1586

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 5, SEPTEMBER 2013

Proportional-Integral Control of First-Order

Time-Delay Systems via Eigenvalue Assignment
Sun Yi, Patrick W. Nelson, and A. Galip Ulsoy

Abstract A new design method is presented for proportionalintegral (PI) controllers of first-order plants in the presence
of time delays. In general, time delays can limit and degrade
the achievable performance of the controlled system, and even
induce instability. Thus, the PI gains should be selected carefully
considering such effects of time delays. Unlike existing methods,
the design method presented in this paper is based on solutions
to delay differential equations, which are derived in terms of
the Lambert W function. PI controllers for first-order plants
with time delays are designed by obtaining the rightmost (i.e.,
dominant) eigenvalues in the infinite eigenspectrum of timedelay systems and assigning them to desired positions in the
complex plane. The process is possible due to a novel property
of the Lambert W function. The controllers designed using
the presented method can improve the system performance
and successfully stabilize an unstable plant. Also, sensitivity
analysis of the rightmost eigenvalues is conducted, and the results
compare favorably with those of a prediction-based method
for eigenvalue assignment. Extension to design of PI-differential
controllers is discussed with examples.
Index Terms Delay differential equation, eigenvalue
assignment, Lambert W function, proportional-integral control,
time delay.

I. I NTRODUCTION

VARIETY of methods have been developed to

design
proportional-integral/proportional-integraldifferential (PI/PID) controllers to meet the specifications for
time delay systems, such as time-domain specifications and/or
disturbance rejection [1]. Such methods use experimental
tuning, pole placement in the complex plane, plant response in
the frequency domains, and/or graphical methods. However,
when time delays exist between the application of inputs
and their resulting effects, stability analysis is not as
straightforward as it is for nondelayed systems. Therefore,
the design of feedback control becomes difficult [2]. The
principal difficulty is due to the fact that the time-delay terms
always lead to an infinite eigenspectrum.

Manuscript received June 8, 2011; revised May 25, 2012; accepted August 9,
2012. Manuscript received in final form August 27, 2012. Date of publication September 27, 2012; date of current version August 12, 2013. This
work was supported by National Science Foundation under Grant 0555765.
Recommended by Associate Editor R. Landers.
S. Yi is with the Department of Mechanical Engineering, North Carolina
Agricultural & Technical State University, Greensboro, NC 27411 USA
(e-mail: syi@ncat.edu).
P. W. Nelson is with the Center for Computational Medicine and Bioinformatics, University of Michigan, Ann Arbor, MI 48109 USA (e-mail:
pwn@umich.edu).
A. G. Ulsoy is with the Department of Mechanical Engineering, University
of Michigan, Ann Arbor, MI 48109 USA (e-mail: ulsoy@umich.edu).
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TCST.2012.2216267

The stability analysis for time-delay systems can be conducted using bifurcation methods (see [3] and the references
therein). Through analysis, the gains in the controllers are
chosen on the basis of the stability regions in the proportional
and integral control gain space [1], often combined with the
Nyquist method [4]. Alternatively, stabilization problems have
been addressed by using an observer-based controller with a
discrete model [5], or H feedback control with a discrete
event-based model to improve robustness to disturbances and
modeling uncertainties [6], the Smith predictor [7] and its
adaptive version [8], nonlinear adaptive controllers [9], and
Pad approximation approaches [10]. PI controllers with the
Smith predictor have been shown to improve the performance
in simulations and experiments [11]. Pole placement for timedelay systems, which is not feasible using traditional control
methods, can be achieved by employing the Smith predictor.
Such a prediction-based method was compared to an experimental tuning method, namely, the classical ZieglerNichols
method, in [7]. Alternatively, an act-and-wait control strategy
can be used for pole placement [12][14]. The method uses a
periodic controller where, for example, its period T is double
that of the time delay h (i.e., T = 2 h). Then, during the
first delay interval h, the control input is zero, and during the
second h the control gain is D (i.e., u(t) = Dx(t)). Then,
for the second half of the period, the governing equation can
be written as an ordinary differential equation (ODE), and
the pole-placement problem can be formulated as a finitedimensional one. The control gain is chosen such that the
transition matrix has eigenvalues of which absolute values
are smaller than 1 (or as small as possible). For a thorough
discussion on the method, refer to [15]. Also, a numerical
stabilization method was developed using the sensitivities of
eigenvalues to the changes in the feedback gain in [16] (also,
refer to [17] for a numerical comparison by the authors).
In this paper, PI feedback controllers for linear timeinvariant (LTI) plants having a time delay are designed through
pole placement. The pole placement is achieved by using
the Lambert W function-based approach without the use of
a predictor. The rightmost (i.e., dominant) eigenvalues in
the infinite eigenspectrum can be identified and assigned
using the Lambert W function [18]. This approach is very
similar to pole placement design of PI controllers for systems
without delay, and does not require the use of a predictor.
Through such assignment, the PI gains are calculated. The
designed controllers can improve performance as well as successfully stabilize unstable systems. The proposed PI design
method is compared with other existing methods available
in the literature for illustrative examples. Also, sensitivity of

1063-6536 2012 IEEE

YI et al.: PI CONTROL OF FIRST-ORDER TIME-DELAY SYSTEMS VIA EIGENVALUE ASSIGNMENT

the rightmost eigenvalue with respect to system parameters

(including delays) is studied analytically and, via this analysis,
the robustness of the controller is shown to compare favorably
to Smith predictor-based controllers. Although the Smith predictor is a long-standing control technique, it is still one of
the most widely used and studied control schemes for time
delay systems [11]. Because it and other similar predictorbased methods are the most effective tools for pole placement
for time-delay systems, the Smith predictor is used here as
a benchmark for performance and sensitivity comparisons.
However, the proposed method is also compared with other
PI control design approaches in subsequent sections. Unlike
systems of ODEs, PID design through direct eigenvalue
assignment has not been feasible for systems of delay differential equations (DDEs). Using the method presented, one
can extend the eigenvalue assignment approach for PI/PID
tuning to time-delay systems. Because many properties of a
closed-loop system depend on the locations of its poles, pole
placement is one of the mainstream methods in control system
design [19]. Thus, the method presented here provides a direct
way to extend well-developed control techniques for ODEs to
DDEs.

