» LECTURE NOTES

An Introduction to Disk Margins

PETER SEILER, ANDREW PACKARD, and PASCAL GAHINET

F
eedback controllers are designed to ensure stability
and achieve a variety of performance objectives, in- Summary

T
cluding reference tracking and disturbance rejection. his article provides a tutorial introduction to disk mar-
Control engineers have developed different types of “safe- gins. These are robust stability measures that account
ty factors” to account for the mismatch between the plant for simultaneous gain and phase perturbations in a feed-
model used for control design and the dynamics of the real back system. The article first reviews the classical (gain-
system. Classical margins account for this mismatch by in- only and phase-only) margins and their limitations. This
troducing gain and phase perturbations in the feedback. motivates the use of disk margins, which are defined using
The classical margins are measures of the gain and phase a set of perturbations that has simultaneous gain and phase
perturbations that can be tolerated while retaining closed- variations. A necessary and sufficient condition is provided
loop stability. to compute the disk margin for a single-input, single-output
This article first reviews classical margins and dis- feedback system. Frequency-dependent disk margins can
cusses several important factors that must be considered also be computed, yielding additional insight. The article
with their use (see “Summary”). First, real systems differ concludes with a discussion of stability margins for multi-
from their mathematical models in both magnitude and ple-input, multiple output (MIMO) feedback systems. A typi-
phase. These simultaneous perturbations are not captured cal approach is to assess robust stability “loop-at-a-time,”
by the classical margins, which only consider gain or phase with a perturbation introduced into a single channel and
perturbations but not both. Second, a small combination of all other channels held at their nominal values. MIMO disk
gain and phase perturbations may cause instability even margins provide a useful extension to consider simultane-
if the system has large gain/phase margins. This can be ous variations in multiple channels. This multiloop analysis
especially important when using automated, computer- can provide a more accurate robustness assessment as
based control design over a rich class of controllers. The compared to the one-loop-at-a-time approach.
optimization process may improve both the gain and phase
margins while degrading robustness with respect to simul-
taneous variations. are defined using a general family of complex perturba-
Third, margin requirements must account for the increase tions that account for simultaneous gain and phase varia-
in model uncertainty at higher frequencies. All design mod- tions. Each set of perturbations, denoted D (a, v), is a disk
els lose fidelity at high frequencies. Typical gain/phase parameterized by a size a and skew v. Given a skew v, the
margin requirements, for example, ! 6 dB and 45°, are suf- disk margin is the largest size a for which the closed loop
ficient only if the corresponding critical frequencies remain remains stable for all perturbations in D (a, v).
within the range where the design model is relatively accu- Theorem 1 gives an easily computable expression for the
rate. Fourth, there are alternative robustness margins that disk margin. The expression originates from a variation of
provide more useful extensions to multiple-input, multiple- the small-gain theorem [1]–[3] and provides a construction
output (MIMO) systems. One such extension (discussed for the “smallest” destabilizing complex (gain and phase)
later in the article) is the “multiloop” disk margin, which perturbation. This complex perturbation can be interpreted
accounts for separate, independent gain/phase variations in as dynamic, linear time-invariant (LTI) uncertainty. This is
multiple channels. useful as the destabilizing LTI perturbation can be incorpo-
The article next introduces disk margins as a tool for rated within higher-fidelity nonlinear simulations to gain
assessing the robust stability of feedback systems. Disk further insight. Frequency-dependent disk margins can
margins address, to some degree, the issues regarding clas- also be computed, which provides additional insight into
sical margins (as previously summarized). These margins potential robustness issues.
The class of disk margins defined using D(a, v) includes
Digital Object Identifier 10.1109/MCS.2020.3005277 several common cases that appear in the literature. First, it
Date of current version: 16 September 2020 includes the symmetric disk margins introduced in [4] and,

78 IEEE CONTROL SYSTEMS » OCTOBER 2020 1066-033X/20©2020IEEE

d licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on November 05,2024 at 06:01:51 UTC from IEEE Xplore. Restrictions apply.
 Real systems differ from their mathematical models in both
magnitude and phase.

more recently, discussed in [5] and [6]. Second, the general However, the real dynamics might have a different gain, as
disk margins include conditions based on the distance represented by gP. It is typically assumed that the gain of the
from the Nyquist curve of the loop transfer function to design model at least has the correct sign. Hence, only posi-
the critical -1 point [7]–[9]. This is related to an interpre- tive variations g 2 0 are of interest. The gain margin speci-
tation of disk margins as exclusion regions in the Nyquist fies the minimum and maximum variation for which the
plane. Third, the general disk margins include condi- closed loop remains stable and well posed, as in Definition 1.
tions based on multiplicative uncertainty models used in
robust c­ ontrol [1], [3]. Definition 1
Finally, the article reviews the use of disk margins for The gain margins consist of an upper limit g U 2 1 and a
MIMO feedback systems. A typical extension of classical lower limit g L 1 1 such that
margins for MIMO systems is to assess stability with a gain 1) the closed loop is stable and well posed for all positive
or phase perturbation introduced in a single channel. This gain variations g in the range g L 1 g 1 g U
“loop-at-a-time” analysis fails to capture the effect of simul- 2) the closed loop is unstable or ill posed for gain varia-
taneous perturbations occurring in multiple channels. Disk tions g = g U (if g U 1 3) and g = g L (if g L 2 0).
margins are extended to account for multiple-loop pertur- The upper gain margin is g U = + 3 if the closed loop
bations. This multiple-loop analysis provides an introduc- remains stable and well posed for all gains g 2 1. Similarly,
tion to more general robustness frameworks, for example, the lower gain margin is g L = 0 if the closed loop remains
structured singular-value n [10]–[15] and integral quadratic stable and well posed for all positive gains g 1 1. The
constraints [16]. reported gain margins are often converted to units of deci-
bels, that is, 20 log 10 ( g), where g is in actual units.
BACKGROUND The phase margin is the amount of allowable variation in
This section reviews background material related to the plant phase before the closed loop becomes unstable. This
dynamical systems and single-input, single-output (SISO) corresponds to phase perturbations f := e -jz with z ! R . The
classical control. This material can be found in standard nominal loop transfer function is given by z = 0 and f = 1.
textbooks on classical control [8], [17]–[19]. The term phase variation arises because + L f ( j~) = + L ( j~) - z;
that is, z modifies the angle (phase) of the dynamics. Phase
Classical Margins variations can occur due to time delays in the feedback loop
Consider the classical feedback system shown in Figure 1. (for example, due to implementation on embedded proces-
The plant P and controller K are both assumed to be LTI sors) or deviations in the plant dynamics. Sufficient phase
and SISO systems. The extension to MIMO systems is con- margin is required to ensure that such delays and model
sidered later. Assume the controller K was designed to sta- variations do not destabilize the system.
bilize the nominal model P. Because this nominal model is It can be shown that the positive and negative phases are
only an approximation for the “real” dynamics of the plant, equivalent in a certain sense: z 2 0 causes instability if and
control engineers have developed various types of safety only if - z causes instability. Specifically, if f = e -jz desta-
factors to account for the mismatch between the plant bilizes the loop, then 1 + fL (s) = 0 for some complex s with
model P and the dynamics of the real system. One way nonnegative real part. This implies 1 + frL (rs) = 0 (because
to account for this mismatch is to introduce the complex- the coefficients of L are real). It follows that fr = e jz also
valued perturbation f in Figure 1. Let L := PK denote the destabilizes the loop since s and sr have the same real part.
nominal loop transfer function. The perturbed open-loop
response is L f := f L, and the nominal design corresponds
to f = 1. As f moves away from one, the closed-loop poles
Lf: = fL
can transition from the open left half-plane (LHP) (stable)
into the closed right half-plane (RHP) (unstable). The clas- r e u y
sical gain and phase margins measure how far f can deviate K f P
from f = 1 while retaining closed-loop stability. −
The gain margin measures the amount of allowable per-
turbation in the plant gain. This corresponds to real pertur-
bations f := g ! R . That is, the model used for design is P. FIGURE 1 A feedback system, including perturbation f.

OCTOBER 2020 « IEEE CONTROL SYSTEMS 79

d licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on November 05,2024 at 06:01:51 UTC from IEEE Xplore. Restrictions apply.
 The phase margin specifies the maximum (positive or The condition for this stability transition is as follows: a
negative) variation for which the closed loop remains sta- gain f0 = g 0 or phase f0 = e -jz0 places a closed-loop pole on
ble and well posed, as in Definition 2. A related time delay the imaginary axis at s = j~ 0 if and only if 1 + f0 L ( j~ 0) = 0.
margin can also be defined. This condition causes the perturbed closed-loop sensi-
tivity S f 0 := 1/( 1 + f0 L) to have a pole at s = j~ 0 . The gain
Definition 2 margin is the smallest factor g (relative to g = 1) that puts
The phase margin consists of an upper limit z U $ 0 such that a closed-loop pole on the imaginary axis and, similarly,
1) the closed loop is stable and well posed for all phase for the phase margin. This condition can be used to com-
variations z in the range -z U 1 z 1 z U pute gain and/or phase margins from the Bode plot of
2) the closed loop is unstable or ill posed for z = z U (if the nominal loop L. It also suggests a bisection method
z U 1 3) . to numerically compute the gain and phase margins.
The phase margin is z U = + 3 if the closed loop remains An example is provided next as a brief review of the clas-
stable and well posed for all phases z U 2 0. Reported sical margins.
phase margins are often converted to units of degrees, that
is, z # 180°/ r, where z is in radians. (Note that complex Example 1
numbers repeat with every 360° = 2r change in phase; Consider a feedback system with the following plant P, con-
that is, e jz = e jz + 2r. The phase margin z U = 180° indicates troller K, and nominal loop L:
the closed loop is stable/well posed for −180° 1 z 1 +180°
but unstable or ill posed for z = 180°. The convention P (s) = 1 , K (s) = 25,
s 3 + 10s 2 + 10s + 10
z U = + 3 is equivalent to stability for all phases in the 
L (s) = 3 25 . (1)
range −180° # z # 180°.) 2
s + 10s + 10s + 10
There is a simple, necessary, and sufficient condition to
compute gain and phase margins. The nominal closed loop The nominal closed loop has poles in the LHP at - 9.33 and
is assumed to be stable. Hence, the poles are in the LHP. - 0.33 ! 1.91j and is, thus, stable. The poles of the closed-loop
The poles may transition from the LHP (stable) to the RHP system remain in the LHP for all gain variations f = g 1 1.
(unstable) due to the gain or phase variation. The smallest Hence, the lower gain margin is g L = 0. However, the closed-
variation that causes the transition from stable to unstable loop poles cross into the RHP for sufficiently large gains
occurs when a closed-loop pole crosses the imaginary axis. g 2 1. The upper gain margin g U = 3.6 marks the transition
This occurs when a gain or phase variation places a closed- as poles move from the LHP (stable) into the RHP (unstable).
loop pole on the imaginary axis at s = j~ 0 . The closed loop is stable for g ! [0, g U). For g = g U, the closed
loop has poles on the imaginary axis s = ! j~ 1 at the critical fre-
quency ~ 1 = 3.16 rad/s. That is, 1 + g U L( j~ 1) = 0, and it can
be verified that the perturbed sensitivity S g U = 1/(1 + g U L)
Bode Diagram has poles on the imaginary axis at s = ! j~ 1 .
The poles of the closed loop also cross into the RHP as
Magnitude (dB)

