctinal Equivalence in aClass of Autonomous One-Dimensional Nonlinear DiscreteTime Systems', Inform k Contl, Vol. 42, No. 2, 1979.
13. Signetics, Inc., 'Programmable Communications Interface(PCI)", Application Notes 2651-1, July 1978.
...... 6.... . _26
TOWIARDS A TEtORY of ASYNCHR.ONOUS. * ULDM CODERS AND THEIRAPPLICATIONS TO DISCRETZ-CONTROL Of CONTINUOUS PROCESSES~ p 34
f byDr. Martin 9. WaiSkand Dr. David C. ViUMe
Department of Electrical and Computer EngineeringNortheastern University
Bostoni, IA 02115 USA*Suppor~ted Ln Part by the United Stares Air Fore* Officeof Scientific %esearch, AFSC, Contract F4962O-42-C-3O83
ABSTRACT ways of extending those real-time coders to all of 12Thispapr dvelps noel f area-tie cder in a manner that allow weil-behaved devices to be
-- a t ap r f eelzed a coneofrteal-hichmp cod- constructed. This paper will approach the problem of--a ypeof gnerlize A/ covertr wich apsa.- specifying the "don't-cares" of these maps so as to
quences of n-tupics of real numbers into the binary ahyint-at (nir)relztnso efud4alphabet (0,I). Coders form an essential Interface Several examples of real-time coders will be given.
In the discrete-control environment between the con-tinuous state plant to be controlled. and the discrete Notation: For X a given set, X+ denotes the setcontroller itself. Theae are partially-specified of all finite-length non-null sequences of elements ofInput/output maps, and finite-state realizations of X.them are considered. Examples of the developed theoryare given. II. Genralizing the Simple Coder Model
Standard coders do not explicitly Imcoqporate time1. Introduction into their structure. One way to introduce time Into
the coder model is by making It an explicit Input toA key element In the theory of discrete-control the coder-its value reflecting the "arrival" ties of
of continuous system is the coder-the (hybrid) the other coder inputs. Doing this transfor= a coderinterface between the continuous-valued part of the C: -(, inonewchmp 3)4-{ l.
*system And the discrete-valued pert of the system However, this new coder is defined only ower a limited(typcaly amicrcomute). (inpy, 98Z~Kalaki subset of RX) + - naely those finite length strings
&dLemons, 1980) o h omTecoder io a kind of generalized analog-to- o h om
dgtlconverter, which In the discrete-tIme case,naps squseces of eloomts of an. for some n?-l, into (rl.,tl) ... (rk~tk)a finite set, which we may take as the binary alphabet wiht t2(.. kineiealasrvsf-(0,I). An extensive study of coder design. using boh wards. *(Due to practical restrictions Inherent inalgebraic methods (draw from standard "finite" auto- phsclIlentiosfcdrtistitymata theory) and linguistic methods (based upon a con- phsclile ntiosfcdrtistitymonotone sequence of time value* is fu~rther constrained:capt of languages defined over real alphabet&) has there exists some real value TI such that, for 2.,already been pursuad by the authors in the synchronous k, tj - t 1-l )- TI. This reflects the fact that inputcase, at cited above. This case, wherein both the changes cannot occur arbitrarily quickly.)transitions of the discrete-time plant, and of the turertohiclsofodse ra-im(finite-state) controller, occur In synchrony, under coders," which me will demote as ATC's.I the regulation of an external clock, is clearly aspea" cu, and many application contexts, whereactions are triggerd, by ,aents occuing asynchroncosy, III. An luesple: A lesl-TIme Discrete Integrator
* do sat fall under the umrella of this model. Thus,In this paper, a theory of coders which are intrinsi- Consider the following "imeegrator-Ilhe" *ssale.cally asceronous, and ich explicitly iWncoprate If the coder input to (C1 ,l) o then the weaput Is 0.the motion of time, Is Initiated. If the coder Input Is (rl~tl)... (rk.tk) with k'1, then
lie sdl asymchronous activity by vun Ideal the output io camputed as follow:$ We form the am:"samples and holds" to record the exect mocent In timeI at which aves occur. this allow us to define wheat rl~t2-tl) + r2Ct3-s) +4... r k-l (tk-t k-I)we 411l call "real-time coders." to their smstelemeataty fans they sode strings of real numbers and if the suin is greeter then 0 the output in 11are m8"s defined from a set S Into (0,l) where S eon- otherwise It to 0.sLats omly of fianie-lengt strins of pairs of the Gee my of Implementing sch a coer Is to extendI. fog (r,t), where "r" Is a real uber (produced by its dom&to a f (%a)+and to then employ pro-som process) and "t" is the time at which "r" Is viously developed techniques (Kaliski and Lemons. 190;"seen." Because time always Increases, the strings In VitpeM. 1I6). The mwet natural my of exteadirtr It IsS are alweys of the fore by just uslng the above rule for arlbitrary strings
CuL~vl) ... (ok,*k). regardless of the salurea of the(rl~t) .. (r~tk)sequence Vl'VI.... owh. That is, If the saieaene to of
where the t-walue fo=m a monoosm iaftasia Sequence length 1, the output is S1 If its length Is greaterand whets the differences t2-tl. 0-t0,..., are always then It we compute the ahewe saw an test Its poaltive-greater than soe apriari lever bound TI. mesa. If it is positive, them the output Is I g other-
capable *9 implmfLag sh ap is te view them 40 Now thia exeded I/0 amp caon be realised by the
22)eeiaload *mspeial retice -oao s t mw maei tucdsdw"Lo era
ecpossible sum of the form: f(AC)-f(BC).
n102-1) + u k-i (vk-v k-I) - ta vk A mucb more powerful "converse" to Lomes I is alsowhe. he uk inthse amLa iesam. he tat- true when 5 has the property that "once a aralaewher th "" In hes sus tothesam. Th strt- "S" it never "sets back in," i.e.. *If A to (01 i
Ing state of the coder is the congruence class of the mot In S, then AB is not In S for any 3 in (0,1).null string. which consists of strings whose sum, Call such S (for Lack of a better expression) "tightlyformed as above, Is 0 and whose WUk (lst term) La 0. structmuted."-4.e text state and output maps are as follows: (Noechs: they' are well-defined.) Linea 2: Let S be a tightly structured subset of
Next state map: (0.1). and P a partition (Pi,...,1PO ofS, for whichthe partition propet-7 holds for f:S -- %0,.1;. Then
S(lv)... (uk.vk) ]. (c~d) -- there exists an excansam Z of f to all of 0,i'+ whichis finite state realizable.
((uL~vi) ... (uk,vk) (c~d) I hs w ae
output cap: Theorem 1. (Fundametal Realizability Theorem
£(ul.vl) ... (uk~vk) 3. (c,d) >frPrilySei e / 'as
POS ( ul(v2-vl) + ... + ua k-1 (wit-v k-1) Lotr S he a tightly structured subset of (0,11..Let f be an 1/0 nap from S into (0,I). Then there
+ uk Cd-wit)) exists an extension S of f to all of (0.1)4 Which Lafinite-state realizable If end only if 5 obeys the
where P05(x Is 1 If and only If x is -0. "partition property" with respect to f:
A natural enbedding of the state set Into RzR "There exists a finite partition P..(fl,.,Pk)*Xiss- of S such that for any jal,...k, and any A
andBIn lj, if C InO.)+Is such that AC(ul~vi) ... (uk~vk) 3 'and BC are both in S, then AC and BC belong
to the same block ft of P as wrell (0 setCul(v2-vl) + ... + u k-i (yb-v k-l)-uk vk~uk) necessarily equal to J). Furthermore.
fCAC)f(Ic)."':slnl this embedding, a two-dimensional reas-
:to= of this coder can be specified as follows:V. Generalizing to 1/0 Maps an RxR
Next state map:Al2 of this Seneralizes to maps on gm, using te
(%.7). (c,d) -- > (x + (y-c)d, c) concept of a finita-autanaton defined over a realalphabet CgImpey. 1982). In particular It generaliess
Output nap* to Real-Tize Coders, where, by the very mature of thedomains of such coders (the set S), the tl*t structure
Cx,,'), (c.d) -> POS(% + yd) criterion La mew. (The Input to & RTC go" "bad" as
Startng satesciently far, oncea it becomes bad It remains bed).gobegin wt nitiiedfate fadtr
Colo)ninistic finite-state reel autematon (ISM. x is a753£ If there exist a finite number of states Qi, ... 6
IV The Kay Theoretical Problem: Structured Qit, and, "aisted with each state Qj * j'.l, ... .kv aExtensions of Partially Specified 1/0 Maps partition PJ of the real numbers. A single state of H
Is labelled as the starting stage of X. aid certainA~s :he above esseple so cleerly Indicates, in states. of M are classified as accepting states. Them
*r~ oefcieyraimra-iecdrad to XIs maual soitdwith a coder CKt at- (.)OMP17 te aread deeloed teor fo synhroous Clearly all of this generalize to Impute that are incoders. ~ ~ ~ ~ ~ ~ ~ xi voms Vago.ytoetn h T rm K-W alac fiaite-stage autoeoo a 2-Its onsraied oeaf toallof RW+.It btsam- dimensional 111£, and, for brevity demote se a device
tension can be wd na wa ht-rwre-teby the Symbols 20153£.intriasic structure of the tIC, then we have effected The motioe of a mom-datermislattc M33, alSOthe development ofage esg ol e etit smeralins In a straightfor'ward member. in pertsarcu-. attention to finite-state extenaions In this paper. If ome *aotiaqe with ach state Oj a moms C4 of a
Towds this @sal, let us begin our -diseussion (or ho thma one hias a am-detaimistie, neupleteLY'4wn the follwing leens dram froM the basnic Ideas specified "Ul. The pwoof that suck a dewie is amys2f automata thory. The pceefs of &em& I an 2 are equivelen:to 1i deteaimiatic MUA is viutaupatraighwforier d n ttd for lea f c" . Identical to th' e one used iS fimits ausemata eory.
