Outline

Course info
Introduction to Time-Delay Systems
Time-delay systems in control applications
e sh

System-theoretic preliminaries

Basic properties

lecture no. 1 Delay systems in the frequency domain

Rational approximations of time delays

Leonid Mirkin
State space of delay systems
Faculty of Mechanical Engineering, Technion—Israel Institute of Technology

Department of Automatic Control, Lund University Modal properties of delay systems

Stability of transfer functions and roots of characteristic equations

General info Syllabus

I ECTS credits: 7.5 1. Introduction
I time delays in engineering applications; system-theoretic preliminaries
I Prerequisite: any advanced linear control course
2. Mathematical modeling of time-delay systems
I Grading policy: homework 100% (solutions are graded on a scale of I frequency domain & modal analyses; state space; rational approximations
0–100, each must be at least e  23:1407, average grade of 5 best out
3. Stability analysis
of 6 assignments must be at least 67)
I stability notions; frequency sweeping; Lyapunov’s method
Homework solutions are to be submitted electronically to mirkin@control.lth.se
4. Stabilization methods
I Literature: I fixed-structure controllers; finite spectrum assignment; coprime factorization
1. My slides.
2. J. E. Marshall, H. Górecki, A. Korytowski, and K. Walton, Time-Delay 5. Dead-time compensation
Systems: Stability and Performance Criteria with Applications. London: I Smith predictor and its modifications; implementation issues
Ellis Horwood, 1992.
6. Handling uncertain delays
3. K. Gu, V. L. Kharitonov, and J. Chen, Stability of Time-Delay Systems.
I Lyapunov-based methods; unstructured uncertainty embedding
Boston: Birkhäuser, 2003.
4. R. F. Curtain and H. Zwart, An Introduction to Infinite-Dimensional 7. Optimal control and estimation (??)
Linear Systems Theory. New York: Springer-Verlag, 1995. I H 2 optimizations
On a less formal side Outline
What this course is about. . . Course info
I system-theoretic and control aspects of delays in dynamical systems Time-delay systems in control applications
I exploiting the structure of the delay element
I giving a flavor of ideas in the field System-theoretic preliminaries
I showing that things are (relatively) simple if the viewpoint is right Basic properties

. . . and what it isn’t Delay systems in the frequency domain

I digging up most general and mathematically intriguing cases Rational approximations of time delays
I answering the very problem that motivated you to take this course
State space of delay systems

Modal properties of delay systems

Stability of transfer functions and roots of characteristic equations

Time delay element Why delays?

I Ubiquitous in physical processes
I loop delays
h t Ch t ! 0 t t!
I process delays
y D Dh u W y.t/ D u.t h/ y Dh u I ...

or I Compact/economical approximations of complex dynamics

h kCh k ! 0 k k!
yN D DN h uN W yŒk D uŒk h yN DN h uN I Exploiting delays to improve performance
Loop delays: steel rolling Loop delays: networked control

Plant

Network

Controller

Sampling, encoding, transmission, decoding need time. This gives rise to

I measurement delays
Thickness can only be measured at some distance from rolls, leading to
I actuation delays
I measurement delays

Loop delays: temperature control Internal delays: regenerative chatter in metal cutting
Waves from
previous pass
y(t-T)
Workpiece
Waves from
X current vibration
Ft h0 y(t)
+
Y

Chip h0 Tool

Fc
a Chip Y

X
y(t-T)
Ft Tool
h0+dy h0
Feed Fc
Tool Direction
y(t)

Deviation from the ideal cut satisfies

R / C c y.t
P / C ky.t / D Kf a h0 C y.t


my.t T/ y.t / ;

where T is time of one full rotation of workpiece.

Everybody experienced this, I guess. . .
Delays as modeling tool: heating a can Delays as modeling tool: torsion of a rod
Transfer function of a heated can (derived from PDE model): Transfer function of a free-free uniform rod at distance x from actuator:
1 k e xhs
Ce .2 x/hs
G.s/ D 0  x  1:
X
G.s/ D q1
C 2 s q 1 ;
s mR J1 .mR/ sC˛2m s L : s 1 e 2hs


J0 R cosh C2m

˛ mD1 ˛ 2

1
Its approximation by Gr .s/e xhs
does capture high-frequency phase lag:
Its approximation by G2 .s/ D .1 sC1/.2 sC1/
e sh is reasonably accurate:
Bode Diagram Bode Diagram
1 0.2 40 40
30 30
0.9 0.1 20 20

Magnitude (dB)

Magnitude (dB)
10 10
0.8 0 0 0
−10 −10
0.7 −0.1
−20 −20
−30 −30
0.6 −0.2
y(t) (°C)

−40 −40

imag
0.5 −0.3 −50 −50
0 0

0.4 −0.4 −720 −720

Phase (deg)

Phase (deg)
0.3 −0.5 −1440 −1440

−2160 −2160
0.2 −0.6
−2880 −2880
0.1 −0.7
−3600 −3600
0 1 0 1
10 10 10 10
0 −0.8 Frequency (rad/sec) Frequency (rad/sec)
0 1000 2000 3000 4000 5000 6000 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
time (s) real