30
the phase increases. The phase margin z U = 29.1° marks the
0
transition as poles move from the LHP (stable) into the RHP
–30 (unstable). The closed loop is stable for z ! (- z U, z U). For
–50 z = z U, the closed loop has a pole on the imaginary axis
100 101 s = j~ 2 at the critical frequency ~ 2 = 1.78 rad/s. Again, this
Frequency (rad/s)
corresponds to 1 + e -jzU L ( j~ 2) = 0, and it can be verified
(a)
that the perturbed sensitivity has a pole on the imaginary
270 axis at s = j~ 2 .
180
Phase (°)

90
0 Data
Limitations of Classical Margins
–90 Model There are several important factors that must be considered
180 when using classical margins.
100 101 1) Real systems differ from their mathematical models in both
Frequency (rad/s)
magnitude and phase: The Bode plot in Figure 2 shows a
(b)
collection of frequency responses obtained from input–
output experiments on hard disk drives (blue). A low-or-
FIGURE 2 The experimental frequency responses from many hard
der model used for control design is also shown (yellow).
disk drives (blue) and a low-order design model (yellow) show-
ing (a) magnitude and (b) phase. These data are provided by The model accurately represents the experimental data
Seagate, and the frequency axis has been normalized for pro- up to 2–3 rad/s. However, the experimental data have
prietary reasons. both gain and phase variations at higher frequencies.

80 IEEE CONTROL SYSTEMS » OCTOBER 2020

d licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on November 05,2024 at 06:01:51 UTC from IEEE Xplore. Restrictions apply.
 These simultaneous perturbations are not captured by
the classical margins, which only consider gain or phase Nyquist Diagram
perturbations but not both. 0.8
2) Small plant perturbations may cause robustness issues 0.6
even if the system has large gain/phase margins: Real sys-
0.4
tems have simultaneous gain and phase perturba-

Imaginary Axis
tions, as noted in the first comment. Moreover, there 0.2
are examples of systems with large gain and phase 0
margins but for which a small (combined) gain/phase
–0.2
perturbation causes instability. See [1, Sec. 9.5] for the
construction of such an example. An extreme example –0.4
is given by the loop (2) in the box at the bottom of –0.6
the page. Figure 3 shows a portion of the Nyquist plot
–0.8
for this loop. The feedback system with L has phase –1 –0.5 0
margin z U = 45° and gain margins [ g L , g U] = [0.2, 2.1]. Real Axis
The points corresponding to the phase margin and
upper gain margin (-1/g U) are marked with green FIGURE 3 The Nyquist plot of the loop L in (2). This loop has large
squares in the figure. The classical margins are large. classical gain and phase margins (green squares) but poor robust-
ness to simultaneous gain and phase perturbations.
However, the Nyquist curve for L comes near to the
-1 point. Thus, small (simultaneous gain and phase)
­perturbations can cause the feedback system to become 4) There are alternative robustness margins that provide more
unstable. The key point is that some care is required useful extensions to MIMO systems: A typical extension of
when using classical gain and phase margins. This did classical margins for MIMO systems is to assess stabil-
not present itself as an issue when controllers were ity with a gain or phase perturbation introduced into
designed primarily with graphical techniques. These a single channel. This analysis is repeated for each in-
classical controllers were typically of limited complex- put and output channel. This “loop-at-a-time” analysis
ity and did not have enough degrees of freedom to get fails to capture the effect of simultaneous perturbations
into this corner. However, this issue can be especially ­occurring in multiple channels. Hence, it can provide
important when using automated, computer-based an overly optimistic view of robustness. Alternative ro-
control design over a rich class of controllers. The op- bustness margins are more easily extended to account
timization process may improve both gain and phase for “multiple-loop” perturbations, as discussed later in
margins while degrading robustness with respect to the article.
simultaneous variations.
3) Margin requirements must account for the increase in model SINGLE-INPUT, SINGLE-OUTPUT DISK MARGINS
uncertainty at higher frequencies: Consider again the hard This section introduces the notion of disk margins for
disk drive frequency responses shown in Figure 2. The SISO systems as a tool to address some of the limitations
design model (yellow) loses fidelity at high frequen- of classical margins. Disk margins are robust stability
cies. As a result, the margins must necessarily be larger measures that account for simultaneous gain and phase
at higher frequencies to ensure stability. Requirements perturbations. They also provide additional informa-
based on simple rules of thumb (for example, 45° of phase tion regarding the impact of model uncertainty at vari-
margin) are insufficient and must account for the expect- ous frequencies.
ed level of model uncertainty. For example, the design
model for the hard disk drive data is relatively accurate Modeling Gain and Phase Variations
at low frequencies. The typical 45° phase margin require- Gain and phase variations are naturally modeled as a
ment might be sufficient if the closed-loop bandwidth complex-valued multiplicative factor f acting on the open
remains below 2–3 rad/s, where the design model has loop L and yielding a perturbed loop L f = f L. This factor
small perturbations. However, this typical phase margin is nominally one, and its maximum deviation from f = 1
requirement will be insufficient if the closed-loop band- quantifies the amount of gain and phase variation. A family
width is pushed beyond 2–3 rad/s. of such models is

7 6 5 4
L(s) := - 47.252s 7- 20.234s -6135.4086s +561.6166s + 804 .6454s 3 + 600.0611s 2 + 59.1451s + 1.888 . (2)
99.8696s + 175.5045s + 673.7378s + 890.5109s 4 + 553.1742s 3 - 49.2268s 2 + 12.1448s + 1

OCTOBER 2020 « IEEE CONTROL SYSTEMS 81

d licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on November 05,2024 at 06:01:51 UTC from IEEE Xplore. Restrictions apply.
 f ! D(a, a, b) = $ 1 + ad : d ! C with ; d ; 1 a . , (3)
1.5 1 - bd

1 where a, b, and a are real parameters that define the set of

perturbations. The sets D(a, a, b) contain f = 1, correspond-
0.5 r ing to d = 0, and are delimited by a circle centered on the
Imaginary

φmax real axis (assuming ; ba ; 1 1). For example, the set D(a, a, b)
0 for a = 0.4, b = 0.6, and a = 0.75 is the shaded disk shown
γmin f=1 c γmax
in Figure 4. Note, the nominal value f = 1 is not necessarily
–0.5 at the disk center c. The real-axis intercepts c max and c min
determine the maximum relative increase and decrease of
–1 the gain. The line from the origin and tangent to the disk
determines the maximum phase variation z max achieved by
–1.5 any perturbation f ! D(a, a, b).
0 0.5 1 1.5 2 2.5 3
Real There are two issues with the family of models in (3).
First, if a = - b, then the set only contains the point f = 1.
FIGURE 4 The set of variations D (a, a, b) for a = 0.4, b = 0.6, and Thus, a + b ! 0 is required to avoid this degenerate case.
a = 0.75 . This is equivalent to D (a, v) for v = 0.2 and a = 0.75 . Second, the set is unchanged when multiplying a, b by some
constant and dividing a by the same constant; that is,
D ^a/|l|, la, lb h = D (a, a, b) for any l ! 0. This suggests
1 further imposing a + b = 1. It is useful to ­parameterize these
constants as a: = (1 - v) /2 and b: = (1 + v) /2, where v ! R is
0.5 a skew parameter. This yields the simplified parameterization
Imaginary

1+ 1-v d
f ! D(a, v) = * : d ! C with|d|1 a 4 . (4)
0
2
1- 1+v d
–0.5 2

Again, the sets D (a, v) are delimited by circles centered on

–1
–0.5 0 0.5 1 1.5 2 2.5 the real axis (assuming ; (1 + v) a/2 ; 1 1). The disk in Fig-
Real ure 4 is defined, in this simplified parameterization, by the
σ = 0, α = 0.67 choices v = 0.2 and a = 0.75. The intercepts on the real axis
σ = –1, α = 0.75 correspond to d = ! a and are given by
σ = –5, α = 0.36

(a) 2 - a (1 - v) 2 + a (1 - v)
c min = and c max = . (5)
2 + a (1 + v) 2 - a (1 + v)
1

0.5
The disk center and radius are
Imaginary

0 c = 1 (c min + c max) and r = 1 (c max - c min) . (6)

2 2

–0.5
The maximum phase variation satisfies sin z max = r/c
when r # c. This follows from the right triangle formed
–1
–0.5 0 0.5 1 1.5 2 2.5 from the origin, disk center, and point where the tangent
Real line intersects D(a, v) . If r 2 c, then D (a, v) contains the
σ = 0, α = 0.67 origin, and z max := + 3.
σ = 1, α = 0.54 There is some coupling between v and a. However,
σ = 5, α = 0.27 it is helpful to think of a as controlling the amount of
(b) gain and phase variation, while v captures the differ-
ence between the amount of relative gain increase and
decrease. First, consider the case v = 0. For this choice,
FIGURE 5 (a) A negative v skews the gain variation left, toward
c max = 1/c min ; that is, the maximum gain increase and
more gain decrease. (b) A positive v skews the gain variation
right, toward more gain increase. The parameter a is selected to decrease are the same in relative terms. This is referred to
maintain the same radius for all disks. as the balanced case.

