Loop-Based and Device-Based Algorithms for Stability

Analysis of Linear Analog Circuits in the Frequency Domain

By Michael Tian, V. Visvanathan, J effrey Hantgan, and Kenneth Kundert
N
egativefeedback techniquesarewidelyusedin analog
and RF design to improve circuit properties such as
variationtolerance, bandwidth, impedancematching,
and output waveform distortion. In practice, un-
wanted local return loops also exist around individual transis-
tors through parasitic capacitance. As the size of transistors
continuesto shrink and thebandwidth of transistorscontinues
tobroaden, theselocal returnloopsdegradecircuit performance
significantly in thehigh-frequency regime. Stability isalwaysa
seriousconcernfor feedback circuits. Self-oscillationshavebeen
foundaboveand/or beyondthebandwidthinsuchdiversifiedcir-
cuitsasoptical fiber systemreceivers, power amplifiers, anddis-
tributed microwaveamplifiers. It iscritical to evaluatestability
andstabilitymargin of afeedback circuit. Such information can
beusedfor optimizationintheearlydesignstageandfor diagno-
sis in physical realization stage.
Single-looptheory/multilooprealityisthestate-of-the-art of
stabilityanalysis[8]. All physical networksinthefrequencyband
of interest are intrinsically multiloop structures, yet it is still
common practiceto assessstability from single-loop theory. In
this article, based on Bodes definition of return ratio with re-
spect toasinglecontrolledsource, theloop-basedtwo-port algo-
rithmanddevice-basedgain-nullingalgorithmareproposedfor
small-signal stabilityanalysis. Thesetwoalgorithmsarecomple-
mentaryin termsof applicability, andtheyproduceaccuratesta-
bility information for single-loop networks. After abrief primer
on feedback and stability, we review Bodes feedback theory,
wherethereturn differenceandreturn ratioconceptsareappli-
cabletogeneral feedback configurationsandavoidthenecessity
of identifying and. Middlebrooksnull double-injection tech-
nique, which providesalaboratory-basedwaytomeasurereturn
ratio, is then discussed in the modern circuit analysis context.
We t hen ext end t he uni l at eral feedback model used i n
Middlebrooksapproach to accommodateboth normal- and re-
verse-looptransmissionandcharacterizethereturnloopusinga
general two-port analysis. This loop-based two-port algorithm
determinesthestabilityof afeedback network in which acritical
wire can be located to break all return loops. The device-based
gain-nulling algorithm is then discussed to evaluate the influ-
enceof thelocal return loopsupon network stability. Thisalgo-
rithm determinesthestability of afeedback network in which a
controlledsourcecan benulledtorender thenetwork tobepas-
sive. Conditions under which these two algorithms can be ap-
plied are discussed, and numerical results are provided.
The Feedback- St abi l i t y Rel at i onshi p
To illustrate the relation between feedback and stability, the
ideal single-loopfeedback network isshown in Fig. 1. Following
theterminology of [6], representsthetransfer function of the
unilateral forward active path and is usually referred to as an
open-loop transfer function, while represents the transfer
function of aunilateral feedback path. Thepresenceof theideal
adder inthediagramindicatesthereisnoloadingeffect at thein-
put. Under theseassumptions, it iseasily shown that theoverall
transfer function, which is often called a closed-loop transfer
function, is given by
A
s
s
o
i

1
,
(1)
where1isreferredtoasthefeedbackfactor. Under thecon-
dition1 0 , afeedback network will function likean oscilla-
CIRCUITS & DEVICES I JANUARY 2 0 0 1 8 7 5 5 -3 9 9 6 / 0 1 / $ 1 0 . 0 0 2 0 0 1 IEEE 3 1 I
I 3 2 CIRCUITS & DEVICES I JANUARY 2 0 0 1
t or . Thi s condi t i on i s
known as the Barkhausen
criterion [2], which states
that thefrequency of alin-
ear(ized) oscillator is de-
termined by the condition
that thephaseshift of is
zer o pr ovi ded t hat t he
magni t ude of equal s
unity. Thelinear oscillator design shouldensurethat thiscrite-
rion is satisfied, while the amplifier design should ensure that
thecircuit operatesaway from thiscritical point. Stability mar-
gin is defined as the distance between and unity.
The zeros of the feedback factor are the natural frequencies
(poles) of thenetwork. However, pole-zeroanalysisisnot preferred
for studyingstability not only becauseit hasnumerical difficulties
with largenetworks but also becauseit does not providestability
margin that isof paramount interest tocircuit designers. Sincein-
put s
i
andoutput s
o
can beeither current or voltage, feedback net-
worksoften can beclassifiedintofour different configurations[3]:
shunt-shunt, shunt-series, series-shunt, and series-series. Thein-
put-output two-port approach[3, 6, 7] modelsthesefour configu-
rations using Y, G, H, and Z parameter-based two-port analysis.
Symbolicexpressionsof loopgain in termsof theseparameters
arethereforederived to providestability information.
It should be noted that the ideal single-loop feedback net-
work shown in Fig. 1isnot an adequaterepresentation of aprac-
tical feedback network. In practice, the active path may not be
strictly unilateral; thefeedback path isusually bilateral, andthe
input and output coupling networks are often complicated. A
general feedback configuration containing an input coupling
and an output coupling network isshown in Fig. 2. In addition,
aspointedout in [6], not everysingle-loopfeedback network can
beclassifiedasbeinginoneof thesefour configurations. It isalso
important to knowthat loop gain (thestability measureof are-
turn loop) is uniquely de-
termined by the loop pa-
rametersi t shoul d not
vary as the type and loca-
t i on of i nput sour ce
changes. However, si nce
network zeros enter the
calculation of loopgain,
i nput -out put t wo-por t
analysis produces inconsistent stability information that does
depend upon thetypeand location of input source[4]. Thus, a
general and moreaccurateapproach to perform stability analy-
sis is desirable.
Bode s Feedback Theor y
Bodes original theory was published in [1] and was later ex-
tendedin [6]. In ageneral feedback network, thequantityof cru-
cial effect upon thenetwork stability isthecontrolled sourceof
theamplifier. Without lossof generality, weassumethissource
isavoltage-controlledcurrent sourcespecified asi xv . In the
subsequent discussion, x refers to this controlled source. The
definition of return difference is given below.
Definit ion 1 [6]. ThereturndifferenceF x ( )of afeedback net-
work withrespect toacontrolledsourcexissaidtobetheratioof
two functional values assumed by network determinant under
(a) thecondition that xassumes its nominal value, and (b) the
condition that xassumes the value zero:
F x
x
( )

