Optimal Allocation of Local Feedback in Multistage Amplifiers via Geometric

Programming
Joel L. Dawson, Stephen P. Boyd, Thomas H. Lee Maria del Mar Hershenson
Stanford University Barcelona Design
Stanford, CA 94305-4070 192 South Murphy St.
Sunnyvale, CA 94086



 We consider the problem of optimally allocating which is a special type of optimization problem. Even com-
local feedback to the stages of a multistage amplifier. The local
feedback gains affect many performance indices in a complicated
and nonlinear fashion, making optimization of the feedback gains
a very challenging problem. We show that geometric program-
ming provides a complete solution.
I I NTRODUCTION
The use of linear feedback around an amplifier stage was pi-
oneered by Black [1], Bode [2], and others. The relation be-
tween the choice of feedback gain and the (closed-loop) gain,
bandwidth, rise-time, sensitivity, noise, and distortion prop-
erties, is well understood (see, e.g., [3]). For a single stage
amplifier, the choice of the (single) feedback gain is a simple
plicated geometric programs can be solved very efficiently,
and globally, by recently developed interior-point methods
(see [5][7]). Therefore we are able to give a complete, global,
and efficient solution to the local feedback allocation problem.
In  II, we give a detailed description of the amplifier stage
models used to analyze the performance of the amplifier.
Though simple, the models capture the basic qualitative be-
havior of a source-degenerated differential pair.  III provides a
brief overview of geometric programming, and an example of
derived amplifier characteristics. A design example is given in
 IV.
II A MPLIFIER STAGE MODELS

problem. In this section we describe the different models of an ampli-

In our work we consider the multistage amplifier shown in fier stage used in our analysis.
1,consisting
figure of open-loop amplifier
 , with local stages
   employed
denoted
feedback gains  A Linearized static model
around the stages. The simplest model we use is the linear static model shown
replacements   in figure 2. The stage is characterized by  "!$#%'& , where #(
+  +
 is the gain of the ) th stage, which we assume to be positive.
- - We use this simple model for determining the overall gain of
  the amplifier, determining the maximum signal swing, and the
sensitivity of the amplifier gain to each stage gain.
* & 
Fig. 1. Block diagram of multistage amplifier. +
# & 
PSfrag replacements -

We assume that the amplifier stages are fixed,

the problem of choosing the feedback gains 
and
 consider
  . The
choice of these feedback gains affects a wide variety of per-
formance measures for the overall amplifier, including gain,
bandwidth, rise-time, delay, noise, distortion and sensitivity 
properties, maximum output swing, and dynamic range. These
performance measures depend on the feedback gains in a com- Fig. 2. Linearized static model of amplifier stage.
plicated and nonlinear manner. It is thus far from clear, given a
set of specifications, how to find an optimal choice of feedback
gains. We refer to the problem of determining optimal val- B Static nonlinear model
ues of the feedback gains, for a given set of specifications on
To quantify nonlinear distortion effects, we use a static non-
overall amplifier performance, as the local feedback allocation
linear model of the amplifier stage as shown in figure 3. We
problem.
assume a nonlinearity of the form

  !,+ - &./!0#  &2143  &57698 - &5:. 

In our work [4], we have shown that the local feedback al-
location problem can be cast as a geometric program (GP), (1)
 Q
3;'& 5
* &
-
 * & 
replacements
+
PSfrag replacements
+
# &  + +
- -
# &  +

 
Fig. 3. Nonlinear static model of amplifier stage. Fig. 5. Static noise model of amplifier stage.

This form is inspired by the transfer characteristic of a source-
coupled pair [8], and is a general model for third-order non-
linearity in a stage with an odd transfer characteristic. The
function + <->= . is called the transfer characteristic of the ) th
stage, and 3? is called the third-order coefficient of the ampli-
fier stage. Note that the gain and third-order coefficient are
related to the transfer characteristic by
III A MPLIFIER A NALYSIS
Having established the models of  II, there are many figures
of merit that are straightforward to derive. For our purposes
we require that such derivations result in posynomials, thereby
enabling the use of geometric programming. To this end, it is
useful to adopt return differences, RES!UTV6WC#( , as our design

#  !0+A@ -CB .  3  !D1 + @ @ @F -EB . 

variables.
(2) This section provides a basic description of geometric pro-
gramming, followed by an example of a derived amplifier char-
We assume that 3?HG B , which means the third-order term is acteristic. A more complete treatment of geometric program-
compressive: as the signal level increases from zero, the non- ming can be found in [9], [7], [5]; an extensive series of deriva-
linear term tends to decrease the output amplitude when com- tions can be found in [4].
pared to the linear model.
A Geometric programming
C Linearized dynamic model
To characterize the bandwidth, delay, and rise-time of the  ?X Y  be aX real-valued
X  Let  . It is calledfunction of real, positive variables
a posynomial function if it has the