82 IEEE CONTROL SYSTEMS » OCTOBER 2020

d licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on November 05,2024 at 06:01:51 UTC from IEEE Xplore. Restrictions apply.
 In Figure 5, an example of a balanced disk with v = 0 The set D(a max, v) is a stable region for gain and phase
and a = 2/3 is shown in both (a) and (b) (blue disk with variations; that is, variations by a factor f inside D(a max, v)
dashed outline). The real-axis intercepts c min = 0.5 and cannot destabilize the feedback loop. Note the set D(a, v) is
c max = 2 are balanced in the sense that they both corre- not necessarily a disk, as demonstrated in Figure 6. Hence,
spond to changing the gain by a factor of two. The disk the term disk, strictly speaking, refers to the disk |d|1 a. If
moves to the right when increasing v from the balanced little is known about the distribution of gain variations, then
case v = 0 and adjusting a to keep the radius constant. v = 0 is a reasonable choice as it allows for a gain increase
This is illustrated in Figure 5(b). This means that v 2 0 or decrease by the same relative amount. The choice v 1 0
models a gain variation that can increase by a larger factor is justified if the gain can decrease by a larger factor than
than it can decrease. Similarly, decreasing v from the bal- it can increase. Similarly, the choice v 2 0 is justified when
anced case v = 0 moves the disk to the left, as shown in Fig- the gain can increase by a larger factor than it can decrease.
ure 5(a). This means that the gain can decrease by a larger An alternative approach is to use D (a, v) to cover known
factor than it can increase, including changing sign. For gain and phase variations, for example, neglected actuator
v = - 1, the disk intercepts are c min = 1 - a and c max = 1 + a ; or sensor dynamics. This approach requires some knowl-
that is, the gain can increase or decrease by the same abso- edge of the plant modeling errors specified in terms of gain
lute amount. and phase variations. Then, a and v are selected to give
These examples clarify the meaning of the term skew for the smallest set D (a, v) that covers these known variations.
the parameter v. For v = 0, the nominal factor f = 1 is the The goal is then to assess the robustness of the closed loop
geometric mean of the range (c min, c max), and it moves off with respect to this set of variations. This second analysis
center when selecting a positive or negative value for v. In approach can be performed by computing the disk margin
summary, a skew v = 0 means that the gain can increase a max associated with the chosen skew v. If a max $ a , then
or decrease by the same factor; that is, it has a symmetric the closed loop is stable for all variations in D (a, v), and
range of variation in decibels. A nonzero skew indicates a the system is robust to the known modeling errors.
bias, on a logarithmic/decibel scale, toward gain decrease There is a simple expression for the disk margin a max .
(v 1 0) or gain increase (v 2 0). As with the classical margins, the nominal feedback sys-
For fixed v, the parameter a 2 0 controls the size of tem is assumed to be stable, and the closed-loop poles
the region D (a, v). This is illustrated in Figure 6 for v = 0. are in the LHP for f = 1. The poles move continuously in
The region is the interior of a disk for a 1 2/|1 + v|. The the complex plane as f ! D (a, v) is perturbed away from
size of the disk increases for larger values of a. The region
becomes a half-plane for a = 2/|1 + v| and the exterior of a
disk for a 2 2/|1 + v|. It can be shown, with some algebra,
that c max - c min = 8a/(4 - a 2 ( 1 + v) 2 ). Thus, if a 2 2/|1 + v|,
then c max 1 c min ; that is, c max becomes the “left” intercept 2 2
on the disk. Equation (6) still provides a valid definition
for the disk center c. However, the disk radius, in this less
0 0
common case, is r = ; c max - c min ;/2. The case a 1 2/|1 + v|
is most relevant in practice since it corresponds to the inte-
rior of a disk with bounded gain and phase variations. –2 –2
However, the case a $ 2/|1 + v| can be used to model situ-
ations where the gain can vary substantially or the phase 0 2 4 0 2 4
is essentially unknown. This qualitative analysis provides (a) (b)
guidance on the effect of the parameters v and a.

Disk Margins: Definition and Computation 2 2

There are two common robustness analyses that can be per-
formed with the set D (a, v) of gain and phase variations. The
0 0
first approach is to select v and compute the largest value of
a for which closed-loop stability is maintained. This yields
a stability margin, formally defined next, that can be used to –2 –2
estimate the degree of robustness for a feedback loop.
–2 0 2 4 –4 –2 0 2
Definition 3 (c) (d)
For a given skew v, the disk margin a max is the largest value
of a such that a closed loop with f L is well posed and sta- FIGURE 6 Increasing a increases the size of D (a, v) , as shown with
ble for all complex perturbations f ! D (a, v). v = 0: (a) a = 0.5, (b) a = 1, (c) a = 2, and (d) a = 4.

OCTOBER 2020 « IEEE CONTROL SYSTEMS 83

d licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on November 05,2024 at 06:01:51 UTC from IEEE Xplore. Restrictions apply.
 Proof of the Disk Margin Condition

I n here, the main technical result used to compute disk mar- The second equality simply groups the state matrix (A - BC)
gins (Theorem 1) is proven. It is assumed for simplicity that for the nominal closed loop with f = 1. The nominal closed loop
L has no feedthrough; that is, D = 0. The results require some is assumed to be stable. Thus, j~ 0 I - (A - BC) is nonsingular.
minor modifications for systems with nonzero feedthrough, for ex- Hence, (S1) is equivalent to
ample, to handle well posedness. First, the stability transition
0 = det ^I + (f0 - 1) ^ j~ 0 I - (A - BC) h-1 BC h. (S2)
condition is stated as a technical lemma with a formal proof
using state-space arguments. Finally, apply Sylvester’s determinant identity [22, Corr. 3.9.5]
to shift around C and obtain
Lemma S1
0 = 1 + (f0 - 1) C ( j~ 0 I - (A - BC)) -1 B = 1 + (f0 - 1) T ( j~ 0). (S3)
Assume the closed loop is stable for a nominal, single-input,
single-output (SISO) loop L. In addition, let ~ 0 be a given Note that T = L /(1 + L) and, hence, (S3) is equivalent to
frequency and assume L ( j~ 0) ! 0 . There is a perturbation 1 + f0 L ( j~ 0) = 0 . The perturbation can be expressed as
f0 ! D (a, v) that causes the closed loop to have a pole at f0 = (2 + (1 - v) d 0) / (2 - (1 + v) d 0) for so me d 0 ! C with
s = j~ 0 if and only if ^S ( j~ 0) + ^v - 1 2hh d 0 = 1 holds for some | d 0 | 1 a (4). Thus, (S3) can be rewritten, after some algebra,
| d0 | 1 a . in terms of the nominal sensitivity S: = 1/ (1 + L) as

` S ( j~ 0) + j d 0 = 1. (S4)
v-1
Proof 2
Let (A, B, C, D = 0) denote a state-space representation for the This final step requires the assumption that L ( j~ 0) ! 0. This
nominal loop L. Let Tf denote the transfer function from refer- ensures S ( j~ 0) ! 1 and d 0 ! (1 + v) /2 so that the correspond-
ence r to output y for the perturbed feedback system in Fig- ing perturbation f0 is finite. 4
ure 1, that is, the complementary sensitivity function. The nota- The main disk margin condition (Theorem 1) is restated be-
tion T with no subscript will refer to the nominal complementary low with a formal proof. This is a variation of a technical result
sensitivity with f = 1. known as the small-gain theorem [1]–[3].
A state-space realization for the perturbed Tf is given by
(A - fBC, B, C, 0) . Hence, the condition for some f0 ! D (a, v) to Theorem 1 (Restated)
cause a closed-loop pole at s = j~ 0 is Let v be a given skew parameter defining the disk margin. As-
0 = det ^ j~ 0 I - (A - f0 BC) h sume the closed loop is well posed and stable with the nomi-
= det ^ j~ 0 I - (A - BC) + (f0 - 1) BC h .
 nal, SISO loop L. The disk margin is given by
(S1)

f = 1. The poles may move into the RHP (unstable closed < G < 3 := max |G ( j~)|. (8)
~ ! R , {+ 3}
loop) if f is varied by a sufficiently large amount from the
nominal value f = 1. The transition from stable to unsta- This is called the H 3 norm for the stable system G, and it
ble occurs when the closed-loop poles cross the imaginary corresponds to the largest gain on the Bode magnitude plot.
axis. The condition for this stability transition is as follows:
a ­perturbation f0 ! D (a, v) places a closed-loop pole on the Theorem 1
imaginary axis at s = j~ 0 if and only if 1 + f0 L ( j~ 0) = 0. Let v be a given skew parameter defining the disk margin.
The definition of D(a, v) (4) i m pl i e s t h a t f0 = Assume the closed loop is well posed and stable with the
(2 + ( 1 - v) d 0)/(2 - (1 + v) d 0) for some d 0 ! C with |d 0|1 a. nominal, SISO loop L. The disk margin is
Thus, the stability transition condition can be rewrit-
ten, after some algebra, i n terms of t he sensit ivit y a max =
1 . (9)
v-1
S := 1/(1 + L) as S+
2 3

Proof
` S( j~ 0) + j d 0 = 1. (7)
v-1
2 Below is a sketch of a proof; see “Proof of the Disk
Margin Condition” for a more formal proof. Let a 0: =
To summarize, some f0 ! D(a, v) causes a closed-loop 1/< S + (v - 1) /2 < 3 and consider any f ! D (a 0, v) with
pole at s = j~ 0 if and only if ^S ( j~ 0) + (v - 1)/2h d 0 = 1 corresponding ; d ; < a 0 . Equation (9) implies the inequal-
holds for some |d 0|1 a. This condition forms the basis for ity ; S (j~) + (v - 1) /2 ; $ ; d 0 ; 1 1 for all ~. Hence, (7) cannot
Theorem 1 regarding the disk margin. The theorem uses hold, and no closed-loop pole can cross the imaginary axis
the following notation for the peak (largest value) gain of while f stays inside D (a 0, v). This shows that a max $ a 0 .
a stable, SISO, LTI system G, Conversely, consider the peak f requency ~ 0, whe r e

