94 Int. J. Mechatronics and Automation, Vol. 3, No.

2, 2013

Robust control stability using the error loop

Enrico Canuto*, Wilber Acuña-Bravo and

Carlos Perez Montenegro
Politecnico di Torino,
Department of Control and Computer Engineering,
Corso Duca degli Abruzzi 24, 10129 Turin, Italy
E-mail: enrico.canuto@polito.it
E-mail: wilber.acunabravo@polito.it
E-mail: carlos.perez@polito.it
*Corresponding author

Abstract: This paper formulates the error loop as a tool for designing robust stability control
systems in the presence of structured and unstructured uncertainties. The error loop indicates that
uncertainties can be accommodated through the design of the noise estimator, which is the unique
feedback channel from plant to control. The real-time model that is embedded in the control unit
and the noise estimator constitute a state predictor. The embedded model consists of a
controllable dynamics plus a disturbance dynamics fed by the noise estimator. It is shown that
causality constraint prevents perfect cancellation of causal uncertainties (unknown disturbance),
but makes the control law which is fed by the state predictor to play a role, thus offering a further
degree of freedom. Employing asymptotic expansions of the closed-loop transfer functions,
simple, explicit design formulae derive from stability inequalities. They relate closed-loop
eigenvalues to model parameters and requirements, and define an admissible frequency band for
the state predictor bandwidth. This paper restricts formulation to the univariate case. A simple
example is provided with simulated and experimental data.

Keywords: robust control; closed-loop; stability; error loop; embedded model control.

Reference to this paper should be made as follows: Canuto, E., Acuña-Bravo, W. and
Perez Montenegro, C. (2013) ‘Robust control stability using the error loop’, Int. J. Mechatronics
and Automation, Vol. 3, No. 2, pp.94–109.

Biographical notes: Enrico Canuto received his degree in Electrical Engineering from
Politecnico di Torino, Turin, Italy, where he joined the staff as an Associate Professor of
Automatic Control in 1983. From 1982 to 1997, he contributed to data reduction of the European
astrometric mission Hipparcos. Technological studies in view of scientific and drag-free space
missions, like Gaia and GOCE, provided the opportunity of applying embedded model control to
drag-free satellites and to electro-optics. He contributed to the conception, design and
implementation of the nanobalance interferometric thrust-stand, capable of sub-micronewton
accuracy. Presently, he is involved in the design of the orbit, formation and attitude control of the
Next Generation Gravity Missions of the European Space Agency. His research interests cover
all the entire field of control problems that are challenging because of complexity, uncertainty
and precision.

Wilber Acuña-Bravo received his degree in Electrical Engineering from Universidad del
Quindío, Armenia, Colombia, and MS in Control Engineering from Universidad de Los Andes,
Mérida, Venezuela. In 2011, he received his PhD in Information and System Engineering from
Politecnico di Torino, Turin, Italy, where he is currently a Research Assistant. His research
interests include electro-hydraulic control systems and control applications.

Carlos Perez Montenegro received his degree in Electrical Engineering from Pontificia
Universidad Javeriana, Bogotá, Colombia. Currently, he is a PhD student in Information and
System Engineering at Politecnico di Torino, Turin, Italy. His main research interests include
space control applications and quadricopter design and control.

Copyright © 2013 Inderscience Enterprises Ltd.

 Robust control stability using the error loop
1 Introduction
1.1. Uncertainty and the error loop
Closed-loop performance is affected by uncertain
discrepancies between model and reality. Three kinds of
discrepancies may be distinguished (Maciejowski, 1989;
Doyle et al., 1992) as in Figure 1.
Causal uncertainty, denoted by clouds, regards
unpredictable actions, referred to as innovation, and
formulated either as a white-noise or a bounded and
arbitrary-signal class corresponding to wd and wu in
Figure 1. They are independent of the model state and
command and only affect performance. The mechanism
through which they affect a plant is not merely
unpredictable, as they accumulate prior to be released.
Causes of this kind are referred to as ‘unknown’ disturbance
and their model, the disturbance dynamics D in Figure 1,
generates the signal class du ∈ Du. The control goal is to
cancel their effect less the innovation, as the latter occurs
while the command is acting on the plant. Innovations wd
and wu are necessary and sufficient for updating the state of
D. They must be real-time estimated by correlating them to
the model error e = y – ym through a noise estimator as in
Canuto et al. (2010). This is the basic mechanism of the
‘disturbance observers’ pioneered by Johnson (1971), and
Hostetter and Meditch (1973), and further studied and
applied by Mita et al. (1998), Bickel and Tomizuka (1999),
Schrijver and van Dijk (2002), Choi et al. (2003), and
Katsura et al. (2007). The embedded model control of
Canuto (2007) inspired by Donati and Vallauri (1984)
develops around the concept that noise is the necessary and
sufficient feedback from plant to control. The ensemble of
noise estimator and design model, to be embedded as a
real-time model in the control unit – hence the name
embedded model – implements a one-step state predictor.
The structured (or parametric) uncertainty h(·), denoted
by a cloudy 3D box, either refers to discrepancies from the
model class in the form of parameter uncertainty or to
neglected relations between model variables (cross
couplings). They must be distinguished from the known
interconnections, that are denoted with h (⋅) and must be
treated as known disturbances. They have been extensively
studied in the literature, as for instance by Chapellat et al.
(1990), Foo and Soh (1993), Ross Barmish (1994),
Calafiore and Dabbene (2002), Chen et al. (2005),
Patre et al. (2008). They may take several forms,
from command-to-state to output-to-state relations: only
state-to-command relations are treated here since they affect
the model eigenvalues. They are included in the so-called
design model and implemented as a numerical simulator,
but they are forcedly neglected in the embedded model,
except for the known part h, and are surrogated by the
unknown disturbance class Du. The advantage is that h can
be made implicitly known by du and therefore cancelled by
u, thus favouring robustness of the model-based design
95
since model-plant discrepancies encoded in e are attenuated.
The difficulty is that h, being output and command
dependent cannot belong to Du, as the latter is defined as a
class of command independent signals. This is one of the
chief instability sources of any feedback control design,
since feedback signals – in this case the estimated
innovations wd and wu – spill components incoherent with
the design model. Appropriate stability conditions need to
be formulated and proved, which are shown to be mainly
related to the noise estimator (and therefore to the state
predictor) design. Specifically, frequency-domain design
suggests to enlarge the state predictor bandwidth (BW) to
allow estimation and cancellation of the dominant low
frequency components.
Unstructured uncertainty, denoted by a cloudy 3D box,
refers to model-class discrepancies and uncertainties, which
are unavoidable due to the embedded model finite order.
They may take several forms as in Maciejowski (1989),
Doyle et al. (1992) and Ross Barmish (1994). The output
fractional form [multiplicative form in Doyle et al. (1992)]
is adopted here as in Canuto (2007) and in Canuto et al.
(2012). They are driven by the model output ym and yield
the model error e (plant minus model) as output, as follows:
e = y − ym = ∂P ( ym ) + wy , (1)
where y is the measured plant output. The dynamic operator
∂P is the (uncertain) ‘neglected dynamics’, a term
preferable to unstructured uncertainty, and wy is the
measurement error, which has been assumed adding to ∂P.
Components of (1) are completely neglected in the
embedded model. Spilling e to the command through
feedback must be restricted, since, for what concerns ∂P, it
can be assimilated to a command-dependent output error. In
addition, wy prevents accurate estimation of the innovations
wd and wu, which difficulty leads to Kalman filter
optimisation (see Kwakernaak and Sivan, 1972). The
uncertainty-based design restricts noise estimation in the
frequency band (usually at higher frequencies below
Nyquist frequency) where ∂P dominates.
As a result in robust design, structured and unstructured
uncertainties must be accommodated by a trade-off in the
design of the state predictor BW: a wider band is required to
cancel structured discrepancies, a narrower band to prevent
spilling of neglected dynamics. The design defines a
stability interval where to place noise-estimator gains or,
equivalently, the state-predictor eigenvalues. Converting
eigenvalues into frequency domain, lower and upper bounds
of the state-predictor BW are obtained. The interval width
depends on the ratio of the predictor and control-law
eigenvalues, and on a stability margin. When the interval is
void, robust design becomes unfeasible, which may be due
to excessive stability margin (conservative design) or large
uncertainty (poor modelling).
96 E. Canuto et al.