Exploiting delays: repetitive control Exploiting delays: preview control

Any T -periodic signal f .t/ satisfies (with suitable initial conditions) If we know reference in advance, we can exploit it to improve performance:
min
1
f .t/ f .t T/ D 0 or, in frequency domain, f .s/ D sT
0:707
1 e
This motivates the configuration, called repetitive control:
sensor
y u e r 11111
00000
00000
11111
P .s/ C.s/
m
e sT

available preview 0 preview length

which is
I generalization of the internal model controllers (including I)
and guarantees (if stable!)
I asymptotically perfect tracking of any T -periodic reference r
Outline Linear systems
We think of systems as linear operators between input and output signals:
Course info
y D G u:
Time-delay systems in control applications
G is said to be
System-theoretic preliminaries
I causal if ˘ G .I ˘ / D 0, 8 , where ˘ is truncation at time 
Basic properties  t!  t!

y ˘ u
Delay systems in the frequency domain

Rational approximations of time delays I time-invariant if it commutes with shift operator: GSh D Sh G , 8h
State space of delay systems h t Ch t ! 0 t t!

y Sh u
Modal properties of delay systems

Stability of transfer functions and roots of characteristic equations I stable if 9  0 (independent of u) such that kG uk  kuk, 8u

LTI systems Transfer functions

Let g.t/ be impulse response of an LTI system G . Then In the s -domain (Laplace transform domain) convolution becomes product:
Z 1 Z t
y.t/ D g.t  /u. /d (convolution integral) y.t / D g.t  /u./d ” y.s/ D G.s/u.s/;
1 0
R1
For causal systems g.t/ D 0 whenever t < 0. Then where G.s/ D Lfg.t /g ´ g.t /e st
dt is called the transfer function of G .
0
Z t
y.t/ D g.t  /u./d Some basic definitions:
1 I Static gain: G.0/
or even I Frequency response: G. j!/ ´ G.s/jsD j!
Z t
y.t/ D g.t  /u. /d I High-frequency gain: lim sup!!1 kG. j!/k
0
I G.s/ is said to be proper if 9˛ > 0 such that sups2C˛ G.s/ < 1, where
if we assume that u.t/ D 0 whenever t < 0.

C˛ ´ fs 2 C W Re s > ˛g D ˛

I G.s/ is said to be strictly proper if lim˛!1 sups2C˛ G.s/ D 0

Transfer functions: causality & L2 -stability Rational transfer functions
A p  q transfer function G.s/ is said to be rational if 8i D 1; p and j D 1; q
Theorem
LTI G is causal iff its transfer function G.s/ is proper. b mij s mij C


C b1 s C b0
Gij .s/ D
anij s nij C


C a1 s C a0
Some definitions: for some nij ; mij 2 ZC .
R1
1=2
I L -norm of f .t/: kf k2 ´
2
0 kf .t /k dt 2
(kf k22 is energy of f .t /)
I G is said to be L2 stable if kG kL2 7!L2 ´ supkuk2 D1 kG uk2 < 1 System G is finite dimensional iff its transfer function G.s/ is rational.
H 1 ´ G.s/ W G.s/ analytic in C0 and sups2C0 kG.s/k < 1
˚
I

Some properties greatly simplified when G.s/ is rational:

Theorem
LTI G is causal and L2 -stable iff G 2 H 1 . Moreover, in this case
I rational G.s/ is proper (strictly proper) iff nij  mij (nij > mij ) 8i; j
I rational G 2 H 1 iff G.s/ proper and has no poles in C0 ´ jR [ C0
kG kL2 7!L2 D kGk1 ´ sups2C0 kG.s/k D sup!2R kG. j!/k: I high-frequency gain of rational G.s/ is kG.1/k

Note that G 2 H 1 implies that G.s/ is proper.

State-space realizations State-space realizations (contd)

Any causal LTI finite-dimensional system G admits state space realization in Assume that x0 D 0 (no history to accumulate). Then:
the sense that there are n 2 ZC and x.t / 2 Rn (called state vector) such that I impulse response of G is g.t / D Dı.t / C C eAt B ,  
A B
(
P
x.t/ D Ax.t / C Bu.t /; x.0/ D x0
I transfer function of G is G.s/ D D C C.sI 1
A/ B µ
C D
y D Gu )
y.t/ D C x.t/ C Du.t /
State-space realization is not unique. There even are realizations of the very
for some A 2 R nn
,B2R nm
,C 2R pn
,D2R pm
, and initial condition same system with different state dimensions. A realization called
x0 . In this case I minimal if no other realizations of lower dimension exist
 and .A; B; C; D/ is minimal iff .A; B/ controllable and .C; A/ observable.
Z
x.t C / D eA x.t / C eA. /
Bu.t C /d ;
0
Any two minimal realizations connected via similarity transformation:
implying that
if we know x.t/ and future inputs, we can calculate future outputs 1
I
   
A B TAT TB
x.t / 7! T x.t / ) 7 ! 1 ;
(i.e., no knowledge of past inputs required). This means that C D CT D
I state vector is history accumulator,
which changes neither impulse response nor transfer function, obviously.
which is the defining property of state vector.
System properties via state-space realizations State-space calculus
An LTI causal finite-dimensional system is These can be verified via simple flow-tracing:
I L2 stable iff its minimal realization has Hurwitz “A” matrix1 . I Addition:
This is independent of the realization chosen.
2 3
    A1 0 B1
A1 B1 A2 B2
C D 4 0 A2 B2 5
C2 D1 C2 D2
C1 C2 D1 C D2
I Multiplication:
The “D ” matrix is informative too:
2 3 2 3
A1 0 B1 A2 B2 C1 B2 D1
 
A B
is strictly proper iff D D 0
  
I G.s/ D A2 B2 A1 B1
C D D 4 B2 C1 A2 B2 D1 5 D 4 0 A1 B1 5

A B
 C2 D2 C1 D1
I High-frequency gain of G.s/ D is exactly kDk D2 C1 C2 D2 D1 C2 D2 C1 D2 D1
C D
I Inverse (exists iff det D ¤ 0):
1
BD 1 C BD 1
   
A B A
D
C D D 1C D 1

1A matrix is said to be Hurwitz if it has no eigenvalues in C0 (closed right-half plane).