84 IEEE CONTROL SYSTEMS » OCTOBER 2020

d licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on November 05,2024 at 06:01:51 UTC from IEEE Xplore. Restrictions apply.
 a max =
1  (9, restated) closed loop is stable for all f ! D (a 0, v) because the poles for
S+ v-1 the nominal system are in the left half-plane and do not cross
2 3
the imaginary axis into the right half-plane (RHP). This can be
Proof formalized with a homotopy argument and proof by contradic-
-1
Define a 0: = S + (v - 1) /2 3 . The proof consists of two steps. tion. Specifically, suppose the closed loop has a pole in the
First, it is shown that there is a destabilizing perturbation on RHP for some f0 ! D (a 0, v) . Consider the equation parameter-
the boundary of D (a 0, v) . The perturbation set D (a, v) con- ized by 0 # x # 1:
tains this destabilizing perturbation for any value a $ a 0 .
Hence, the disk margin satisfies a max # a 0. Second, it is
0 = det ^sI - ^ A - f (x) BC hh where f (x): = 1 + x (f0 - 1). (S5)
shown that the closed loop is stable and well posed for all
perturbations f ! D (a 0, v) . It follows from these two steps
that a max = a 0 . For each value of x, this is a polynomial in s whose
Fo r t h e f i r s t s te p, l e t ~ 0 b e t h e fr e q u e n c y w h e r e roots correspond to the poles of the closed loop with per-
S + (v - 1) /2 achieves its peak gain. Define the per- turbation f (x) . For x = 0 , this corresponds to the nominal
t u r bation d 0: = (S (j~ 0) + (v - 1) /2) -1 ! C. By construction, feedback system (f = 1), and all roots are in the LHP by
(S ( j~ 0) + (v - 1) /2) -1 d 0 = 1 and, hence, by Lemma S1, the assumption. For x = 1, this corresponds to the perturbed
corresponding f0 places a closed-loop pole at s = j~ 0 . More- feedback system f 0 , and there is a root in the RHP by as-
over, | d 0 | : = a 0. Hence, the corresponding f0 is on the bound- sumption. Note that f (x) remains in the disk D (a 0, v) for
ary of D (a, v) . One technical detail arises if L ( j~ 0) = 0 . This all 0 # x # 1. The roots of a polynomial equation are con-
corresponds to a max = 2 / ; v + 1 ;, which implies the closed loop tinuous functions of the coefficients. Hence, there must be
retains stability for any perturbation in the half-plane shown some x ! [0, 1] for which (S5) has a root on the imaginary
in Figure 6(c). axis. This implies that the closed loop with perturbation
Next, show the closed loop is stable and well posed for all f (x) ! D (a 0, v) has a pole on the imaginary axis. However, it
perturbations f ! D (a 0, v) . Each such perturbation can be ex- has been shown that no perturbation can cause the closed
pressed as f = (2 + (1 - v) d) / (2 - (1 + v) d) for some | d | 1 a 0 . loop to have roots on the imaginary axis. Thus, the original
The bound | d | 1 a 0 implies that (S ( j~) + 0.5 (v - 1)) d ! 1 for assumption that f 0 causes an RHP root is false. That is,
all ~ . It follows, again, by Lemma S1 that the closed loop has the poles of the closed loop must remain in the LHP for all
no poles on the imaginary axis for any f ! D (a 0, v) . Hence, the perturbations in D (a, v) .  4

; S ( j~ 0) + (v - 1) /2 ; = < S + (v - 1) /2 < 3 . T he n, d 0 defined eval­­uate the frequency response of S( j~ 0) and define

d 0 := ^ S ( j~ 0) + (v - 1)/2 h . The corresponding perturbation
-1
by (7) causes a stability transition at s = j~ 0 and satisfies
; d 0 ; = a 0 . Hence, a max # a 0 .  4 f 0 = (2 + (1 - v) d 0)/(2 - (1 + v) d 0) causes the closed loop to
The margin a max decreases as S + (v - 1)/2 3 increases; be unstable (if ~ 0 finite) with a pole on the imaginary axis
that is, large peak gains of S + (v - 1)/2 correspond to small at s = j~ 0 or ill posed (if ~ 0 = 3). If this construction yields
robustness margins. Several special cases are often consid- d 0 = 2/(v + 1), then f 0 = 3. This occurs when S( j~ 0) = 1
ered in the literature. The disk margin condition for the bal- and L( j~ 0) = 0. This corresponds to the trivial case where
anced case (v = 0) can be expressed as a max = ; (S - T ) /2 ; -31 . D(a max, v) is a half-space (a max = 2/|1 + v|), and the closed
This is known as the symmetric disk margin [4]–[6] because loop retains stability for any perturbation in this half-space.
the disks D(a max, v = 0) are balanced in terms of the rela-
tive gain increase and decrease. If v = -1 or v = +1 , then Example 2
Consider again the loop L (s) = 25/ ^s 3 + 10s 2 + 10s + 10h in­­
-1
the disk margin condition simplifies to a max = T 3 , and
-1
a max = S 3 , respectively. These special cases are called troduced in Example 1. The feedback system with this loop
T-based and S-based disk margins. is nominally stable. By Theorem 1, the symmetric disk mar-
Efficient algorithms are available to compute both the gin for v = 0 is given by a max =; (S - T ) /2 ; -31 . The peak gain
peak gain of an LTI system and the corresponding peak of (S - T ) /2 is 2.18 at the critical frequency ~ 0 = 1.94 rad/s.
frequency [20], [21]. T hese ca n be used to compute This yields a symmetric disk margin of a max = 0.46. The
S + (v - 1)/2 3 and, thus, the disk margin. The formal corresponding symmetric disk D (a max, v = 0) has real-axis
proof of Theorem 1 also provides an explicit con- intercepts at c min = 0.63 and c max = 1.59.
st ruction for a destabilizing perturbation f0 on the The closed loop is stable for all gain and phase perturba-
boundary of D (a max, v). First, compute the frequency tions in the interior of this disk. However, there is a desta-
~ 0, where S + (v - 1)/2 achieves its peak gain. Next, bilizing perturbation on the boundary of D (a max, v = 0).

OCTOBER 2020 « IEEE CONTROL SYSTEMS 85

d licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on November 05,2024 at 06:01:51 UTC from IEEE Xplore. Restrictions apply.
 This construction yields d 0 = 0.212 - 0.406j and the destabiliz- Connections to Gain and Phase Margins
ing perturbation f0 = 1.128 - 0.483j . The closed loop with Disk margins are related to the classical notion of gain and
this perturbation is unstable, with a pole at s = j~ 0 . Figure 7 phase margins, but they provide a more comprehensive
shows the closed-loop sensitivities for the nominal f = 1 assessment of robust stability. Specifically, the uncertainty
(blue solid) and destabilizing perturbation f0 (red dashed). model D (a, v) accounts for simultaneous changes in gain
The perturbed sensitivity has infinite gain at the critical fre- and phase, whereas the classical margins only consider
quency ~ 0 due to the imaginary axis pole. 3 variations in either gain or phase. The disk margin frame-
The destabilizing perturbation 0 is a complex number
f work models gain and phase variations as a multiplicative
with simultaneous gain and phase variation. This critical factor f, taking values in D (a, v). Perturbations on the unit
perturbation causes an instability with a closed-loop pole circle ^|f|= 1h correspond to phase-only variations, while
on the imaginary axis at the critical frequency ~ 0 . This perturbations on the real axis ^ f ! R h correspond to gain-
complex perturbation f0 can be equivalently represented only variations.
as an LTI system with real coefficients. Specifically, there is The disk margin a max can be used to compute guaran-
a stable, LTI system ft0 such that 1) ft0 ( j~ 0) = f0 and 2) ft0 ( j~) teed gain and phase margins, denoted (c min, c max) and z m ,
remains within D (a max, v) for all ~. This LTI perturbation respectively, as shown in Figure 8. Recall that closed-loop
ft0 can be used within higher-fidelity nonlinear simulations stability is maintained for all f in the open set D(a max, v).
to gain further insight. Details on this LTI construction are Specifically, the closed loop is stable for the portions of the
provided in “Linear Time-Invariant Perturbations.” unit circle and real axis that intersect the disk D(a max, v).
This provides lower estimates (c min, c max) and (-z m, z m) for
the admissible classical gain-only and phase-only varia-
Bode Diagram tions. The real-axis intercepts correspond to d = ! a max and
40 are given by
Nominal Sensitivity
30 Perturbed Sensitivity
Critical Frequency 2 - a max (1 - v) 2 + a max (1 - v)
c min = and c max = . (10)
2 + a max (1 + v) 2 - a max (1 + v)
Magnitude (dB)

20

10 To determine z m, note that the unit circle intersects the

boundary of D(a max, v) at cos z m + j sin z m . Consider the
0
(possibly oblique) triangle formed by this intersection point,
–10 the origin, and the center c of D (a max, v). Apply the law of
cosines to this triangle to obtain r 2 = 1 + c 2 - 2c cos z m . This
–20 yields the following expression for z m:
10–1 100 101
Frequency (rad/s) 2 2 1 + c min c max
cos z m = 1 + c - r = . (11)
FIGURE 7 The Bode magnitude plot of sensitivities for nominal
2c c min + c max
f   = 1 and destabilizing perturbation f0 = 1.128 - 0.483j .
If D (a max, v) fails to intersect the unit circle [for example,
D (a max, v) entirely contains the unit disk], then the right
side of (11) will have magnitude greater than one. In such
1 cases, z m : = + 3, and the feedback system is stable for any
phase variation.
φm
Note that (c min, c max) and (-z m, z m) are safe levels of
0.5
gain-only and phase-only variations. Each value of v yields
Imaginary

a new pair of such estimates, and we can vary v to refine

0 these estimates. This is of limited practical value, however,
γmin γmax
since we can directly compute the classical margins and
D (α max, σ ) varying v amounts to make assumptions about the gain
–0.5
variations that may not hold for the real system.
More importantly, disk margins can be used to quan-
–1 tify the effect of combined gain and phase variations that
0 0.5 1 1.5 2 occur in any real feedback loop. This can, again, be done
Real using simple geometry. First, consider a given level c of
gain variation, as shown in Figure 9(a). The intercepts of
FIGURE 8 The guaranteed gain and phase margins from largest the line y = x tan z with the bounding circle of D(a max, v)
disk D (a, v) maintaining stability. determine the safe range (-z, z) for phase variations

86 IEEE CONTROL SYSTEMS » OCTOBER 2020

d licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on November 05,2024 at 06:01:51 UTC from IEEE Xplore. Restrictions apply.
 Linear Time-Invariant Perturbations