L~al:Lea fU -2 0 1) ee gien I/ map we omit 2ite.tlycryeerf*reLemm It ot fs --. (o0 bea $I" 1/ maLImes I and 2Imdaeycm* a ~
Weined on a subset S of 0,1.Suppose& that thare tine coders. as their domains are tightly structured.Is a finite stt nehso X hs re"a&se me sn M gI Specifically, lot us formally defines a RM~ (This$ of f to to'*+-. Then there to a finite partition definition complemens the WaoM&s Noe give% InP1,. ..,Ph) of I osbat the fellawing property to settle& Xl ao"e.)
9 rue, for all strins A~ gmdAin i,.Dsfinuition: A Real-flag Coder (UQC S Wn aop
If A "d At*eImS e Sftblakof P. A C I& fm lass(0.0 bwer'is be fllow&* submit of
Ifs af Uedgsoty te Oe& 46at A 46d a we WI) ad I a (VwIO) -MACt~t) such OhatA2 -
uhere -AX is some fixed real constant. (Q2,0) Q~43 (412.1) -->Q1
Sots that S Is tightly Structured. There is (Q3,0) Q2 (Q3.1) -->Q3nothing in the proofs of !a= I and 2 that dependsupon the fact that the inu-s to the automata are 0 or1. Thus rhe Lams generalize to the reel-tine coder Vt. Conclusionscase and, thus. ve have:
The process of "cretively" selecting "don't care"Theorem 2: V?''adc:al Fialte-State Realiz- values so as to provide for meaningful extensions ofi.±: s~r f :r ?aa.-e dara) real-tine coders is a fundamental open research problem
of this theory. Mfany more general coder types exist."Let JMC be a gi7on reel-time coder. ?ben RTC such as the shift-finitary coders (VIRey. 1930).is finite-state realizable If and only If there Current work, then. to seein to extend the aboveexists a fiats partition (P,... ,Pk) of its preliminary results to these more general code"., asdomain S auth that far ay J-l,. . .,k, anwd well as to look at other vays of modelling asynchrony&=y A. 3 in Pi. If C !n U1i is such that AC In coder design.ead 3C are bjoth in S then AC and SC are both
necessarilyr equal to J) and 2TC(AC)-ftTC(BC)." VII. Referencs
76-ere is an explicit techecI&= for constructing Kaliski, N. 1. and Lemons. K.. "Discrete-Coding. oft'-s !~Izce-s:&:* realizatica whea the partition Continuous Valued Signals," 1980 Conference. on Inform-pro~arzy holds (as in :he ;rnof of Lome 2.) tion Sciences end Systess. Princeton University,
Princeton. NJ.
An ExapI& of a vi=4t&ay leal-Tine Coder Vinpey. D. G.,* "linite-State Control of Discrete-TineContinuous Processes: An Aatomta-otivated Approach."
Z* gIev here a senpie exnpIs based upon Vinpey Ph.D. Thesis, KIT Department of Ilectrical Engineering
~192) to Illustrate; the conceptsaboe A Unitary and Computer Science, June 1962.
real-ti=e Integrator considered earlier Iq definitely Vinpey, D. G., Johnson, T. L., and Calisi, M. 3.,n2c fUnitary.. "Realization of AID and D/A Coders." 1L981 .MCC, 3
Czzse;!er the f*22owIng real-time coder. Charlottesville, VA, June 1931. I
ZC(rltl) -0 if C1l))-0; If not
0 If (ek+th) >0 0 and the numbher ofmoz-negazira suams r4l..r -l + t k- Iis even or zero. i
!s i: f!Initesc&ae realizable? And, If so, whatIs a realization for It? We proceed to answer these;-.:s:±czs.
!ue doz-aft S of this coder satisfies the partitionprpryfor the following partition {Pl. P2, W3s
79. a s-.r~zgs haw-ng am odd numer of non-negatIve rme wita the last sun negative)
72 - istring having en odd umbor of uenegative sums with the last im nnnegatiM)
sad
P3 - strings baving aa even numer (or stwo)mon-inegtive $WIms
~~1 17 our developments Above, them, 2C Is MUSiTaY.tfar:- It is stay to directly construct a I2WUU fer
based upon the Unas atvve. lthe, then do a_I~tyit is clear that, as with finitaVY. We 0tdyeE
;,"Mtisor WAd a finite state aohies (gum, 1302)IMe qU02gise' Q Outputs 0 ItfJ4& uito 3W 41 ethe
val : stuts I. The finite state m be K be*Woe sztas oemepa to n1, 1i. awn itrnQ1, Q2, and 413. 9) ise the stasting 0""t. ase
areepting Veates are 91, d4 Q). ase tgmnitime weas folows:
E ORAFDU[-: JAI;LAnY, 1983
TOUA7DS A THEOPY OF FINITARY ASYiiCI!RO[OUS CODERS
1) Generalizing the Simple Coder Eodel
The standard coders previously considered by us did notexplicitly incorporate time into their structure. Rather,tiue was implicit, synchronous, and "regular". It madesense to tall: about the kth input or output (or state), andthe value of k automatically stepped through the integers.
About the simplest way to introduce time into the coder...odel is by making it an explicit input to the coder -- itsvalue reflecting the "arrival" time of the other coderInputs. Doin- this transforms a coder C:R -->(O,1} into onewhich uaps (Rxr)+-->0,1}. However, this new coder (call itETC for "real-time coder") is defined only over a limitedsubset of (ExR)+ -- namely those finite length strings ofthe form:
(rl,tl) ... (rk,tk)
with tl < t2 < ... < tk, since time always movesfcrwards.
Due to practical restrictions inherent in physicalitiplementations of coders, this strictly monotone sequenceof time values is further constrained: there exists somereal value TI (for "Temporal Input" constraint) such that,for J=2,...,k, tj - t J-1 >= TI.
(Theoretically, at least, such constraints are ergodic-theoretic in nature -- a notion to be explored in greaterdepth during the remainder of this contract year.)
The output of the RTC may still be regarded as a binarynumber b = 0, or 1, and may be viewed as occuring within TOseconds after the last input arrives. (TO for "TemporalOutput" constraint). Note that various racing problems, andhazards, and such, must be dealt with- to physicallyimplement such a device. For example, TO may have to besmaller than TI to assure that the correct output is read.
2. An Example: A Real-Time Discrete Integrator
Consider the following "integrator-like" example. Ifthe coder input is (rl,tl), then the output is 0. If thecoder input is (rl,tl) ... (rk,tk), with k>1, then theoutput is computed as follows: We form the sum:
ri(t2-tl) + r2(t3-t2) + ... r k-1 (tk-t k-i)
L ___________________
Page 2
If the su r is Zreater than 0 the output is 1;otherwise it is 0.
One way of implementing such a coder is to extend itsdorain to all of (RxE)+ and to then employ the standardtechniques we have already developed. The most natural wayof extending it is by just using the above rule forarbitrary strings (ul,vl) ... (uk,vk), regardless of thenature of the sequence vl,v2,...,vk. That is, if thesequence is of length 1, the output is 0; if its length isgreater than 1, we compute the above sum and test itspositivity. If it is positive, then the output is 1;otherwise it is 0.
.ow this extended I/O map can be realized by thecustcmary techniques of Nerode equivalence. Let us see whatthe 'erode equivalence relation is in this case. First,consider two strings of length greater than 1. Two strings(ul,vl)...(uk,vk) and (pi,ql)...(pm,qm) will be equivalentif and only if the two sums below are equal for any(al,bl).. .(aj,bj):
ul(v2-vl) + ... + u k-1 (vk-v k-1) + uk(bl-vk)
+ al(b2-bl) + ... + a J-l(bj-b J-1) and
pl(q2-ql) + ... + p m-1 (qm-q m-i) + pm(bl-qm)
+ al(b2-bl) + ... + a J-1(bj-b J-i)
(11ote that, strictly speaking , we want POS of each sumtc be equal, where POS is the map POS(x)=1 if and only if x
is positive; otherwise it is 0. It should be clear,hovever, that if the sums differ, we can append yet anotherterm (a j+1 b J 1) which can cause differences in signs ofthese new sums, and thus differences in the PO-values.)
If j=1 wse may ignore the second line in the twoexpressions above, and, in fact, may ignore its role in theequivalence condition completely, as it is identical in bothexpressions. Thus the two original strings will beequivalent if and only if the two first lines above areequal, for any bl whatsoever. Setting bl to 0, inparticular, forces equality of the following twoexpressions:
ul(v2-vl) + ... u k-1 (vk-v k-i) - uk vk and
pl(q2-ql) + ... + p m-1 (qm-q m-1) - pm qm
Let us denote this common value by A. Then we mustalso have that A~uk bl a A n- li, for any real number bl.Since bl may be 1, it immediately follows that uk must infact equal pm.