Schur complement Outline

If det M11 ¤ 0,
Course info
I M111 M12
     
M11 M12 I 0 M11 0
M D D ; Time-delay systems in control applications
M21 M22 M21 M111 I 0
11 0 I
System-theoretic preliminaries
while if det M22 ¤ 0,
Basic properties
I M12 M221
     
M11 M12
22 0 I 0
M D D ;
M21 M22 0 I 0 M22 M221 M21 I
Delay systems in the frequency domain
where
Rational approximations of time delays

11 ´ M22 M21 M111 M12 and
22 ´ M11 M12 M221 M21
State space of delay systems
are Schur complements of M11 and M22 , respectively. Consequently,
Modal properties of delay systems
det M D det M11 det
11 D det M22 det
22 ;
Stability of transfer functions and roots of characteristic equations
provided corresponding invertibility holds true.
Time delay element: continuous time Transfer function of Dh
By the time shifting property of the Laplace transform:
h t Ch t ! 0 t t!
sh
y D Dh u W y.t/ D u.t h/ y Dh u
y.t / D u.t h/ ” y.s/ D e u.s/

Thus
Causality follows by Dh .s/ D e sh
;
Dh ˘ D ˘ Ch Dh ) ˘ Dh .I ˘ / D ˘ .I ˘Ch /Dh : which is irrational.

Since ˘ ˘Ch D ˘ , we have that

Transfer function of Dh is
˘ Dh .I ˘ / D .˘ ˘ /Dh D 0 I entire (i.e., analytic in whole C)
Time-invariance follows by
I bounded in C˛ for every ˛ 2 R
Hence,
Dh S D DhC D S Dh e sh
2 H1

Impulse response is clearly ı.t h/

Time delay element: discrete time Outline

Course info
h kCh k ! 0 k k!
yN D DN h uN W N D uŒk
yŒk N h yN DN h uN Time-delay systems in control applications

System-theoretic preliminaries

By the time shifting property of the ´-transform: Basic properties

h
N D uŒk
yŒk N h ” y.´/
N D´ N
u.´/ Delay systems in the frequency domain

Thus Rational approximations of time delays

1
DN h .´/ D h ;
´ State space of delay systems
which is rational (of order h, all h poles are at the origin, no finite zeros).
Modal properties of delay systems

Stability of transfer functions and roots of characteristic equations

Frequency response Effects of I/O delay on rational transfer functions
Obviously, Consider the simplest interconnection:
sh j!h
e jsD j! D e D cos.!h/ j sin.!h/
sh
L.s/ D Lr .s/e for some rational Lr .s/:
Has
I unit magnitude (je j!h j  1) and In this case L. j!/ D Lr . j!/e j!h , meaning that
I linearly decaying phase (arg e j!h D !h, in radians if ! —in rad/sec)
jL. j!/j D jLr . j!/j and arg L. j!/ D arg Lr . j!/ !h:
Bode Diagram Nichols Chart
2
In other words, delay in this case
Magnitude

does not change the magnitude of Lr . j!/ and

0
0
I

adds phase lag proportional to ! .

Open−Loop Gain (dB)

−2 I

0 −4

−180
Phase (deg)

−6

−360
−8

−540
0 1
10 10 −540 −360 −180 0
Frequency (rad/sec) Open−Loop Phase (deg)

Effects of I/O delay: Bode diagram Effects of I/O delay: Nyquist diagram
Magnitude (dB)

Imaginary Axis
0
Phase (deg)

−1 0 1
10 10 10 −1 0
Frequency (rad/sec) Real Axis
Effects of I/O delay: Nichols chart It’s not always so easy
If delay not I/O, frequency response plots might be much more complicated.
Consider for example the (stable) system:
1 e 2s 1.t / 1.t 2/
G.s/ D ; with impulse response g.t / D ;
2s 2
Open−Loop Gain (dB)

0
Bode Diagram Nyquist Diagram
0
0

Magnitude
−20

−0.2
−40

Imaginary Axis
0
−0.4

Phase (deg)
−90

−0.6

−540 −360 −180 0 −180

Open−Loop Phase (deg) 10
−1
10
0 1
10 −0.2 0 0.2 0.4 0.6 0.8 1
Frequency (rad/sec) Real Axis

Outline Why to approximate

Delay element is infinite dimensional, which complicates its treatment. It is
Course info
not a surprise then that we want to approximate delay by finite-dimensional
Time-delay systems in control applications (rational) elements to
I use standard methods in analysis and design,
System-theoretic preliminaries I use standard software for simulations,
Basic properties I avoid learning new methods,
I ...
Delay systems in the frequency domain