Figure 1 Block-diagram of embedded model and uncertainties (see online version for colours)

Figure 2 Block-diagram of the error loop (see online version for colours)

wu

Disturbance du
wd dynamics
D

Structured
uncertainty
h ( ⋅)
Control unit e
h (⋅) Dynamics d y State
Known ey
predictor e −
disturbance
M Sm
eˆy
wy

y Neglected e State eˆ y ey
− −
dynamics ∂P predictor
− ym
Vm Causality
S c ML w
Uncertainties

Stability inequalities are obtained in Section 3 from error e defined in (1), the tracking error ey and the prediction
equations driven by model uncertainties in Figure 1. They error eˆ y defined by
generate a loop passing through different errors, and
therefore designated as the ‘error loop’ in Canuto (2007). ey = y − ym
Three main errors enter the loop in Figure 2: the model error (2)
eˆ y = ym − yˆ m ,
 Robust control stability using the error loop
where y is the reference signal driving the loop and yˆ m is
the one-step prediction of the model output ym. None of the
above errors is measurable as they include the model output
ym, which is only available either as a signal class or from
simulated runs. The corresponding measurable errors that
are indicated as the output of the control unit in Figure 2,
are made available by replacing ym either with yˆ m or y in
(2). They correspond to the measured model error e , to the
measured tracking error eˆy and to the control error ey. They
are defined as follows
e = y − yˆ m = e + eˆ y
eˆy = y − yˆ m = ey + eˆy (3)
e y = y − y = eˆy − e = ey − e.
1.2 Content of this paper
In Section 3, this paper concentrates on the derivation and
properties of the error loop. The main result is the
following: structured and unstructured uncertainties may be
accommodated by the state predictor sensitivity Sm and its
complement Vm as shown in Figure 2. The result seems
partly departing from most of the literature, where either the
control law is dedicated to the purpose as in Solihin et al.
(2011) or no distinction between control and state predictor
is made. On the contrary, structured uncertainties must be
real-time estimated so as to update the embedded model.
This is obtained by shaping the lower-frequency part of the
sensitivity Sm. Unstructured uncertainties must be blocked
from spilling into the embedded model in order to prevent
instability. This is achieved by shaping the higher-frequency
part of Vm. Actually, as shown in Section 3, because of
causality, which prevents cancelling innovation wu in
Figure 1, – wu and wd can only be causally estimated –, the
state predictor is affected by the correction Sw = ScMLwSm
in Figure 2, right bottom, which depends on the control-law
sensitivity Sc, on the controllable dynamics M and on the
noise estimator Lw. Causality adds a degree of freedom and
shows how state predictor and control law intertwine in the
overall control sensitivity S and its complement V.
Shaping Sm and S is done by approximating them and
their complements Vm and V with low- and high-frequency
asymptotes, respectively, within the Nyquist frequency
band. Asymptotes can be related to noise estimator and
control gains and to state predictor and state feedback
eigenvalues. The intersections of the asymptotes with
the zero dB line are interpreted as the bandwidths of
(asymptotic) sensitivity and complement. The main result
affirms that a robust design accommodating structured and
unstructured uncertainty defines an admissible region Δfm of
the BW fm of the state predictor sensitivity Sm. The
admissible region in turn depends on the control ratio γ = fm
/ fc, fc being the BW of Sc, and on the stability margin η–1 >
1. Machinery and proof of the asymptotes are carried out
step by step throughout the paper.
Main contributions are credited to be:
97
1 the modelling procedure pivoting on the disturbance
dynamics D and on the driving noise as the sole
plant-to-model feedback
2 the performance equation (63) including all the
uncertainty contributions and their closed-loop
accommodating filters
3 the derivation of (63) in Section 3
4 the asymptotic inequalities in Section 3 and their
derivation throughout this paper.
More sophisticated algorithms, simulation and in-field
tuning may refine the resulting robust design.
Formulation employs discrete-time transfer functions
denoted as y(z) = M(z)u(z). Sometimes z is dropped. Since
this paper aims to be introductory, the univariate case is
treated.
2 The embedded model and the uncertainty
2.1 The embedded model and its realisation
2.1.1 Generalities
A discrete-time dynamics is considered, which is associated
to the time unit T, to the Nyquist frequency fmax = 0.5/T and
to discrete times iT. The model output ym is the response to
the command u and to the disturbance d as follows (see
Figure 1):
ym ( z ) = M ( z ) ( u ( z ) + d ( z ) ) = M ( z )u ( z ) + d y ( z )
(4)
d ( z ) = du ( z ) + wu ( z ) + h ( xc ) + h ( x c ) ,
where
M = Cc ( zI − Ac ) Bc = Cc N ( z ) / φ ( z ).
−1
(5)
is irreducible and of order nc. To be simple, CcN(z) is
assumed to be Hurwitz. The input signals u and d are scaled
to the output units; dy is the output disturbance to be used in
Section 3. The disturbance d is decomposed into the
following components:
1 The unknown disturbance du(z) = D(z)wd(z) is driven by
a noise vector wd of size nw through the irreducible
dynamics D = Mc(zI – Ad)–1Gd of order nd, with state
vector xd. The disturbance must be rejected by the
control action.
2 The unpredictable noise wu(z) cannot be cancelled due
to causality, since any sample wu(i) occurs while the
command sample u(i) actuates plant and model.
3 The non-linear feedback h(xc ) + h (xc ) depends on the
state xc of M; the unknown part h(xc) is treated as a
structured uncertainty to be estimated and rejected by
the control action.
The disturbance dynamics D plays the role of the weighting
functions in Maciejowski (1989) and Doyle et al. (1992),
but is designed to be parameter-free and explicitly coded in
98 E. Canuto et al.

the control unit. A disturbance adding to the command is the neglected dynamics, as well as amplifier dynamics and
simplest case, which is referred to as the collocated case. LuGre friction dynamics (Canudas de Wit et al., 1995).
The not collocated case is treated in Canuto (2007). Signals Transfer functions and h(·)in (4) hold
wd and wu may be treated as white noise, but when statistics
is unknown they may be also interpreted as bounded, ⎡ ym = θ ⎤
M ( z ) = ( z − 1) −2 , xc = ⎢ ⎥
arbitrary signals, whose average converges to a null value. ⎣ Δθ ⎦
Null value is therefore their best prediction. ⎡ wa ⎤ ⎡a ⎤
All the state variables of M and D are observable by the D( z )w d ( z ) = ( z − 1) −1 ⎣⎡1 ( z − 1) −1 ⎦⎤ ⎢ ⎥ ( z ), xc = ⎢ ⎥ (9)
w
⎣ s⎦ ⎣s⎦
model output ym. Low- and high-frequency asymptotes of M
for f < fmax are as follows: h ( xc ) = − ( sgn(Δθ ) A (| Δθ |) − mgl cos ( ym ) ) T 2 / J

lim f →0 M ( jf ) = ( f 0 / f )
m0 h ( xc ) = 0.
(6)
lim f →∞< fmax M ( jf ) = ( f ∞ / f )
m∞
, Parameters in (6) can be proved to hold
m0 = m∞ = 2, d 0 = d ∞ = 4
where m0 > 0 and m∞ > 0 imply a pole-zero excess of at (10)
least one. Since D is sized 1 × nw, it is post-multiplied times f 0 = f ∞ = f d 0 = f d ∞ = (2πT ) −1.
the arbitrary transfer vector Ld(z), nw × 1, in a way that no
The block diagram with input, state and output variables is
zero-pole cancellation occurs, and
in Figure 3. The noise size is nw = 3 and the embedded
lim z →1 L d ( z ) = Ld 0 < ∞ model size parameters are n = d0 = d∞ = 4. The
(7)
lim z →∞ L d ( z ) = Ld ∞ < ∞. block-diagram includes the reference dynamics to be
explained in Section 2.3. Boxes marked with Σ represent
Then, given Ld(z) satisfying (7), we assume discrete-time integrators having the generic state equation
lim f →0 M ( jf )D( jf )L d ( jf ) = ( f d 0 / f )
d0
x(i + 1) = x(i ) + u (i ), x(0) = x0 . (11)
(8)
lim f →∞< fmax M ( jf )D( jf )L d ( jf ) = ( f d ∞ / f )
d∞
, Initial states are marked by a superimposed arrow.