T he main disk margin result (Theorem 1) provides a con- This technical lemma can be applied to obtain the LTI desta-
struction for a destabilizing perturbation f0. This perturba- bilizing perturbation from the disk margin analysis. Let f0 denote
tion is a complex number with simultaneous gain and phase a destabilizing complex perturbation in D (a max, v) with critical
variations. The perturbation can be equivalently represented frequency ~ 0. This destabilizing perturbation is constructed
as a linear time-invariant (LTI) system with real coefficients. from a corresponding d 0 ! C with | d 0 | = a max . By Lemma S2,
This equivalence is based on the following technical lemma. if ~ 0 is finite and nonzero, then there is a stable LTI system dt 0
such that dt 0 ^ j~ 0h = d 0 . If ~ 0 = 0 or 3, then d 0 will be real, and
Lemma S2 a constant system can be selected (that is, dt 0 = d 0) . In either
Let a finite frequency ~ 0 2 0 and a complex number d 0 ! C case, the dynamic perturbation dt 0 can be chosen as a constant
be given. There exists a stable, LTI system dt 0 such that or first order. In addition, the dynamic perturbation has a norm no
dt 0 ( j~ 0) = d 0 and || dt 0 | | 3 # | d 0 | . larger than the given uncertainty (that is, || dt 0 || 3 # | d 0 | = a max) .
Finally, define the following LTI perturbation:
Proof
2 + (1 - v) dt 0
The basic idea is that, if b 2 0 , then H (s) : = (s - b) / (s + b) is sta- ft0 = . (S6)
2 - (1 + v) dt 0
ble with magnitude | H ( j~) | = 1 for all ~ . This is called an all-pass
system. Moreover, the phase of H goes from 180° down to 0° with If (1 + v) a max /2 1 1, then this perturbation ft0 is stable (by the
increasing frequency. Similarly, - H (s) is stable and all pass, and small gain theorem) and on the boundary of D (a max, v) for all
it has a phase that goes from 360° down to 180°. Thus, a transfer frequencies. The system dt 0 has, at most, one state, and a mini-
function of the form ! c (s - b) / (s + b) is given, where c > 0 can mal realization of ft0 will also have, at most, one state. More-
match any desired magnitude and phase at a given frequency. over, ft0 (j~ 0) = f0 , and, thus, ft0 (j~ 0) causes the closed loop to
The remainder of the proof provides details for the construction. be unstable with a pole at s = j~ 0 . The LTI perturbation ft0 can
If d 0 ! R, then simply select the (constant) system dt 0 : = d 0 . be used within higher-fidelity nonlinear simulations to gain fur-
Consider the alternative where Im {d 0} ! 0 . In this case, d 0 = ! ce jz ther insight.
for some c > 0 and z ! (0, r) . Specifically, if Im {d 0} 2 0 , then
ce jz is the polar form for d 0 . If Im {d 0} 1 0 , then it has phase Example S1
+d 0 ! (- r, 0) . Hence, +d 0 = z - r for some z ! (0, r) , and d 0 The symmetric disk margin was computed in Example 2 for
has the polar form ce j (z - r) = - ce jz . the loop L (s) = 25/ (s 3 + 10s 2 + 10s + 10). The disk margin is
Next, note that, for b 2 0, the phase of H (s) = (s - b) / (s + b) a max = 0.46 with critical frequency ~ 0 = 1.94 rad/s. In addi-
is tion, the destabilizing perturbation f0 = 1.128 - 0.483j was con-
structed from d 0 = 0.212 - 0.406j . The complex number d 0 has
+H ( j~) = + ( j~ - b) - + ( j~ + b)
magnitude 0.458 and phase –1.089 rad. Hence, it can be ex-
= ; r + tan -1 c mE - ; r - tan -1 c mE = 2 tan -1 c m .
b b b
2 ~ 2 ~ ~
pressed as d 0 = - ce jz, with c = 0.458 and z = 2.052 rad. Select
b = ~ 0 tan ^z/2 h = 3.226 . Based on the proof for Lemma S2, the
As mentioned, the phase of H goes from r rad down to zero as the first-order system dt 0 (s) : = - 0.458 (s - 3.226) / (s + 3.226) is sta-
frequency increases. Thus, b can be selected to achieve the phase ble, with dt 0 ( j~ 0) = d 0 and || dt 0 || 3 = a max . Equation (S6) with v =
z ! (0, r) at the specified frequency ~ 0. Select b = ~ 0 tan ^z/2 h 0 yields the LTI perturbation ft0 = (0.627s + 3.226) / (s + 2.2024).
so that H ( j~ 0) = e jz . Finally, define dt 0 (s) : = ! c (s - b) / (s + b) It can be verified that ft0 ( j~ 0) = f0 . Hence, the perturbed closed-
with this b and the appropriate sign for ! c . Then, dt 0 is sta- loop sensitivity S: = 1/ (1 + f0 L) is unstable, with a pole on the
ble, with dt 0 ( j~ 0) = d 0 and || dt 0 || 3 = | d 0 |.  4 imaginary axis at s = j~ 0 .

c­ oncurrent with the gain c. By the law of cosines, the value rent gain variations. Again, by the law of cosines, the values
of z satisfies r 2 = c 2 + c 2 - 2cc cos z. This can be equiva- - +
c and c are the roots of (12) with the phase variation
lently expressed as z given.
The locus of (c, z) solutions delimits the “safe” variations
2
c - c (c min + c max) cos z + c min c max = 0. (12) as shown in Figure 10 in units of (decibels, degrees). The same
bounding curve is obtained from the perturbations f corre-
This expression with gain level c = 1 simplifies to the sponding to d = a max e ji with i ! [0, r]. This parameterizes
previous relation for z m (11). Next, consider a given level the bounding curve as (c, z) = (| f |, angle( f )) with
z of phase variation, as shown in Figure 9(b). The inter-
cepts of the line y = x tan z with the bounding circle of 2 + (1 - v) a max e ji
f= , i ! [0, r]. (13)
D(a max, v) determine the safe range (c -, c + ) for concur- 2 - (1 + v) a max e ji

OCTOBER 2020 « IEEE CONTROL SYSTEMS 87

d licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on November 05,2024 at 06:01:51 UTC from IEEE Xplore. Restrictions apply.
 The classical gain-only and phase-only margin esti-
mates correspond to the boundary points (0, z m) and
1 (20 log 10 c min, 20 log 10 c max). This assumes the standard case,
where the real-axis intercepts satisfy 0 1 c min # 1 # c max 1 3.
0.5 Recall that the maximum phase variation z max of any per-
φ turbation in D(a max, v) satisfies sin z max = r/c when r # c.
Imaginary

For the balanced case v = 0, the peak phase variation

0 occurs at c = 1 (phase-only variation), and, hence, z max = z m
γmin γ γmax
for this case. For nonzero v, the peak z max is not achieved
for phase-only variation and requires some amount of gain
–0.5
D (α max, σ ) variation. The safe region in Figure 10 fully quantifies how
the disk margin a max translates into safe levels of gain-only,
–1 phase-only, and combined gain/phase variations.
0 0.5 1 1.5 2
Real Example 3
(a) The classical gain-only and phase-only margins for L(s) =
25/(s 3 + 10s 2 + 10s + 10) were previously computed in
1 Example 1 as g L = 0, g U = 3.6, and z U = 29.1°. Recall also
that the symmetric disk margin for this loop was com-
puted in Example 2 as a max = 0.46. The symmetric disk
0.5 provides guarantees that the classical gain margins are
at least g L # c min = 0.63 and g U $ c max = 1.59. These sym-
Imaginary

φ
0 metric margins are ±4.05 dB; that is, they are symmet-
γ– γ+ ric as multiplicative factors from the nominal gain of
γmin γmax
one. The symmetric disk also guarantees classical phase
D (α max, σ )
–0.5 margins of at least i U $ z m = 25.8°. The gain-only and
phase-only guarantees from the symmetric disk margin
are conservative relative to the actual classical margins.
–1
However, it is important to emphasize that the symmet-
0 0.5 1 1.5 2
ric disk margin provides a stronger robustness guaran-
Real
tee. Specifically, it ensures stability for all simultaneous
(b)
gain and phase variations in the disk D(a max, v = 0).

FIGURE 9 The geometry of (a) admissible phase variations for a Nyquist Exclusion Regions
given gain variation c and (b) admissible gain variations for a Disk margins have an interpretation in the Nyquist plane.
given phase variation z . To simplify the discussion, consider the typical case
where D(a max, v) is the interior of a disk with real inter-
cepts satisfying 0 1 c min 1 1 and 1 1 c max 1 3. The disk
margin analysis implies that 1 + f L( j~) ! 0 for all pertur-
Stable Regions for bations f ! D(a max, v) and all frequencies ~ ! R , {+3}.
Combined Gain and Phase Variations Rewrite this stability condition as L( j~) ! - f -1. The set
50
{- f -1 ! C: f ! D(a max, v)} is a disk with real-axis intercepts
Phase Variation (°)

40 1
(-c -min 1
, -c -max). Thus, the condition L ( j~) ! - f -1 can be
30 ­interpreted as a Nyquist exclusion region; that is, the Nyquist
plot L( j~) does not enter the disk {- f -1 ! C : f ! D(a max, v)}.
20
This exclusion region contains the critical point (-1, 0) and
10 is tangent to the Nyquist curve of L at some point -1/f 0 .
0 Varying the skew v produces different exclusion regions
–15 –10 –5 0 5 10 15 with different contact points.
Gain Variation (dB) The exclusion regions can be related to common disk
σ=0 σ=1 σ = –1 margins used in the literature. If v = -1 , then the disk mar-
gin is a max = < T < -31. This margin is related to the robust sta-
bility condition for models with multiplicative uncertainty
FIGURE 10 The safe combinations of gain and phase variations for of the form P (1 + d) [1], [3]. The real-axis intercepts for this
a max = 0.75 . T-based margin are c min = 1 - a max and c max = 1 + a max . The