0
.
(2)
Return difference F x ( ) is the most important quantity en-
countered in feedback theory for anumber of practical reasons.
First, stability of thenetwork dependsupon thezerosof F x ( ) in
the complex frequency plane, which are natural frequencies of
the network. Second, F x ( ) is considered as a generalization of
thefeedback factor 1intheideal feedback network shownin
Fig. 1. Third, thesensitivity function S x ( ) of theamplifier with
respect totheelement xiscloselyrelatedtothereturndifference
F x ( )andin many practical cases, S x ( ) isapproximately equal
to the reciprocal of return difference.
Return ratio T x ( ) is defined as
T x F x
x
x
( ) ( )

1
0
0

.
(3)
In theengineering context, thenegativeof return ratio T x ( ) is
referredtoasloopgain, whichisalsoageneralizationofinthe
ideal single-loopfeedback network shown in Fig. 1. Correspond-
ingly, theBarkhausencriterionintermsof returndifferenceand
return ratio is F x ( ) 0and T x ( ) 1, respectively.
ReferringtoFig. 3, return differenceandreturn ratiocan be
physicallyinterpretedasfollows. First, deactivateall network in-
put i
1
, K, i
N
; second, replace the input of controlled source x
with an external voltagesourceof 1V, thusxnolonger depends
s
i
s
o

1. The ideal single-loop feedback network.

R
s
I
s
Input
Coupling
Network
N

Output
Coupling
Network
R
L
2. A general feedback network with input and output coupling [6].
All physical net wor ks in t he fr equency band
of int er est ar e int r insically mult iloop
st r uct ur es, yet it is st ill common pr act ice
t o assess st abilit y fr om single- loop t heor y.
CIRCUITS & DEVICES I JANUARY 2 0 0 1 3 3 I
upon v
i
. This1Vinput pro-
duces a current of x am-
peres at the output of the
cont r ol l ed sour ce. The
transmitted voltage to the
original input of the con-
trolled source (v
i
) is then
the negative of the return
ratio T x ( ). Also, thereturn difference F x ( ) issimply thediffer-
encebetweentheexcitationof 1Vappliedat theinput of thecon-
trolled sourceand thetransmitted voltageto theoriginal input
of thecontrolledsource. Fromtheproportionalitypropertyof a
linear circuit, thereturn ratio issimply v v
i ext
/ in casetheex-
ternal sourcev
ext
isnot 1V. Thephysical interpretation of return
ratioT x ( )isimportant becauseit isuponthisinterpretationthat
Middlebrook based thelaboratory-oriented measurement tech-
nique to obtain the return ratio of a feedback network.
For thesubsequent discussion, weconsider thespecial case
that the controlled source of interest is a one-port element.
Insteadof relatingavoltageandacurrent at different ports, xre-
latesthevoltageand current at thesameport. Thedefinition of
return differenceandreturn ratiocan still beused. Theoriginal
circuit equation is
M
L L
M
M
M
M
M
Y v i
jj j j

1
]
1
1
1

1
]
1
1
1

1
]
1
1
1
,
(4)
where the determinant of circuit matrix is.
Nowwereplacetheone-port element byan independent cur-
rent sourceof xamperes. Meanwhile, weremoveall network ex-
citations, and then the circuit equation becomes
M
L L
M
M
M
Y x v x
jj j

1
]
1
1
1

1
]
1
1
1

1
]
1
1
1
~
0
0
,
(5)
wherethedeterminant of circuit matrixnowis
x 0
, andthere-
sponse
~
v
j
due to the current excitation xis simply
~
v x
j
jj
x

0
,
(6)
where
jj
is the co-factor of theY
jj
term.
Since
~
v
j
isthevoltageresponseduetothecur-
rent sourcexat the same port, then
~
v
x Y
j jj
x