where d0 – m0 > 0 and d∞ – m∞ > 0 imply that DLd possesses Figure 3 Block diagram of the embedded model of the example
at least a unitary pole. The lim f →∞< fmax accounts for (see online version for colours)

asymptotes in the region f < fmax: they are due to poles in

excess of the type (z – 1 + β)–1 with 0 ≤ β < 1, and thus [ rad ] [ rad ] y (i )
[ rad ]
Δθ ( i ) e (i )
exclude pure delays. u (i ) ym ( i ) = θ ( i ) −
∑ ∑
The noise wd is multivariate, since D is the composition
of different stochastic signals, and each component is driven
by different noise components. Moreover, M and D are [ rad ] [ rad ]
assumed to be strictly causal, which implies that the relevant a (i ) s (i ) ws ( i )
∑ ∑
output ym and du are linear combinations of state vectors wa ( i )
only. D is usually stylised from experimental spectral wu ( i )
densities by whitening their perturbing sources as in Canuto
and Rolino (2004), and Canuto (2008). The embedded
model control implements the control unit around equation
Δθ ( i )
(4) written in the form of a state equation. This entails the u (i ) au ( i ) y (i )
∑ ∑
need of estimating the noise vectors wd(i) and wu(i) for
updating controllable and disturbance states. As in Kalman
filtering (see Kwakernaak and Sivan, 1972), the only way to Similar design model and considerations apply to a ball and
the purpose is to extract the causal estimates w and wu beam device (Keshmiri et al., 2012), when ball position is
from the model error e defined in (1). uncontrolled. The ball control would require a more
complex dynamics than (9).
2.1.2 Example
2.2 Noise estimation and output prediction
Consider a balanced robot arm moving, in a vertical plane, a
mass m that is distant l from the rotation centre. The 2.2.1 Generalities
counter-clockwise rotation θ with respect to the horizontal
Noise estimation forces the embedded model to update and
plane is denoted with ym. The angular rotation ω = y m is instantiate the signals ym(i) and du(i) into the realisations
replaced by the angular increment Δθ = ωT. The arm is yˆ m (i ) and dˆu (i ). The mark ^ denotes one-step prediction as
driven by a DC motor and is subject to gravity torque and
friction. The total inertia on the gear output shaft is J. Gear the relevant signals only depend on the past measures
backlash and torsional deformations are confined into y(i – k), k > 0. On the contrary, the noise estimates w d (i )
 Robust control stability using the error loop 99

and wu , and the actual (or measured) model error embedded model are one-step anticipated as indicated by
e (i ) = y (i ) − yˆ m (i ) are barred, meaning they depend on the the boxed z.
present measure y(i). Thus, one must distinguish between
Figure 4 Embedded model and noise estimator (see online
the a priori embedded model (4) and the a posteriori model version for colours)
which is implemented in the control unit and is forced by a
realisation e of e. Consequently, unknown prediction errors from plant
y (i )
establish between model signals and their prediction as
follows: u (i ) Controllable yˆ m ( i ) e (i )
dynamics
eˆ y = ym − yˆ m d̂ ( i ) M
(12)
eˆd = du + h ( xc ) − dˆu . z
Known cross Noise
coupling estimator
Defining the noise error as ew = wu − wu , a first expression h ( ⋅)
relating ê to the errors in (12) is z
wd (i )
dˆu ( i ) Disturbance
eˆ y ( z ) = M ( z ) ( eˆd + ew ) ( z ). (13) dynamics
Ld
D
To prove it with the help of (4), rewrite the former equation z wu ( i )
in (12) as follows: Lw
Embedded model
ym − yˆ m =M ( u + du + wu + h ( xc ) + h ( xc ) )
(14)
(
−M u + dˆu + wu + h ( xc ) . ) The coefficients of Ld and Lw, i.e., the noise estimator gains,
are fixed by the state-predictor closed-loop eigenvalues, i.e.,
Equation (14) yields (13) with the help of the second
by the poles of Sm. The closed-loop spectrum is denoted as:
equation in (12).
A further expression derives from the linear and Λ m = {1 − γm1 , ..., 1 − γmk , ..., 1 − γmμ } , (20)
time-invariant noise-estimator, which is driven by the
measured model error e and is partitioned as follows: where μ ≥ n = nc + nd is the order of the state predictor.
Eigenvalues in (20) are written in terms of the
w d ( z ) = L d ( z )e ( z ) complementary eigenvalue γmk. When the latter is real and
(15)
wu ( z ) = L w ( z )e ( z ), 0 < γmk ≤ 1, provides the frequency fmk [Hz] through
γmk = κm2πfmkT, where π–1 ≤ κm ≤ 1 and κm → 1 as γmk → 0.
where Ld and Lw are suitable transfer functions having finite
Asymptotic inequalities in Section 3 will be obtained by
gains for z → 1 and z → ∞, i.e.,
assuming equal and real eigenvalues, i.e.,
lim z →1 L w ( z ) = lw0
(16) γmk = γm = 2πf mT , k = 1, ..., μ, (21)
lim z →∞ L w ( z ) = lw∞ .
having set κm = 1, i.e., γm << 1. The frequency fm is referred
Specifically Lw must be all-pass for estimating the white to as the state-predictor frequency.
noise wu, whereas Ld may be low-pass. Replacing (15) in The low-frequency asymptote of Sm follows from (8)
(13) and dropping z yields and holds
eˆ = M ( h ( xc ) + du + wu + h ( xc ) ) − M ( DL d + L w ) e S m 0 ( jf ) = lim f →0 S m ( jf )
(17)
= d y − MHeˆ − MHe,
= lim f →0 (1 + M ( DL d + L w ) )
−1

where the output disturbance dy in (4) has been used

1 (22)
together with the overall feedback H = DLd + Lw and the = lim f →0
1 + ( f d 0 / f ) + lw 0 ( f 0 / f )
d0 m0
key error equality
d0
e = e + eˆ y . (18) ⎛ f ⎞
=⎜ ⎟ ,
Assumptions on Ld and Lw imply H to be proper. Inserting ⎝ f m0 ⎠
the sensitivity Sm = (1 + MH)–1 and the complement where fm0 = fd0. The high-frequency asymptote of Vm
Vm = 1 – Sm into (17), yields the second expression of eˆ, follows from (6) and holds
namely
lim f →∞< fmax Vm ( jf ) = lim f →∞< fmax M ( DL d + L w )
eˆ y ( z ) = − Vm ( z )e( z ) + S m ( z )d y ( z ), (19) m∞
⎛ f ⎞
which is illustrated in Figure 2. = ⎜ m∞ ⎟ (23)
⎝ f ⎠
Figure 4 shows the noise estimator driving the
f m∞ = f ∞ ( lw∞ )
1/ m∞
disturbance dynamics. All the output variables of the .
100 E. Canuto et al.