J ",- m . ---- ,m m m i nm n a
_ -------- - . -- °
. Page 3
Summarizing, then, a necessary and sufficient conditionfor two strings (ul,vl)...(uk,vk) and (F1,qi)... (pm,qn) oflength greater than one to be Nerode equivalent is thatuc:p2 and that
ui(v2-vl) + u k-1 (vk-v k-1) - uk vk
pl(q2-ql) + ... p m-1 (qm-q n-1) - pni qm
::ow let us turn to strings of length 1. First supposethat we have two strings, both of length 1. Call themCul,vl) and (pl,ql). It must be that for any
(al,bl)...(aj,bj), J>=1,
u1(bi-vl) + ai(b2-bl) + a j-1(bj-b j-1)
pi(bl-ql) + ai(b2-bl) + . a j-i(bj-b j-1)
where the terms on the right are not present if j=1.-t is immediate that we must have -ul v1 = -pl q1, andu =p1.
For one string of length 1, and the other of lengthgreater than one, it can similarly be deduced that therequisite condition for equivalence is that, calling thestrings (ul,vl) and (pl,ql)...(pm,qm):
u1=pn and
-ul v1 = p1(q2-ql) + ... + p m-1 (qm-q m-i) - pm qm
As for the null string NL, a similar condition isq derived: fL is equivalent to (ul,vl).. .(uk,vk), k>=1 if and
only if:
uk=O and
ul(v2-vl) + ... u k-1(vk-v k-i) = 0
All of this suggests that the Nerode state-reducedmodel of this real time coder is the following. There isone state for each possible sum of the form:
ul(v2-vi) + ... + u k-1 (vk-v k-1) - uk vk
where the "uk" in these sums is the same. The startingstate of the coder is the congruence class of NL, the nullstring, which consists of strings whose sum, formed asabove, is 0 and whose "uk" (last term) is 0. The next stateand output maps are as follows: (Note that they arewell-defined)
Next state map:
• , i-- . . .|- ..... .
Page 4
[ (ul,vl) ... (uk,vk)], (c,d) --- >
C (ul,vl) ... (uk,vk) (c,d) ]
Output map:
[ (ul,vl) ... (uk,vk)], (c,d) --- >
POS( ul(v2-vl) + . + u k-1 (vk-v k-i)
+ uk (d- vk) )
A natural embedding of the state set into RxR exists --
[ (ul,vi) ... (uk,vk) ] --- >
( ul(v2-vl) + + u k-1 (vk-v k-1)-uk vk,
uk)
Using this embedding, a two-dimensional realization ofthi4s coder can be specified as follows:
Next state map:
--------------
(x,y), (c,d) --- > (x + (y-c)d, c)
Output map:
(x,y), (c,d) --- > POS(x + yd)
Starting state:
(0,0)
A diagram of this coder is shown in Figure 1. Notethat it is assumed in this diagram that the inputs areconnected to event-driven ideal samples and holds. Eventsoccur at the input times, and the samples and holds serve to"capture" the inputs at these times, and the timesthemselves. In this diagram we assume that the output isread TO seconds after the input occurs.
p. I . .
• "; : • ' I I i l " ' m
V Page
The Key Theoretical Problem: Structured
Extensions of Partially Specified I/O Iaps
As the above example so clearly indicates, in order toeffectively realize real-time coders, and to employ thealready developed theory for synchronous coders, we mustfind a good way to extend the ETC from its constraineddoizain to all of (RxR)+. If this extension can be made in away that "preserves" the intrinsic structure of the ETC,then we have effected the development of a useful designtool.
Towards this goal, let us begin our discussion with thefollowing lemmas , drawn from the basic ideas of automatatheory:
LemLa 1: Let f:S -- > {0,1) be a given I/O map definedon a subset S of {0,1)+. Suppose that there is a finitestate machine 1 which realizes an extension g of f to{0,11+. Then there is a finite partition P={P1,...,Pk} of Ssuch that the following property is true, for all strings Aand B in S:
If A and B are in the same block of P, and C in {0,1}+s any string for which both AC and BC are in S, then AC and_C are also in the same block of P (which is not necessarilythe block that A and B are in), and f(AC)=f(BC).
Proof of Lemma 1: We may assume without loss ofgenerality that V is a Moore machine; call its starting
state Qi. Call the remaining states of H Q2,...,Qk. Definethe partition P as follows: Block Pj of P, j=1,...,kconsists of those strings of S which take state QI to stateQJ. Clearly (P1,...,Pk) partitions S. If A and B arestrings in S that are in the same block of P, say PJ, thenfor any C whatsoever, AC and BC are in the same block of P.In are in S. Furthermore, by the very definition of "11realizes g" it must be that g(AC)=g(BC) as AC and BC lead Mto the same state. When AC and BC are both in S the g inthe above equality may be replaced by f since g is anextension of f. QED
A much more powerful "converse" to Lemma 1 is also truewhen S has the property that *once a string leaves S" itnever "gets back in", i.e. if A in (0,1)+ is not in S, thenAB is not in S for any B in (0,1)+. Call such S (for lackof a better expression) "tightly structured".
Lemma 2: Let S be a tightly structured subset of(0,1)+ and P a partition {P1,...,Pk) of S, , for which thepartition property holds for f:S -- > (0,1). Then thereexists an extension g of f to all of (0,1)+ which is finitestate realizable.
Pag e
Proof of Lemra 2: Consider the non-deterministicfinite-state 1Voore machine M defined as follows. Thestarting state of ; will be called QO. There is a specialstate of (where strings not in S will take 1i) that will becalled QiCS. The remaining states of I correspond to thenumber of blocks in 2 in the following way. For j=1,...,k,look at Pj. 1. all the strings in Pj have that propertythat f maps them to 0, then create just a single state QjO;if all the strings have the property that f maps them to 1,then create just a single state QJ1. Otherwise create twostates QjO and Qjl. Thus V! contains at least k+2 states,and at most 2k+2; in any case it is finite state.
The accepting states of 1 will be the QJ1's, ifPresent, for j=1,...,k. The transitions of 1 are defined asfollows: First let's look at QO. If 0 is in S, then 0 isin some block of P, say Pj. Direct a 0-arrow from QO to QjOif f(O)=O; otherwise direct the 0-arrow to Qjl. If 0 isnot in S then direct the 0-arrow to QNS. The 1-arrow fromQO can similarly be constructed, based upon whether or not 1is in S, and, if so, what the value of f(1) is. Next weconsider U.S. This will be considered to be a "dead state"in the sense that once entered, it cannot be left. Thusbeth the 0-arrow, and the 1-arrow, go from QNS to QNS. (Aswe will see below, all strings that are not in S lead toQI:S; by the tight structure constraint on QNS, once QNS isentered it is appropriate to get trapped there.)
Finally we turn to the transitions for the remainingstates. Consider, for a given J, the states QJO and Oil,or, if just one is present, whichever it happens to be.Consider the problem of drawing the O-arrow. Now thestrings in Pj divide into two classes -- those which, whenfollowed by 0, are still in S, and those which, whenfollowed by 0 are no longer in S. By the partition propertycn S, all of the strings so "continuable" by 0 are elementsof the same block of P, say Pm, and the continued stringshave the same f-value. We can use the above remarks to draw0-arrows from QjO and/or QJl as follows: Suppose that thereis a state QjO. This means that there are strings in PJ forwhich f is 0. If any of these strings are "continuable" by0, by the above remark, they all have the f-values whencontinued, and all belong to Pm, when continued. Call thef-value of the continued strings "b" (b=0,1). Clearly, byour very construction, state Qmb must be present. Draw a0-arrow from QJO to Qmb. If Pj also contains strings forwhich f is 0, but which are not continuable, draw a 0-arrowfrom QJO to QNS. Thus one or two 0-arrows will emanate fromQJO, going to either QNS, Qmb, or both. A similar procedurecan be used to draw 0-arrow(s) from QJ1. Furthermore thesame procedure can be used to draw the 1-arrows from QJO andQJ1.
p
Page 7
It is easy to see that if al,...,an is a string in Sthat ('ay the tight structure of S), each of the strirt-s al,al a2, ... , al .... .,a n-1 is in S an that there is a pathi n ., starting at CO, and remaining thereafter among the
's. 1' i3 also easy to see that accepts al,...,an in Sif and only if f(al ,...,an) -. ;rinrsthat are not in S, in general, but the only strings in Sthat are accepted are those for which f is 1.
I'ow 11 can be converted to an equivalent deterministic.oore nachine I1' which accepts exactly the same strings asdoes i.. Let us define g:{0,1}+ -- > [0,1} by g(A)=1 if andonly if V', accepts A. By our remarks in the precedingparagraph, g is an extension of f. Obviously g is finite-ztate realizable (by 1I'). The proof is complete. QED
SUe nay summarize the above two Lemmas with thefollowing Theorem, which we call the Fundamentalrealizability Theorem for Partially-Specified I/O 1aps:
Theorem 1: (Fundamental Realizability Theorem forPartially Specified I/O Maps)
Let S be a tightly structured subset of [0,11+. Let fbe an I/O map from S into {0,1}. Then there exists ane~xtension g of f to all of 10,11+ which is finite-staterealizable if and only if S obeys the "partition property"1ith respect to f:
" There exists a finite partition P={PI,...,Pk}
of S such that for any j=1,...,k, and any A and B
in P1, if C in {0,11+ is such that AC and BC are
both in S, then AC and BC belong to the same block
of Pm of P as well (m not necessarily equal to k).
Furthermore f(AC):f(BC)."