88 IEEE CONTROL SYSTEMS » OCTOBER 2020

d licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on November 05,2024 at 06:01:51 UTC from IEEE Xplore. Restrictions apply.
 disk of perturbations is centered at the nominal f = 1, and Similar to Theorem 1, this frequency-dependent margin is
a max is the radius. The gain can increase and decrease by given by
the same absolute amount. However, the corresponding
1 1 v-1 -1
Nyquist exclusion disk has intercepts (- c -min , - c -max ), and a max (~) = S( j~) + . (15)
2
this exclusion disk is skewed (that is, its center is offset rela-
tive to -1). Moreover, the actual disk margin a max is equal to the small-
If v = +1, then the disk margin is a max = < S < -31. The real- est of all of the frequency-dependent disk margins,
axis intercepts for this S-based margin are c min = (1 + a)-1
and c max = (1 - a)-1. The disk of perturbations is skewed, a max = min ~ ! R , {+ 3} a max (~). (16)
wit h t he center offset f rom t he nom i nal f = 1. T he
corresponding Nyquist exclusion disk has intercepts A plot of a max (~) versus ~ provides more information about
1 1
(- c -min , - c -max) = (-1 - a, -1 + a). This Nyquist exclusion the feedback loop than just its smallest value a max . For
disk is centered at -1, with a as the radius. The S-based example, such a plot can identify frequency bands where
marg i n a max def i nes t he dista nce f rom t he Nyqu ist the disk margin is weak. The margins in these frequency
curve of L to the critical -1 point. Specifically, if v = +1, bands can then be compared with the expected level of
then a max = min ~|1 + L( j~)|. Based on this interpretation, model uncertainty. Frequency-dependent margins may
the S-based margin has also been called the vector gain also reveal robustness issues away from the gain crossover
margin [7], [8] and modulus margin [9]. frequency, for example, near a resonant mode that has not
Finally, if v = 0, then the disk margin is given by been sufficiently attenuated. This motivates the case for
a max = ; (S - T ) /2 ; -31 . This symmetric disk margin was intro- plotting disk margins versus frequency or, for easier inter-
duced in [4] and, more recently, discussed in [5] and [6]. pretation, plotting the equivalent gain-only and p ­ hase-only
The center of the perturbation disk is offset from the margins (c min, c max) and z m as a function of ­frequency. The
1
nominal f = 1 but is balanced in the sense that c max = c -min . formulas obtained earlier for (c min, c max) and z m [(10) and
The gain variation can increase or decrease by the same (11)] can be used with a max replaced by a max (~).
relative factor. Moreover, the corresponding Nyquist
1 1
exclusion disk has intercepts (- c -min , - c -max). This Nyquist
3
exclusion disk also has the center offset from -1. How-
ever, the exclusion disk is, again, balanced in the sense 2 L ( jω)
that the real-axis intercepts are the same relative fac- σ =0
tor from -1. Thus, for v = 0, both the perturbation and 1 σ =3
σ = –3
Nyquist exclusion sets are symmetric (balanced) disks.
0
Example 4
–1
Figure 11(a) shows the Nyquist plot and three exclu-
sion regions for L (s) = 25/(s 3 + 10s 2 + 10s + 10). Each ex­­
–2
clusion region is the disk {- f -1 ! C: f ! D (a max, v)} with
-1
a max = < S + (v - 1) /2 < 3 . Figure 11(b) is zoomed more tightly –3
on the exclusion regions. Note that each exclusion region –3 –2 –1 0 1 2
is tangent to the Nyquist curve of L at some point. These (a)
tangent points correspond to - f -0 1, where f 0 is the destabi-
L ( jω)
lizing perturbation for the given skew v. 0.6 σ =0
σ =3
0.4 σ = –3
Frequency-Dependent Margins
The disk margin for a given skew v is the largest value 0.2
of a such that the closed loop remains well posed and 0
stable for all perturbations in D (a, v). The perturbations
–0.2
are parameterized as f (d) with |d|1 a. Computing the
disk margin amounts to finding the smallest d such that –0.4
1 + f (d) L( j~) = 0 at some frequency ~. This problem can be –0.6
considered at each frequency. That is, define the disk mar-
gin at the frequency ~ as –0.8
–1.5 –1 –0.5 0
(b)
a max (~) := min {|d|: 1 + f (d) L( j~) = 0}. (14)

This specifies the minimum amount of gain and phase FIGURE 11 (a) The Nyquist exclusion regions based on disk margins
variation needed to destabilize the loop at this frequency. with different skews and (b) a zoomed-in view of the exclusion regions.

OCTOBER 2020 « IEEE CONTROL SYSTEMS 89

d licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on November 05,2024 at 06:01:51 UTC from IEEE Xplore. Restrictions apply.
 Example 5 plot corresponds to the weaker of the two gain margins;
Consider the loop transfer function that is, c m := min (1/c min, c max). At each frequency, the gain
margin value indicates the minimum amount of relative
6.25 (s + 3) (s + 5) gain variation needed to destabilize the loop at this fre-
L(s) = . (17)
s(s + 1)2 (s 2 + 0.18s + 100) quency, that is, cause a closed-loop pole to cross the imagi-
nary axis at this frequency.
The Bode plot for this loop is shown in Figure 12(a). This The frequency-dependent phase margin plot has a simi-
loop has a resonance near 10 rad/s. Figure 12(b) plots the lar interpretation. The frequency where these margins
frequency-dependent gain-only and phase-only margins are smallest is the critical frequency and corresponds to the
computed from the symmetric disk margin. The gain-only frequency that minimizes a max (~). This pinpoints the fre-
quency band where stability is most problematic and, typi-
cally, lies near the crossover frequency. The plot may also
Bode Diagram highlight other problematic regions. For example, the disk-
20 based margins in Figure 12 are weak in a wide band around
Magnitude (dB)

0 the crossover but also near the first resonant mode. Also note
–20 that c m " 3 and z m " 90° past 10 rad/s because a max (~) " 2,
–40 and, thus, the stable region D(a max (~), v = 0) approaches the
–60 half-plane Re( f ) $ 0.
–80
–90 MARGINS FOR MIMO SYSTEMS
–135 This section briefly reviews two different margins for MIMO
Phase (°)

–180
feedback systems. The first analysis is loop at a time. This
–225
introduces perturbations in a single channel while holding
–270
–315 all other channels fixed. This can be overly optimistic, as it
10–1 100 101 fails to capture the effects of simultaneous perturbations in
Frequency (rad/s) multiple channels. The second analysis considers the effects
(a) of such simultaneous perturbations in multiple channels.

Gain Margin γm
Gain Margin (dB)

40
Loop-at-a-Time Margins
Loop-at-a-time analysis is a simple extension of classical
20 margins to assess the robustness of a MIMO feedback sys-
tem. The procedure is illustrated for a 2 # 2 MIMO plant,
0
10–1 100 101 as shown in Figure 13. A scalar (gain, phase, or disk) per-
Frequency turbation f 1 is introduced at the first input of the plant P.
The other loop is left at its nominal (unperturbed) value.
Phase Margin (dB)

Phase Margin φm First, break the loop at the location of the perturbation, as
80 shown in Figure 14(a). Next, compute the transfer function
60 from the scalar input z 1 to the scalar output u 1 (with the
40 other loop closed as shown). Denote this SISO open-loop
transfer function as L 1 . The subscript of L 1 reflects that the
10–1 100 101
loop was broken at the first channel at the input of P. The
Frequency
perturbation f 1 closes the loop from u 1 to z 1 .
(b)
Hence, the MIMO feedback with perturbation at the
first input of P can be redrawn as the SISO feedback s­ ystem
FIGURE 12 The (a) open-loop response for L and (b) corresponding
frequency-dependent disk gain and phase margins for v = 0. shown in Figure 14(b). The (gain, phase, or disk) margin
associated with this loop can be computed using the SISO
methods discussed previously. This gives the margin asso-
u1 z1 ciated with the first input of P. Note that L 1 is the transfer
f1 function from z 1 to u 1 . Hence, Figure 14 is in positive feed-
K P
u2 = z2 back. The margins must be evaluated using - L 1 because
the standard convention assumes the loop is in negative
feedback. The margins can be computed similarly at the
second input of P as well as at both outputs of P.
FIGURE 13 A multiple-input, multiple-output feedback system with In general, loop-at-a-time margins are computed by
perturbation in the first input channel of P. breaking one loop, with all other loops remaining closed.

90 IEEE CONTROL SYSTEMS » OCTOBER 2020

d licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on November 05,2024 at 06:01:51 UTC from IEEE Xplore. Restrictions apply.
 If the plant is n y # n u, then this gives n u margins at the
inputs of P and n y margins at the outputs of P. Unfortu- u1 z1
nately, the loop-at-a-time margins can be overly optimis-
tic. Specifically, a MIMO feedback system can have large K P
loop-at-a-time margins and yet be destabilized by small u2 = z2
perturbations acting simultaneously on multiple channels.
Example 6 demonstrates this situation. This motivates the
development of more advanced robustness-analysis tools. (a)
z1
Example 6 f1 L1
Consider a feedback system with the following plant and
controller with a = 10:
(b)
P := 2 1 2 = G and K := -; E . (18)
s - a 2 a (s + 1) 1 0
s + a - a ( s + 1) s - a 2 0 1 FIGURE 14 (a) A multiple-input, multiple-output feedback system
with a loop broken at the first input channel of P. (b) A single-input,
This example is taken from [10]. The dynamics represent a single-output feedback system with perturbation f1 and loop L1 ob-
simplified model for a spinning satellite. Additional details tained at input 1 of the plant.
can be found in [3, Sec. 3.7] or [1, Sec. 9.6]. Breaking the loop
at the first input of P, with the other loop closed, yields the restricted to a set D(a, v) (4) defined for a given skew v.
SISO open-loop transfer function L 1 = -( 1/s). This loop Symmetric disks of perturbations (v = 0) are a common
(when in a positive feedback, as in Figure 14) has no 180° choice. The multiloop disk margin is a single number a max
phase-crossover frequencies. Thus, the classical gain mar- defining the largest generalized disk of perturbations f 1
gins are g L = 0 and g U = 3. This loop has a single gain and f 2 for which the closed loop in Figure 15 is well posed
crossover at ~ = 1 rad/s, which gives a classical phase mar- and stable. It is emphasized that the perturbations f 1 and
gin of z U = 90°. f 2 are allowed to vary independently; that is, they are not
Finally, the SISO loop L 1 corresponds to the sensi- necessarily equal.
t ivit y S 1 = s/(s + 1) a nd complementar y ­s en sit ivit y More generally, if the plant P is n y # n u, then there will
T1 = 1/(s + 1). T h e s y m m e t r ic d i sk m a r g i n (v = 0) is be n u perturbations introduced at the plant input. The mar-
a max = ; (S 1 - T1) /2 ; -31 = 2. This corresponds to a disk cover- gin for this configuration is called the multiloop input disk
ing the entire RHP; that is, stability is maintained for any margin. Alternatively, n y perturbations can be introduced
combination of gain/phase, such as at Re{ f1} 2 0. These at the plant output. This is referred to as the multiloop
results demonstrate that the MIMO feedback system is output disk margin. Finally, (n y + n u) perturbations can
very robust to perturbations at the first input of P, assum- be introduced into both the input and output channels to
ing all other inputs/outputs remain at their nominal val- obtain the multiloop input/output disk margin.
ues. Breaking the loop at the second input of P or either In the most general case, multiloop margins can be defined
output of P yields the same open-loop transfer function, with perturbations introduced at arbitrary points in a feed-
for example, L 2 = -(1/s) at the second plant input. Thus, back system. This general formulation corresponds to a
the loop-at-a-time analysis demonstrates that the MIMO feedback system with a collection of complex perturbations
feedback s­ystem is very robust to perturbations at any ( f 1, f, fn). The multiloop margin is the largest value of a such
single input or output of P, assuming all other inputs/out- that the feedback system remains well posed and stable for
puts remain at their nominal values. all perturbations ( f 1, f, fn) in the set D(a, v) ­specified for a
Consider the following small simultaneous perturbation at given skew v. The next two examples illustrate various types
both input channels of the plant: f1 = 0.9 and f 2 = 1.1. These of multiloop margins. The theory required to compute such
simultaneous perturbations to both input channels destabi- multiloop margins is reviewed in the “Computing Multiloop
lize the MIMO feedback system. The loop-at-a-time margins Disk Margins” section.
fail to capture such simultaneous variations in multiple chan-
nels. As a consequence, the loop-at-a-time margins provide an
overly optimistic assessment of the system robustness. u1 z1
f1
Multiloop Disk Margins K P
u2 z2
Multiloop disk margins capture the effects of simultane- f2
ous perturbations in multiple channels. Figure 15 illus-
trates the use of multiloop disk margins for a 2 # 2 MIMO
plant P. Scalar perturbations f1 and f 2 are introduced at
the two input channels of the plant. The perturbations are FIGURE 15 The multiloop input disk margins for a 2 × 2 plant P.