1587

GS ( s )

1 e

sh

KP

KI
s

G P ( s )

e-sh

GP(s)

Gc

Smith Predictor
(a)

GS ( s )

KP

KI
s

e-sh

GP(s)

(b)

GL ( s )

KP

KI
s

e-sh

GP(s)

(c)
Fig. 1. Block diagrams for (a) a closed-loop system with the Smith predictorbased control, (b) its equivalent system when the response of the plant is
predicted perfectly [20], and (c) closed-loop system with the Lambert W
function-based control.

II. PI C ONTROL FOR S YSTEMS W ITH T IME D ELAYS

A first-order plant with a pure time delay can be described
by
KM
esh
(1)
G(s) = G P (s)esh =
M s + 1
where M is the time constant, h is the time delay, and K M
represents the steady-state gain.
Consider the PI controller
K I
G S = K P +
.
(2)
s
Many controllers in industrial processes only have PI action
[1] and such controllers are widely used, for example, in
automotive controllers [8]. A primary goal of this paper is
to choose the gains K P and K I such that a stable closedloop system with improved transient response (e.g., reduced
rise time) is obtained. In this paper, barred variables ()
denote gains selected by using the Smith predictor approach to
distinguish them from gains selected by using the Lambert W
function approach. It has been pointed out that PI controllers
are sufficient for all systems that have first-order transfer functions to obtain zero steady-state error and adequate transient
responses [1], [19], and they are widely used for controlling
numerous industrial processes. However, it is well known
that the longer the time delay, the more difficult it is to
stabilize the system. Moreover, the delay term in the closedloop characteristic equation complicates the stability analysis
and the design of the controller to guarantee stability [11].
A. Use of the Smith Predictor
The Smith predictor in Fig. 1(a) results in a delayed
response of a delay-free system by moving the time delay
outside the feedback loop only when the model in the Smith
is assumed to be exactly the same
predictor, i.e.,G P (s) and h,

as the plant, i.e., G P (s) and h. Then, the controller G S (s)

in (2) can be designed considering only the delay-free plant
G P (s) [see Fig. 1(b)]. This is the main advantage of the Smith
predictor control.
For example, in order to meet given time domain specifications, the desired eigenvalues can be chosen from the desired
natural frequency n and the desired damping ratio , as

(3)
d = n n 1 2 i = d i.
Assuming no time delay, K P and K I are chosen as
2 M
K I = n ,
KM

2 n M 1
K P =
KM

(4)

by substituting the desired values into the closed-loop characteristic equation of the system in Fig. 1(b)


K K I
1 + K M K P
s2 +
s+
= 0.
(5)
M
M
This means that the controller G S (s) can be designed considering only the nondelayed part G P (s) of the plant ignoring the
time delay esh . This method is, however, based on pole-zero
cancellation and, thus, the stability is vulnerable to uncertainty
in system parameters [21]. Careful modeling and parameter
identification are crucial for its successful application [22].
Furthermore, the Smith predictor cannot handle disturbances
and nonzero initial conditions. Problems caused by parameter
mismatches were studied in [23], and its shortcomings have
been discussed in [11]. It is well known that Smith predictorbased controllers are sensitive to uncertainties in the delay
[24]. This is discussed in more detail in Section III. Various
extensions and modifications to the Smith predictor have been
proposed to address its limitations (e.g., [25] and [26]).

1588

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 5, SEPTEMBER 2013

B. Use of the Lambert W Function-Based Approach

In this subsection, as an alternative to the Smith predictor,
a design approach for the PI controller is developed via rightmost eigenvalue assignment using the Lambert W function.
The open-loop transfer function with a PI controller is


KM
KI
y
esh K P +
(6)
G open =
=
M s + 1
s
e
where e = y + r . Then, the closed-loop system, as in
Fig. 1(c), in the time domain becomes
1
KM KP
KM KI
y
y (t h)
y(t h)
M
M
M
+ {K M K P s + K M K I } r (t h).

y =

(7)

By defining x 1 y, x 2 y , this equation can be rewritten as

x1 = x 2
1
KM KP
KM KI
x2
x 2 (t h)
x 1 (t h)
M
M
M
(8)
+ {K M K P s + K M K I } r (t h).

x2 =

Then we obtain the closed-loop system in state-space form,

where we ignore the reference input to focus on stability, as
follows:

0 1
0
0
1 x(t) K M K I K M K P x(t h).
x (t) =
0

M
M

Ad

(9)
From the roots of the characteristic equation of (9), the eigenvalues of the system are obtained. However, due to the time
delay esh , the system is infinite dimensional and, thus, there
exist an infinite number of eigenvalues. The principal difficulty
in analyzing and controlling systems with time delays arises
from this transcendental character, and the determination of
this eigenspectrum typically requires numerical, approximate,
or other approaches [2]. Obtaining and controlling the entire
infinite eigenspectrum is not as straightforward as for systems
of ODEs. Instead, for DDEs, such as (9), it is desired to
locate the dominant eigenvalues, which are rightmost in the
complex plane, and to assign them to the desired positions.
The Lambert W function-based approach is an efficient tool
for doing this, and in the subsequent section the approach is
applied to (9) to design the PI controller and compared with
the Smith predictor. With the coefficient matrices A and Ad
defined in (9), the solution matrix S0 is computed as
1
W0 (Ad hQ0 ) + A
h
where the unknown matrix Q0 is obtained by solving
S0 =

W0 (Ad hQ0 )eW0 (Ad hQ0 )+Ah = Ad h.