Generalizing to I/0 maps on RxR:
In the pages that follow we will show that all of thisgeneralizes to maps on RxR, using a form of our earlierdefinition of finite-automaton defined over real alphabets.In particular this will generalize to Real-Time Coders,where, by __ ;ory nature of the domains of such coders (theset S), the tight structure criterion is met. (The input toa RTC goes "bad" as soon as the time-coordinate does notadvance sufficiently far; once it becomes bad it remainsj bad".--
Page 8
'-e will culminate, therefore, in the fcllowing variantof Theorem 1:
Theoren 2: (Fundamental Realizability Theorem for2eal-Time Coders)
Let RTC be a real-time coder. Then RTC is finite-state realizable (in the finitary coder sense) if and onlyif its domain obeys the partition property.
We beGin we an intuittve definition of a deterministicfinite-state real automaton (FSRA). 11 is a FSRA if thereexist a finite number of states Qi,...,Qk, and, associatedwith each state Qj, j=1,...,k, a partition Pj of the realnumbers. A single state of 14 is labelled as the startingztate of 1', and certain states of Y are classified asaccepting states. Thus 14 is naturally associated with acoder C1: 19 -- > {0,11. Clearly all of this generalizes toinputs that are in RxR. We call such a finite-stateautomaton a 2-dimensional FSRA, and, for brevity denote suchL device by the symbols 2DFSRA.
The notion of a non-deterministic FSRA also generalizesin a straightforward manner. In particular if oneassociates with each state Qj a cover C of R (or RxR) thenone has a non-deterministic, completely specified FSRA. Theproof that such a device is always equivalent to adeterministic FSRA is virtually identical to the one used infinite automata theory. .e sketch it below.
Cne creates a deterministic FSRA by initially defininga state for each of the non-empty subsets of [QI,.. .Qk).The starting state of this new machine is the same state(singleton subset) as that of the non-deterministic one.The accepting states are any states whose correspondingsubsets contains at least one accepting state of thenon-deterministic machine. To define the transitions of a3iven state we look at each state in the associated subset.
Ile pick a real number, say r, and ask to what states in theoriginal non-deterministic machine r takes us. Ile take theunion of all of these r-mapped states (for each state in theassociated subset) and that subset of {Q1,...,Qk) is wherewe send r to. We repeat this for each real number, therebyassociating a partition of the reals (or of RxR in thetwo-dimensional case) with each subset of {Q1,...,Qk). Therest of the proof is clear. We can easily demonstrate thatthe deterministic machine and the non-deterministic oneaccept exactly the same set of strings.
The proofs of Lemmas 1 and 2 immediately carry over forreal time coders, as their domains are tightly structured.Specifically, let us formally define a RTC:
Page 9
Definition: A Real-Time Coder (RTC) is any map from Sinto {0,11 ihere S is the following subset of RxR:
S {rl,tl) .... (rk,tk) such that
t2-t1>TI,., tk- t k-1 > TI I
where TI is some fixed real constant.
lUote that S is tightly structured. There is nothing inthe proofs of Lemmas 1 and 2 that depends upon the fact thatthe inputs to the automata are 0 or 1. Thus the Lemmas
* generalize to the real-time coder case and, thus, 'rem 2holds. We repeat its statement below and give, i. ,ection
* 2.1, a detailed example.
"Let RTC be a given real-time coder. Then RTC is
finite-state realizable if and only if there exists
a finite partition [P1,...,Pk} of its domain S such
that for any J=1,...,k, and any A,B in Pj, if C
in 2xR is such that AC and BC are both in 3 then AC
and BC are both in the same block Pm of the partition
(m not necessarily equal to J) and RTC(AC)fRTC(BC).*
There is an explicit mechanism for constructing thefinite-state realization when the partition property holds.
4. An Example of a Finitary Real-Time Coder
le give here a simple example based upon Wimpey (1982)to illustrate the concepts above.
Consider the following real-time coder.
FTC((rl,t1)) 2 0 if (r1 tl)>=O; 1 if not
RTC((r1,tl) .... (rk,tk)) z
0 if (rk+tk)>x 0 and the number of
non-negative sums rl+tl,...,
r k-1 + t k-1 is even or zero.
1 otherwise
Is it finite-state realizable? And, it so, what I a
realization for it? We proceed to answer these questions.
T ______.. .
Page 10
The domain S of this coder satisfies the partitionproperty for the following partition {P1,P2,P3}:
P1 { strings having an odd number of non-negative
sums with the last sum negative)
P2 = strings having an odd number of non-negative
sums with the last sum non-negative} and
P3 { strings having an even number (or zero)
non-negative sums}
Note that if strings A and B in Pi are such that AC andBC are in P1 then RTC(AC)=RTC(BC)=1; if they are in P2 thenFTC(AC)=RTC(BC)=0. If AC and BC are in P3 thenRTC(AC)=RTC(BC)=1. Sicilarlj if A and B are in P2 then ifAC and EC are in P1 the common output value is 1; if theyare in P2 the common output value is 0. If they are in P3the output is 1. Finally if A and B are in P3 then thecommon outputs are (for termination in P1,P2, and P3respectively) 1,0,and 1.
By our developments above, then, RTC is finitary. Infact it is easy to directly construct a 2DFSRA for RTC,based upon the arguments above. Rather than do so directlyit is clear that, as with finitary, ordinary coders, RTC mayOe realized as a cascade of a simple quantizer and a finitestate machine:
The quantizer Q outputs 0 if rj+tj is >=0; otherwiseit outputs 1. The finite state machine X has three statescorresponding to P1,P2, and P3. Call them Q1,Q2, and Q3.Q3 is the starting state. The accepting states are q1, andP'3. The transitions are as follows:
(QI,0) -- > Q3 (Q2,1) -- >Q1
(Q2,0) -- > Q3 (Q2,1) -- >Q3
(Q ,0) - Q2 (Q3,1) -- >Q3
i..I. -.;-A THEORY OF ORBITAL BEHAVIOR IN A CLASS OF NONLINEAR
SYSTEMS: "CHAOS" AND A SIGNATURE-BASED APPROACH
by M.Kaliski, Northeastern University
Boston, MA *
S.Yunkap Kwankam, Universite do Yaounde,
Yaounde, CAMEROUN
P. Halpern, Northeastern University,
Boston, MA
(s This author's work has been supported in part by tue
United States Air Foros Offioe of Scientift Research,
AFSC, Contract Number FP9620-82-C-0080.)
M ailing M&dress: Dr. Martin E. KaliskiProfessor and Director of Ciputer EngineeringNortheastern University
; Boston I NA
02115
1A,
-- - - - -- - - -
Page 2
Abstract
A theory of orbital behavior in certain autonomous
one-dimensional nonlinear systems is pursued, using a
approach based upon the concept of orbital signature.
Particular attention is paid to the fixed point structure
of such systems with the ultimate aim of using the
signature repertoires of these systems to characterize
fixed-point orders and the presence of *chaotic regimes* .
A system-theoretic approach is pursued here -- an
approach which complements other recent studies of a more
analytical nature (employing ergodic theoretic methods.)
Chaotic behavior in a certain subclass of these systemsais -
completely characterized in terms of the first two
iterates of a specific known point in the range of the
system transition function.
r Page3
Table of Notation
-symbol meaning
k.fpq, f a generic unimodal or subbell
function with breakpoint p and peak
value q
fk pq, fk ror k>=O, the kth Iterate of fpq or
f
sigt~x),Sig(x) the signature (infinite) of x under
the mapping f
sigfk(x),aigk(x) the k-signature of x under the map
f
ls8W, rs(x)' the left, right signatures of x
lak(x), rsk(x) the left, right k-signatures of x
aOzk that
Sk tbe k-sigzatmre repertoire of a
-A
Page 4
given map
S the (infinite) signature repertoire
of a given map
sequence oonoatenation
set interseotion
a**n the sequence a a a ... a (n times)
U set union
B'J the Jth left rotation of s
Ak, I the signature bins of a given
string a
Lj(a) the jth left shift of a
[ ... ] closed Interval
subset. of
0> not equal to
Page 5
I. Introduotion
There has been great interest in recent years in the
autonomous behavior of nonlinear one-dimensional systems
defined over the unit interval [0,11 (Kaliski and Klein,
1982; Klein and Kaliski, 1979; Collet and Rokmann, 1980;
Li and Yorke, 1975; Parry, 1966; Baillieul , Brockett, and
Washburn, 1980; and Feigenbaum, 1980, to name several)
There is, In particular, interest in the *chaotic
behavior' of the equilibrium (fixed) points of such
systems. This *random' behavior arises even in processes
describable by simple first-order difference equations of
the form:
x k+1 a f(xk)
where f:[0,1J-->[0,1].
Various approaches have been used to describe systems
of the above form, ranging from graphical methods (Hay and
Oster,1976) , to purely analytical techniques (Li and
Yorke,1975) and ergodio theoretic appoaches (Parry,1966).
A system theoretic approach, Introduced by Klein and
Kalisk-i, and cited above, Is pursued In this paper. This
approaoh treats the tuactieo t as the state transitie
function of an autonomous oae-diasnsional nonlinear
t , , . . . ... - . . a
Page 6
discrete-time system. The ooncept of Isignature' (defined
below) is used to describe the orbit of any given starting
state.