The frequencies fm0 and fm∞ are referred to as the sensitivity φm ( z ) = ( z − 1)4 + l0 ( z − 1)2 + l1 ( z − 1) + l2 . (27)
and complement BWs, and are proportional to fm through
the coefficients κm0 and κm∞ that depend on the spectrum Λm Since the polynomial in (27) lacks the third degree
in (20). They hold coefficient, no set {l0, l1, l2} exists capable of stabilising Sm.
The only alternatives are either to add a fictitious noise
f m 0 = κm 0 f m on the angular increment Δθ in Figure 3, as it is done by
f m∞ = κm∞ f m (24) Kalman filters, or to employ a dynamic estimator capable of
lw 0 = κ w0 f mnc , respecting noise layout and size. As an alternative
interpretation of the dynamic estimator, we may think of
where the polynomial relation of lw0 in (16) has been added. replacing the missing input noise with a further output q
which is linearly independent of e . In fact, any dynamic
2.2.2 Example filter, in the limit a delay, makes output independent of
input. As a result, the number of available feedback
Since the noise size nw = 3 is less than the order n = d0 = channels from model error to noise double, from three to
d∞ = 4 of the noise-to-output transfer function MD, a six. It can be shown that stability can only be recovered by
dynamic noise estimator is mandatory for stabilising the tuning the six gains together with the parameters of the
closed-loop state predictor (embedded model and noise dynamic filter from the error e to the output q.
estimator), as in Canuto et al. (2010). To prove it, assume Assume a first-order filter q = ( z − 1 + β ) −1 e , because
first a static noise estimator in (15), namely:
the noise deficiency is just ne = n – nw = 1. The following
⎡ wa ⎤ ⎡la ⎤ noise estimator replaces (25):
⎢ w ⎥ ( z) = ⎢l ⎥ e ( z) (25)
⎣ a⎦ ⎣ s⎦ ⎡la ⎤ ⎡ ma ⎤
wu ( z ) = lu e ( z ). L d ( z ) = ⎢ ⎥ + ⎢ ⎥ ( z − 1 + β ) −1
⎣ ls ⎦ ⎣ ms ⎦ (28)
The corresponding loop transfer function MH = M(DLd + L w ( z ) = lu + ( z − 1 + β ) −1 mu .
Lw) in (17) is found to be
l0 ( z − 1) 2 + l1 ( z − 1) + l2
M ( z )H( z ) = , (26)
( z − 1) 4

and the denominator of Sm in (19) holds

Figure 5 Block diagram of the embedded model plus noise estimator and control law (see online version for colours)
 Robust control stability using the error loop 101

The seven gains in (28) must be explicitly related to the 2.3 Reference dynamics and tracking error
closed-loop eigenvalues, or, that is the same, to the poles of
Performance refers to a class of reference signals y which
Sm, whose cardinality is μ = n + ne = 5. Equal eigenvalues
are assumed, namely Λm = {1 – γm, …, 1 – γm}. Since the is assumed to satisfy the same dynamics as in (4), but
seven gains in (28) satisfy the following five equations corrupted by a known disturbance du as follows:
β = 5γm y ( z ) = M ( z ) ( u ( z ) + du ( z ) ) . (31)
lu = lw0 = 10γm2 = κ w0 f m2
Here, du is assumed to be known, which is not the case
lu β + mu + la = 10γm3 (29)
when it is recorded from measurements. The reference
la β + ma + ls = 5γm4 signal class is shaped by the open-loop command u , which
ms = γm5 , is real-time computed by a reference generator (not treated
their selection may be optimised. Equation (29) shows that here, see Figure 5) capable of matching operator requests
the filter gain β is essential for recovering stability, as and technology limits. Typically, model predictive control
already mentioned. Moreover, either the pair (ma, mu) or (Camacho and Bordons, 2003) is concerned with it. The
reference state is denoted with x.
(la, ls) can be set to zero. The latter solution, i.e., la = ls = 0,
is preferable, since forcing them to be zero reduces the Control performance is assessed by means of the
contribution of e which is noisier than q. Using (29) and tracking error ey = y − ym , satisfying
(24), the asymptote parameters in (22) and (23) become
ey ( z ) = M ( z ) ( u + du − u − du − h(⋅) − wu − h (⋅) ) ( z )
(32)
( β / ms )1/ 4
γ ey ( z ) = Cc ( x − xc ) ( z ) = Cc e ( z ).
d 0 = 4, f m 0 = = 4 5 m = κm 0 f m
2πT 2πT (30)
In the example, du = 0, and the reference block-diagram is
l 10γm
m∞ = 2, f v∞ = u = = κ m∞ f m . shown in Figure 3, bottom. The reference state vector is
2πT 2πT
denoted by
Figure 5 shows the block-diagram of the embedded model
in Figure 3 plus the dynamic noise estimator (28) and the ⎡y = θ⎤
x=⎢ ⎥. (33)
control law to be outlined in Section 3.1.2. Gains are ⎣ Δθ ⎦
denoted with circles. Dashed circles correspond to zero
gains.
2.4 Model errors and uncertainties
Figure 6 Bode plots (magnitude) of the sensitivity transfer 2.4.1 Generalities
function and of the complement (see online version
for colours) Following the literature (Maciejowski, 1989), two kinds of
2 model errors are treated in addition to innovations wd and
10
wu.
Structured uncertainty is defined by the uncertain
0
‘static’ feedback h(xc) connecting xc to u as in (4). Being
10 unknown, it cannot be explicitly cancelled by the control
law in Section 3. The inclusion in (4) of the disturbance du
Magnitude

allows to surrogate h(·), but at the cost of closed-loop

-2
10 degradation since a feedback link is replaced by a
command-independent signal as du. Closed-loop stability
Complement V, γ=0.4, f m=1.9 Hz must be guaranteed. h(·) is assumed differentially bounded.
10
-4 The bound is obtained expanding h around the reference
Sensitivity S
state x as follows:
Predictor complement Vm