(10)

(11)

Equation (11) can be solved numerically to obtain the matrix

Q0 for the principal branch (k = 0) using nonlinear solvers
(e.g., fsolve in MATLAB). Then, by substituting the matrix
Q0 into (10) we obtain S0 and its eigenvalues. Then, one
can set the desired locations for eigenvalues equal to those
of S0 , i.e., i (S0 ) = i,desired for i = 1, . . . , n, where

i (S0 ) is i th eigenvalue of the matrix S0 . The gains K P and

K I are then obtained by solving the equation numerically.
For a detailed explanation of the Lambert W function-based
approach, including a method for eigenvalue assignment,
refer to [18]. Also, open-source software for the Lambert W
function method based on MATLAB is available at [27]. As
mentioned in [17], depending on the structure or parameters
of a given system, there exists a limitation on the rightmost
eigenvalues and they cannot be arbitrarily assigned. If nonfeasible values of i,desired are selected, then the above approach
does not yield a solution. To resolve the problem, one may try
again with fewer desired eigenvalues or different values of the
desired rightmost eigenvalues. Then, the solution is obtained
numerically for a variety of initial conditions by an iterative
trial-and-error procedure.
The desired positions for the rightmost eigenvalues, i.e.,
i,desired , can be chosen from the desired natural frequency
and damping ratio using (3). In the assignment of eigenvalues,
the principal branch of the Lambert W function is used to
find the rightmost eigenvalues. For first-order scalar DDEs,
it has been proven that the rightmost eigenvalues are always
obtained by using the principal branch (see [28]). For general
systems of DDEs, such a proof is not available. It has been
observed that, if the coefficient Ad does not have repeated zero
eigenvalues, the rightmost eigenvalues are obtained with the
principal branch [29]. If Ad has repeated zero eigenvalues, they
are obtained by using the branches k = 0 and k = 1 [30].
III. I LLUSTRATIVE E XAMPLES AND
S ENSITIVITY A NALYSIS
In this section, the Lambert W function-based approach is
applied to two different cases: an open-loop unstable plant,
and an open-loop stable plant.
A. Open-Loop Unstable Plant
Consider the unstable system
1
e0.2s
(12)
0.5s 0.2
and the PI controller in (2). For such unstable plants, tuning algorithms based on ZieglerNichols methods [31] and
approximated analytical solutions [32] can be used to obtain
stabilizing PI gains. Also, PI controllers for such unstable
plants with time delays have been designed on the basis of
bifurcation methods. That is, purely imaginary variables are
substituted into the characteristic equation (5) for the characteristic root (i.e., s = j ). Then, the characteristic equation is
divided into an imaginary and a real part. Solving the two
equations simultaneously using numerical nonlinear solvers
yields the ranges for K I and K P that stabilize the unstable
plants as in (12). Such bifurcation methods were combined
with the HermiteBiehler theorem, which is a theorem on
stability of quasi-polynomials, and the regions of K I K P
are derived (see [33] and [34]). For example, using the steps
in [1, Ch. 7] for the system in (12), the range of K P is
0.2 < K P < 3.8. If one picks K P = 2 in the range, then the
corresponding range of K I is 0 < K I < 5.74. Any values in
G(s) = G P (s)esh =

YI et al.: PI CONTROL OF FIRST-ORDER TIME-DELAY SYSTEMS VIA EIGENVALUE ASSIGNMENT

6
5

Response, y

these ranges guarantee stable poles of the closed-loop system

and stabilize the unstable plant. These methods have been
extended to design of robust controllers [35] with the transfer
functions from sensor noise to output. The bifurcation method
was used to find stability boundaries in terms of PID gains only
with frequency responses without knowing transfer functions
[36], [37]. Also, the gains obtained by using such bifurcation
methods can be tuned through graphical methods (e.g., Nichols
charts [38] and Nyquist Plots [39]) to improve the response
in the frequency domains. The PI controllers can be designed
by using the small gain theorem and H norms of transfer
functions [40]. For the unstable plant in (12), the method in
[40] yields ranges of stabilizing gains 0.2 < K P < 2.5. If one
picks K P = 2 in the range, then the corresponding range of
K I is 0 < K I < 4.61, which is conservative compared to that
of the bifurcation method in [1]. The method was extended
to design resilient controllers that can reduce sensitivity to
perturbations in the controller coefficients [41].
On the other hand, stabilizing gains can also be obtained by
assigning the positions of the poles of the closed-loop system.
Because many properties of a closed-loop system depend on
the location of its poles, pole placement is one of the mainstream methods in control system design [19]. Pole assignment
is especially useful when time-domain specifications need to
be considered (e.g., rise time, maximum overshoot, etc., in
various applications [42]). If it is not feasible to assign all
poles, the behavior of a system can often be characterized
with a few dominant poles. Thus, one can attempt to place a
few dominant poles [43]. The Smith predictor is a well-known
method for pole placement. The gains of the PI controller,
using the approach in Section II-A, are found to be K I =
3.1250 and K P = 1.4500 when the desired natural frequency
is n = 2.5 and the desired damping ratio is = 0.5. With
those gains, one can assign the eigenvalues of the system
to be 1.2500 2.1651i , which are stable, theoretically.
However, due to the initial conditions, disturbance, and errors
in simulation, this control leads to instability as seen in
Fig. 2 (dashed). If (1) is unstable, the characteristic equation
of the closed-loop system with the Smith predictor in Fig. 1(a)
retains the unstable pole of the open-loop system. Therefore,
the Smith predictor cannot stabilize the system [21]. Fig. 2
shows responses simulated using Simulink. Even though there
is no disturbance or initial condition mismatch, due to errors
in numerical integration the Smith predictor cannot successfully stabilize the unstable plant (12). Using the act-and-wait
control method, pole placement can also be achieved [15]. For
example, for the time constant = 1/1.25, the corresponding
PI gains are K I = 0.8953 and K P = 2.0527. Those two
methods (the Smith predictor and the act-and-wait control)
make the number of eigenvalues finite and assign them. On the
other hand, the Lambert W function-based approach in Section
II-B does not reduce the number of eigenvalues. Instead, it
assigns the rightmost eigenvalues of the infinite spectrum to
the desired positions. For example, the resulting gains for the
rightmost eigenvalues, 1.2500 2.1651i (n = 2.5 and
= 0.5) are K I = 1.4309 and K P = 1.2232. Table I shows
the PI gains corresponding to other values of n . In designing
PI control for systems of ODEs, an increase in n induces