This paper serves to characterize the fixed point
s'.ructure of such systems through the use of signature
repertoires. Our development is somewhat detailed because
of the need to formalize and doflne our 'vorking tools*;
nonetheless, our results are of Intrinsic interest
because:
(i) they characterise "chaotic behavior" (the presence
of fixed points of all periods) in a subclass of such
systems strictly In terms of the first two iterates of a
specific point in the range of the system transition
function.
(ii) they demonstrate the utility of the signature
concept, a concept which complements alternate approaches
based upon #rgodic theory and -other measure theoretic
concepts.
Muoh of this material appeared in somewhat different
format In one of the authors' doctoral dissertations.
(Ewankan, 1979J
XI. Basic Cooopts. 31b#9lo# "Sanatureo aad Gra Coo
A* @rd r . ,
-. . . . .p - ,' -
Page T
As much of this development in based upon the concepts
Introduced by the earlier work of Caliski and Klein, we
merely summarize their' basic Ideas In the paragraphs
below.
11,1 We begin by describing the types ot functions
considered In this paper.
A subbell function is a continuous map f:[0,1J -
[0,1] for which (Figure 1)
(I) f(0) =f(l) s0,
(ii) f has a unique maximum q at some point p in
(0,1)
and
(111) f is strictly monotone on [0,p] and on (p,11.
We sallrefr top a th brekpont f f nd enoe-bVe sall efe to as he reakoint of anddenoe b
fpq a subbell map f with breakpoint p and peak value q.
Similarly, for k a positive integer, fk pq denotes the
kth Iterate of 'tpq (i.e. the k-told oomposition of fpqI
A subbell ftmo-tion Is a '#Veolal type, of tuaimodel"
PTvf
Page8
function:
A unimodal function in a continuous may f: [0,1] -
(0,1] tar whIcoh conditions (ii) an d (iii) above hold, but
not necessarily condition (i). CWe extend the notation
tpq and tk pq to unimodals in general. ) (Figure 2)
11,2 We introduce the concept of.(orbital) signature next. --
Let x in (0,1] be given, k a positive integer. The
k-signature of a sequence under the mapping fpq, denoted
by sigfk(x) ,is the length k string:
bO bi b2 ... b k-i
where for in Ot ... #k-1#
hi a0 if 0 <= fi pq(x) < p
- if fi pq(x) z p
and 31 If p < fi pq(x) <= I
(where tO pq(x) *xO by convention)
s a bO bi b2 ... b k-i will be oalled regular if, for
all I >a 0, bi a 0 or I but ntotr- . If a Is not regular
It will be called Irregular . In a similar fashion one
can define* ths (Infinite) signature sigf(;) of z by simply
Page 9
letting Ok" range over all the positive integers. The
notions of regularity and irregularity generalize in a
straightforward way. In the sequel when the word signature
is used without further qualification it may mean either
finite length or infinite signature, according to the
context of the discussion. Further, when the specific
function fpq is clear from this context we will often
simply write OfO and drop the f from sigft. When a point
x has a regular signature x is said to be regular;
otherwise it is said to be irregular. Note that if x is
regular then for all k, every k-signature of x is regular,
and conversely.
* 1Z,3 We can define a total ordering relationship upon
binary strings (not necessarily regular signatures) which
the reader will recognize as Gray Code order:
Let al = bO bt ... and s2 a dO dl ... denote two given
binary sequences of equal finite length, or both of
infinite length. Then al < s2 if
(i) a1 is not equal to &2, and
(Ii) if bit position j Is the first one at which al
and s2 differ, then
bO.... b J-1 0 and
1..... _ __.... ..
17
Page 10
dO ... d J-1 1
where * is the Exclusive OR funotion.
This ordering is fundamental in the theory of signatures,
in that all unimodal funotions obey a Omonotonioity of
signatures property" (monotone with respect to this
ordering.) This is explored below in section 11,5.
11,4 Let a be an irregular infinite signature of some
point x under a given unimodal map. An instance of sis
any binary string obtained from s by arbitrarily inserting
O's or 1's for each - in s. Ve refer to the least (in
the < ordering) instance of s as the left-signature of s,
la(x); we similarly define the right-signature of s,
rs(x), as the greatest Instance of a. The notion of left
and right signature trivially extends to regular points x,
where ls(x) = rs(x) = sig(x). Note that these definitions
hold equally well for finite signatures, where we write
the left and right k-signatures of x as lak(x), and
rsk(x), respectively.
We define the conoept of the exploded k-signature
ropertoire,denoted by 8k, to be
k a (rak(x) N x In [0,1]) U ilak(x) \ x In [0,11)
Thus the exploded k-signature repertoire of f
-ii
: !: Page 1 1
consists of all k-signatures of regular points, and the
greatest and least k- signatures of all irregular points
(hence the irregular k-signatures are "exploded".) A
similar interpretation holds for the (k=infinity) exploded
signature repertoire of f, denoted by S.
11,5 As indicated above all unimodal functions obey a
fundamental monotonicity property with respect to their
signatures. We cite the property without proof. Its
proof may be found elsewhere (Klein and Kaliski, 1979.)
Let f be any unimodal map. Let x, y in [0,1]
be given, x < y. Then, in the above defined
Gray code order, rsk (x) <= lak (y), for all
k, and rs (x) <X 1s (y).
III. A "Roadmap" for the Technical Developments That
Follow
To aid the reader In the developments to come let
us sketch out a "roadmap" of the remainder of the paper.
Section IV examinea certain recursions for calculating the
exploded k-aignature repertoires of unizodals and
eubbells. The role of the left signature of the peak
value q in these recursion formulas is oentral. In
Section V we examine certain necessary and sufficient
oonditiona for the existence of fixed points having given
7;. .
Page 12
finite or infinite signatures. This is followed in
Section VI by an examination of the *orders* of fixed
points and of the presence of chaotic regimes in unimodal
and subbell functions. Necessary background material is
introduced as needed, and the reader is often referred to
the existing literature for proofs of many of the basic
results.
IV. Recursions for Determining Signature Repertoires
We begin by discussing two recursions for obtaining
the realizable k-bit signatures of a given unimodal
function. By *realizable" signature we mean a signature
of a regular point or the left or right signature of an
irregular point, i.e. a signature in the exploded
repertoires of f. The first is drawn from the cited
references. Recall that the exploded k-signature
repertoire is denoted by 8k.
Theorem 1: (Klein and Kaliski, 1979) For all k >31,
S k+1 u ((0 . Sk) Cl[ k.1(0), la k+l(p)]}
U ((1 M Sk E[rs kel(p), rs k+1(1)1)
where . denotes conoatenation, denotes interseotion,
and the square brackets represent closed intervals In the
4 ordering on the (k.l)-bit sequences.
a
Page 13
Since the k-bit signatures of the points 0 and 1 under a
subbell function are respectively, 00k and 10*k-1, the
recursion is greatly simplified when f is a subbell. This
simplification leads to an alternate method of determining
realizable signatures for subbell mappings:
Theorem 2: Suppose that f is a subbell. Then for k>=1 :
(i) S1 = (0,11 and,
(ii) S k+1 = 0. S'k U 1. S'k
where S'k consists of those strings s in Sk for which s
is <= lsk(q).
Proof: We need to cite the following Lemma without proof:
Lemma 1: When f Is a subbell the exploded repertoires 3k -
are given by:
S k 1 m 0 . (Sk [O*k, Isk(q)])
U" 1. • k [ [099k, lsk(q)]}
1 __e
Page 114
(The proof of Theorem 2 resumes) Clearly, 31 is as given.
As for the expression for S k+1, observe that 3k " [O&k ,
lsk(q)] a 3'k. The expression for 3 k+1 then follows
immediately from the Lemma. QED
The last theorem underscores the significance of the left
signature of the peak q.In *signature space* a subbell
becomes a one-parameter family of sequences determined
completely by the left signature of the peak value.
V. On The Existence of Fixed Points of fk pq Having Given-
Finite and Infinite Signatures
In this section we are concerned with the existenoe
of fixed points of fk pq for various values of k, having
given finite and infinite signatures.Thes* points are
points xO for which fk pq(xO) a xO and their
characterisation is essential if one is to obtain a theory
of the orbital behavior of the function f in question.
Fixed points eventually map into themselves. We are thus
motivated to oonsider rotations of finite signatures. Our
disoussion begins here.
K
Page 15
V,1 Let a be a k-bit sequence . Let .sJ be the sequence
obtained by rotating a circularly j bits to the left,
0<=j< k. We call s'J the jth left rotation of s. We view
s 0 to be equal to a. We say that sfJ is a rotation
maximal of a, denoted ru(s), if sei <= sfj for ± =
0,1,...,k-1. Note that the existence of a rotation
maximal of any finite binary sequence is guaranteed since
<= is a linear ordering of finite binary sequences of
equal length. The value of J may not be unique however
(for example in the string s= 011011).
V92 Let us now examine sets which are such that any two
points in it have the same k-signature, for some fixed k,
under a. &.Lvo.nLtmtda 3 map,,ni. ( If the points are
irregular then they are in the set if either their left or
right signatures is the signature in question) . From a
system theoretic point of view the common k-eignatures may
be seen as defining a type of isomorphiss. It turns out
that such points form a continuum; i.e.,the sets are
intervals. This stems from the smonotonicity of
signatures' property cited earlier. In faot, the
Intervals are closed. (Klein and Kaliski, 1979).