-6
Predictor sensitivity Sm h ( xc ) = h ( x ) + Δh ( x )( x − xc )
10 -1 (34)
10 10
0
10
1
10
2
x = x − α ( x − xc ) , 0 ≤ α < 1.
Frequency [Hz]
If x and the state tracking error e = x − xc are bounded,
Figure 6 shows typical Bode plots (magnitude) of the also x is bounded. Using (5) and (32), the first order
predictor transfer functions (dashed lines), and of the overall expansion in (34) is written as
transfer functions to be derived in Section 3.
102 E. Canuto et al.
Δh ( x ) N ( z )
Δh ( x )( x − xc ) ( z ) = e ( z ),
Cc N ( z )
and the parametric error eΔh is defined by
ΔH ( x , z )
eΔh ( z ) = M ( z )Δh ( x )( x − xc ) ( z ) =
φ ( z)
where φ has been defined in (5). ΔH ( x , z ) = Δh ( x ) N ( z ) is
a polynomial having degree mh = degΔH < degφ, and
bounded coefficients
Δhk ( x ) ≤ Δhk ,max ( x ) , k = 0, ..., mh − 1.
The same polynomial having the bounds in (37) as
coefficients is denoted with ΔH max ( x, z ).
Unstructured uncertainty is defined as an uncertain
dynamic operator ∂P(ym, …) from the model output ym to
the plant output y. Input-output stability and linear,
time-invariance are assumed, implying that the model error
(1) has the expression
e( z ) = ∂P ( z; p) ym ( z ) + wy ( z ), ∂P ( z; p) ∈ ∂P.
where ∂P is the uncertainty class defined by a parameter
vector p ∈ Π belonging to a bounded set Π, and wy is the
measurement noise. The transfer function ∂P results from
the sampled-data transform of a continuous-time dynamics
enclosed between the plant zero-order hold and the output
sampler. Since ∂P is dropped from the embedded model, it
may affect the overall stability, and calls for robust stability
conditions. Here, we assume that there exists a worst-case
element ∂Pmax(z; pmax) in ∂P such that
max f < fmax ∂Pmax ( f ) = max ∂P max f < fmax ∂P ( f )
0 < arg max f < fmax ∂Pmax ( f ) = f ∂P
= min ∂P arg max f < fmax ∂P ( f ) < f max
i.e., the peak of the worst case is the highest peak in the
class ∂P, and the peak frequency f∂P is the lowest in the class
∂P.
2.4.2 Example
Using (9), the transfer function in (36) is found to be
ΔH ( x , z )
=
( )
Δh1 Δθ ( z − 1) + Δh0 θ ( ), m
h
φ ( z) ( z − 1) 2
where the bounds of the dimensionless coefficients in (40)
are given in terms of the frequencies fg and f Δθ (Δθ ), i.e.,
( )
Δh1 Δθ
= sgn(Δθ )dA ( Δθ ) / d | Δθ | T / J ≤ 2πf v ( Δθ ) T (41)
Δθ =Δθ
( ) ( )
Δh0 θ = mglT 2 sin θ / J ≤ ( 2πf g T ) .
2
The frequency square in (41) is imposed by nc = degφ = 2. fg
(35) is the ‘pendulum’ natural frequency of the arm, whereas
2πf Δθ (Δθ ) is the friction/inertia pole, depending on the
reference angular rate Δθ , and such that f Δθ (Δθ ) ≤ f Δθ (0).
Neglected dynamics is restricted to gear backlash and
e ( z ), (36) torsional deformation. Due to backlash, dynamics is
non-linear. It can be linearly approximated for |e| ≥ emin, the
lower bound corresponding to the backlash width. Then
∂P ( z )  − ( z − 1) ( ( z − 1) 2 + β t ( z − 1) + α t )
2 −1
(42)
αt  ( 2πft T ) , βt  α t (α t + 2ζ t ) , pT = [ ft ζ t ] ∈ Π,
2
(37)
expresses the fractional error between the motor-gear-arm
dynamics having flexible transmission and the rigid body
(9). The parameters in p are the uncertain natural frequency
ft and the damping coefficient ζt. Typical Bode diagrams,
including a ball and beam device, are in Canuto et al.
(2012). The response peak and the peak frequency in (39)
hold
f ∂P = ft ,min = min ft ∈Π ft
(43)
max p∈Π ∂Pmax ( z ) = ( 2ζ t ,min ) .
−1
(38)
3 Control law and the error loop
3.1 Control law and tracking error
3.1.1 Generalities
Given (4) and (31), in order to respect causality, the control
law must depend either on the state variables or on their
combinations like ym and du, but not on wu. In addition, the
control law must guarantee that the tracking error
ey = y − ym is bounded, and that the mean value tends
(39)
asymptotically to zero. The control law
u ( z ) − u ( z ) − du ( z ) = C( z ) ( e ( z ) + eˆ( z ) )
(44)
−du ( z ) − h ( xc ) + eˆd ( z )
can be shown to bound ey under appropriate conditions on
C, and on the prediction errors eˆ y and eˆd . In (44), C(z) is
usually improper with a bounded low-frequency gain
c0 = limz→1C(z) ≠ 0 and
= 1, nc = 2 (40)
lim f →∞< fmax C( jf ) = c∞ f m∞ −1 . (45)
Inserting (44) into (32), simple manipulations yield
( I + M ( z )C( z ) ) ey ( z ) = −M ( z )C( z )eˆy ( z )
(46)
−M ( z ) ( eˆd + ew ) ( z ) − M ( z ) wu ( z ).
Then, by replacing (13) in (46), the tracking error equation
follows:
ey ( z ) = −eˆ y ( z ) − S c ( z )M ( z ) wu ( z ). (47)
In (47), the control sensitivity Sc = (1 + MC)–1 makes its
appearance, and the sign of eˆ y is due to a different sign of
 Robust control stability using the error loop 103

ym in the errors. The asymptotes of ScM and Vc = 1 – Sc can Performance only depends on the non-rejected noise wu
be found to be and initial conditions.

lim f →0 S c ( jf )M ( jf ) = lim f →0 (1 + MC) −1 M 2 Anti-causal law: it corresponds to include wu in (44),

which simplifies (47) to the ideal equality
( f 0 / f )m 0
1
= lim f →0 = ey = −eˆ y . (52)
1 + ( f0 / f )
m 0
c0 c0
(48)
Vc∞ ( jf ) = lim f →∞< fmax Vc ( jf ) = lim f →∞< fmax MC Moreover, adding eˆd = 0 and ew = 0 (model-based
assumptions), tracking error becomes zero, less the
c∞ ( f ∞ )
m∞
f c∞ free-response. Anti-causal laws cannot be realised since
= = .
f f u(i) depends on y(i – 1), whereas wu (i ) depends on
y(i). In terms of transfer functions, it would mean
The coefficients of C derive from the poles of Sc, are of the
Sc → 0 for f < fmax (the open switch in Figure 2 below,
order κ = nc ≥ m∞, and are designed to stabilise Sc. Likewise
ScMLw), which is unrealisable because of the Bode’s
in (20), the closed-loop spectrum is denoted with
integral theorem (Maciejowski, 1989; Wu and
Λ c = {1 − γc1 , ..., 1 − γck , ..., 1 − γcκ } , (49) Jonckheere, 1992).

and equal eigenvalues are assumed, γck = γc = 2πfcT, Clearly, (47) is a combination of the ideal equalities (51)
k = 1, …, κ. The control-law frequency is fc. Likewise in and (52).
(24) we can write Figure 7 shows the whole control unit that is built
around the embedded model: the control law interfaces
f c ∞ = κ c∞ f c model and plant, the noise estimator interfaces plant and
(50)
c0 = κc 0 f cnc model, and the reference generator interfaces operator/plant
and model. The delay z–1, from control law to plant,
Equation (47) may be better understood if one restricts to a balances the model one-step prediction. Digitisation
pair of ideal conditions: converts the computed command to a digital signal, and
may include a non-linear inversion. The opposite is denoted
1 Model-based control law: it corresponds to eˆd = 0,
with D/A and provides the embedded model with the same
ew = 0, to eˆ y = 0 in (13), and to the ideal equality in command dispatched to the plant.
(47)
ey = −S c Mwu . (51)

Figure 7 The control unit built around the embedded model (see online version for colours)
to plant from plant
y (i )

u (i ) Controllable yˆ m ( i ) e (i )
−
Digitisation z −1 D/A dynamics
d̂ ( i ) M
−
z
Known cross-
coupling
e ( i ) + eˆ ( i ) h ( ⋅) Noise estimator
−
C z
wd (i )
dˆu ( i ) Disturbance
Ld
dynamics
− D
z
wu ( i )
Lw
z

z
Reference y (i ) from
dynamics Reference plant/operator
du ( i ) M generator
Control law
u (i )