0
1
,
(7)
whereYistheadmittanceof theone-port network
that x faces. Return ratio T x ( ) for aone-port ele-
ment xis then given by
T x v
x
Y
j
( )
~
.
(8)
This equation will be used
in thesubsequent sections
for the development of re-
turn-loop models.
Stability margin is de-
fined as the distance between return ratio and 1, or equiva-
lently, thedistancebetween loop gain and unity. Notethat the
return ratio is a complex function of frequency. For the com-
pletenessof thefeedback theory, thedefinitionsof gain margin
and phase margin are given below.
Definit ion 2 [6]. Thegain margin isdefinedtobetheamount
of magnitudein decibelsof thereturn ratiobelowthe0dBlevel
at the frequency for which the phase is180 degrees.
Definit ion 3 [6]. Thephasemargin isdefinedtobethephase
difference in degrees of the return ratio above 180 degrees at
the frequency for which the gain is 0 dB.
Nul l Doubl e- I nj ect i on Techni que
In hisoriginal work [9], Middlebrook proposedtheso-callednull
double-injectiontechniquetomeasurereturn ratiobasedon its
physical interpretation. In thissection, wereviewthisapproach
in the modern circuit analysis context.
Recalling the physical interpretation of return ratio in the
last section, it involvesdeactivatingnetwork input andbreaking
thecontrolledsourceat theinput side. Thisbreakpoint concept
can beextended, without lossof accuracy, tobreak thefeedback
loop at any location on either the active path or the feedback
path, as long as the dc impedance characteristics are not dis-
turbed. To demonstratethisconcept, thesignal flowgraph of a
broken feedback loop is illustrated in Fig. 4.
By assuming the feedback loop has unilateral signal flow
from eto f, without loss of generality, wecan model thewhole
feedback loop with an impedance on the input side and a con-
trolledsourcewith an impedanceon theoutput sidein either of
the two configurations (essentially, Norton equivalent and
Thvenin equivalent) shown in Fig. 5. Both configurations
model the same loop, thus constants k
1
and k
2
must satisfy
k kYY
e f 1 2
. s
e
ands
f
must beof thesamephysical quantities, so
the return ratio is either v v
f e
/ or i i
f e
/ .
If weviewthereturn ratioasthevoltageratiov v
f e
/ , wecan
substitutetherepresentation of Fig. 5(a) into thefeedback loop
i
1
Y
11
+ +


xv
i
Y
NN
i
N
v v
3. The physical interpretation of return difference and return ratio.
Null double inject ion has been widely
applied in bot h labor at or y- based
measur ement and cir cuit analysis.
I 3 4 CIRCUITS & DEVICES I JANUARY 2 0 0 1
in Fig. 4. At theoutput sideof thefeedback loop, notethat the
network isloadedwithY
e
sincethedcbiasof theoriginal circuit
is not disturbed, we have
k v
v
Y Y
e
f
e f
1
0 +
+
,
(9)
thus the return ratio is
T
v
v
k
Y Y
f
e e f

+
1
.
(10)
Goingback tothealternativeviewof thereturn ratioasasys-
temthat hascurrentsasitsinput andoutput signals, weplugin
the representation of Fig. 5(b) into the feedback loop in Fig. 4
and have the following equation:
k i
Y Y
i
e
e f
f 2
1 1
+

_
,

.
(11)
Basedon therelation between k
1
andk
2
, weconcludeonceagain
that
T
i
i
k
Y Y
f
e e f

+
1
.
(12)
Wecan alsomakeuseof Eq. (8). At thenormal circuit opera-
tion, nodeeandf areconnectedtogether, thusFig. 5(a) isequiv-
alent to aone-port element (discussed in theprevious section)
with x k
1
andY Y Y
e f
+ . Directly following Eq. (8), we have
T
x
Y
k
Y Y
e f

+
1
.
(13)
Fi gure 5 i s t he fundament al ret urn-l oop model used i n
Middlebrooksapproach, andEq. (13) istheessential formulato
calculate the return ratio under this model.
Toanalyzethestability of thefeedback loop, it doesnot mat-
ter whether thedesigner viewsvoltageor current asthesignal of
interest; both producethesameanswer, aslongasthedcimped-
ancesarenot disturbedwhen theloopisbroken. In practice, this
requires careful selection of the break-point location. For the
voltagedrivingcase, thebreak point shouldbelocatedwherethe
impedanceY
f
looking backwardfrom the break point is suffi-
ciently smaller than theimpedanceY
e
lookingforwardfromthe
break point; theoppositecondition,Y Y
f e
<< isnecessaryfor the
current drivingcasetogiveacorrect result. Generally, for an ac-
tual network, it maynot bepossibletofindabreak point that sat-
isfies either of these extreme conditions.
By performing two signal-injection-based return ratio mea-
surements, theinfluenceof break point on dcimpedancecan be
canceledout such that thelocation of thebreak point can bese-
lected anywhere on the loop. This double-injection configura-
tion isshown in Fig. 6. At thenormal circuit operation, both i
inj
andv
inj
areset tozero; thus, thelooppropertiesareundisturbed.
For thefirst signal injection, i
inj
andv
inj
areprovidedsimulta-
neously to null i
f
. Wehavethefollowing equation at theoutput
side of the feedback loop:
k v Y v
e f f 1
0 + , (14)
andthenull voltagereturnratioT
v
n
isdefinedandcalculatedas
T
v
v
k
Y
v
n f
e f
i
f

0
1
.
(15)
For the second signal injection, i
inj
and v
inj
are provided si-
multaneously to null v
f
. Wehavethefollowing equation at the
output side of the feedback loop:
i k
i
Y
f
e
e
1
,
(16)
andthenull currentreturnratioT
i
n
isdefinedandcalculatedas
s
e
s
f
Loop Circuitry
4. The signal flow graph of a broken feedback loop.
v
e
v
f
+ +

Y
e
Y
e
k v
1 e
k i
2 e
Y
f
Y
f
i
e
i
f
(a)
(b)
5. The Norton and Thvenin equivalents of the feedback loop.
CIRCUITS & DEVICES I JANUARY 2 0 0 1 3 5 I
T
i
i
k
Y
i
n f
e
v
e
f

0
1
.
(17)
According to Eq. (13) and manipulating thealgebra, wehave
thereturn ratio
T
T T
T T
v
n
i
n
v
n
i
n