V3 Let f be a given unimodal map. Also let a a bO bl...
be an arbitrary Infinite binary sequence. The k-sIgnature
bin of a (kfl), with respect to t# Is the set Ak a
Page 16
{x\lsk(x) and/or rak(x) a bO bl ... b k-1). The signature
bin of a with respect to f is the set A = (x\ls(x) and/or
ra(x) = s). As stated in V,2 , above, the sets Ak, if
non-empty, are closed intervals in [0,1]. It is also true
that A is a closed interval, and in fact is equal to " Ak
(taken over all values of k.) (Klein and Kaliski, 19791.
V,4 Infinite Sequences and Fixed Points of fk pq:
When considering infinite binary sequences, the concepts
of rotation and rotation maximality need to be replaced by
their infinite sequence analogues, shift and shift
maximality, respectively. Let al and s2 be infinite
binary sequences. We call s2 the jth left shift of &1
denoted LJ(sl) , if 32 is obtained from 31 by deleting the
leftmost j bits. Let S be the set of all left shifts (or
tails) of the infinite binary sequence s. Then, the shift
maximal of a is sup(S), with respeot to the < order. We
denote the shift maximal of a by sm(s). Note that, In
general, sm(s) is not an element of S.
We now turn to three key questions that arise in
exploring the Issue of tixed point existence.
QCISTIOI 1: When Is as Iatinite binary sequenoe s In the
Page 1T
exploded signature repertoire S of a given subbel. mapping
f pq?
Theorem 3: Let a be an infinite binary sequenoe. Then a is
in the exploded signature repertoire 3 of a subbell fpq if
* and only if
iiLi(s) <z 1s(q), for J>u.
Proof (-)): Suppose a is in S, and is the left and/or
right signature of x in [0,1]. Then every shift of s Is In
S , and is the left and/or right signature of a point in
the orbit of x. Suoh points are, from the form of f pq,
Fat most q. If q is not in the orbit of x then, from the
monotonioity of signatures principle, Lj(s) <= ls(q) for
all J>:l, and we are done. If q Is in the orbit of x
It Is easy to argue that all shifts of a corresponding to
Iterates of x equal to q are equal to le(q); all others,
by monotonioity of signatures, are less than or equal to
ls(q). The oonclusion follows.
(<-): .To prove the oconverse, we shall need the following:
Loma 3: Let a be a given iatinte binary sequence. Then a
Is In the exploded signature repertoire 3 of a unimodal r
pq if and only If for all k every k-trunscation of a Is In
Page 18
Sk. CBy the k-trunoation or a we mean Its first k-bits.)
Proof of Lonma 3: The necessity of the condition Is
obvious. Sufficiency requires some disocussion. Suppose
every k- truncation of a Is In Sic. Lot (LIE) ,kui,2,...
be the k-signature bins based an a. Thus the Ak are
nested and closed, and non- empty by hypothesis. Ve have
noted that A s Ak. Thus A Is non-empty also, I.e. a Is
in S. QED.
(The proof of Theorem 3 resumes) We exploit the fact that
f pq is a subbell by Invoking Theorem 2. Suppose Uj(s) <=
le(q) for .1 1,2.... Let a a bO bi.. Since bO Is
either 0 or 1, bO is in 31. Similarly bi is In S1. And
as bi is the first bit of Li(s) and Li(s) <a 3s(q),by
hypothesis,it follows that bi <a lsl(q). It is immediate
that hO hi Is in 32# by Theorem 2. Now coider bO-bi b2.
It Is easy to prove that hi b2 is In 32 by an argument
similar to that given above; further hi b2 represent the
first two bits of Li(s). Therefore < sLis a l2s(q),
bl b2 <= 1sV(q.Thua it is readily shown that bO bl b2 is
In S3, again by Theorem 2 .by repeated use of the
arguments above 'we - an, shiny that ,hO- hi ...b k-.1 the
k-truncation of a. Is a realizable finite signature of f
pq flor all k. ueaca, by hona 3, V is In S. OM3
Ve turn next to; t- 1
- ~- - -- ... -- ~T
1 -Page 19
QUESTION 2: When is an infinite signature a in S the
signature of a fixed point of a unimodal fk pq?
We have the following result:
Theorem 4: Let a u bO bl b2 ... be a given infinite binary
sequenoe in S , for some unimodal f pq . Let k>=l be
given. If a is periodic with period k then there is a
fixed point x of f pq of period k for which ls(z) and/or
rs(z) = a.
(Note that the Theorem applies, in particular, to
subbells, as well.)
Proof: We state the following lemma without proof.
Lemma* 4: Let (Bi) , izl,2,... be a collection of subsets
of 0,1] and r:[0,1 -> [0,1] a given map. Then
r( AB ) E r(Bi)
(where the intersection is taken over all 1)
(The proof of Theorem 4 esumes) Lot (Al) , Il,2,.i. be
the i-signature bins based on a. Then A "Ai. is a
nonempty olosed antepVal-.as are-the Al , aAd A a (y\la(y)
and/or ra(y) a )}. Note that Al 1,,A2 I A ... Aad beeaV6
-7
Page 20
of this nesting Amk zA also, mz1,2,...
NownotethatfkCAmk) E A (m-l)k, for all a )u 2,
because of the periodicity of a. 3o afk(Ank) ^ A (m-l)k,
where the first intersection begins with mal; the second
with m=2. but the latter is (beginning with mini) OAmk
A,so we have, ^fk(Auk) I A. But by Leoas 4, fk(aAak) E
^fk(Ank) . Therefore fk(A) ufk(Amk) E A and it is
readily shown that f' has a fixed point x of period k In A.
Clearly this point has ls(x) and/or ra(x) = a. QED
In the case where the unimodal is, In fact, a subbell an
even more powerful statement can be made.
Theorem 5: Let k>zl and finite, and s an infinite binary
sequence and fpq ,a subbell, be given. Then there is a
fixed point of fk pq with left or right signature s if
(1) am(s) <= ls(q.), an4
(ii) a is periodic, with period n such that-divide k.
Proof: The proof we give Is based on the fact that-every
periodic Infinite binary sequenoe, contains Its shift.
maximal. Suppose s is periodic with period. n which
divides k# ad am(s) (o.s(q). As. s(s) to esatalned. &s at
Lis als(q), torjs1, 24. RUeno,by_ Theorem 3. a Is
in the signare, prepevte~vse S.* *fpq s ad'thea'esu2I-
follows Imed*atSy';fi* Theeea s..
Page 21
V,5 Finite Sequences and Fixed Points of fk pq.
Before posing *QUESTION 38 we need to address several
issues. Let us begin with the following finite sequence
analogue of Theorem 3:
Theorem 6: Let f be a subbell function with peak value q.
Let a be a k-bit sequence such that rm(s) <= lak(q). Then
every rotation of a is in Sk.
Proof: In order to prove the theorem, we shall need the
following result, stated without proof: (Kwankau, 19791
Lemma 5: Let f be a subbell function with break point p.
Let a be a binary sequence of length k, such that for 3.
Op ... ,pk-1
(i) a'j <X lak(p), if sfj begins with 0,
(it) stj >= rsk(p), if stj begins with I
Then s*J Is in 3k for j 0,1,2,...k-1
(The proof of Theorem 6 resumes.) Let f be a subbellwfunotion with breakpOi~nt pe sad peak Yalue q. V. Show by a
Ut
rage 2
reductio-ad-absurdum method that if a is as defined in the
Theorem, every rotation al, of s, satisfies the following
conditions:
31 <= lsk(p), if al begins with 0
and
51 ) rsk(p), if al begins with 1.
We will then invoke Lemma 5 to complete the proof.
Assume, then, that there is some rotation &l, of s for
which either al 0 bl...b k-1 > lk(p) or sl I bl...b k-1
< rsk(p).As p 'maps into q, we know that lsk(p) = 0 is
•k-l(q) and rsk(p) a 1 Is k-l(q) Thus in the first ease 0
bl ... b k-1 > 0 Is k-1(q) and so bl ... b k-1 > is
k-i(q). Note that it z1 and z2 are binary sequenoes of
equal length with s1 > x2, then for any binary sequenoes
z3 and z4 of equal length, z1 s3 > z2 %4. Thus bl ... b
k-1 0 a s121 > lsk(q). (In the above notation al a bl
*b k-i, :2% is k-l(q), 83=0, and z4a the kth bit of ls(q).)
In the scoond ease we have that I bl ... b k-1 < I I&
k-i(q) and so bl... b-k-1 > I k-l(q.). Therefore, by the
note bl ... b k-1 I &I1l > lskCq).
"ae ths hev& tM '.1tha :forl %h& , case 81.i >'
Page 23
lsk(q), which cannot be since s1*1 <a ru~sl) <a Isk(q).'
*Therefore the original assumptions aunt be false, I.e.
every rotation sl of a must be such that al <a lsk(p) If
it begins with a 0; otherwise it must be >z rak(p). By
Lemma 5, then, every rotation of a is a realizable
k-signature of f. QED
Remark 1: Let s be a binary sequence of finite length.
Then rm(s'0 n) =rm(s)"On.
We shall need the following lemma as well in the
developments below.
Lemma 6: Let xO be a fixed point of a unimodal fk pq.
Then either ls(xO) and/or re(xO) is periodic of period k.
Let a denote the first k bits of this periodic signature.
Then there exists x1 In the orbit of xO whose left
k-signature is equal to ru(s). Furthermore, rm(s)9*2
<= lo2k(q), and, when It is equal to la2k(q), then ls(q)
Is, in fact, periodic, and equal to ru(s) rm(s)..