Embedded model

Control unit
104 E. Canuto et al.
The final expression relating ey to the model error e and to
the output disturbance dy in (4) is obtained as follows.
Firstly, by replacing (15) and (18) in (47) one finds
ey = − ( I + S c ML w ) eˆ y − S c ML w e.
Then, by replacing the prediction error eˆ y through (19) and
by defining the overall ‘control sensitivity’ S and the
complement V = 1 – S as follows:
S = Sm + Sw
V = 1 − S = Vm − S w
S w = S c ML wS m ,
the final expression is found to be
e ( z ) = V ( z )e( z ) − S( z )d y ( z ).
The following alternative expression of V will be employed
hereafter:
V = Vm − S c ML wS m = Vm − S c M ( H − DL d ) S m
= Vm Vc + S c MDL d S m .
The high-frequency asymptotes of the two terms in the last
part of (56) descend from (8), (23) and (48), and hold
( f m∞ )m ∞
f c∞
lim f →∞< fmax Vm Vc = = ⎜ v∞ ⎟
f m∞ +1
lim f →∞< fmax S c MDL d S m = ( f d ∞ / f )
Thus, if d∞ > m∞ + 1 and fv∞ ≥ fd∞, the former asymptote in
(57) dominates and the high frequency asymptote of V
simplifies to
m∞ +1
⎛ f ⎞
V∞ = lim f →∞< fmax V ( jf ) = ⎜ v∞ ⎟
⎝ f ⎠
f v∞ = κv∞ f m / γ1/ ( m∞ +1) , κv∞ = κ m∞ κc∞(
1/ m∞ +1)
The overall functions S and V combine the predictor
transfer functions Sm and Vm with the causality correction
Sw. The latter is imposed by wu not being rejected. Equation
(55) is the basic ‘error loop’ equation, to be further
elaborated by making explicit e and dy.
3.1.2 Example
C is a proportional plus derivative compensator, i.e.,
C( z ) = c∞ ( z − 1) + c0 .
Assuming that the poles of Sc are equal and collected in
Λc = {1 – γc, 1 – γc}, the gain equations are c∞ = 2γc,
c0 = γc2 = κc 0 f c2 . The overall block diagram of the control
unit is in Figure 5. The errors
eˆθ (i ) = eθ (i ) + eˆθ (i )
eˆΔθ (i ) = eΔθ (i ) + eˆΔθ (i )
are the ‘measured’ tracking errors already defined in (3).
3.2 Stability and performance versus structured and
unstructured uncertainty
3.2.1 Robust stability
(53) Input-output stability of (55) occurs if and only if:
1 S is asymptotically stable, which in turn requires the
stability of Sm and Sc, and
2 the input signals e and dy are bounded and causally
independent of ey .
(54) The latter condition only occurs when structured and
unstructured uncertainties are zero: h = ∂P = 0 (the open
switches in Figure 2), i.e.,
e( z ) = w y ( z )
(55) (61)
d y ( z ) = M ( z ) ( D( z )w d ( z ) + wu ( z ) ) ,
and imply that the tracking error is only forced by an
‘unpredictable’ noise as in the linear quadratic Gaussian
control (Kwakernaak and Sivan, 1972).
(56)
On the contrary, making explicit structured and
unstructured uncertainties in (55) by means of (38) and (4),
and dropping z, the implicit ‘error loop’ equation follows:
( )
ey = V∂P y − ey − SMh ( xc ) + Vwy
⎛ f ⎞
m∞ +1
(62)
−SM ( Dw d + wu ) .
⎝ f ⎠ (57)
Equation (62) is made explicit in ey through (34) and (36),
d∞
.
which yields the stability equation
(1 + V∂P + SΔH ( x ) / φ ) ey = V ( ∂Py + wy )
(63)
−SM ( h ( x ) + Dw d + wu ) .
, f ≥ f v∞ Equation (62) is graphically represented in Figure 2. There,
(58) the uncertainties in Figure 1 and the control unit are
. combined so as to build the ‘error loop’. The control unit,
detailed in Figure 7, has been transformed by the above
derivation into a combination of the ingredients of S and V,
namely, M (controllable dynamics), Sm (state predictor), Sc
(control law and controllable dynamics) and Lw (noise
estimator of wu).
Observe that:
1 the unstructured uncertainty ∂P is filtered by V, which
being a low-pass filter, implies that ∂P should become
significant only at higher frequencies
(59) 2 the structured uncertainty ΔH is filtered by S, which
being a high-pass filter, implies that ΔH should become
significant only at lower frequencies.
A sufficient stability condition in terms of the frequency
response may descend from the ‘small-gain theorem’ in
Desoer and Vidyasagar (1975), and leads to
(60) max f < fmax V ( f ) ∂P ( f ) + S( f )ΔH ( x , f ) / φ ( f )
(64)
≤ η < 1,
 Robust control stability using the error loop
where η-1 is a stability margin. Inequality (64), which is the
key result, is similar but not equal to inequalities in Doyle
et al. (1992), where stability and performance inequalities
are combined.
3.2.2 Performance
Inequality (64) allows to rewrite (63) as
ey ≤ (1 − η )
−1
( )
V ∂Py + wy − SM ( h ( x ) + Dw d + wu ) , (65)
which becomes the performance inequality, |.| being a
suitable norm. It is out of the paper aim to discuss and apply
(65). It suffices to point out that it can be split in two
components:
1 the ‘deterministic’ term V∂Py − SMh( x ) depending on
the reference trajectory
2 the ‘random’ term Vwy − SM (Dw d + wu ) depending
on noise components.
Given η and the eigenvalue range derived from (64) in
Section 3.3, the first term in (65) fixes the slew rate of the
reference trajectory, whereas the random part fixes sensor
and actuator noise. It has been already remarked that ey is
only available through simulated runs, and is related to the
measured control error ey – defined as reference minus
measure - through the model error e as in (3). Therefore, if
the tracking error ey is made bounded by (65) and
| ey | << | e | – which are mandatory requirements – the
control error ey becomes the opposite of the model error e.
Alternative and measurable conditions derive from the
former decomposition of ey in (3). In the case that | eˆy |→ 0,
which corresponds to approach the anti-causal law (52), the
control error becomes the opposite of the measured model
error e , a condition that can be easily verified in practice.
3.3 Asymptotic stability inequalities
Explicit expressions of (64) and (65) are obtained under the
following assumptions.
1 State predictor and control law have uniform but
distinct eigenvalues as already stated in (21) and (49).
Their ratio
γ = γm / γc = f m / f c
is referred to as the control frequency ratio.
2 The components of the inequality (64) reach their
maximum values within separate frequency domains:
the sensitivity frequency band, f ≤ fs0, and the
complementary sensitivity band, f ≥ fv∞, to be defined
below.
Consider the former component in (64) due to parametric
uncertainty, and replace S(f) with the low-frequency
asymptote S0(f). The latter, derived from (22), (48) and (24),
holds
105
S 0 ( jf ) = lim f →0 S( jf ) = lim f →0 S m (1 + S c ML w )
= ( f / fs0 ) 0 , f ≤ fs0
d
(67)
f s 0 = f m 0 / (1 + lw0 / c0 )
1/ d 0
= κ m 0 f m / (1 + κ s γ nc )
1/ d0
< fm0 ,
where κs = κw0 / κc0. Equation (67) affirms that the overall
sensitivity BW fs0 is narrower than the state predictor fm0,
because Sc is not zero. Since, as stated in Section 2.4,
ΔH (x ) / φ is strictly proper and nc = degφ < d0, the
parametric component SΔH / φ in (64) achieves its
maximum close to fs0. Then, replacing the coefficients of
ΔH (x ) with their bound in (37), the following inequality
results
max f < fmax S 0 ( f )ΔH ( x , f ) / φ ( f )
(68)
≤ (1 + εs ) ΔH max ( x, f s 0 ) / φ ( f s 0 )
where εs << 1 accounts for the actual |S(f)| in the
neighbourhood of fs0.
Maximisation of the second component in (64) employs
the asymptote (58). Considering (39) and assuming f∂P ≥ fv∞,
a local maximum of |V∞(f)∂P(f)| occurs for fv∞ ≤ f ≤ f∂P,
since |V∞(f)| in (58) is monotonically decreasing. Having the
Example of Section 2.1.2 in mind, we assume
max f < fmax V ( f )∂P ( f )
(69)
≤ (1 + εv )( f v∞ / f ∂P ) ∞ ∂Pmax ( f ∂P ; p max ) ,
v
where εv < 1 plays the role of εs in (68).
Finally, (64) splits into the asymptotic inequalities
( )
1/ v∞
f v∞ ≤ f ∂P η / ∂Pmax ( f ∂P ; p max ) , η <1
(70)
φ ( f s 0 ) ≥ ΔH max ( x, f s 0 ) / η, η < 1,
where εs in (68) and εv in (69) have been absorbed by the
stability margin 1/η.
Given η, (70) can be solved for fv∞ < fmax and fs0 < fmax.
The first inequality in (70) provides an upper bound to fv∞.
The second is a polynomial inequality, that, since degφ >
degΔHmax(·), provides a lower bound as fs0 → ∞. Bounds in
(70), if they are feasible, can be transformed into bounds on
the state-predictor frequency fm, and on the control ratio γ in
(66), by replacing the expressions (67) and (58) of fs0 and
(66)
fv∞, respectively. At the end, (70) can be replaced by
f m,min (γ, η) ≤ f m ≤ f m ,max (γ, η) < f max . (71)
Inequality (71) proves the following.
1 A robust design is driven by the predictor frequency fm,
which is bounded from above by the neglected
dynamics and bounded from below by the parametric
uncertainty.
2 The admissible band Δfm = fm,max / fm,min ≥ 1 depends on
the control ratio γ > 0 and on the stability margin
106 E. Canuto et al.