+
,
(18)
and
Y
Y
T
T
f
e
i
n
v
n
.
(19)
Comparedwith other double-injection techniques, null double
injectionisnumericallystableevenwhenreturnratioTapproaches
zero. Today, thistechniquehasbeen widelyappliedin both labora-
tory- based measurement and circuit analysis.
Eval uat i on of Tr ue Ret ur n Rat i o
Thereturn-loopmodel usedin Middlebrooksapproach implic-
itlyassumesthat signalsflowthrough thefeedback loopunilat-
erally. Thisisareasonableassumption for most low-frequency
applications. However, aswepointedout before, for most prac-
tical designs, theactivepath maynot bestrictlyunilateral, and
thefeedback path isusually bilateral. Thus, in addition to nor-
mal loop transmission, reverseloop transmission also exists
aroundthefeedback loop. Thesignal level of reversetransmis-
sion, which is ignored by thenull double-injection technique
andmost other approaches, can beaslargeasnormal transmis-
sion for microwaveapplications[8]. Toaccurately evaluatethe
stability of afeedback network, both normal- and reverse-loop
transmission should beconsidered. In this section, weextend
the unilateral return-loop model shown in Fig. 5 to the bilat-
eral return-loopmodel, andwecalculatethetruereturnratio,
which is proven to be the sum of normal return ratioand re-
versereturnratio.
Figure 7 shows the bilateral return-loop model configured
for thedouble-injection technique. Recallingthat at thenormal
circuit operation i
e
equalstoi
f
andv
e
equalsv
f
, Y
e
, Y
f
, k
1
, andk
3
areconnected in parallel. Compared with theone-port casedis-
cussedabove, weknowthat x k k +
1 3
,Y Y Y
e f
+ ; then, for the
bilateral model, we have
T
x
Y
k k
Y Y
e f

+
+
1 3
.
(20)
Thebilateral return-loopmodel can becharacterizedusing
ageneral two-port analysis. Let i
inj
andv
inj
bethedriving(input)
signals and i
f
and v
e
be the dependent (output) signals. Since
thefeedback network islinear, theonly constraint tobeplaced
upon each dependent signal, by superposition, is the linear
sum of the two values it would have due to each input signal
alone. Thus
i
v
A B
C D
i
v
f
e

1
]
1

1
]
1

1
]
1
inj
inj
.
(21)
The A, B, C, and Dparameters are characterized by a dou-
ble-injection technique that is essentially two small-signal ac
analyses: for thefirst small-signal ac analysis, weset i
inj
0 and
v
inj
1, then
B i D v
f
i v
e i v



inj inj
inj inj
and
0 1
0 1
,
,
;
(22)
for the second smal l -si gnal ac anal ysi s, we set i
inj
1 and
v
inj
0, then
A i C v
f i v e i v


inj inj inj inj
and
1 0 1 0 , ,
.
(23)
Herethedouble-injectiontechniqueisperformedtocharacter-
ize the return loop. Although a similar technique is used in
i
e
i
f
v
inj
i
inj
Y
e
k v
1 e
Y
f v
e
v
f
6. The illustration of double-injection technique.
i
e
i
f
v
inj
i
inj
Y
e
k v
3 f
k v
1 e
Y
f
v
e
v
f
7. Double-injection technique based on the bilateral return loop
model.
I 3 6 CIRCUITS & DEVICES I JANUARY 2 0 0 1
Middlebrooksapproach, thepurposethereisto
calculatethevoltageandcurrent returnratios.
Once the return loop is characterized, the
relationshipbetween thek
1
, k
3
,Y
e
,Y
f
, andA, B,
C, D parameters needs to be derived. Con-
sidering Fig. 7, circuit equations at the input
side and output side of the feedback loop are
+ + Y v i i k v v
e e f e inj inj 3
0 ( ) , (24)
k v i Y v v
e f f e 1
0 + + ( )
inj
. (25)
Compared with Eq. (21), wehavetheexpres-
sions for A, B, C, and Din terms of k
1
, k
3
, Y
e
,
andY
f
:
A
k Y
k k Y Y
B
YY k k
k k Y Y
f
e f
e f
e f

+ + +


+ + +
1
1 3
1 3
1 3
, ,
C
k k Y Y
D
k Y
k k Y Y
e f
f
e f

+ + +

+
+ + +
1
1 3
3
1 3
, .
Solving these four equations, we have
k
AD BC A
C
1


,
(26)
k
AD BC D
C
3

+
,
(27)
Y
AD BC A D
C
e

+ + 1
,
(28)
Y
BC AD
C
f


.
(29)
ReplacingEqs. (26), (27), (28), and(29) intoEq. (20), thefor-
mula that accurately calculates the true return ratio is
T
AD BC A D
BC AD A D

+
+ +
2
2 1
( )
( )
.
(30)
Recalling that theonly differencebetween theunilateral model
and thebilateral model isthat theunilateral model ignoresthe
reversetransmission factor k
3
, by replacing Eqs. (26), (28), and
(29) intoEq. (13), thenull double-injectiontechniquecalculates
the normal return ratio as
T
AD BC A
BC AD A D
normal


+ + 2 1 ( )
.
(31)
Accordingly, by ignoringk
1
, thereversereturn
ratio is calculated as
T
AD BC D
BC AD A D
reverse