Rational approximations of time delays

State space of delay systems

Modal properties of delay systems

Stability of transfer functions and roots of characteristic equations

What to approximate: bad news What to approximate: good news
On the one hand, Yet we never work over infinite bandwidth. Hence, we
I phase lag of the delay element is not bounded (and continuous in ! ). I need to approximate e sh
in finite frequency range
On the other hand, or, equivalently,
I rational systems can only provide finite phase lag. I approximate F .s/e sh
for low-pass (strictly proper) F .s/.
Therefore, phase error between e and any rational transfer function R.s/
sh

is arbitrarily large. Moreover, for every R.s/ there always is !0 such that This can be done, since
arg e j!h arg R. j!/ continuously decreasing function of ! , 8!  !0 .
I I phase lag of delay over finite bandwidth is finite
Hence there always is frequency !1 such that and
R. j!1 / I magnitude of F . j!/e j!h decreases as ! increases,
j!1 h
arg e arg R. j!1 / D  2k i.e.,
which implies that at frequencies where the phase lag of F . j!/e j!h large,
the function effectively vanishes.
j!1 h
e

This, together with the fact that je j!h j  1, means that

I rational approximation of pure delay, e sh
, is pretty senseless
as there always will be frequencies at which error2 is  1 (i.e.,  100%).
Also, we may consider h D 1 w.l.o.g., otherwise s ! s= h makes the trick.
2 Thus, we never can get better approximation than with R.s/ D 0 . . .

Truncation-based methods Truncation-based methods: naı̈ve approach

General idea is to Note that
e s=2
I truncate some power series, e s
D
es=2
which could give accurate results in a (sufficiently large) neighborhood of 0.
and truncate Taylor series of numerator and denominator. We could get:
Pn 1 i
s i D0 iŠ2i . s/
e  Pn 1 i
iD0 iŠ2i s

This yields:

n 1 2 3 4
sh sh s 2 h2 sh s 2 h2 s 3 h3 sh s 2 h2 s 3 h3 s 4 h4
sh 1 1 1 2 C 8 1 2 C 8 48 C 384
e  2
1C sh
2 8
2 2 2 2
48
3 h3 2 2 s 3 h3 4 h4
2 1C 2 C s 8h
sh
1C 2 C s 8h
sh
C s 48 1C 2 C s 8h
sh
C 48 C s384

For n D 2 called Kautz formula. But

I becomes unstable for n > 4 !
Truncation-based methods: Padé approximation Example: Œ2; 2-Padé approximation
s 2 q1 sCq0
Consider approximation In this case RŒ2;2 .s/ D s 2 Cq1 sCq0
and Taylor expansions are

s Pm .s/ s2 s3 s4
e  µ RŒm;n .s/; e s
D1 sC C



Qn .s/ 2 6 24
2q1 2q 2 2.q13 q1 q0 / 3 2.q14 2q12 q0 /
where Pm .s/ and Qn .s/ are polynomials of degrees m and n, respectively. RŒ2;2 .s/ D 1 s C 21 s 2 s C s4



3
Taylor expansions at s D 0 of each side are q0 q0 q0 q04

s s2 s3 from which
e s
CD1 C


q1 2 1
1Š 2Š 3Š q0 D 2q1 and D
4q1 6
0 00 000
RŒm;n .0/s RŒm;n .0/s 2 RŒm;n .0/s 3
RŒm;n .s/ D RŒm;n .0/ C C C C


and then q1 D 6 and q0 D 12, matching 5 coefficients.
1Š 2Š 3Š

Thus, Œ2; 2-Padé approximation is

The idea of Œm; n-Padé approximation is to find coefficients of RŒm;n .s/ via
2
I matching first n C m C 1 Taylor coefficients s s 2 6s C 12 1 2s C 12s
e  2 D :
of these two series. If n D m, it can be shown that Pn .s/ D Qn . s/. s C 6s C 12 s2
1 C 2s C 12

Truncation-based methods: Padé approximation (contd) Padé approximation: example

1
General formula for Œn; n-Padé approximation is Consider Padé approximation of sC1 e s . This can be calculated by Matlab
Pn Pn function pade(tf(1,[1 1],’InputDelay’,1),N).
n .2n i/Š .2n i/ŠnŠ
s/i s/i


s i D0 i .2n/Š
. iD0 .2n/Š.n i /Š i Š .
e  Pn n .2n i/Š

D Pn .2n i/ŠnŠ
iD0 i .2n/Š
si iD0 .2n/Š.n i/Š iŠ si e s
and its 2nd and 4th order approximations Corresponding approximation errors
sC1
Bode Diagram Bode Diagram

This yields: 0 0

−5

Magnitude (dB)
−10 −20
n 1 2 3 4 −15
sh s 2 h2 s 2 h2 s 3 h3 s 2 h2 s 3 h3 s 4 h4
sh 1 1 sh 1 sh2 C 10 1 sh2 C 28 84 C 1680

−20 −40
e 1C sh
2 2 12
s 2 h2 s 2 h2
120
3 h3 s 2 h2 3 3 4 h4

Magnitude (dB)
2 1C sh
2 C 12 1C sh
2 C 10 C s120 1C sh
2 C 28 C s 84
h
C s1680 −25

−30 −60
0

Using Routh-Hurtitz test one can prove that −180

−360 −80
Phase (deg)

−540
I Œn; n-Padé approximation stable for all n. −720
−100
−900

−1080

−1260 −120
−1 0 1 −1 0 1
10 10 10 10 10 10
Frequency (rad/sec) Frequency (rad/sec)
Padé approximation: example (contd) Padé approximation: example (contd)
We may also compare step responses and Nichols charts Increasing approximation order improves the match between step responses
1
s
of sC1 e s and its Padé approximation:
e
sC1 and its 2nd and 4th order approximations
Step Response Nichols Chart
s
1 10
e
sC1 and its 50th order approximation e s and its 50th order approximation
5 Step Response Step Response
0.8
1 1.2
0