Proof: Clearly *ig(xO) is* periodic with period k. if
sig(xo) is regular then a is equal to sigk(zO) and it is
equally apparent that the rotation maximal of a occurs as
a subsequence of sig~xO). If this subsequence begins at
position Jin *L4(xO), J)*O, then ra(s) is the k-signature
Page 24
of xl =fj pq(xO). Furthermore sig(xl) =ru(s) ru(s)
Note that when sig(xo) is regular p and hence q cannot
occur in the orbit of xO, and thus all points in the orbit
of xO are less than q. It is immediate, from the
uonotonicity of signatures principle, that r3(&)*2(
ls2k(q). If rm(s)6"2 is < ls2k(q) there is nothing left
to prove. Assume then that ru(s)662 is equal to ls2k(q).
For ease of notation, write rm(s) as 81. Thus sig(xl) val
sl .... Again, from the montonicity of signatures
principle, sig~xl) <= 13(q). Assume that it is less than"
ls(q). We know by hypothesis that its first 2k bits agree
with those of ls(q). It must be, then , that 13(q) is of
the form slyln z, where n>=2 and where the Infinite
sequence z does not begin with sl and obeys sl**n z > al
31.. Suppose that aig(xl) and ls(q) differ for the
first time at the nlc+j th bit, with O<J<k. Let anc, If n
is even, and (n-l)k if n is odd. Consider L'm(a1 *I..
and L'm(ls(q)). The former Is again al al ... ,# whereas
the latter is z if a i.s *ven, and al z if n is odd. Note
thatr In both cases (n even, n odd) an even number of alls
were deleted, and hence al sl ... < L'm(la~q)). Since
these two sequences differ within the first 2k bits (since
a does not begin with W1 It must be that the first 2k
bits of L'm(lo(q)) are greater than sI sI w ls2k~q). So
L'u(ls(q)) > 18(q).. This Is not possible by Theorem 3,
for Js(q) is In 3. The conalusioa then is that lo(q) Is
equal to sIg(x1) and is thus periodic and equal to ru~e)
Page 25
rm(s) ...
When xO is irregular we argue as follows. p and q must
occur in the orbit of xO, and no point greater than q is
in its orbit. It is easy to demonstrate that ls(q) will be
periodic of period k and that either la(xO) or rs(xO) will
be also, depending on the parity of the initial portion of
sig(xO) before the first "-" occurs. (After this 0-8 the
remainder of both ls(xO) and rs(xO) is ls(q).) So s is the
initial k-bits of this periodic signature and rm(s) is
lsk(q), by the monotonicity of signatures principle.
Furthermore the point xl having rm(s) as its left
k-signature is q itself, and thus ls(xl) is equal to
s(q). With ls(q) periodic it is immediate that rm(s)**2
is equal to ls2k(q) and 1s(q) = rm(s) rm(a) ... QED
We are now in a position to address the final question of
section V:
QUESTION 3: When is a given k-bit sequence a the k-bit
left and/or right signature of a fixed point of fk pq?
Let us refer to the k-signature a of Lemma 6 (i.e. the
initial k bits of ls(xO) and/or rs(xO), whichever is
periodic) as the *periodic" k-signature of xO.
Theorem 7: Let a be a given k-bit binary sequenoe and fpq
Page 26
I l n
a subbell function. Then there is a fixed point of fk pq
with Operiodic" k-signature s if and only if rm(s) "*2 <=
ls2k(q), with ls(q) = rm(s) ru(s) ... if rm(s) "2 a
ls2k(q).
Proof: (-)) This is immediate from Lemma 6.
(<-) Conversely, write rm(s) as z1 to simplify the
discussion below. Consider the sequence z2 = a **n, for
any n > =2. From Remark 1, rm(z2) = z1 **n. Now, by
hypothesis, z1 902 <a la2k(q), with ls(q) = z1 z1 ... in
the case of equality. Thus, rm(z2) <= lsnk(q). By
Theorem 6, then, z2 is a realizable finite signature of
fpq . Thus every truncation of z2 is also a realizable
finite signature of fpq. Note that this is true for all n
> 1. Therefore by Lemma 3, the infiniteNsequence s s ...,
made up of repetitions of a is a realizable signature of
fpq. As a 3 ... is periodic with.'eriod k (or possibly a
factor of k), Theorem 4 implies that there is a fixed
point x, of fk pq with 1s(x) and/or rs(x) a a a s .... s
is clearly its "periodic" k-signature. QED
Signature-Distinct Uniuodals and Subbells:
A unimodal Fn fpq is said to be signature-distinot
if sig(x) = sig(y) implies that x u y. A wide-class of
signature distinot unimodals have been shown to exist
Page 27
(Klein and Kaliski, 1979) and include all those unimodals
which are "piecewise strictly expansive", i.e. for which
there exists E > I such that for all x,y in [Op],
IF(x)-F(y)I >= E Ix-y-1, and similarly for all x,y in
[p,11. A consequence of signature-distinctness is that if
x is not equal to y then no instance of sig(x) is equal to
an instance of sig(y), either. (Klein and Kaliski, 1979)
When attention is restricted to the
signature-distinct subbells a more powerful form of
Theorem 7 can be proven:
Theorem 8: Let s be a given k-bit binary sequence and f
pq a signature-distinct subbell function. Then there is a
fixed point of fk pq with "periodic" k-signature s if and
only if rm(s)0"2 <= ls2k(q) with the firther proviso,in
the case of equality, that q be a fixed point of fk pq.
1 ".i.
Proof:
(-)) If such a fixed point exists then by Theorem 7
rm(s)'*2 < a ls2k(q). In the case of equality ls(q) is
periodic with period k and equal to rm(s) rm(s) ... So, by
Theorem 4 there exists a fixed point x of fk pq whose left
and/or right signature is equal to ls(q). Since f pq is
signature-distinct x must, in fact, be equal to q.
jU-) If rm(s)e*2 is < ls2k(q) the result is immediate from
Page 28
Theorem 7. If equality holds and q is a fixed point of fk
pq, it is easy to see that ls(q) must be periodic with
period k. Since ls(q) begins with ra(s) ru(s) it must, in
fact, be the case that lo(q) a ru(s) ru(s) ... The result
is immediate from Theorem 7 then too. QED
VI On Fixed Points and Their Orders
Our attention turns in this final section to understanding
the nature of the fixed point regimes of unimodals and
subbell functions. To simplify the notation in this
section we omit the wpqw from fpq and fk pq, when
convenient to do so.
Let us begin our development by discussing the notion of
the order of a fixed point. First let us define the order
of a sequence.
A finite sequence s is of order k if there exists k
such that a can be written as al al ...81, where &l Is a k
bit sequence ,and where al cannot be similarly decomposed.
Thus a sequence of only ones (or xeros) is of order one.
The five bit sequence 10110 is of order five, whereas the
six bit sequence 010101 Is of order two and the eight bit
sequence 10111011 is of order four.*
.. . . . ..... . ......... . ....... . .... . .. ... .. ... ! L ! _ . :
Page 29
xO is a fixed point of order k of the mapping f if,
(i) fk (O) xO
and
(ii) fj (xO) <> xO, for j 1,2,...,k-1
i.e., a fixed point of order k first maps into itself
after k iterations under the given mapping. Note that if
x is a fixed point of fk for some k and if the Operiodic e
k-signature of x is of order k then x Is a fixed point of
order k.
Certainly every uniaodal map f with peak value
greater than its break point has a regular fixed point
with one-bit signature 1. This is clearly a fixed point
of order one which we call the nontrivial fixed point of
*order one . Thus the k-signature of such a point is 1'k,
and is in Sk. We are naturally motivated to ask if a
k-bit sequence containing but a single zero might be too.
Our next result shows that 101"0k-2 is, indeed, a
realizable signature provided that f obeys the
restrictions below.
Theorem 9: Let f be a unimodal map with peak value q and
break point p such that 0 and 1 are regular and such that
p
'.L
Page 30
(i) 132(q) z 10
(ii) sig2(o) - 00
and (iii) sig3(1) * 100
Then 1Ol'k-2 is a realizable k-signature of f, for k >v 2.
Proof: Let f, p, and q be as stated. Clearly the Theorem
is true for k=2. Assume then that k > 2. With 1a2(q) s
10, then q>p, Is(p) a 010... and rs(p) a 110... Now
1ek-2 is a realizable signature, by our remarks above.
Therefore, since sig k-1(0) < 01ek-2 <z Is k-1(p) we have
that 010"k-2 is too, by Theorem 1. Next onsider 101"*k-2.
Since rsk(p) < 1010ek-2 < sigk(1),.by the sane reoursion
101ek-2 is a realizable signature of f. QED
A subell satisfies all three conditions of the
theorem provided 1*2(q) a 10. Therefore,for such
subbells,101*0k-2 Is always a realizable k-signature. It
is also easily veri.fied that this sequence is of order k.
Let us restrict our attention to these subbella In the
remainder of this section, referring to then as
weoll-structured* subbells.
Theorem 10: Let f be a Ovoll-struotured* subbell. Then f
Page 31
has fixed paints at order k, k odd, and >= 3, if and only
if it has a fixed point, xO, of fk, whose *periodic*
k-signature is equal to =1Ol'0k-2.
Proof: The proof of this theorem hinges on three
properties of the sequence 101'**c-2. First, as already
nated,it is at order k. Secondly, it is Its rotation
maximal. The third property is stated as a lemma, also
given without proof. (Kwankam, 1979)
Lemma 7: Let s be a k-bit sequence other than the trivial
sequences 09*k and 1"*k. If k is odd, then
ru(s) >= 1O1'*k-2
(The proof of Theorem 10 resumfes:)
(C)By our earlier remarks .4tdoncerning orders of
sequences and orders of fixed :pOints, the result is
immediate, since 1019"k-2 is of order k.