1589

4
Smith predictor

3
2
1

Lambert W function
0

10
Time, t

15

20

Fig. 2. Simulated responses of the unstable system (12) controlled with the
Smith predictor (dashed line) and the Lambert W function-based approach
(solid line) using Simulink. Because the Smith predictor is based on unstable
pole-zero cancelation, it fails to stabilize the system.
TABLE I
G AINS K I AND K P OF PI C ONTROLLER O BTAINED BY U SING THE
L AMBERT W F UNCTION A PPROACH VIA R IGHTMOST E IGENVALUE
A SSIGNMENT FOR THE U NSTABLE P LANT IN (12)

Dominant pole

PI gains obtained by
using the Lambert W
function


d = n n 1 2 i

KI

KP

1.0

0.5

0.5000 0.8660i

0.3646

0.6509

1.5

0.5

0.7500 1.2990i

0.7156

0.8604

2.0

0.5

1.0000 1.7321i

1.0908

1.0525

a decrease in rise time while not changing the overshoot. As

seen in Fig. 3, by adjusting the values of n , one can tune the
gains to meet time-domain specifications in a similar way to
ODEs.
Because the Lambert W function-based approach does not
require pole-zero cancellation, the designed controller safely
stabilizes the system (see Fig. 2). Another prediction-based
approach, i.e., finite spectrum assignment, also has unsolved
problems regarding integral approximation and, thus, can fail
to stabilize unstable systems [2]. It has also been shown in
[17] that the Pad approximation does not guarantee stability
of the controlled system because of its inherent inaccuracy.
B. Open-Loop Stable Plants
When the time constant M is 0.5 and the time delay
h is 0.2 with K M = 1 for the plant in (1), the gains
for PI control, for several desired natural frequencies and
damping ratios, are obtained by using the Smith predictorbased approach and the Lambert W function-based approach,
which are given in Table II. The Smith predictor moves the
time delay outside the feedback loop as seen in Fig. 1(b)
and makes the number of poles finite (in this case 2) by
canceling the other (infinite number of) eigenvalues. On the

1590

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 5, SEPTEMBER 2013

1.8

150

1.6

100

1.2

1.2500 2.1651i

50

1
0.8

Imag()

Response, y

1.4

increasing n
(1.0, 1.5 and 2.0)

0.6

0
50

0.4
0.2
0

100
0

Time, t

150

Fig. 3. Simulated responses of the unstable system (12) controlled with

PI controllers designed by using the Lambert W function-based approach.
Increase in n from 1 to 1.5 to 2.0, induces a decrease in the rise time of the
transient response.
TABLE II
G AINS K I AND K P OF THE PI C ONTROLLER O BTAINED BY U SING
THE

25

20

15
10
Real()

Fig. 4. Eigenspectrum of the closed-loop system in Fig. 1(c). By using

the Lambert W function-based approach in the subsection II-B, the rightmost
eigenvalues of the infinite eigenspectrum are assigned to the desired positions,
1.2500 2.1651i.

L AMBERT W F UNCTION A PPROACH VIA R IGHTMOST

E IGENVALUE A SSIGNMENT

1.2
1

Lambert W function

K I

K P

KI

KP

1.0

0.3

0.5000

0.7000

0.6214

0.5143

1.1

0.5

0.6050

0.4500

0.6884

0.2439

1.5

0.5

1.1250

0.2500

1.1751

0.0155

2.5

0.5

3.1250

0.2500

2.5629

0.6013

Response, y

Smith predictor [from(4)]

0.8
0.6

Lambert W Function
0.2
0

other hand, the Lambert W function-based approach, without

the cancellation, obtains the rightmost eigenvalues and assigns
them to the desired positions. For example, when the desired
natural frequency n = 2.5 and the desired damping ratio is
= 0.5 (thus, the desired positions for the rightmost eigenvalues are 1.2500 2.1651i ), the eigenspectrum of the closedloop system depicted in Fig. 4 is obtained. The rightmost
eigenvalues are placed exactly at the desired positions, which
are obtained from (3), and others are to the left of them. The
response of the closed systems by using the Smith predictor
[Fig. 1(a)] and the Lambert W function-based approach are
shown in Fig. 5. Although the systems in Fig. 1(c) have an
infinite number of eigenvalues, by assigning the rightmost
(thus, dominant) eigenvalues to the desired positions, one can
meet time-domain specifications [44] assuming that the other
eigenvalues that are not controlled are sufficiently far from the
rightmost eigenvalues (e.g., Fig. 4). However, if such subdominant eigenvalues get closer to the dominant ones (e.g., as time
delays increase [45]), it is more difficult to improve transient
responses via the presented eigenvalue assignment method.
As for robustness, if (1) is open-loop stable, the Smith predictor works as desired and does not have apparent mismatch
problems unlike the example in Section III-A. However, as
shown in Fig. 1, the Lambert W function-based controller has a
simpler form than the Smith predictor-based one. Also, for the
Smith predictor-based control in Fig. 1(a), the feedback signal
for the controller, G S (s), uses y esh , which is the output y
predicted one time-delay ahead. Therefore, it is reasonable to

Smith Predictor

0.4

Desired (Nondelay)
0

0.5

1.5

2
2.5
Time, t

3.5

Fig. 5.
Responses of systems controlled by using the Smith predictor
[Fig. 1(a)] and the Lambert W function [Fig. 1(c)] when the desired rightmost
eigenvalues are 1.2500 2.1651i for M = 0.5, h = 0.2, and K M = 1
in (1).