0.6 1

Open−Loop Gain (dB)

−5 0.8
Amplitude

0.8
0.4 −10
0.6
0.6

Amplitude

Amplitude
−15
0.2
0.4 0.4
−20

0 0.2
−25 0.2

0
−0.2 −30
0 0.5 1 1.5 2 2.5 3 3.5 4 −720 −630 −540 −450 −360 −270 −180 −90 0 0
Time (sec) Open−Loop Phase (deg) −0.2

−0.2 −0.4

From loop shaping perspectives,

0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4
Time (sec) Time (sec)

I approximation performance depends on crossover requirements.

Not true for the approximation of the pure delay e s
!

Outline State equation with input delay

Course info if input delayed by h
P / D Ax.t / C Bu.t /
x.t ! P / D Ax.t / C Bu.t
x.t h/
Time-delay systems in control applications Solution then becomes:
System-theoretic preliminaries
Z t C
A
x.t C / D e x.t / C eA.tC /
Bu. h/d
t
Basic properties Z tC
!
A
De x.t / C eA.t /
Bu. h/d
Delay systems in the frequency domain t
!
Z t hC
Rational approximations of time delays De A
x.t / C e A.t h /
Bu. /d
t h
State space of delay systems Z t Z t hC
!
A A.t h / A.t h /
De x.t / C e Bu. /d C e Bu. /d
Modal properties of delay systems t h t

Stability of transfer functions and roots of characteristic equations It depends on initial “state” x.t /, future inputs over Œt; t h C   (if  > h),
and past inputs over Œt h; t  (or Œt h; t h C  if  < h).
“In my country there is problem. . . ” (B. Sagdiyev) Checking intuition on discrete-time case
. . . and that problem is: Consider
I x.t / no longer accumulates the history. N C 1 D ANxŒk
xŒk N C BN uŒk
N h:
This, in turn, implies that x.t/ can no longer be regarded as the “state”. This system can be thought of as serial interconnection

xN ´ h uN uN
PN0 .´/ DN h .´/

Intuitively, the “true” state vector at every time instance t should contain where
I x.t/ 2
0 0



I 0 0
3
I u.t C / for all  2 Œ h; 0—denoted u .t / 60 I



0 0 07
6 :: ::
:: : : :::: 7
6 7
AN BN
 
PN0 .´/ D and DN h .´/ D 6 :
6 :
: : : :7
I 0
7
60
6 0 0


I 07 7
40 0 0


0 I5
I 0 0


0 0

Checking intuition on discrete-time case (contd) Checking intuition on discrete-time case (contd)
To recover the state vector, write state equation:
xN ´ h uN uN
PN0 .´/ DN h .´/ 3 2 N
N C 1
xŒk A BN 0 0 0


N
xŒk 0
2 32 3 2 3
N
6 uŒk h C 1 7 60 0 I 070



N
6 uŒk h 7 607
Thus, for PNh .´/ ´ PN0 .´/DN h .´/ we have:
6 7 6 76 7 6 7
6N h C 2 7



6 uŒk 6N h C 1 7
7 60 0 0 07I7 6 uŒk 7 607
6 6 7
:: D
7 6 :: :: :: ::
:: 7 :: :: C N
7 6 :: 7 uŒk
AN BN
6 6
:
2 3
0 00


0 6 : 7 6: : : :7: 6 : 7 6:7
0 I 0


0 0
2 3 6 7 6 76 7 6 7
60 0 I 007


0 N
4 uŒk 1 5 4 0 0 0 0


I 5 4 uŒkN 2 5 4 0 5
60 0 I


0 07 6 7
N
uŒk 0 0 0 0


0 N
uŒk 1 I
7 6 0 0 0 I07


0
7
6 : :::: : : :: :: 7 6
N BN 6
6 ::

A : : : : :7 6: :: :: ::
:: 7:: ::
PNh .´/ D
„ ƒ‚ … „ ƒ‚ …
7D6
6 :: : : ::7
7 : : xN a ŒkC1 xN a Œk
I 0 60
6
0 0


I 07
60
7 6 0 0 0


I 07
6 7
40 0 0


0 I5 6 Thus, the state vector at time k , xN a Œk, indeed
0 0 0


0 I 5
7
40
I 0 0


0 0 I includes both xŒk
N and whole input history uŒi
N in k hi k 1
I 0 0 0


0 0
State equation with input delay (contd) State equation with output delay
Thus, the state vector of Consider
P / D Ax.t / C Bu.t /
x.t
P / D Ax.t / C Bu.t
x.t h/
and assume that that we measure delayed x , i.e., y.t / D x.t h/.
at time t is .x.t/; u .t// 2 .R ; fŒ h; 0 7! R g/. This implies (among many
n m Discrete counterpart looks like this:
other things) that
´ h xN xN uN
DN h .´/ PN0 .´/
I initial conditions for this system are .x.0/; u .0//,
which is also a function. For example,
with
I zero initial conditions should read 2 3
0 I 0