+
+ + 2 1 ( )
.
(32)
It isobviousthat thetruereturnratioisthesum
of normal returnratioandreversereturnratio.
Theevaluation of Eq. (30) isreferredtoasa
loop-basedtwo-port algorithmin this article.
In summary, this algorithm consists of two
steps. First, deactivate the network input and
performtwosmall-signal acanalysestocharac-
terizethe A, B, C, and Dparameters; then the
true return ratio is calculated using Eq. (30).
Thisloop-based algorithm appliesdefinitely to
asingle-loopfeedback network andamultiloop
feedback network in which a critical wire can
be located to break all return loops. In the
multiloop case, model parameter k
1
, k
3
, Y
e
, and Y
f
are lumped
from all return loops.
It shouldbepointedout theproposedloop-basedtwo-port al-
gorithm differs fundamentally from the input-output two-port
analysis described in [3, 6, 7]. The latter approach models the
whole feedback network using two-port analysis based on the
ideal single-loopfeedback network shown in Fig. 1. Both thein-
put andoutput of thefeedback network can beeither current or
voltage, thus the specific feedback configuration needs to be
pre-identified to perform input-output two-port analysis. While
our loop-based two-port algorithm only models the feedback
loopanditsassociatedloadingeffects, all feedback networksfall
into the same loop model; since the driving signal and return
signal are of the same physical quantity, the voltage case and
current case are essentially equivalent. Furthermore, the loop
model in our approach is purely determined from the loop pa-
rameters, which provides auniquestability measure, whilein-
put-output two-port analysis produces results dependent upon
thetypeandlocation of network input [4]. Thisishighlightedby
a detailed example in the Numerical Results section.
The Devi ce- Based Gai n- Nul l i ng Al gor i t hm
Our loop-basedtwo-port algorithmrequiresthat circuit design-
ershaveaccesstothereturn loop, such that abreak point can be
placedtocharacterizetheloop. Therearetwocategoriesof appli-
cationsthat areof muchinterest tomicrowavecircuit designers.
In theRFregime, thedeviceparasiticeffect becomessignificant.
TaketheMOSFETtransistor asanexample: thesignal flowsback
fromtheinternal drain totheinternal gatethrough parasiticca-
pacitanceC
gd
. Under this scenario, thelocal returnloopexists
around an individual transistor. Self-oscillation is afrequently
encounteredphenomenoninhigh-frequencydesign, soit isvery
important toevaluatethestabilityof an individual transistor. In
this case, one design objective is to ensure that an individual
transistor isstablewith areasonablestabilitymargin beyondthe
operating bandwidth. On theother hand, theobjectiveof linear
R
f
I
in
Q1
R
c
V
cc
8. A simple bipolar feedback circuit.
CIRCUITS & DEVICES I JANUARY 2 0 0 1 3 7 I
oscillator design isto ensurethereturn ratio goesthrough the
critical oscillating point (Barkhausen criterion) at the desired
frequency. Some oscillators depend on the internal local loops
alone to start, others require an external loop or sometimes a
trigger isneeded. In all cases, return ratioisauseful design and
analysistool. However, sincethefeedback loop ishidden inside
thedevice, our loop-basedalgorithmcan nolonger beappliedto
these types of applications.
Toaddressthisissue, let usgoback totheoriginal definition
of return differenceand return ratio. Therearetwo factorsthat
prohibit thedirect useof theminpractice. First, direct measure-
ment of anetwork determinant isalmost impossible; second, the
original definition isbased on asinglecontrolled source, while
the actual return loops are far more complicated. Our loop-
based algorithm lumpsthecomplicated loop structureinto the
bilateral return loop model.
In circuit analysis, the calculation of the network determi-
nant israther simple. Given an N N matrix A, after LUdecom-
position, we have
A L U , (33)
whereL isthelower triangular with all diagonal ele-
mentsequal toone, andUistheupper triangular. The
determinant function of Ais simply

U
ii
i
N
1
.
(34)
Numerically, thedeterminant of alargematrixcan
be off the limit of the floating-point representation.
Oneway tohandlethisproblemistoorder thediago-
nal elements before the return ratio computation,
that is
T x
U
U
U
U
ii
i
N
ii
i
N
ii
ii i
N
( )
~
~