[ Z \ X Ja `Eb X YJ c'b X JLd b

overall amplifier, we use the linearized dynamic model shown form
in figure 4. Here the stage is represented by a simple one-pole
transfer function with time constant I  (which we assume to be  - X 
 
 
 
 
  X  ."! \] _^ === (3)
positive).
* where ^fe G B and #  eUg R. When hWk!i j 6,T ,l X  m isX Y called
+ & M KEJLNPO K  
 monomial function.  Thus,  B
p  m X Y . o q is a monomial. Posynomials 5 6 X;nY o a5
is posynomial and l - X
for example,
replacements -
are closed under sums, products, and nonnegative scaling.
A geometric program (GP) has the form

 subject to
Fig. 4. Linear dynamic model of amplifier stage.
D Static noise model
 in figure 5, which
Last, we have the static noise model shown
includes a simple output-referred noise Q  . Our noise model
is characterized by the RMS value of the noise source, which
we denote Q  . We assume that noise sources associated with
 More complicated noise models can also be handled by our method.
different stages are uncorrelated.
 n - XX . 
ry  - X .tsuT  v) !wT  l E>z%x4 
minimize
X  /- { .VB !U T )v!w)/T w
!  T l   (4)
l
where   are posynomial functions and y  are monomial func-
tions. Geometric programs were introduced by Duffin, Peter-
son, and Zener in the 1960s [9].
The most important property of geometric programs for us
is that they can be solved, with great efficiency, and globally,
using recently developed interior-point methods [7], [5]. Ge-
ometric programming has recently been used to optimally de-
sign electronic circuits including CMOS op-amps [10], [11],
and planar spiral inductors [12].
B Example Derivation A Trade-offs among bandwidth, gain, and noise
Here we examine the static nonlinearity of a cascade of Consider a three-stage amplifier, all stages identical, with pa-
stages. This derivation would be useful for detemining the rameters
spurious-free dynamic range, or for evaluating intermodulation

#  !DT B  I  !UT B ?<  Q A  !  B jr

distortion products.
We begin by deriving the closed-loop third-order coefficient (11)

model of  II. The output  is related to the input * through the

of a single feedback amplifier stage, using the static nonlinear
The required closed-loop gain is 23.5dB. We maximized the

|!}+ - * 1W?~. 
relation bandwidth, subject to the equality constraint on closed-loop
(5) gain, and a maximum allowed value of input-referred noise.

Differentiating both sides with respect to * leads to the familiar

Figure 6 shows the optimal bandwidth achieved, as a function
of the maximum allowed input-referred noise. As it must, the
result from elementary feedback theory: optimal bandwidth increases as we relax (increase) the input-

a@ -CB ./! T76+ @-C+B . E- B . ! # R !# 

referred noise limit. Figure 7 shows the optimal values of the
feedback gains as the input-referred noise limit varies.
@
(6)

6
x 10
Differentiating again yields 2.1

a@ @ -CB ./! + @ @R -EB . ! B 

2

5
(7)
1.9

and, once more, 1.8

p Y F
Bandwidth (rad/sec)

! 1 R3 
A@ @ @ -EB ./! + @ @ @ -CB .PR1 R q + @ @ C- B . w
1.7

(8)
1.6

F
using + @ @ @ -CB .!1 3 and + @ @ -EB .! B from the previous equa- 1.5

tion. This equation shows that the third-order coefficient of the 1.4
closed-loop transfer characteristic is given by

3 !  @ @ @F -CB . ! R 3 
1.3
PSfrag replacements
(9) 1.2

1.1
4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8
This is the well-known result showing the linearizing effect of Maximum Allowed Input-referred Noise (Vrms) 7
x 10
(linear) feedback on an amplifier stage.
More generally, the third-order coefficient of a cascade of Fig. 6. Maximum bandwidth versus limit on input-referred noise.
stages can be expressed as [4]
 ? E  
[  These curves roughly identify two regions in the design
3 ! ]  \]  (# 5\ 3   # e
space. In one, the noise constraint is so relaxed as to not be
e ] O
(10)
 an issue. The program identifies the optimum bandwidth so-
lution for the given gain, which is to place all of the closed
This formula gives the relation between the local return differ- a 
loop poles in the same place. In the other, the tradeoff between
is strong. The noise contribution of Q
ences and the third-order coefficient of the overall amplifier. bandwidth and noise 
is independent of R , but the noise contributions
IV D ESIGN E XAMPLE
 of the follow-
ing stages can be diminished by making R (and therefore  )