0 0 0
0 I 0


0 0
2 3
x.0/ D 0 and u. / D 0; 8Œ h; 0: 60 0 I


0 0 07
60 0 I


0 07 6
6: :: :: : : : ::
7
:: 7
6 ::
6
:::: : : :: :: 7 AN BN
7  6 :: : : : :: : :7
PNh .´/ D 6 : : : : : :7 6
D6
7
0


I
6
7
I 0 60 0 0 07
60 0 0


I 07
7
60 0 0


0 I 07
6 7
There is more consistent (and elegant) way to reflect all this via state-space
6 7
40 0 0


0 I
AN BN 5
5
0


0
6 7
40 0
description, using semigroup formalism. See (Curtain and Zwart, 1995) for I 0 0


0 0
details. I 0 0 0


0 0

State equation with output delay (contd) State equation with output delay (contd)
Returning to continuous-time case, state vector of

N
xŒk h C 1 0 I 0


0 0 N
xŒk h 0
2 3 2 32 3 2 3 (
P / D Ax.t / C Bu.t /
x.t
6N h C 2 7 6 0 I


0 N h C 1 7 6 0 7
6 xŒk 0 0 7 6 xŒk
y.t / D x.t
7 6 7 6 7 6
6 :: 7 6 :: :::: : : :: :: 7 6 :: 7 6 :: 7
7 h/
: 7 6: : : : : : 76 :
7D6 7 C 6 : 7 uŒk
N
6 7 6 7 6 7
at time t is x .t / 2 fŒ h; 0 7! Rn g (may be convenient to write .x.t /; x .t //).
6
6 N 0 0


I N
6 xŒk 1 7 7 60 0 7 6 xŒk 2 7 6 0 7
6 7 6 7 6
7
4 N
xŒk 5 40 0 0


0 I 5 4 xŒk N 1 5 4 0 5
N C 1
xŒk 0 0 0


0 AN N
xŒk BN
„ ƒ‚ … „ ƒ‚ … The
xN a ŒkC1 xN a Œk
I initial condition is then the function x .0/
and zero initial conditions would mean

x./ D 0; 8 2 Œ h; 0:
Thus, the state vector at time k , xN a Œk,
I includes whole history of xŒi
N  in k hi k 1
State delay equations Homogeneous LTI state equations: classification
If we use a P controller u.t/ D Ky.t /, the closed loop system becomes (Lumped-delay) retarded equation:
r
P / D Ax.t / C BKx.t
x.t h/: X
P /D
x.t Ai x.t hi /; 0 D h0 < h1 <


< hr D h
This kind of equations called retarded functional differential equation. iD0

(Lumped-delay) neutral equation:

r
X r
X
If we use a D controller u.t/ D K y.t
P /, the closed loop system becomes P
Ei x.t hi / D Ai x.t hi /; 0 D h0 < h1 <


< hr D h
iD0 i D0
P
x.t/ D Ax.t/ C BK x.t
P h/ or x.t
P / P
BK x.t h/ D Ax.t /
Distributed-delay retarded equation:
This kind of equations called neutral functional differential equation. Z 0
P /D
x.t ˛./x.t C /d
h

If ˛./ D Ai ı.t C hi /, we have the lumped-delay equation above.

P
i

I In all cases the “true” state is x .t / 2 fŒ h; 0 7! Rn g.

Adding inputs and outputs The same in s domain (with zero initial conditions)
(Still not the most) general form: (Still not the most) general form:
„ Z 0 Z 0 Z 0
„  Z 0  Z 0 Z 0
s s
P C
x.t/ P C /d D
./x.t ˛. /x.t C  /d C ˇ. /u.t C /d s IC ./e d X.s/ D ˛./e dX.s/ C ˇ. /es d U.s/
hx hx hu hx hx hu
Z 0 Z 0 Z 0 Z 0
s
y.t / D . /x.t C /d C ı. /u.t C  /d Y.s/ D ./e dX.s/ C ı. /es dU.s/
hx hu hx hu

with the state “vector” .x .t /; u .t // 2 fŒ hx ; 0 7! Rn g; fŒ hu ; 0 7! Rm g . with zero initial conditions x./ D 0 ( 2 Œ hx ; 0), u./ D 0 ( 2 Œ hu ; 0).

Important special (lumped-delay) case: Important special (lumped-delay) case:

†X rx rx
X ru
X
† rx
X rx
X ru
X
shi shi shi
P
Ei x.t hi / D Ai x.t hi / C Bi u.t hi / s Ei e X.s/ D Ai e X.s/ C Bi e U.s/
iD0 iD0 iD0 iD0 iD0 i D0
Xrx Xru Xrx Xru
shi shi
y.t / D Ci x.t hi / C Di u.t hi / Y .s/ D Ci e X.s/ C Di e U.s/
iD0 i D0 iD0 iD0

with E0 D I and 0 D h0 < h1 <


< hmaxfrx ;ru g D h. with E0 D I and 0 D h0 < h1 <


< hmaxfrx ;ru g D h.
Outline Characteristic equation
Distributed-delay equation
Course info
 Z 0  Z 0 Z 0
Time-delay systems in control applications s IC s
./e d X.s/ D ˛. /es dX.s/ C ˇ. /es dU.s/
hx hx hu

System-theoretic preliminaries
(or its lumped-delay counterpart) can be rewritten as
Basic properties Z 0 ru
X
X.s/ D
1 .s/ ˇ./es dU.s/ or X.s/ D
1 .s/ Bi e shi
U.s/;
Delay systems in the frequency domain hu iD0
where
Rational approximations of time delays Z 0 rx
X

.s/ ´ s.I C . // ˛./ es d or
.s/ ´ shi


.sEi Ai / e
State space of delay systems hx i D0

Modal properties of delay systems called the characteristic matrix. Then equation

Stability of transfer functions and roots of characteristic equations detŒ
.s/ µ .s/ D 0

called characteristic equation.