1
0
1
0
1
1 1,
(35)
where
~
U
ii
and
~
U
ii
0
areordered to bemonotonically in-
creasingor monotonicallydecreasing. Thisshouldal-
ways work since the return ratio T x ( ) itself has a
reasonable value.
Although thesingle-controlled-source-baseddefi-
nition cannot be applied to a general sophisticated
feedback network, thedeterminant-basedreturnratio
calculation can beperfectly applied to assessthesta-
bility of an individual transistor with local return
loops. This approach, referred to as a device-based
gain-nullingalgorithm, calculatesreturn ratiobased
on individual transistors at each frequency: the nor-
mal network determinant and the network determi-
nant wi t h cont rol l ed source x nul l ed are fi rst
calculated, then thereturn ratio associated with this
specific transistor isdetermined by Eq. (3). Here, the
controlled source x is assumed to be the dominant controlled
source(g
m
in most cases) in thetransistor under investigation.
Unlikeour loop- basedalgorithm, thegain-nullingalgorithmis
device-based, and there is no need to access the return loop.
Appl i cabi l i t y of Loop- Based and
Devi ce- Based Al gor i t hms
Theloop-based two-port algorithm produces stability informa-
t i on for a speci fi c r et ur n l oop, whi l e t he devi ce-based
gain-nullingalgorithmproducesstabilityinformation for aspe-
cific controlled source. Before proceeding further, we should
first investigatetheconditionunder whichthesetwoalgorithms
produceconsistent stability information for asingle-loop feed-
back network.
Our loop-basedanddevice-basedalgorithmsproduceconsis-
tent stability information for asingle-loop feedback network as
longasthefollowingconditionissatisfied: all controlledsources
of thenetwork appear in thenetwork determinant in asimple
product form [6]. This implies the nulling of any controlled
R
f
R
f
I
probe
I
probe
I
in
R
x
R
x
R
in
R
in
v
i
n
v
i
n
V
in
G V
m in
G V
m in
R
out
R
out
v
o
u
t
v
o
u
t
(a)
(b)
9. A single-loop feedback circuit, R
x
200k, R
f
100k, R
in
k 50 ,
R M
out
1 , g m
m
1 .
I 3 8 CIRCUITS & DEVICES I JANUARY 2 0 0 1
sourcein thenetwork rendersthewholenetwork to bepassive.
In other words, thereturn ratio with respect to any controlled
source is the same.
This condition can be further extended to multiloop cases:
for afeedback network that consistsof multipleloops, if acritical
controlled source can be located such that the nulling of this
sourcerendersthewholenetwork passive, return ratio with re-
spect to thiscontrolled sourceistheaccuratestability measure
of the network.
Asimplecircuit shown in Fig. 8 can beused to illustratethe
applicabilityof our loop-basedanddevice-basedalgorithms. Two
factors affect the applicability of this circuit: the local return
loop inside the transistor and the reverse-loop transmission
around the return loop. There are four cases:
1. If both the local loop and reverse transmission can be ig-
nored, both our loop-basedanddevice-basedalgorithmsproduce
accurateand consistent stability information for thenetwork.
2. If the reverse transmission can be ignored, only the de-
vice-based algorithm produces accurate stability information
for the network. If the network is a multiloop network, the
loop-basedalgorithmcan nolonger beappliedsincenowirecan
be found to break both global and local loops.
3. If thelocal return loopcan beignored, onlytheloop-based
algorithm produces accurate stability information for the net-
work. Theloop-based algorithm producesaccuratestability in-
formation for amultiloop network aslong asacritical wirecan
be found to break all loops. Nulling the normal transmission-
controlled sourcedoesnot render thewholenetwork to bepas-
sive, thus the device-based algorithm cannot be applied
4. If none of these two factors can be ignored, this corre-
sponds to general multiloop networks. Neither our loop-based
nor our device-based algorithms can be applied.
It isobviousthat our loop-basedanddevice-basedalgorithms
arecomplementary in termsof applicability for single-loopnet-
works. When thelocal return loop is theonly feedback loop in
thenetwork, theloop-basedalgorithmcan nolonger beapplied,
and our device-based algorithm produces accuratestability in-
formation. When the single-loop network consists of multiple
controlled sources, and they arenot in thesingleproduct form
in network determinant, our loop-basedalgorithmproducesac-
curatestabilityinformation sinceawirecan alwaysbefoundun-
der thismultipledevicesituation. Case2andCase3aretheexact
conditions under which our loop-based and device-based algo-
rithms can be applied to multiloop applications.
Numer i cal Resul t s
The proposed loop-based two-port algorithm and device-based
gain-nullingalgorithmhavebeenimplementedinaSpectrecircuit
simulator. In thissection, numerical resultsaregiven toillustrate
the accuracy advantage over input-output two-port analysis and
thenull double-injection technique. Theconditionsunder which
our loop-based and device-based algorithms can be applied to
multiloop applicationsarealso demonstrated with real circuit de-
signs. Although return ratiohasbeen usedthrough thetheoretical
V
dd
R1
I
probe
C
f
R
f
R2
V
in
M1
CL
10. A single MOSFET feedback circuit.
20 200
0
150
20
40
60
80
100
120
140
50
160 100
L
o
o
p

G
a
i
n

(
d
B
)
L
o
o
p

G
a
i
n

(
D
e
g
)
w/o Local Loop
w/o Local Loop
w/ Local Loop w/ Loop Loop
Loop Based Alg. Loop Based Alg.
10
0
10
0
10
2
10
2
10
4
10
4
10
6
10
6
10
8
10
8
10
10
10
10
10
12
10
12
100
50
0
Frequency Domain (Hz)
(a)
Frequency Domain (Hz)
(b)
11. The loop-gain waveform of the single MOSFET circuit.
devel opment i n previ ous
secti ons, l oop gai nthe
negativeof return ratioand
widely used in the engi-
neering contexthas been
produced as the output of
stability analysis.
Single-Loop Feedback Circuit
This example is taken from [4] and is used to demonstrate the
consistencybetween our loop-basedtwo-port algorithmandthe
device-based gain-nulling algorithm. Our loop-based and de-
vice-based algorithms produce consistent loop-gain results for
single-loop circuits and do not depend upon thetypeand loca-
tion of input sources. In contrast, input-output two-port analy-
sis, which does depend upon the type and location of input
sources, could produce yet another loop-gain result.
Thecircuit diagramisshown in Fig. 9. Notethat thecircuits
in Fig. 9(a) andFig. 9(b) areexactlythesameexcept thetypeand
location of the input source is different. Figure 9(a) shows a
shunt-shunt feedback configuration, whileFig. 9(b) showsase-
ries-shunt feedback configuration.
This circuit contains a single return loop and a single con-
trolled source, and it corresponds to Case 1 discussed above.
Current probe I
probe
(equivalent to the zero-valued voltage
sourcein Spectre) isinserted into thereturn loop to character-
izetheloop. I
probe
isusedtoaddv
inj
andi
inj
on-the-flyfor two-port
analysis. As expected, both our loop-based and device-based al-
gorithmsproducethesameloop-gain result of 35.09 for either
the shunt-shunt or series-shunt configurations.
In contrast, the input-output two-port analysis produces a
loop-gain result of 25.71 for theshunt-shunt configuration and
66.85for theseries-shunt configuration [4]. Notethat thesetwo
numbers differ by more than a factor of two. This observation
demonstrates that even for a typical single-loop circuit, in-
put-output two-port analysis produces inconsistent loop-gain
results as the type and location of the input source changes.
The difference of loop-
gain results between our
algorithms and input-out-
put two-port analysis can
beexplained asthat thein-
put-output two-port analy-
sis is purely based on the
si mpl i fi ed uni di recti onal
feedback model. For thiscircuit, theactivepath isstrictlyunidi-
rectional, while the feedback path is not.
Single MOSFET Feedback Circuit
Figure 10 shows a single MOSFET feedback circuit that is
usedtoillustratethestabilityanalysisof afeedback circuit oper-
atingat such ahigh frequencythat thelocal return loopscannot
be neglected.
For a MOSFET transistor, local return loops exist between
thedrain and gatethrough parasitic capacitances. Asfrequency
CIRCUITS & DEVICES I JANUARY 2 0 0 1 3 9 I
Q3
V
cc
Q4
I
probe
Q5
R1
Q1 Q2
V
in
I
bias
RL CL
12. A bipolar operational amplifier.
40
30
20
10
0
10
20
30
50
40
100
50
150
60 200
L
o
o
p