consider the sensitivity with respect to mismatch of parameters

and compare the robustness of the two different control
methods. Sensitivity is discussed in the next two subsections.
In the literature, it has been shown that estimation of time
delay in continuous LTI systems is more difficult than other
system parameters [46]. Also, it is well known that control
using the Smith predictor is sensitive to delay mismatches
[24]. Thus, the sensitivity analysis is focused on the time delay.
However, the analysis for other parameters can be conducted
in a similar way and those results are also discussed.
C. Sensitivity of the Smith Predictor
In this subsection, the sensitivity of the closed-loop eigenvalues to small variation in the time delay h is considered for
the Smith predictor-based controller.
The transfer function of the closed-loop system with the
Smith predictor in Fig. 1(a) can be represented as
G S G P esh
y


=
.
r
1 + 1 es h G S G P + G S G P esh

(13)

YI et al.: PI CONTROL OF FIRST-ORDER TIME-DELAY SYSTEMS VIA EIGENVALUE ASSIGNMENT

1591

TABLE III
S ENSITIVITY C OMPARISON W ITH R ESPECT TO h. T HE L AMBERT W F UNCTION -BASED A PPROACH S HOWS
S MALLER VALUES OF THE S ENSITIVITY AND I MPROVEMENT IN ROBUSTNESS
n

Sensitivity,
h

Rightmost

Smith predictor (S1)

Lambert W function (S2)

Re(S1)Re(S2)
100
Re(S1)

1.0

0.3

0.3000 0.9539i

0.8697 0.1245i

0.7431 0.3423i

14.59

1.1

0.5

0.5500 0.9526i

1.2191 0.4722i

0.9606 0.0662i

21.20

1.5

0.5

0.7500 1.2990i

1.8135 0.9045i

1.4319 0.4189i

21.04

2.5

0.5

1.2500 2.1651i

4.0803 2.6198i

3.2842 2.4937i

19.51

Assuming that all parameters except the time delay are well
matched, i.e., G P = G P , the denominator will be


D E(s) 1 + 1 es h G S G P + G S G P esh

= 1 + G S G P G S G P es h + G S G P esh
= 1 + G S G P G S G P esh (es 1)
(14)
where h h. Then substituting the terms
GP

Improvement (%)

eigenvalue,

BP
KM
,
=
AP
M s + 1

GS

K p s + K I
BC
=
AC
s

By substituting the eigenvalues for s in (19), one can

get the sensitivity of the eigenvalues with respect to small
variances in the time delay. The sensitivity in (19) represents
an incremental change in the positions of the eigenvalues
corresponding to an incremental change in the time delay.
Numerical values for several cases, with the same parameter
set as in Section III-B, are given in Table III.
D. Sensitivity of the Lambert W Function Approach

(15)
yields

D E(s) = 1 + G S G P G S G P es h + G S G P esh
B P BC
B P BC s h
B P BC sh
= 1+

e
+
e .
A P AC
A P AC
A P AC

From the state equation (9) of the closed-loop system in

Fig 1(c), the sensitivity of the eigenvalue with respect to the
time delay, h, is obtained in a way similar to the previous
derivation. The characteristic equation of (9) is
s2 +

(16)
Setting D E(s) = 0 yields the closed-loop characteristic
equation

Ch(s) = A P AC + B P BC B P BC es h + B P BC esh


= ( M s + 1) (s) + K M K P s + K I



K M K P s + K I es h

 sh
+K M K P s + K I e
= 0.
(17)
Differentiating both sides of the characteristic equation with
respect to the time delay h yields
s
s
s

Ch(s) = 2 M s
+
+ K M K P
+ K M K I
h
h
h
h


s

+ K I es h
K M K P
h
 s



K M K P s + K I es h h
h


s
+ K I esh
+K M K P
h



 sh
s

+K M K P s + K I e
s .
(h)
h
(18)
Because (/h)Ch(s) = 0, we get the sensitivity of s with
respect to the time delay h as


K M K P s + K I esh (s) K M K I
s
=
.
(19)
h
2 M s + 1 + K M K P

1
K M K P sh K M K I sh
s+
se
+
e
= 0.
M
M
M

(20)

In a similar way to the previous section, by differentiating both

sides the sensitivity of the eigenvalues with respect to change
in the time delay, h, is given by
s
=
h

2s M

s 2 K M K P esh + s K M K I esh
.
+ 1 + K M K P esh (K M K P s + K M K I )esh h
(21)

The numerical values of the sensitivity with respect to h

for the Lambert W function-based approach and the Smith
predictor are compared in Table III. As seen from the table,
for several rightmost eigenvalues, which are arbitrarily chosen,
the Lambert W function-based approach shows smaller values
of the sensitivity and, thus, improvement in robustness. For
comparison, improvement is calculated as the ratio of decrease
in real parts of the sensitivities to the real part of the sensitivity
of the Smith predictor (i.e., (Re(S1) Re(S2))/Re(S1)100).
Because in the Smith predictor control the feedback signal for
the controller [G S (s) in Fig. 1(a)] uses the predicted output
y (e.g., yesh ) by canceling signals [11], it is sensitive to
infinitesimal delay mismatches [24]. On the other hand, for
the Lambert W function-based approach, the feedback signal
for the controller [G L (s) in Fig. 1(c)] is not predicted and does
not require any cancelation. This may result in improvement
in robustness, as seen in Table III.
For other parameters in (1), sensitivity analysis can be
conducted in a way similar to the time delay h. For example,
for K M and M the obtained sensitivities are summarized

1592

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 5, SEPTEMBER 2013

TABLE IV
C OMPARISON OF THE S ENSITIVITY. W ITH R ESPECT TO K M , THE L AMBERT W F UNCTION -BASED A PPROACH S HOWS
S MALLER VALUES OF THE S ENSITIVITY AND AN I MPROVEMENT IN ROBUSTNESS . O N THE O THER H AND , W ITH
R ESPECT TO M , U NLIKE h AND K M , THE S ENSITIVITY S HOWS M IXED R ESULTS
n

Sensitivity,
KM
Smith predictor
Lambert W function

Improvement

(%)

Smith predictor

Lambert W function

(%)