OCTOBER 2020 « IEEE CONTROL SYSTEMS 91

d licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on November 05,2024 at 06:01:51 UTC from IEEE Xplore. Restrictions apply.
 Example 7 satellite feedback system is sensitive to small perturba-
Consider the spinning satellite discussed in ­Example 6. tions occurring at both inputs to the plant. The multiloop
The multiloop input margin is computed for this 2 # 2 output margin is the same for this system.
feedback system using symmetric disks (v = 0). This Multiloop margins can also be defined with perturba-
yields a max = 0.0997, corresponding to the disk with c max = tions introduced (simultaneously) at the two inputs and
-1
( 1 + 0.5a max)/(1 - 0.5a max) = 1.105 a n d c min = c max = 0.905. two output channels. For the spinning satellite, this mul-
Hence, the plant can tolerate independent perturbations f 1 tiloop input/output margin is a max = 0.0498, correspond-
and f 2 at the plant inputs with gain-only variations in ing to (c min, c max) = (0.941, 1.051). Details on this example
(0.905, 1.105). These margins indicate that the spinning including the corresponding code can be found in the

az
Step az

+– In1 Out1 +
In1 Out1 Fin Deflection
+
az Control
q Control
q

Airframe Model

az Response
(a)

Incidence and Airspeed

Alpha
U, w
7 V
V

Mach
q
Mach 6

Thrust 3
V
Attitude 2
+
+ Fx Attitude
Fx
Alpha q 3
q

a qdot 4
2
qdot
a Fz Fz
Xe, Ze 1
1 Rho Xe, Ze
Rho U, w

4 Fin M M
Ax, Az 5
Fin
Ax, Az
Aerodynamics
(b)

FIGURE 16 (a) A Simulink diagram for a longitudinal aircraft controller and (b) one subsystem containing the aerodynamics for the air-
frame model.

92 IEEE CONTROL SYSTEMS » OCTOBER 2020

d licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on November 05,2024 at 06:01:51 UTC from IEEE Xplore. Restrictions apply.
 Matlab example “MIMO Stability Margins for Spinning with (c min, c max) = (0.442, 2.263). Hence, the classical mar-
Satellite” (https://www.mathworks.com/help/robust/ug/ gins are at least g L # 0.442, g U $ 2.263, and z U $ 42.3°.
mimo-stability-margins-for-spinning-satellite.html). Next, the disk margins are computed using all three
analysis points. The multiloop margin with symmetric
Example 8 disks (MM3) is a max = 0.428. Hence, the feedback system
Consider the Simulink diagram for the aircraft longitudi- remains well posed and stable for independent perturba-
nal controller shown in Figure 16. Figure 16(a) shows blocks tions at the three analysis points that remain in the disk, with
for the airframe dynamics, inner-loop pitch rate (q) control, (c min, c max) = (0.648, 1.544).
and outer-loop vertical acceleration (a z) control. Figure 16(b)
shows one subsystem containing the aerodynamics for the Computing Multiloop Disk Margins
airframe model. This Simulink model is part of a Matlab Consider a feedback system with n complex perturbations
example in “Stability Margins of a Simulink Model” (see ( f 1, f, fn) introduced at arbitrary points. It is assumed that
https://www.mathworks.com/help/robust/ug/stability- the feedback system is well posed and stable if all perturba-
margin-of-a-simulink-model.html). The model is modified tions are at their nominal value, f i = 1, for all i. The multiloop
to include three complex perturbations inserted at vari- disk margin, denoted a max , was defined in the “Multiloop
ous points. One perturbation is inserted at the plant input Disk Margins” section. It is the largest value of a such that the
[the red dot in Figure 16(a)]. Two other perturbations are feedback system remains well posed and stable for all pertur-
inserted in the aerodynamics subsystem [the red dots in bations ( f 1, f, fn) in the set D(a, v) with a given disk skew v.
Figure 16(b)]. These are inserted on signals for the verti- The condition for SISO disk margins (Theorem 1) can be
cal force Fz and pitching moment M. These two additional generalized for the multiloop case. The starting point for
perturbations can be used to model, for example, the dis- the SISO disk margin result was the following condition:
crepancy in the modeled and actual aerodynamics for this f ! D(a, v) places a closed-loop pole at s = j~ if and only if
force and moment. 1 + f L( j~) = 0. The next step was to express the perturbation
Figure 17 shows the Matlab code to compute two dif- f in terms of |d|1 a. This led to the stability condition (7):
ferent disk margins for this example. The linio com-
mand specifies the analysis points. The model is nonlinear. 1 - d ` S( j~) + v - 1 j = 0. (19)
2
Hence, the dynamics must first be linearized around an
operating point. This is done using the linearize com- This has the form 1 - dM ( j~) = 0, where M := S + (v - 1)/ 2.
mand. The symmetric disk margin is computed at the plant This is the stability condition for a feedback system with d in
input (DMi). Note that linearize returns the loop transfer positive feedback with M. Similarly, each ­perturbation in a mul-
function assuming positive feedback, while diskmargin tiloop analysis can be expressed as follows for some |d i|1 a
assumes negative feedback. This symmetric disk margin at
the plant input is a max = 0.774. This corresponds to a disk f i = (2 + ( 1 - v) d i)/(2 - (1 + v) d i) .(20)

% Open simulink model from Matlab example

open_system('airframemarginEx.slx')

% Specify analysis point at plant input

aPoints(1) = linio('airframemarginEx/q Control',1,'looptransfer');

% Specify analysis points inside aerodynamic model

blk = ['airframemarginEx/Airframe Model/' ...
'Aerodynamics & Equations of Motion/Aerodynamics'];
aPoints(2) = linio(blk,3,'looptransfer');
aPoints(3) = linio(blk,4,'looptransfer');

% Linearize and compute disk margin at plant input

Li = linearize('airframemarginEx',aPoints(1) );
DMi = diskmargin(-Li)

% Linearize and compute disk margins at three analysis points

L3 = linearize('airframemarginEx',aPoints);
[DM3,MM3] = diskmargin(-L3)

FIGURE 17 The code for aircraft multiloop margins.

OCTOBER 2020 « IEEE CONTROL SYSTEMS 93

d licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on November 05,2024 at 06:01:51 UTC from IEEE Xplore. Restrictions apply.
 In this way, the multiloop margin analysis involving per- the imaginary axis at j~. The MIMO case requires an
turbations f i is mapped to an equivalent M–T positive feed- appropriate generalization for the connection between the
back loop, as shown in Figure 18. Here, M: = S + (v - 1) /2 $ I n, “gain” of the system M and the existence of small, desta-
where S is the closed-loop sensitivity function at the points bilizing perturbations (d 1, f, d n). First, let T 1 C n # n denote
of perturbation (red dots in Figure 16). M is an n # n sta- the set of diagonal, complex matrices and define the norm
ble system. Moreover, T ! C n # n is the diagonal matrix of for any T ! T by < D < := max i = 1, f, n|d i|. In other words, the
complex perturbations T := diag (d 1, f, d n). The multiloop norm is given by the largest (magnitude) of the diagonal
margin is equivalent to the largest value of a such that the entries. Note that all perturbations f i are in the set D (a, v)
positive feedback system with M and T := diag (d 1, f, d n) is if and only if |d i|1 a, that is, if and only if < D < 1 a.
well posed and stable for all complex perturbations |d i|1 a Next, define the function n : C n # n " [0, 3) by
(i = 1, f, n). Additional details on this M–T modeling frame-
n(M 0) := ` min < D < : det (I - M 0 T) = 0 j . (21)
-1
work can be found in [1]–[3].
T!T
The nominal perturbation corresponds to T = 0 with
nominal system M. The assumption of nominal stability, By definition, n(M ( j~)) is large if and only if there is a
thus, implies the poles of M are in the LHP. The perturba- “small” T 0 ! T such that det (I - M ( j~) T 0) = 0. By the ear-
tion T causes the closed-loop poles to move continuously lier discussion, this perturbation causes the M–T system
in the complex plane away from their nominal values. The to have a pole on the imaginary axis. This function μ is
poles may move into the RHP (unstable closed loop) if T known as the structured singular value or, simply, “mu”
is varied by a sufficiently large amount from the nominal [10]–[15]. The structured singular value can be used to
value T = 0. The transition from stable to unstable occurs assess the robust stability and performance of systems
when the closed-loop poles cross the imaginary axis. As in with more general types of uncertainties, including real,
the SISO case, it is useful to have a condition that character- complex, and dynamic LTI uncertainties.
izes this stability transition, that is, a condition that charac- The version in (21) is a special instance of this more
terizes the existence of imaginary axis poles. general framework adapted for multiloop disk margins. It
It can be shown that the M–T system has a pole on the is difficult to exactly compute n (M 0) for a given complex
imaginary axis at j~ if and only if det (I - M ( j~) T) = 0. matrix M 0 and uncertainty set T. However, there are effi-
To sketch a simplified derivation, consider the case where cient algorithms to compute the upper and lower bounds
M has no direct feedthrough (D = 0). Let (A, B, C, D = 0) on n(M 0). The following theorem provides a condition for
be a state-space realization for M. The poles of the M–T the multiloop disk margin using this function μ. The nota-
system are given by the eigenvalues of the state matrix tion for the peak value of μ across all frequencies is
A cl := A + BTC. There is a pole on the imaginary axis at j~
if and only if det ( j~I - A cl) = 0. Stability of M implies that < n (M ) < 3 := max n ^ M ( j~) h . (22)
~ ! R , {+ 3}
j~ is not an eigenvalue of A. Hence, ( j~I - A) has a nonzero
determinant, and its inverse exists. Thus, det ( j~I - A cl) = 0
is equivalent to det ^I - ( j~I - A)-1 BTC h = 0. Finally, apply Theorem 2
Sylvester’s determinant identity [22, Corollary 3.9.5]: Assume the nominal feedback loop is well posed a nd
stable. Then the multiloop disk margin is given by
0 = det (I - ( j~I - A)-1 BDC) = det ( I - M ( j~) T). a max = < n (S + (v - 1) /2 $ I n) < 3, where S is the closed-loop
sensitivity function at the locations where the gain and
If there is only one perturbation (n = 1), then the determi- phase are perturbed.
nant condition simplifies to 1 - M ( j~) $ d = 0. This is the
same condition that appeared in the proof for the SISO Proof
small-gain result [(7) and rewritten in (19)]. Recall that the closed-loop system with perturbations fi
In this SISO case, if the gain |M ( j~)| is large, then there given by (20) is equivalent to an M - T feedback loop
is a small perturbation d = M ( j~)-1 that causes a pole on with M: = S + (v - 1) /2 $ I n and D = diag (d 1, f, d n). The
proof consists of two steps. First, it is shown that there is
a destabilizing perturbation on the boundary of the disk
d1 e1
|D|1 < n (M ) < -31. Let ~ 0 be the frequency (possibly infinite)
∆ where n(M ( j~)) achieves its peak. By definition, there is
a perturbation T 0 such that 1) det ( I - M ( j~ 0) T 0) = 0 and
2) < D 0 < = < n (M ) < -31. The M–T system is either ill posed (~ 0
M infinite) or unstable with an imaginary axis pole (~ 0 finite).
v e2 d2 Any open disk with a radius larger than < n (M ) < -31 contains
this destabilizing perturbation. Hence, the multiloop disk
FIGURE 18 An M–Δ feedback system for multiloop margins. margin is # < n (M ) < -31.