G
a
i
n

(
d
B
)
L
o
o
p

G
a
i
n

(
d
e
g
)
10
5
10
5
10
6
10
6
10
7
10
7
10
8
10
8
10
9
10
9
10
10
10
10
Frequency Domain (Hz)
(a)
Frequency Domain (Hz)
(b)
200
150
100
50
0
Normal Loop Gain Normal Loop Gain
13.99 dB
314.88 MHz 314.88 MHz 62.7 MHz
True Loop Gain True Loop Gain
54.04 deg
62.7 MHz
13. The loop gain waveform of the bipolar operational amplifier.
Our loop- based and device- based algor it hms
ar e complement ar y in t er ms of applicabilit y
for single- loop net wor ks.
I 4 0 CIRCUITS & DEVICES I JANUARY 2 0 0 1
moveshigher, theeffect of theselocal loopsupon circuit perfor-
mancecannot beneglected. To conduct theexperiment, wefirst
turn off the local return loops by setting the following bsim3v3
model parameters: xpart 1, capmod 0, cgdo 0, cgdl 0, and
cf 0. Thiscircuit containstworeturn loopsthrough R
f
andC
f
,
respectively; however, current probeI
probe
can beplacedasshown
in Fig. 10 to break both loops. In addition, thereisonly onecon-
trolledsource(g
m
of M1) that existsinthenetwork; thusbothour
loop-based and device-based algorithms produce consistent and
accurateloop-gain resultsasplotted by thesolid linein Fig. 11.
Next, we turn back the local return loops by setting the
bsim3v3 model parameters as follows: xpart 0, capmod 2,
and resetting parameterscgdo, cgdl, and cf to default such that
theyarenot forcedtozero. Duetotheexistenceof local loops, no
break point canbefoundtobreak all loops; however, thereisstill
onlyonecontrolledsource(g
m
of M1) that existsin thenetwork,
andnullingthiscontrolledsourceactually rendersthenetwork
tobepassive. ThiscorrespondstoCase2discussedabove. Thus,
our device-basedalgorithmstill producesaccurateloop-gain re-
sultsasplottedinthedashedlineinFig. 11. Our loop-basedalgo-
rithm producesresultsthat areinaccuratein high frequencies,
(plotted as the dotted line for comparison).
Bipolar Operational Amplifier
Figure12showsabipolar operational amplifier circuit with dif-
ferential input stage and single feedback loop. This circuit is
usedtodemonstratethedifferencebetweenour loop-basedalgo-
rithm that calculatesthetrueloop gain and Middlebrooksnull
double-injectiontechniquethat calculatesthenormal loopgain.
Figure 13 plots the loop-gain waveforms produced by our
loop-based algorithm and null double-injection technique. It is
obvious that the loop-gain waveforms overlapped very well ex-
cept at high frequency. This difference increases as the fre-
quency goes higher. The reason behind this is that the null
double-injection techniqueonly modelsthenormal loop trans-
missioninthereturnloop, whileour loop-basedalgorithmmod-
els both normal and reverse transmission, and the reverse
transmission increases as frequency goes higher.
Thestability marginscan beeasily obtained from theloop-gain
waveform, and our loop-based algorithm produces gain margin of
13.99 dBat 314.88 MHz and phasemargin of 54.04 degreeat 62.7
MHz. For astablecircuit or transistor, both gain margin andphase
marginshouldbepositive, or thephaseplot stayspositivewithinthe
operating bandwidth. This circuit corresponds to Case3 discussed
above. Nullingtheg
m
in either Q2or Q5doesnot render thewhole
network tobepassive, andour device-basedgain-nullingalgorithm
doesnot produceaccurateloop-gain resultsfor thiscircuit.
Linear Oscillator Circuit
Figure14 shows alinear oscillator circuit referenced from [7].
This circuit is designed to oscillate at a frequency of 10 MHz.
Figure15plotsboth thedBplot andphaseplot of loopgain on
onegraph. Theapplicationof stabilityanalysisonthelinear oscilla-
tor design is to ensure that both the dB plot and
phaseplot of loopgainarecloseenoughtozeroat the
desired oscillating frequency and that the phase of
loopgainchangesrapidlyintheneighborhoodof the
oscillation frequency for good frequency stability.
I t i seasyt overi fyfromFi gure15t hat at 10.6
MHz, t he l oop gai n i s 0.04 dB and 1.3 degrees
and t hat t hephaseof l oop gai n changesrapi dl y
from 90 degrees t o 90 degrees around t he os-
ci l l at i ng frequency.
Dr. Michael Tianiscurrentlywith CadenceDesign
Systems, where he works on signal integrity and
analog and RF CAD. He has previously worked at
Lucent Technologies Bell Laboratories and the
University of Iowa (e-mail: mtian@cadence.com).
Dr. V. Visvanathan was until recently with Ca-
denceDesign Systems. Heiscurrentlywith Texas
Instruments (India) whereheworks in theASIC
Product Development Center as the Chief Con-
V
cc
L1 C1
C2
I
probe
Q1 Q2
R1 C3
I
ee
VP
R2 C4
14. An emitter-coupled pair oscillator.
dB 20 (Loop Gain)
Phase (Loop Gain)
100
80
60
40
20
0
20
40
60
80
100
L
o
o
p