1.0

0.3

0.8797 0.6350i

0.5495 0.6061i

37.53

0.4525 1.0170i

0.4930 0.6794i

8.96

1.1

0.5

0.6824 0.8857i

0.3845 0.7864i

43.66

1.0714 0.9286i

0.9675 0.5443i

9.70

1.5

0.5

0.5823 1.0599i

0.2354 0.9663i

59.56

1.4258 1.4201i

1.4364 0.8225i

0.74

2.5

0.5

0.4085 1.6487i

-0.2070 1.6364i

49.33

2.1361 3.0292i

2.9950 1.9965i

40.21

300

200

The set of the rightmost eigenvalues

by the principal branch (k=0)

Imag()

100

Imag()

Sensitivity,

Improvement

100

200

300
30

25

20

15

10

Real()

3
2.5

1.5

0.5

Real()

Fig. 6. Eigenvalues of the system in (24). The eigenvalues obtained by using the principal branch (k = 0) are rightmost. By adjusting the gains in (22), the
eigenvalues are assigned to the desired position (s = 1 1.5i).

in Table IV. The sensitivity with respect to K M for the

Lambert W function also has smaller values than the Smith
predictor. Thus, the Lambert W function-based approach
enhances robustness. However, this does not mean that the
presented approach always leads to less sensitive controllers.
In Table IV, for the sensitivity with respect to M , the
results are mixed. Thus, use of the Lambert W function does
not always reduce sensitivity of the rightmost eigenvalues.
Therefore, for stable first-order plants, after comparing sensitivity with respect to parameters (especially the ones having
larger variance, for example, due to difficulty in estimation),
one can choose a more suitable method to design PI controllers, which is more robust against variance in system
parameters.
IV. E XTENSION TO PID C ONTROL
The matrix Lambert W function can be used for general
systems of retarded DDEs. Thus, the presented method for PI
control can be extended to more general cases, such as higher
order plants and/or PID design.
Consider the PID controller


KI
+ K Ds
(22)
G L (s) = K p +
s
and the second-order plant with a pure time delay
G(s) =

KM
esh .
(m s + 1)(s + a)

(23)

The system of DDEs representing the closed-loop system is

given in matrix-vector form as

0 1
0
x(t)
1
x(t)

= 0 0
0 am (ma+1)
m

0
0
0
x(t h).
0
0
0
(24)
+
KM KI
KM KP
KM KD
m m m
Assuming the desired positions of the two rightmost eigenvalues are s = 1 1.5i , through the design method using
the Lambert W function, the resulting gains are K P =
2.0248, K I = 1.9061, and K D = 0.3769 for m = 0.5,
K M = 1, a = 1, and h = 0.2. Fig. 6 shows eigenvalues of
(24). The eigenvalues obtained by using the principal branch
(k = 0) are rightmost. By adjusting the gains in (22), the
eigenvalues are assigned to the desired positions. For the
3 3 system of DDEs in (24), the solution matrix in (10)
has three eigenvalues. Thus, three rightmost eigenvalues can
be simultaneously assigned. For example, the desired three
eigenvalues are s = 2 i, 1.5, and the corresponding
gains are K P = 2.8128, K I = 2.2009, and K D = 0.8353 (see
Fig. 7). The simulation results using Simulink are shown in
Fig. 8 for the two PID gain sets obtained above. The solid
line, whose eigenvalues are shown in Fig. 7, shows a shorter
rise time and a smaller overshoot than the dashed line, whose
eigenvalues are shown in Fig. 6.
In the examples, two and three of the rightmost eigenvalues
are assigned and the gains can be obtained successfully.

YI et al.: PI CONTROL OF FIRST-ORDER TIME-DELAY SYSTEMS VIA EIGENVALUE ASSIGNMENT

300

1.5

200

The set of the rightmost eigenvalues

by the principal branch (k=0)

1
Imag()

100

Imag()

1593

0
100

0.5
0
0.5
1

200

1.5

300
25

20

15

10
Real()

2
2.5

1.5
Real()

0.5

Fig. 7. Eigenvalues of the system in (24). The eigenvalues obtained by using the principal branch (k = 0) are rightmost. By adjusting the gains in (22), the
eigenvalues are assigned to the desired position (s = 2 i and s = 1.5).

1.4

systems with time delays can be successfully stabilized, as was

shown in Section III. Also, because prediction of the plant is
not required, the obtained controller has a simpler form, as
shown in Fig. 1, and is more robust in the presence of delay
mismatches.
As shown with an example in Section IV, this methodology
can be extended to more general cases. For example, it is
applicable to higher-order systems with time delays and to
design more general controllers (e.g., PID control). Further
theoretical studies, as well as experimental implementation and
evaluations, are currently being conducted.

Response, y

1.2
1
0.8
0.6
0.4
0.2
0
0

Fig. 8.

0.5

1.5

2
Time, t

2.5

3.5

Responses of the system in (23) with the PID controller in (22).

However, as the dimension of systems and, thus, number of

unknowns in (11) increases, computational load also increases
correspondingly, in general. For example, solving (11) for Qk
takes 0.20.3 s for 2 2 examples. But it takes 34 s for
3 3 examples in this paper using a PC with 2.40 GHz
CPU and 4 GB RAM. The nonlinear solver fsolve was used,
and computing time was measured using cputime in MATLAB.
Also, increased number of control gains adds more computing
time. Thus, the presented method is not practically feasible for
arbitrarily high-dimensional systems. Besides, the relationship
between the number of control gains and the number of the
eigenvalues that can be assigned should be studied further in
connection with the controllability of systems of DDEs [47].
V. C ONCLUDING R EMARKS AND F UTURE W ORK
PI controllers are the most commonly used controllers for
first-order time-delay systems. In this paper, we have studied a
new approach to the design of PI controllers for such systems.
Choosing the gains in the PI controllers corresponding to
the desired system performance is essential to achieve the
control goals. The presented approach based on the Lambert
W function enables one to choose the gains by assigning the
rightmost (i.e., dominant) eigenvalues to the desired positions,
which are obtained from the desired natural frequency and
damping ratio. Unlike prediction-based methods, the approach
is not dependent on pole-zero cancellation and, thus, unstable