94 IEEE CONTROL SYSTEMS » OCTOBER 2020

d licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on November 05,2024 at 06:01:51 UTC from IEEE Xplore. Restrictions apply.
 Next, it is shown that the M–T feedback system is sta- Control System Technology Award, and the Berkeley Cita-
ble and well posed for all perturbations in the interior of tion Award. He passed away on September 30, 2019, after a
< D < 1 < n (M ) < -31. It follows from the definition of μ that the long battle with cancer. He contributed substantially to this
M–T system is well posed and has no imaginary axis poles work, including the drafting of the article.
for any perturbation < D < 1 < n (M ) < -31. Hence, the closed Pascal Gahinet has been with MathWorks since 1996 as
loop is stable for all < D < 1 < n (M ) < -31 because the poles do a scientist and architect for the Control System Toolbox and
not cross the imaginary axis into the RHP. This can be for- Robust Control Toolbox. His focus has been on numerical
malized with a homotopy argument. 4 algorithms for classical and robust control. His early con-
Additional details on computing disk margins using the tributions include the LMI Control Toolbox. His recent con-
structured singular value can be found in [23] and [24]. The tributions include the PIDTUNE and SYSTUNE algorithms
structured singular value can be used to extend the results for automated control system tuning.
in this article for assessing robust stability and performance
with more general classes of parametric and dynamic uncer- REFERENCES
[1] K. Zhou, J. Doyle, and K. Glover, Robust and Optimal Control. Englewood
tainty. The integral quadratic constraint framework [16] Cliffs, NJ: Prentice Hall, 1996.
is even more general and can be used to assess the impact [2] G. Dullerud and F. Paganini, A Course in Robust Control Theory: A Convex
of nonlinearities. Approach. New York: Springer-Verlag, 2000.
[3] S. Skogestad and I. Postlethwaite, Multivariable Feedback Control: Analysis
and Design, 2nd ed. Hoboken, NJ: Wiley, 2005.
CONCLUSIONS [4] M. Barrett, “Conservatism with robustness tests for linear feedback con-
This article provided a tutorial introduction to disk margins. trol systems,” Ph.D. dissertation, Univ. of Minnesota, 1980.
[5] J. Blight, R. Dailey, and D. Gangsaas, “Practical control law design for
These are robust stability measures that account for simul- aircraft using multivariable techniques,” Int. J. Control, vol. 59, no. 1, pp. 93–
taneous gain and phase perturbations in a feedback system. 137, 1994. doi: 10.1080/00207179408923071.
They can also be used to compute frequency-dependent [6] D. Bates and I. Postlethwaite, Robust Multivariable Control of Aerospace
Systems. Amsterdam, The Netherlands: IOS Press, 2002.
margins, which provide additional insight into potential [7] O. Smith, Feedback Control Systems. New York: McGraw-Hill, 1958.
robustness issues. Disk margins were also described for the [8] G. Franklin, J. Powell, and A. Emami-Naeini, Feedback Control of Dynamic
multiple-loop analysis of MIMO systems. This multiloop Systems, 8th ed. London, U.K.: Pearson, 2018.
[9] A. Falcoz, C. Pittet, S. Bennani, A. Guignard, C. Bayart, and B. Frapard,
analysis provides a more accurate robustness assessment “Systematic design methods of robust and structured controllers for satellites,”
than one-loop-at-a-time analysis. These multiloop disk mar- CEAS Space J., vol. 7, no. 3, pp. 319–334, 2015. doi: 10.1007/s12567-015-0099-8.
gins also provide an introduction to more general robust- [10] J. Doyle, “Robustness of multiloop linear feedback systems,” in Proc.
IEEE Conf. Decision and Control, 1978, pp. 12–18. doi: 10.1109/CDC.1978.267885.
ness frameworks, for example, structured singular-value μ [11] M. Safonov, Stability and Robustness of Multivariable Feedback Systems.
and integral quadratic constraints. Cambridge, MA: MIT Press, 1980.
[12] J. Doyle, “Analysis of feedback systems with structured uncertainties,”
Proc. Inst. Elect. Eng. D, Control Theory Appl., vol. 129, no. 6, pp. 242–250, 1982.
ACKNOWLEDGMENTS doi: 10.1049/ip-d.1982.0053.
The authors thank Christopher Mayhew, Raghu Venkata- [13] J. Doyle, “Structured uncertainty in control system design,” in Proc. IEEE
raman, and Brian Douglas for helpful suggestions. The Conf. Decision and Control, 1985, pp. 260–265. doi: 10.1109/CDC.1985.268842.
[14] A. Packard and J. Doyle, “The complex structured singular value,” Au-
authors also gratefully acknowledge Brian Douglas for tomatica, vol. 29, no. 1, pp. 71–109, 1993. doi: 10.1016/0005-1098(93)90175-S.
the creation of tutorial videos (see https://www.youtube [15] M. Fan, A. Tits, and J. Doyle, “Robustness in the presence of mixed
.com/watch?v=XazdN6eZF80 and https://www.youtube parametric uncertainty and unmodeled dynamics,” IEEE Trans. Autom.
Control, vol. 36, no. 1, pp. 25–38, 1991. doi: 10.1109/9.62265.
.com/watch?v=sac_IYBjcq0) corresponding to this article. [16] A. Megretski and A. Rantzer, “System analysis via integral quadratic
Finally, the authors thank Seagate for providing the hard constraints,” IEEE Trans. Autom. Control, vol. 42, no. 6, pp. 819–830, 1997. doi:
disk drive frequency responses shown in Figure 2. 10.1109/9.587335.
[17] N. Nise, Control Systems Engineering, 6th ed. Hoboken, NJ: Wiley, 2010.
[18] R. Dorf and R. Bishop, Modern Control Systems, 13th ed. London, U.K.:
AUTHOR INFORMATION Pearson, 2016.
Peter Seiler (pseiler@umich.edu) is a professor at the Uni- [19] K. Ogata, Modern Control Engineering, 5th ed. London, U.K.: Pearson, 2009.
[20] S. Boyd, V. Balakrishnan, and P. Kabamba, “A bisection method for
versity of Michigan in the Department of Electrical En- computing the H 3 norm of a transfer matrix and related problems,” Math.
gineering and Computer Science. From 2004 to 2008, he Control Signals Syst., vol. 2, no. 3, pp. 207–219, 1989. doi: 10.1007/BF02551385.
worked at Honeywell Research Labs on various aerospace [21] N. Bruinsma and M. Steinbuch, “A fast algorithm to compute the H 3
norm of a transfer function matrix,” Syst. Control Lett., vol. 14, no. 4, pp. 287–
and automotive applications. He has since worked on ­robust 293, 1990. doi: 10.1016/0167-6911(90)90049-Z.
control theory, with applications to wind turbines, flexible [22] D. S. Bernstein, Scalar, Vector, and Matrix Mathematics: Theory, Facts, and
aircraft, and disk drives. Formulas, 3rd ed. Princeton, NJ: Princeton Univ. Press, 2018.
[23] G. Deodhare and V. Patel, “A “modern” look at gain and phase margins:
Andrew Packard, deceased, was the Fanuc Chair Pro- An H 3 /μ approach,” in Proc. AIAA Conf. Guidance, Navigation and Control,
fessor of Mechanical Engineering at the University of Cali- 1998, pp. 325–335. doi: 10.2514/6.1998-4134.
fornia, Berkeley. He was an author of the Robust Control [24] D. G. Bates, R. Kureemun, and I. Postlethwaite, “Quantifying the ro-
bustness of flight control systems using Nichols exclusion regions and
Toolbox distributed by MathWorks. He was a Fellow of the the structured singular value,” IMechE J. Syst. Control Eng., vol. 215, no. 6,
IEEE and received the campus Distinguished Teaching pp. 625–638, 2001. doi: 10.1243/0959651011541355.

Award, the 1995 Donald P. Eckman Award, the 2005 IEEE

OCTOBER 2020 « IEEE CONTROL SYSTEMS 95

d licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on November 05,2024 at 06:01:51 UTC from IEEE Xplore. Restrictions apply.