η–1 > 1. The asymptotic approximations that have been Figure 8 Admissible state predictor frequency fm for η = 0.5
adopted to obtain (71), tend to be accurate for γ ≤ 1. (see online version for colours)

Because of the assumptions made at the beginning of this 0.6 fm,max @ η=1/2
section, the closed-loop eigenvalues entailed by (71) must 10
fm,min @ η=1/2
be kept as first trial, to be refined by simulation and 0.5
10
in-field. The admissible band Δfm(γ) ≥ 1 provides
degrees-of-freedom to satisfy the performance inequality 0.4
10
(65) as in Canuto (2007, 2008).

fm [Hz]
0.3
10
3.4 Example: robust stability design
0.2
10
Using (9), (28) and (59), the low-frequency asymptote of S
and the high-frequency asymptote of V can be written as 0.1
10
S 0 ( z ) = ( z − 1) 4 β (1 + lu / c0 ) / ms
(72)
V∞ ( z ) = ( z − 1) −3 lu c∞ . 10
-1
10
0 1
10
φ [Hz/Hz]
They correspond to the asymptotes of the Bode plots in
Figure 6. Expressing fs0 and fv∞ in terms of fm and γ as in Figure 9 The admissible frequency band Δfm versus γ and η
(67) and (58), yields (see online version for colours)

(
f s 0 ( f m , γ ) = f m / 5 (1 + 2γ 2 ) )
1/ 4
η=0.2
(73)
f v∞ ( f m , γ ) = f m (20 / γ)1/3 . η=0.25
0 η=0.33
10
Δfm [Hz/Hz]

The stability inequalities in (70) become η=0.5

f v ( Δθ ) f g2
f s20 ( f m , γ ) − fs0 ≥
η η (74)
f v∞ ( f m , γ ) ≤ ft ,min ( 2ηζ t ,min ) .
1/3

They have been solved for fm and γ in the Table 1 using the
parameters of the same table. -2 0 2
10 10 10
Three alternative eigenvalues, denoted as cases 1, 2 and γ [Hz/Hz]
3, have been selected: they correspond to the square marks
on the line fm,max(γ) in Figure 8 and to γ = {1, 0.4, 0.2}. As The design has been restricted to the upper bound fm,max(γ)
shown by Figure 9, γ = 1 corresponds to the widest and to the region γ ≤ 1 in order to force the control error ey
admissible band Δfm, whereas γ = 0.2 is a lower threshold to stay within the target bounds in Figure 10. The frequency
since uncertainty cannot be more accommodated for γ < 0.2. bounds fm,max and fm,min are shown in Figure 8 versus the
In practice, given the state predictor frequency fm, the eigenvalue ratio γ for a stability margin η–1 = 2. The margin
control law frequencies fc and fv∞ in (73) become too large, is sufficient since the worst-case uncertainties enter the
and the first inequality in (70) (accommodating the bound expressions (74).
neglected dynamics) cannot be guaranteed.

Table 1 Parameters and design

Parameter Symbol Unit Value Comments

Gravity frequency fg Hz 0.11
Friction frequency fv Hz 0.25 Δθ ≥ 0.75 mrad

Transmission frequency ft,min Hz 22

Damping ζt,min 0.025
Backlash emin mrad ≤0.4
Predictor frequency: case 1 fm,max Hz 2.2 @ η = 0.5, γ = 1
idem, case 2 fm,max Hz 2.6 @ η = 0.5, γ = 0.4
idem, case 3 fm,max Hz 1.3 @ η = 0.5, γ = 0.2
 Robust control stability using the error loop 107

Figure 9 shows the admissible frequency band Δfm. The
horizontal dashed line corresponds to Δfm = 1 (the band has
no width). The largest achievable stability margin is close to
4. The widest admissible band occur for γ ≅ 1.
slower in the set point achievement, as anticipated above.
The backlash effect cannot be cancelled, being smaller than
the encoder quantisation. It couples with sensor quantisation
and gives rise to a limit cycle around the set point not larger
than 1 bit.

4 Simulated and experimental results Figure 11 Set point achievement for two different designs
(see online version for colours)
The alternative designs in Table 1 have been tested through
simulation and a ball and beam device leaving the ball 2.002
Reference
uncontrolled. Their performance has been compared
Measure: γ=0.4

Angular position [rad]

through the control error ey defined in (3). The angular
position is measured by an incremental encoder mounted on 2.001 Measure: γ=0.2
the gear output shaft. Encoder quantisation is of the same
order of the backlash in Table 1. Up and down motion is
driven by a suitable reference generator that accounts for 2
speed and current limits, which latter impose motion
duration. Figure 10 shows the control error during up and
down motion and intermediate halt intervals. Bounds to 1.999
control error have been set larger during motion (dashed
lines in Figure 10). 4 5 6 7 8 9
Time [s]
Figure 10 Simulated control error for two different designs
(see online version for colours) The trapezoidal angular position of the gear pivoting the
-3 beam is measured by the gear encoder and is shown in
x 10
Figure 12. The difference between reference and measured
4 Tolerance
angle – the control error – can be perceived from the
Tolerance
3
enlargement in Figure 13, when the reference angle reaches
Control error: reference-measure γ=0.2
the maximum value. The measured position reaches the
Control error: reference-measure γ=0.4
2 maximum value after a damped oscillation ending in a limit
cycle (the square wave overlapping the constant reference)
Angle [rad]

1 imposed by backlash, friction and encoder quantisation.

Figure 12 shows the ball stroke (both simulated and
0 experimental), that, being uncontrolled, swings between the
beam extremes with some delay owing to friction.
-1
Figure 12 Experimental and simulated gear and ball motion
-2 (see online version for colours)
Gear Angular Positions
-3
2 4 6 8 10 12 14
Time [s] 0.8
0.6
Performance improves by increasing γ from 0.2 (case 3) to
Angular position [rad]