R EFERENCES
[1] G. J. Silva, A. Datta, and S. P. Bhattacharyya, PID Controllers for TimeDelay Systems (Control Engineering). Boston, MA: Birkhuser, 2005.
[2] J. P. Richard, Time-delay systems: An overview of some recent
advances and open problems, Automatica, vol. 39, no. 10, pp. 1667
1694, 2003.
[3] G. J. Silva, A. Datta, and S. P. Bhattacharyya, PI stabilization of firstorder systems with time delay, Automatica, vol. 37, no. 12, pp. 2025
2031, 2001.
[4] W. Krajewski, A. Lepschy, and U. Viaro, Designing PI controllers for
robust stability and performance, IEEE Trans. Control Syst. Technol.,
vol. 12, no. 6, pp. 973983, Nov. 2004.
[5] J. D. Powell, N. P. Fekete, and C. F. Chang, Observer-based air fuel
ratio control, IEEE Control Syst., vol. 18, no. 5, pp. 7283, Oct. 1998.
[6] L. Mianzo, H. Peng, and I. Haskara, Transient air-fuel ratio H
preview control of a drive-by-wire internal combustion engine, in Proc.
Amer. Control Conf., 2001, pp. 28672871.
[7] S. Nakagawa, K. Katogi, and M. Oosuga, A new air-fuel ratio feed
back control for ULEV/SULEV standard, in Proc. SAE World Congr.,
2002, no. 2002-01-0194.
[8] Y. Yildiz, A. Annaswamy, D. Yanakiev, and I. Kolmanovsky, Adaptive
air fuel ratio control for internal combustion engines, in Proc. Amer.
Control Conf., Seattle, WA, 2008, pp. 20582063.
[9] R. Turin and H. Geering, Model-reference adaptive A/F-ratio control
in an SI engine based on Kalman-filtering techniques, in Proc. Amer.
Control Conf., 1995, pp. 40824090.
[10] L. Guzzella and C. H. Onder, Introduction to Modeling and Control of
Internal Combustion Engine Systems. Berlin, Germany: Springer-Verlag,
2004.
[11] Q. C. Zhong, Robust Control of Time-Delay Systems. London, U.K.:
Springer-Verlag, 2006.
[12] P. Gawthrop, Act-and-wait and intermittent control: Some comments,
IEEE Trans. Control Syst. Technol., vol. 18, no. 5, pp. 11951198, Sep.
2010.
[13] T. Insperger, Act-and-wait concept for continuous-time control systems
with feedback delay, IEEE Trans. Control Syst. Technol., vol. 14, no. 5,
pp. 974977, Sep. 2006.
[14] T. Insperger, P. Wahi, A. Colombo, G. Stepan, M. Di Bernardo, and S.
J. Hogan, Full characterization of act-and-wait control for first-order
unstable lag processes, J. Vibrat. Control, vol. 16, nos. 78, pp. 1209
1238, 2010.

1594

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 5, SEPTEMBER 2013

[15] T. Insperger and G. Stepan, Semi-Discretization for Time-Delay Systems:

Stability and Engineering Applications. New York: Springer-Verlag,
2011.
[16] W. Michiels, K. Engelborghs, P. Vansevenant, and D. Roose, Continuous pole placement for delay equations, Automatica, vol. 38, no. 5, pp.
747761, 2002.
[17] S. Yi, P. W. Nelson, and A. G. Ulsoy, Eigenvalue assignment via
the Lambert W function for control for time-delay systems, J. Vibrat.
Control, vol. 16, nos. 78, pp. 961982, 2010.
[18] S. Yi, P. W. Nelson, and A. G. Ulsoy, Time-Delay Systems: Analysis and
Control Using the Lambert W Function. Singapore: World Scientific,
2010.
[19] K. J. strm and T. Hgglund, Advanced PID Control. Triangle Park,
NC: ISA, 2006.
[20] O. J. M. Smith, Closer control of loops with dead time, Chem. Eng.
Progr., vol. 53, no. 5, pp. 217219, 1957.
[21] T. Furukawa and E. Shimemura, Predictive control for systems with
time-delay, Int. J. Control, vol. 37, no. 2, pp. 399412, 1983.
[22] L. Guzzella and A. Amstutz, Control of diesel engines, IEEE Control
Syst. Mag., vol. 18, no. 5, pp. 5371, Oct. 1998.
[23] K. Yamanaka and E. Shimemura, Effects of mismatched Smith controller on stability in systems with time-delay, Automatica, vol. 23,
no. 6, pp. 787791, 1987.
[24] W. Michiels and S. I. Niculescu, On the delay sensitivity of Smith
predictors, Int. J. Syst. Sci., vol. 34, nos. 89, pp. 543551, 2003.
[25] T. Liu, Y. Z. Cai, D. Y. Gu, and W. D. Zhang, New modified Smith
predictor scheme for integrating and unstable processes with time delay,
IEE Proc.-Control Theory Appl., vol. 152, no. 2, pp. 238246, Mar.
2005.
[26] S. I. Niculescu and A. M. Annaswamy, An adaptive Smith-controller
for time-delay systems with relative degree n  2, Syst. Control Lett.,
vol. 49, no. 5, pp. 347358, 2003.
[27] S. Duan. (2010). Supplement for Time-Delay Systems: Analysis
and Control Using the Lambert W Function [Online]. Available:
http://www-personal.umich.edu/ulsoy/TDS_Supplement.htm
[28] H. Shinozaki and T. Mori, Robust stability analysis of linear timedelay systems by Lambert W function: Some extreme point results,
Automatica, vol. 42, no. 10, pp. 17911799, 2006.
[29] S. Yi, P. W. Nelson, and A. G. Ulsoy, Delay differential equations via
the matrix Lambert W function and bifurcation analysis: Application to
machine tool chatter, Math. Biosci. Eng., vol. 4, no. 2, pp. 355368,
2007.
[30] S. Yi, P. W. Nelson, and A. G. Ulsoy, Eigenvalues and sensitivity
analysis for a model of HIV-1 pathogenesis with an intracellular delay,
in Proc. ASME Dynamic Syst. Control Conf., Ann Arbor, MI, Oct. 2008,
pp. 19.
[31] A. M. De Paor and M. OMalley, Controllers of ZieglerNichols type
for unstable process with time delay, Int. J. Control, vol. 49, no. 4, pp.
12731284, 1989.
[32] V. Venkatashankar and M. Chidambaram, Design of P and PI controllers for unstable first-order plus time delay systems, Int. J. Control,
vol. 60, no. 1, pp. 137144, 1994.
[33] L. Ou, W. Zhang, and L. Yu, Low-order stabilization of LTI systems
with time delay, IEEE Trans. Autom. Control, vol. 54, no. 4, pp. 774
787, Apr. 2009.
[34] G. J. Silva, A. Datta, and S. P. Bhattacharyya, New results on the
synthesis of PID controllers, IEEE Trans. Autom. Control, vol. 47, no. 2,
pp. 241252, Feb. 2002.
[35] M. Bozorg and F. Termeh, Domains of PID controller coefficients
which guarantee stability and performance for LTI time-delay systems,
Automatica, vol. 47, no. 9, pp. 21222125, 2011.
[36] T. Lee, J. M. Watkins, T. Emami, and S. Sujoldzic, A unified approach
for stabilization of arbitrary order continuous-time and discrete-time
transfer functions with time delay, in Proc. 46th IEEE Conf. Decision
Control, Dec. 2007, pp. 21002105.
[37] S. Sujoldzic and J. M. Watkins, Stabilization of an arbitrary order
transfer function with time delay using a PID controller, in Proc. 45th
IEEE Conf. Decision Control, Dec. 2006, pp. 846851.
[38] E. Poulin and A. Pomerleau, PID tuning for integrating and unstable
processes, IEE Proc.-Control Theory Appl., vol. 143, no. 5, pp. 429
435, Sep. 1996.
[39] Y.-G. Wang and H.-H. Shao, Optimal tuning for PI controller, Automatica, vol. 36, no. 1, pp. 147152, 2000.
[40] A. N. Gndes, H. zbay, and A. B. zgler, PID controller synthesis
for a class of unstable MIMO plants with I/O delays, Automatica,
vol. 43, no. 1, pp. 135142, 2007.