G
a
i
n

(
d
B
,

d
e
g
)
10
5
10
6
10
7
10
8
Frequency Domain (Hz)
15. Loop gain of the linear oscillator calculated by our loop-based algorithm.
sultant. Hehaspreviously workedat Bell Labsandhasbeen on
thefaculty of theUniversity of Maryland, CollegePark, andthe
Indian Instituteof Science, Bangalore.
Dr.J effreyHantganiscurrentlytheengineeringdirector for An-
alogsimulation andRFCADdevelopment with CadenceDesign
Systems, Inc. Hehaspreviously worked at AT&T/Lucent Tech-
nologies Bell Laboratories and has been on the faculty of The
StateUniversityof NewYork at StonyBrook andCornell Univer-
sity (e-mail: jch@cadence.com).
Ref er ences
1. H.W. Bode, NetworkAnalysisandFeedbackAmplifier Design. NewYork:
Van Nostrand, 1945.
2. W.K. Chen, ActiveNetwork andFeedback Amplifier Theory. New York:
McGraw-Hill, and Washington, DC: Hemisphere, 1980, pp. 185-246.
3. P.R. Gray and R.G. Meyer, AnalysisandDesignof AnalogIntegratedCir-
cuits. 3rd Ed. New York: Wiley, 1993, pp. 535-591.
4. P.J. Hurst, Acomparison of twoapproachestofeedback circuit analysis,
IEEE Trans. Educ., vol. 35, pp. 253-261, Aug. 1992.
5. P.J. Hurst and S.H. Lewis, Determination of stability using return ratios
in balancedfullydifferential feedback circuits, IEEETrans. Circuits Syst.
II, vol. 42, pp. 805-817, Dec. 1995.
6. E.S. Kuh and R.A. Rohrer, Theory of Linear ActiveNetworks. San Fran-
cisco, CA: Holden-Day, 1967, pp. 520-605.
7. K.S. Kundert, TheDesignersGuidetoSPICE andSpectre. Norwell, MA:
Kluwer, 1995, pp. 67-107.
8. D.J.H. Maclean, Broadband Feedback Amplifiers. Chichester, NY: Re-
search Studies Press, 1982.
9. R.D. Middlebrook, Measurement of loopgain in feedback systems, Int. J .
Electronics, vol. 38, pp. 485-512, Apr. 1975.
CIRCUITS & DEVICES I JANUARY 2 0 0 1 4 1 I
CDI
Soci et y New s
CAS NAMES NEW EDITOR TO MAGAZINE
Chung-Yu Wu wasrecently appointed thenewCASeditor
for IEEE Circuits & Devices Magazine, replacing Dr. Mona
Zaghloul. Dr. Wu wasborn in 1950. Hereceived hisM.S. and
Ph.D. degrees from the Department of Electronics Engi-
neering, National Chiao-TungUniversity, Taiwan, in 1976and
1980, respectively.
From1980to1984hewasan associ ateprofessor at Na-
ti onal Chi ao-Tung Uni ver-
si ty. Duri ng 1984-1986, he
wasavi si ti ngassoci atepro-
fessor i n theDepartment of
El ect r i cal Engi neer i ng,
Portl and State Uni versi ty,
Oregon. Si nce1987, hehas
been a professor i n the Na-
ti onal Chi ao- Tung Uni ver-
si ty. He was a reci pi ent of
theIEEEThi rdMi l l enni um
Medal ; theOutstandi ng Ac-
ademi c Awardby theMi ni s-
try of Educati on i n 1999;
the Outstandi ng Research Award by the Nati onal Sci ence
Counci l i n 1989-90, 1995-96, and 1997-98; theOutstand-
i ng Engi neeri ng Professor by theChi neseEngi neer Asso-
ci at i on i n 1996; and t he Tung-Yuan Sci ence and
Technol ogy Award i n 1997. From 1991 to 1995, hewasro-
tated to serve as di rector of the Di vi si on of Engi neeri ng
andAppl i edSci encei n theNati onal Sci enceCounci l . Cur-
rently, heistheCentennial Honorary Chair Professor at the
National Chiao-Tung University.
Hehaspublishedmorethan200technical papersininter-
national journalsandconferenceproceedings. Healsohas18
patents, includingnineUSpatents. Since1980, hehasserved
asaconsultant tohigh-tech industry andresearch organiza-
tions. He has built up strong research collaborations with
high-tech industries. His research interests focus on low-
voltage low-power mixed- mode circuits and systems for
multimediaapplications, hardwareimplementation of visual
andauditoryneural systems, RFcommunicationcircuitsand
systems, biochips, and bioelectronics.
Dr. Wu served on the Technical Program Committees of
IEEE ISCAS, ICECS, APCCAS. He served as the VLSI Track
Co-Chair of theTechnical ProgramCommitteeof ISCAS99. He
served as a Guest Editor of the Multimedia Special Issue for
IEEET-CSVTin August-October 1997. Healsoservedasan as-
sociateeditor for theIEEE TransactionsonVLSI Systemsand
IEEETransactionsonCircuitsandSystems-Part II. Heserved
as General Chair of IEEE APCCAS92. Currently, he serves as
theassociateeditor of theIEEETransactionsonVLSI Systems
and IEEE Transactions on Multimedia, and as the Distin-
guishedLecturer of IEEECASSociety. Heisoneof theCASrep-
resentativesintheIEEENeural NetworkCouncil andIEEETAB
Nanotechnology Committee. Healso servesasthemember of
the international advisory committee of IEEE APCCAS and
ICECS. He is a member of Eta Kappa Nu and Phi Honorary
Scholastic societies, and heisaFellowof IEEE. CDI