Example 1 Example 2
Consider     Consider 
  
0 1 0 0 0 0 0 1
P /D
x.t x.t/ x.t h/ P /C
x.t P
x.t h/ D x.t /
0 1 k1 k2 k1 k2 0 1
Then Then
               
1 0 0 1 0 0 sh s 1 1 0 0 0 sh 0 1 s 1

.s/ D s C e D sh sh
.s/ D s Cs e D sh sh
0 1 0 1 k1 k2 k1 e s C 1 C k2 e 0 1 k1 k2 0 1 sk1 e s C 1 C sk2 e

and and
2 sh
.s/ D s C s C .k2 s C k1 /e .s/ D s 2 C s C .k2 s 2 C k1 s/e sh
Example 3 Quasi-polynomials
Consider 
   General form:
0 1 k1 0
P /D
x.t x.t/ x.t h/ rx
!
0 1 0 k2 X
s hQ i
X
shi
det .sEi Ai /e D P .s/ C Qi .s/e
Then iD0 i

s C k1 e sh for polynomials P .s/ 6 0 and Qi .s/ (9j , Qj .s/ 6 0) and delays hi > 0.
       
1 0 0 1 k1 0 sh 1

.s/ D s C e D sh
0 1 0 1 0 k2 0 s C 1 C k2 e
Classification by delay pattern:
and 1. single-delay: P .s/ C Q.s/e sh
2 sh 2sh
.s/ D s C s C .k1 C k2 /s C k1 e C k1 k2 e


2. commensurate-delay: P .s/ C sih
P
i Qi .s/e
hi
3. incommensurate-delay: if at least one of hj
is irrational

Classification by principal degrees of s :

1. retarded: deg P .s/ > deg Qi .s/, 8i
2. neutral: deg P .s/  deg Qi .s/, 8i , and 9j s.t. deg P .s/ D deg Qj .s/
3. advanced: 9j s.t. deg P .s/ < deg Qj .s/

Roots of quasi-polynomials Roots of quasi-polynomials: where are they located

Apparently, the simplest example is Some fundamental properties:
sh
1. there is a finite number of roots within any finite region of C
1 C ke ; k > 0:
(meaning there are no accumulation points for roots of quasi-polynomials)

It is readily seen that it has 2. roots for large values of jsj belong to a finite number of areas
I infinite number of roots, ˇ ˇ
ˇRe s C ˇi lnjsjˇ <
those at
ln k .1 C 2i / for some > 0 and ˇi 2 R. For
sD Cj ; i 2Z
h h I retarded quasi-polynomials ˇi > 0
N 0 iff k  1 and in C n C
(all these roots are in C N 0 iff 0 < k < 1). I neutral quasi-polynomials ˇi D 0
I advanced quasi-polynomials ˇi < 0
3. retarded quasi-polynomials have a finite number of roots in C˛ for all ˛
This is generic property, i.e.,
I quasi-polynomials have infinite number of roots.
Rightmost root: retarded case Rightmost root: neutral case
Consider Consider
rx
X
P
x.t/ D Ai x.t hi /; x .0/ D . / rx
X rx
X
iD0 P
Ei x.t hi / D Ai x.t hi /; x .0/ D . /
iD0 iD0
and let .s/ be its characteristic quasi-polynomial. Define
and let .s/ be its characteristic quasi-polynomial. Define
r ´ maxfRe s W .s/ D 0g
r ´ supfRe s W .s/ D 0g
Theorem
For any  > r there is a  > 0 such that Theorem
For any  > r there is a  > 0 such that
kx.t /k   et max k. /k; 8t 2 RC
 2Œ h;0
P
kx.t /k   et max .k. /k C k./k/; 8t 2 RC
2Œ h;0
for all continuous initial conditions  .
for all continuous and differential initial conditions  .

Outline Simple special case

Special (SISO single-delay) case:
Course info
N0 .s/
Time-delay systems in control applications G.s/ D ;
M0 .s/ C Mh .s/e sh

System-theoretic preliminaries for some real polynomials M0 .s/, Mh .s/, N0 .s/ such that deg M0  deg Mh
and deg M0  deg N0 . This transfer function
Basic properties
I is proper
Delay systems in the frequency domain I is analytic on C n , where  is set of roots of M0 .s/ C Mh .s/e sh
D0
I has only poles as its singularities
Rational approximations of time delays
To further simplify matters, assume also that there is
State space of delay systems I no pole / zero cancellations in G.s/