C-)Conversely, suppose f has fixed points of order k.
Let yO be one 'such fixed point. Assume that Its
*periodic" k-signature s is not equal to 1010 *k-2. (If It
Is, we are done). Then by Lemma 7, ru(s) > 11*k-2.
Now, by Theorem 7# ra(s)*92 Is less than or equal to
1s2k(q). Thus, It In Immediate that 1O1*sk-2 1O10*k-2 <
ls2k(q). Using Theorem 7again, the result follows. ORD
kp
Page 32
We have seen, in the foregoing, conditions governing
the existence of fixed points of odd orders for
*well-structured* aubbells. We shall now show how the
existence of fixed points of order k, implies the
existence of fixed points of order k + 1, k + 2 , and k -
1, as long as k is odd and greater than one. Formally, we
have:
Lemma 8: Let f be a "well-structured" subbell. If f has a
fixed point of order k, k odd and k > 1, then f has fixed
points of orders k + 1, k + 2, and k - 1.
Proof: Let k, an odd integer greater than one be given.
Assume f has a fixed point of order k. Then by Theorem
10, there is a fixed point xO, of fk whose *periodic"
k-signature is 101ek-2. From Theorem 7, it follows that
1010#k-2 1010k-2 <a ls2k(q)
Consider the sequence 101ek-1 which we denote
by *I. It In of order k + 1. Moreover, it is Its
rotation maximal. It we can show that al a1< Is 2k+2(q) ,
then by Theorem 7, there is a fixed point of f k+1 with
'poriodio' (k.1) -SLgnature i. Ixamine the leftmost k +
2 bits of l *I and of (10106k.2)(lO1tkk-2). With k odd,
10leek-1 1 < 101Ok-2 lO.Thus the.lettmoet 1k bits of s1
I are < (10100k-2)(COi*k-2) cad hence C slk(q). so &I
Page 33
So, itel Theore 7, n mntined there Is a fixed point at
fk whose 'periodic3 k-signature is 1O1"k-1. Since this
is of order k+1 it is immediate that the fixed point In
also.
Consider ntext the sequence 10104k ,which we denote by s2.
It is at order k+2, and is its own rotation maximal also.
By considering the leftmost k+.2 bit& of s2 s2 and those of
101##k-2 1O10*k-2 and by arguing as we have done for al,
jit can be readily shown that &2 s2 < 1s 2k+4(q). Henoe f
has fixed points of order k + 2.
Finally consider the sequence 10109k-3 which we denote by
93. It is of order k - 1, and it is rotation maximal.
This time, we examine the leftmost k + I bits of s3 83 and
of (1O1**k-2)(1O100k-2) By arguments similar to -those
presented above for sl and a2, it can be shown that s3 33
< 1s 2k-2(q). Theictore, f has fixed points-of order kc -
* 1. This completes the proof. QED
Therefore when a *wejll-struatured" subbell ha~s fixed
points of o"d .ordev kc, kc > 1, i:t also has fixed points or.
orders kc - 1, kc *., and I. 2., Note that If kc SA 0*44, so
is kc + 2. ftb emistnacwot tixfed potafts of ordt k* Il
thus implies the eith.of' ftix04 40ints vt 0#40118 I' + 3and k + c 4 is, 16*r o. Alia ~c gaet bib
Page 34
repeated, ad infinttum. This leads, naturally, to theI following result:Theorem 11: Let f be as above. Then if f has a fixed
point of order kO, with k.0 > 1 and kG odd, then f has
fixed points of every order k >= (kG 1).
We conclude, from this theorem, that a
woll-structured" subbell with fixed points of
order three
has fixed points of all orders. (Every subbell has,&
fixed point of order one.) This conclusion is very such
like that of Li and Yorke (LI and Yorke, 1975). If, we
restrict our attention to signature distinct
vwell-struoturedO subbella, we can propose a method of
determining the existence of fixed points of all orders
which is considerably easier than the method implied by Li
and Yorke's theorem.
Theorem 12: Lot f be a aignature-distinct
*well-structured* subbell. Then f has fixed points of all
orders if and only if r33(q)zl00.
Proof:(<-) The sufficiency of the condition in readily
proved. Suppose r$3(q) Is 10&. If $i83(q) has no ftsh in
it, then 1s3(q) io also 100. Therefore as 101,101(
156(q), we have by ?hoores 7, that fS has a -fix.dpointt
with three-bit Uperio#1e' signature 101. This L oloarly
Bob - -----..-.-
Pass 35
*a fixed point of ordor three, which by fteorea I t means f
has fixed points of all ardoir greati' than one. As every
* subbell has a fixed point ot order one, f has fixed points
of all ordera.
If sig3Wq has a da init he ig3(q) skuit be 10-
or else 1.2(q) <> 10, a inedeaty for 'vell-utruaturedO
subbells. Theiietore- q is a, fixed point of f3
Furthermore, q It a fixed point' of- ordor 3. Thus f has
fixed points of all orders.
(->) To prove that ra3,6) a t",O is -a "necessary oontudito,
i~t ins ufficient t6 show' that If it is not not" therwe is
at least one drder oif ixed poiji~f -bich Is not-part- O~f
the 10 fixed' oit, repike '4Wr -of f-I souch or~der IS-
three. Since 100 is the Itrgesi th)-if -bit iaeno., It
must, be that rs3,(q) < 100. Since ls2(q)a10, It follows
that as2(q) is 10 and, rs3(q) and 1.3(q) must both-btV
101. Thus siC3(q) does. not contain a dash. Therefore q
.1 ot a fixed, point off3. Now, from Theorem Sp lo(q) Is
shift maia. v ia x )-Uz 101t, ti eay,** pio,'
that toe(q) 4V iM4tti
the. rr 1*.e~i~i~ia leti 04tOWSIltta-60 Li
at~ Mwe 691416 Sirt& ~ U4e~~.L ~S~~
-Ot t441 VdMP4,4tsitr ~~ .&ar zt s *e-
ip
A -1 .-
Page 36
immediate that f has no fixed points of order three from
Theorem 10. This completes the proof. QED
We therefore need examine only the right signature of
the peak value of a "well-struotured" signature distinct
subbell to determine whether or not the subbell has fixed
points of all orders. What is more important --from
considerations of ecomomy of computation-- is that we need
look at only its first three bits i.e., we need to compare
with p the values of q, f(q) and f2(q)1
The family of "symmetric tent' maps shown in Figure 3 will
exhibit this 'chaotic" behavior (fixed points of all
orders) provided that a>= (1+sqrt(5))/4, a fact that Is
readily demonstrated by showing that for such values of a,
and only for such values, will rs3(q)= 100.
VII. Conclusions
This paper has presented a variety of results
concerning the fixed point structure of certain mape on
the unit interval. The underlying common factor of the
developed theory has been that of the signature of a point
and of its role in characterizing the map's orbital
behavior. From an expositional point-of-viev the
signature-basod, theory is appealing due to its minimal
dependence on advanoed measure-theoretio concepts
A
Page 3T'
-typically found in the current literature.
iv.77
References
Kalisici, M. and Klein, Q. 'B ehavior at a Class of
Nonlinear Discrete Time Systems.*, submitted
to Mathematical Systems Theory, 1982
Klein, Q. and Kaliski, M. "Functional Equivalence in a
Class of Autonomous One-Dimensional Nonlinear
Discrete-Time Systems', Information a-ad
Control,J42,2, 1979
Collett,P. and Eckmann,J. "Iterated Maps on the Interval
as Dynamical Systems', Birkhausr, 1980.
Li,T. and !orke,J. "Period Three Implies Chaos', American
Mathematical Monthly, 82, 1975.
Parry, W. "Symbolic Dynamics and Transformations of the
Unit Interval*, Trans. American Mathematical
Society,122, 1966.
Ballieul,J., Brockett,R. and Wash%.ur., R. 'Chaotic Motion
in Nonlinear Feedback Systems*, IEEE Trans. on Circuits
* and Systems, CAS-27,11,1980.
Felgenbau,4. 'Universal Behavior in Nonlinear Systems',
Page 39
Los Alamos Science, 1980.
May, R. and Oster,G. "Bifurcations and Dynamic Complexity
in Simple Ecological Models", American Naturalist, 1976.
Kwankam, S. *A Structure Theory for the Fixed Points of
the Iterates of Certain One-Dimensional Nonlinear
Functions Defined Over the Unit Interval", Ph.D.
Dissertation, Northeastern University Dep't of
Electrical Engineering,Boston, MA, 1979.
4
At 0 17 A YNCIF NS DI SRI If COMIRO[ Uf CONT INIJIUS PROCI(I)NOR1 A;T RAN INI V BIOSTON MA M I KAIL SK I I I At
10 J H It AtOFFIR 3 87 F 49620 F82 0081
I lt ( 1 1,1 ,11 1 DI , I N
F --
A Lit
lj 2''' - "
Ut> K MW-
J.
U /k
A~~~~~n -.,., .~~
- - ~Wwo- w
zt
~-..~--- ~ - - -~ - - - - - - -.------ - - -~- - -----
IA - - -
I
'I; e
II I
I II
I
I V I
9.
.- ~
1! ~>i7i Kj <:~~4II.