0.4 (case 2) and to 1 (case 1), as Figure 10 shows. The result 0.4
is reasonable since a lower γ in Figure 8 narrows the 0.2
predictor BW and consequently increases the norm of the
second term in the right-hand side of (65). This clearly 0
occurs during up and down motion, say from 9 s to 12 s in -0.2
Figure 10, when friction, assumed unknown, cannot be fully
-0.4 Experimental
cancelled by the predicted dˆu in (44). When the set-point is
-0.6 Simulated
reached, control error is dominated by backlash and no Ball stroke [m]: simulated
significant difference appears between the different cases. -0.8 Ball stroke [m] : measured
The case 1, not shown in Figure 10, behaves more or less as
the case 2, except that the control error becomes noisier, due 12 14 16 18 20 22
Time [s]
to a larger control BW. Therefore, case 2 looks a good
trade-off: the relevant transfer functions have been reported The transient oscillation in Figure 13 is the same as in
in Figure 6. Figure 11 and would disappear by adding an appropriate
Figure 11 shows the achievement of a set point for the known term h in the control law. To prove this, the
cases 2 and 3. One may observe that the case 3 is much
108 E. Canuto et al.
measured angular position of the simulated and
experimental runs are compared in Figure 13. The
experimental measurement is free of the transient oscillation
because an appropriate h has been added to the control law
for accounting viscous and static friction. The result
highlights the different frequency bands where dˆu and
h (xˆ c ) counteract the plant uncertainty. Assuming that the
‘anti-causal limit’ holds, dˆu is estimated within the
sensitivity BW fm in (30). On the contrary, h (xˆ c ),
depending on the controllable state xˆ c , is estimated within
the control law BW fc, which, in absence of command
limitations, may approach the Nyquist frequency
fmax = 0.5/T = 100Hz. The limit cycles in Figure 13 are
different in their period because of a different static friction.
Experimental measurements in Figure 13 certify that the
design model is a faithful description of the real plant.
Figure 13 Gear set point achievement (enlargement of
Figure 12) (see online version for colours)
Gear Angular Positions
Experimental
0.79 Simulated
Angular position [rad]
0.785
0.78
0.775
12 14 16 18 20 22
Time [s]
5 Conclusions
The error loop, i.e., the loop from model to tracking error,
points out which is the key design tool for accommodating
uncertainties. The tool is the noise estimator, which is
responsible for the noise estimates that update the
disturbance state in the embedded model. Actually,
causality constraint adds a further degree-of-freedom to the
design, allowing the control law to play a role in making the
design feasible, especially when contrasting stability and
performance inequalities need to be satisfied. Employing
asymptotic expansions of the closed-loop transfer functions,
simple and explicit design formulae relating closed-loop
eigenvalues to model parameters and requirements have
been derived. Refinement may be pursued through
simulation and in field.
Acknowledgements
The paper is a revised and expanded version of a paper
entitled ‘The error loop and robust closed-loop stability’,
presented at IEEE ICMA 2012 Conference, Chengdu,
5–8 August 2012.
References
Bickel, R. and Tomizuka, M. (1999) ‘Passivity-based versus
disturbance observer based robot control: equivalence and
stability’, ASME J. Dynamic Systems, Measurement, and
Control, Vol. 121, No. 1, pp.41–47.
Calafiore, G. and Dabbene, F. (2002) ‘A probabilistic framework
for problems with real structured uncertainty in system and
control’, Automatica, Vol. 38, No. 8, pp.1265–1276.
Camacho, E.F. and Bordons, C. (2003) Model Predictive Control,
Springer Verlag, London.
Canudas de Wit, C., Olsson, H., Åström, K.J. and Lischinsky, P.
(1995) ‘A new model for control of systems with friction’,
IEEE Trans. on Automatic Control, Vol. 40, No. 3,
pp.419–425.
Canuto, E. (2007) ‘Embedded model control: outline of the
theory’, ISA Transactions, Vol. 46, No. 3, pp.363–377.
Canuto, E. (2008) ‘Drag-free and attitude control for the GOCE
satellite’, Automatica, Vol. 44, No. 7, pp.1766–1780.
Canuto, E. and Rolino, A. (2004) ‘Multi-input digital frequency
stabilization of monolithic lasers’, Automatica, Vol. 40,
No. 12, pp.2139–2147.
Canuto, E., Acuna Bravo, W., Molano, A. and Perez, C. (2012)
‘Embedded Model Control calls for disturbance modeling and
rejection’, ISA Transactions, Vol. 51, No. 5, pp.584–595.
Canuto, E., Molano, A. and Massotti, L. (2010) ‘Drag-free control
of the GOCE satellite: noise and observer design’, IEEE
Trans. on Control Systems Technology, Vol. 18, No. 2,
pp.501–509.
Chapellat, H., Dahleh, M. and Bhattacharyya, S.P. (1990) ‘Robust
stability under structured and unstructured perturbations’,
IEEE Trans. on Automatic Control, Vol. 35, No. 10,
pp.1100–1108.
Chen, S-H., Chou, J-H. and Zheng, L-A. (2005) ‘Stability
robustness of linear output feedback systems with
time-varying structured and unstructured parameter
uncertainties as well as delayed perturbations’, J. Franklin
Institute, Vol. 342, No. 2, pp.213–234.
Choi, Y., Yang, K., Chung, W.K., Kim, H.R. and Suh, I.H. (2003)
‘On the robustness and performance of disturbance observers
for second-order systems’, IEEE Trans. on Automatic
Control, Vol. 48, No. 2, pp.315–320.
Desoer, C. and Vidyasagar, M. (1975) Feedback Systems:
Input-Output Properties, Academic Press, New York.
Donati, F. and Vallauri, M. (1984) ‘Guaranteed control of
‘almost-linear’ plants’, IEEE Trans. on Automatic Control,
Vol. 29, No. 1, pp.34–41.
Doyle, J.C., Francis, B.A. and Tannenbaum, A.R. (1992) Feedback
Control Theory, Macmillan Pu. Co., New York.
 Robust control stability using the error loop
Foo, Y.K. and Soh, Y.C. (1993) ‘Robust stability bound for
systems with structured and unstructured perturbations’, IEEE
Trans. on Automatic Control, Vol. 38, No. 7, pp.1150–1154.
Hostetter, G. and Meditch, J. (1973) ‘On the generalization of
observers to systems with unmeasurable, unknown inputs’,
Automatica, Vol. 9, No. 6, pp.721–724.
Johnson, C. (1971) ‘Accommodation of external disturbances in
linear regulator and servomechanism problems’, IEEE Trans.
on Automatic Control, Vol. 16, No. 6, pp.635–644.
Katsura, S., Matsumoto, Y. and Ohnishi, K. (2007) ‘Modelling of
force sensing and validation of disturbance observer for force
control’, IEEE Trans. on Industrial Electronics, Vol. 54,
No. 1, pp.530–538.
Keshmiri, M., Jahromi, A.F., Mohebbi, A., Amoozgar, M.H. and
Xie, W-F. (2012) ‘Modelling and control of a ball and beam
system using model based and non-model based control
approaches’, Int. J. on Smart Sensing and Intelligent Systems,
Vol. 5, No. 1, pp.14–35.
Kwakernaak, H. and Sivan, R. (1972) Linear Optimal Control
Systems, J. Wiley & Sons, New York.
Maciejowski, J. (1989) Multivariable Feedback Design, Addison
Wesley, Wokingham, Berskshire, UK.
109
Mita, T., Hirata, M., Murata, K. and Zhang, H. (1998) ‘H∞ control
versus disturbance-observer-based control’, IEEE Trans. on
Industrial Electronics, Vol. 45, No. 3, pp.488–495.
Patre, P.M., MacKunis, W., Makkar, C. and Dixon, W.E. (2008)
‘Asymptotic tracking for systems with structured and
unstructured uncertainties’, IEEE Trans. on Control Systems
Technology, Vol. 16, No. 2, pp.373–379.
Ross Barmish, B. (1994) New Tools for Robustness of Linear
Systems, Macmillan Pu. Co., New York.
Schrijver, E. and van Dijk, J. (2002) ‘Disturbance observers for
rigid mechanical systems: equivalence, stability, and design’,
ASME J. Dynamic Systems, Measurement, and Control,
Vol. 124, No. 4, pp.539–548.
Solihin, M.I., Akmeliawati, R. and Legowo, A. (2011) ‘Robust
feedback control design using PSO-based optimization: a case
study in gantry crane control’, Int. J. of Mechatronics and
Automation, Vol. 1, No. 2, pp.121–131.
Wu, B-F. and Jonckheere, E.A. (1992) ‘A simplified approach to
Bode’s theorem for continuous-time and discrete-time
systems’, IEEE Trans. on Automatic Control, Vol. 37, No. 11,
pp.1797–1802.