[41] H. zbay and A. N. Gndes, Resilient PI and PD controller designs

for a class of unstable plants with I/O delays, Appl. Comput. Math.,
vol. 6, no. 1, pp. 1826, 2007.
[42] K. H. Ang, G. Chong, and Y. Li, PID control system analysis, design,
and technology, IEEE Trans. Control Syst. Technol., vol. 13, no. 4, pp.
559576, Jul. 2005.
[43] Q. G. Wang, Z. Zhang, K. J. Astrom, and L. S. Chek, Guaranteed
dominant pole placement with PID controllers, J. Process Control,
vol. 19, no. 2, pp. 349352, 2009.
[44] S. Yi, P. W. Nelson, and A. G. Ulsoy, Robust control and time-domain
specifications for systems for delay differential equations via eigenvalue
assignment, J. Dynamic Syst., Meas. Control, vol. 132, no. 7, pp. 17,
2010.
[45] T. Li and E. K. W. Chu, Pole assignment for linear and quadratic
systems with time-delay in control, Numer. Linear Algebra Appl., 2011,
to be published.
[46] J. Tuch, A. Feuer, and Z. J. Palmor, Time delay estimation in continuous
linear time-invariant systems, IEEE Trans. Autom. Control, vol. 39,
no. 4, pp. 823827, Apr. 1994.
[47] S. Yi, P. W. Nelson, and A. G. Ulsoy, Controllability and observability
of systems of linear delay differential equations via the matrix Lambert
W function, IEEE Trans. Autom. Control, vol. 53, no. 3, pp. 854860,
Apr. 2008.

Sun Yi received the B.Sc. degree in mechanical

and aerospace engineering from Seoul National University, Seoul, Korea, in 2004, and the M.Sc. and
Ph.D. degrees in mechanical engineering from the
University of Michigan, Ann Arbor, in 2006 and
2009, respectively.
He is currently an Assistant Professor with the
Department of Mechanical Engineering, North Carolina A&T State University, Greensboro. His current research interests include time-delay systems,
and analysis and control of dynamic systems with
applications to mobile robots and satellites.
Dr. Yi is the Publication Chair of the 2012 ASME DSCC/MOVIC Conference.

Patrick W. Nelson received the B.S. degree in mathematics from Arizona State University, Tempe, and
the M.S. and Ph.D. degrees in applied mathematics
from the University of Washington, Seattle, in 1994,
1995, and 1998, respectively.
He is currently a Research Assistant Professor
with the Department of Computational Medicine and
Bioinformatics, University of Michigan, Ann Arbor.
His current research interests include delay differential equations and their applications to biology and
engineering.
Dr. Nelson was a recipient of the Burroughs Welcome Career Award at the
Scientific Interface in 2002, and was on the Board of Directors of the Society
of Mathematical Biology from 2006 to 2011. He is an Associate Editor of
the Journal of Diabetes and Metabolism.

A. Galip Ulsoy received the B.S. degree in engineering from Swarthmore College, Swarthmore, PA,
the M.S. degree in mechanical engineering from
Cornell University, Ithaca, NY, and the Ph.D. degree
in mechanical engineering from the University of
California at Berkeley, Berkeley, in 1973, 1975, and
1979, respectively.
He is the C.D. Mote, Jr. Distinguished University
Professor of mechanical engineering and the W.C.
Ford Professor of manufacturing with the University of Michigan, Ann Arbor. His current research
interests include the dynamics and control of mechanical systems.
Dr. Ulsoy was a recipient of numerous awards, including the Rufus T.
Oldenburger Medal from the ASME in 2008. He is a member of the National
Academy of Engineering and a fellow of the ASME, SME, and IFAC.