small. It follows that  is the greatest of the feedback gains,
We find that complicated problems of feedback allocation
followed by  Y and  .
5
can be solved, globally and efficiently, using geometric pro-
gramming. We can take as an objective any posynomial per- We can also examine the optimal trade-off between band-
width and required DC gain. Here we impose the fixed limit
formance measure, and apply any combination of posynomial
on input-referred noise at TT B
? V rms, and maximize
constraints. We can also compute optimal trade-off curves by
varying one of the specifications or constraints over a range, the bandwidth subject to a required closed-loop gain.
computing the optimal value of the objective for each value of Figures 8 and 9 show the maximum attainable bandwidth and
the specification. the optimal feedback gain allocation as a function of the re-
 0.9 0.9

0.8 0.8

0.7 0.7
Feedback Voltage Gain (V/V)

Feedback Voltage Gain (V/V)

0.6 0.6

0.5 f3 0.5

0.4 f2 0.4
f2 f3
replacements 0.3
PSfrag replacements 0.3

0.2 0.2
f1

0.1
f1
0.1

0 0
4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 20 25 30 35 40 45 50 55 60
Maximum Allowed Input-referred Noise (Vrms) 7
x 10 Required Gain(dB)

Fig. 7. Optimal feedback allocation pattern, for maximum bandwidth with Fig. 9. Optimal feedback allocation pattern for maximum bandwidth versus
limit on input-referred noise. Gain = 23.5dB. required closed-loop gain. Maximum Input-referred noise = 4.15e-7 Vrms.

quired closed-loop gain. Again we see two regions in the de- unambiguously detected.
sign space caused by the noise constraint.
R EFERENCES
6
x 10
[1] H. S. Black. Stabilized feedback amplifiers. Bell System Technical
1.6
Journal, 13:118, 1934.
[2] H. W. Bode. Network Analysis and Feedback Amplifier Design. Van
Nostrand, New York, 1945.
1.4 [3] J. K. Roberge. Operational Amplifiers: Theory and Practice. John
Wiley and Sons, Inc., New York, New York, 1975.
[4] J. L. Dawson, S. P. Boyd, M. Hershenson, and T. H. Lee. Optimal
1.2 allocation of local feedback in multistage amplifiers via geometric pro-
gramming. IEEE Transactions on Circuits and Systems I, 2000. In
Bandwidth(rad/s)

press.
1 [5] S. Boyd and L. Vandenberghe. Introduction to convex opti-
mization with engineering applications. Course Notes, 1997.
http://www.stanford.edu/class/ee364/.
0.8
[6] J. Ecker. Geometric programming: methods, computations and appli-
cations. SIAM Rev., 22(3):338362, 1980.
[7] Y. Nesterov and A. Nemirovsky. Interior-point polynomial methods in
convex programming, volume 13 of Studies in Applied Mathematics.
replacements 0.6
SIAM, Philadelphia, PA, 1994.
[8] P. R. Gray and R. G. Meyer. Analysis and Design of Analog Integrated
Circuits. John Wiley and Sons, New York, NY, 1993.
0.4
20 25 30 35 40 45 50 55 60 [9] R. J. Duffin, E. L. Peterson, and C. Zener. Geometric Programming
Required Gain(dB) Theory and Applications. Wiley, 1967.
[10] M. Hershenson, S. Boyd, and T. H. Lee. GPCAD: A tool for CMOS op-
Fig. 8. Maximum bandwidth versus required closed-loop gain. Maximum amp synthesis. In IEEE/ACM International Conference on Computer
input-referred noise = 4.15e-7 Vrms. Aided Design, pages 296303, San Jose, CA, 1998.
[11] M. Hershenson, S. P. Boyd, and T. H. Lee. Automated design of folded-
cascode op-amps with sensitivity analysis. In International Conference
on Electronics, Circuits and Systems, volume 1, pages 11, September
V C ONCLUSION 1998.


[12] M. Hershenson, S. Mohan, S. Boyd, and T. H. Lee. Optimization of in-
ductor circuits via geometric programming. In Design Automation
In our work we have demonstrated that the local feedback Conference, pages 994998, 1999.
allocation problem is globally solvable by the use of geometric
programming. We emphasize the advantages of this method
over most general methods of nonlinear optimization: there is
no danger of getting trapped in a local extremum; there is
no need for a user-supplied starting point; infeasibility can be