Modal properties of delay systems

Stability of transfer functions and roots of characteristic equations

Asymptotic poles location j a j < 1
Mh
Let a ´ Pa .1/, where Pa ´ M0
(j a j is the high-frequency gain of Pa ). In this case, G.s/ has at most finitely many unstable poles. Moreover, the
following result can be formulated:
Theorem
lnj a j
As jsj ! 1, poles of G.s/ are asymptotic to points with Re s D h
. Lemma
N 0.
Let j a j < 1. Then G 2 H 1 iff G.s/ has no poles in C
Proof.
Poles are solutions of the characteristic equation M0 .s/ C Mh .s/e sh
D 0 or Proof (outline).
N 0 , it does not belong to H 1 .
If G.s/ has a pole in C
esh D Pa .s/ D a C O 1


s If G.s/ has no poles in C N 0 , there is a bounded S  CN 0 such that in C
N0nS
1. M0 .s/ has no roots
By Rouché’s arguments, as jsj ! 1 roots approach solutions of esh D a , ˇ N .s/ ˇ
i.e., sh D ln. a / C j2k , k 2 Z. Thus, 2. jPa .s/j < b and ˇ M0
0 .s/
ˇ < c for some j a j  b < 1 and c > 0.
( Thus, jG.s/j is bounded in S (no poles and the set is bounded) and
lnj a j C j2k if a < 0
sh ! jN0 .s/=M0 .s/j c
lnj a j C j.2k C 1/ if a > 0 jG.s/j D < < 1; N 0 n S:
8s 2 C
j1 C Pa .s/e j
sh 1 b
which for a ¤ 0 approaches vertical line with Re s D lnj a j and for a D 0
Hence, G.s/ is analytic and bounded in C0 .
approaches Re s D 1.

j a j > 1 j a j D 1: example 1
1
In this case, G.s/ has infinitely many unstable poles. Hence, the following Let G.s/ D sC1C.sC2/e s . This transfer function
result can be formulated: I has infinitely many poles in C0 , hence unstable.
Lemma
Let j a j > 1. Then G 62 H 1 .

Proof. To see this, consider its characteristic equation at s D  C j! :

Obvious. s
 C j ! C 2 . C 2/2 C ! 2
e e j! D H) e D
 C j! C 1 . C 1/2 C ! 2

Right-hand side here > 1 iff  > 23 . Hence, whenever  > 32 , e > 1 or,
equivalently,  > 0. Since roots accumulate around  D 0, there might be
only a finite number of roots in   32 .
j a j D 1: example 2 j a j D 1: example 2 (contd)
Let G.s/ D 1
. This transfer function
s
N 0 of poles of G.s/ satisfying
We know that there is a sequence fsk g 2 C n C
sC1Cs e
I N 0.
has no poles in C

sk
sk C 1 C sk e D 0; with limk!˙1 jsk j D 1 W
To see this, consider its characteristic equation at s D  C j! :

 C j! C 1
ˇ ˇ ˇ ˇ
 j! 
ˇ 1 ˇˇ ˇˇ  ˇ
e e D H) e D ˇ1 C ˇ  ˇ1 C 2 ˇ
Then
 C j!  C j! ˇ  C! ˇ
2
1 1
G. sk / D D D 1 C sk ;
If  D 0, 1 D 1 C 1=j!j: unsolvable. If  > 0, e 
 1: contradiction. 1 sk sk e sk 1 sk C sk2 =.1 C sk /
so that jG. sk /j ! 1 C j2k C 1j . Thus, G.s/ unbounded on f sk g 2 C0 , so
I G 62 H 1 and hence unstable.

j a j D 1: example 3 j a j D 1: on the safe side

1
Let G.s/ D .sC1/.sC1Cs e s/ . In this case Thus, we saw that in this case G.s/ might be stable. Yet this
1
I G2H I stability is fragile (extremely non-robust).
(in fact, kG.s/k1 D 2). Hence, Indeed,
I G.s/ is stable. I infinitesimal increase of jPa .1/j leads to instability.
It is thus safe to regard such systems as practically unstable.

In general, if j a j D 1, the transfer function N0 .s/

M0 .s/CM1 .s/e sh 2 H 1 only if3

deg M0 .s/  deg N0 .s/ C 2:

3 For proof and further details see “H 1 and BIBO stabilization of delay systems of neutral

type,” by Partington and Bonnet, Systems & Control Letters, 52, pp. 283–288, 2004.
Internal stability and high-frequency gain Summary
So far we learned that
wu
Pr .s/e sh
I N 0,
if j a j < 1, G 2 H 1 iff it has no poles in C
I if j a j > 1, G 62 H 1 because it has infinitely many poles in C0
u C.s/ y wy I if j a j D 1, G.s/ practically unstable

 wy 
This feedback system called internally stable if transfer matrix u ,
y
wu 7!

Pr .s/e sh
 
1
1
2 H1 Important point:
1 Pr .s/C.s/e sh
C.s/ Pr .s/C.s/e sh
I N 0 ” stability criterion might fail
classical “no poles in C
Its .1; 1/ entry, the sensitivity function, is of the form
1 1 ML .s/ N0 .s/
1 Pr .s/C.s/e sh µ 1 Lr .s/e sh D ML .s/ NL .s/e s µ M0 .s/ M1 .s/e s ;

for which deg M0 .s/ D deg N0 .s/ (6 deg N0 .s/ C 2). Thus, if jLr .1/j D 1,
I this system cannot be internally stable.

Slight modification doing the trick

Theorem
Transfer function
N0 .s/
G.s/ D deg M0  max deg Mh ; deg N0 ;
˚
;
M0 .s/ C Mh .s/e sh

N ˛.
is (practically) stable iff 9˛ < 0 such that G.s/ has no poles in C

This result says that in time-delay systems (both retarded and neutral)
I poles play essentially the same role as in the rational case,
N0 !C
we just should slightly redefine stability region (C N ˛ , for some ˛ < 0).
This, in turn, makes it possible to
I extend classical stability analysis methods to time-delay systems.

N 0 , yet make sure that

We still may use C
I no pole chain is asymptotic to Re s D 0.