Helmut E.

Graeb

Analog Design
Centering
and Sizing

 
Analog Design Centering and Sizing
ANALOG DESIGN CENTERING
AND SIZING

HELMUT E. GRAEB
Institute for Electronic Design Automation
Technische Universitaet Muenchen, Germany
A C.I.P. Catalogue record for this book is available from the Library of Congress.

ISBN 978-1-4020-6003-8 (HB)

ISBN 978-1-4020-6004-5 (e-book)

Published by Springer,
P.O. Box 17, 3300 AA Dordrecht, The Netherlands.

www.springer.com

Printed on acid-free paper

All Rights Reserved

c 2007 Springer
No part of this work may be reproduced, stored in a retrieval system, or transmitted in any form or by
any means, electronic, mechanical, photocopying, microfilming, recording or otherwise, without written
permission from the Publisher, with the exception of any material supplied specifically for the purpose
of being entered and executed on a computer system, for exclusive use by the purchaser of the work.
Contents

List of Figures xi
List of Tables xix
Preface xxi
1. INTRODUCTION 1
1.1 Integrated Circuits 1
1.2 Analog Circuits 4
1.3 Analog Design 6
1.4 Analog Sizing 13
2. TOLERANCE DESIGN: EXAMPLE 19
2.1 RC Circuit 19
2.2 Performance Evaluation 20
2.3 Performance-Specification Features 20
2.4 Nominal Design/Performance Optimization 20
2.5 Yield Optimization/Design Centering 22
3. PARAMETERS & TOLERANCES,
PERFORMANCE & SPECIFICATION 27
3.1 Parameters 27
3.2 Parameter Tolerances 28
3.3 Range-Parameter Tolerances 29
3.4 Statistical Parameter Distribution 31
3.5 Univariate Normal Distribution 32
3.6 Multivariate Normal Distribution 34
3.7 Transformation of Statistical Distributions 39
3.8 Generation of Normally Distributed Sample Elements 43
vi ANALOG DESIGN CENTERING AND SIZING

3.9 Global and Local Parameter Tolerances 44

3.10 Performance Features 45
3.11 Numerical Simulation 46
3.12 Performance-Specification Features 47
4. ANALOG SIZING TASKS 49
4.1 Sensitivity-Based Analysis 49
4.1.1 Similarity of Performance Features 51
4.1.2 Similarity of Parameters 51
4.1.3 Significance of Parameters 51
4.1.4 Adjustability of Performance Features 52
4.1.5 Multiple-Objective Behavior 52
4.1.6 Exercise 52
4.1.7 Sensitivity-Based Analysis of Tolerance Objectives 53
4.2 Performance-Sensitivity Computation 53
4.2.1 Simulator-Internal Computation 53
4.2.2 Finite-Difference Approximation 53
4.3 Scaling Parameters and Performance Features 54
4.3.1 Scaling to a Reference Point 54
4.3.2 Scaling to the Covered Range of Values 54
4.3.3 Scaling by Affine Transformation 55
4.3.4 Scaling by Equalization of Sensitivities 55
4.4 Nominal Design 56
4.5 Multiple-Objective Optimization 56
4.5.1 Smaller or Greater Performance Vectors 57
4.5.2 Pareto Point 59
4.5.3 Pareto Front 59
4.5.4 Pareto Optimization 61
4.6 Single-Objective Optimization 61
4.6.1 Vector Norms 63
4.6.2 Performance Targets 63
4.7 Worst-Case Analysis and Optimization 64
4.7.1 Worst-Case Analysis 64
4.7.2 Worst-Case Optimization 68
4.8 Yield Analysis, Yield Optimization/Design Centering 70
4.8.1 Yield 70
4.8.2 Acceptance Region Partitions 72
4.8.3 Yield Partitions 75
Contents vii

4.8.4 Yield Analysis 75

4.8.5 Yield Optimization/Design Centering 76
4.8.6 Tolerance Assignment 80
4.8.7 Beyond 99.9% Yield 81
5. WORST-CASE ANALYSIS 85
5.1 Classical Worst-Case Analysis 86
5.1.1 Classical Worst-Case Parameter Vectors 88
5.1.2 Classical Worst-Case Performance Values 89
5.1.3 Discrete Parameters 89
5.1.4 Corner Worst Case 90
5.2 Realistic Worst-Case Analysis 90
5.2.1 Realistic Worst-Case Parameter Vectors 93
5.2.2 Realistic Worst-Case Performance Values 93
5.3 Yield/Worst-Case Distance – Linear Performance Feature 93
5.4 General Worst-Case Analysis 95
5.4.1 General Worst-Case Parameter Vectors 98
5.4.2 General Worst-Case Performance Values 98
5.5 Yield/Worst-Case Distance – Nonlinear Performance Feature 99
5.5.1 Yield Approximation Accuracy 100
5.5.2 Realistic Worst-Case Analysis as Special Case 105
5.6 Exercise 105
6. YIELD ANALYSIS 107
6.1 Statistical Yield Analysis 107
6.1.1 Monte-Carlo Analysis 109
6.1.2 Importance Sampling 110
6.1.3 Yield Estimation Accuracy 111
6.2 Tolerance Classes 115
6.2.1 Tolerance Interval 115
6.2.2 Tolerance Box 116
6.2.3 Tolerance Ellipsoid 118
6.2.4 Single-Plane-Bounded Tolerance Region 121
6.2.5 Corner Worst Case vs. Realistic Worst Case 126
6.3 Geometric Yield Analysis 128
6.3.1 Problem Formulation 128
6.3.2 Lagrangian Function 131
6.3.3 First-Order Optimality Condition 132
6.3.4 Second-Order Optimality Condition 133
viii ANALOG DESIGN CENTERING AND SIZING

6.3.5 Worst-Case Range-Parameter Vector 135

6.3.6 Worst-Case Statistical Parameter Vector 135
6.3.7 Worst-Case Distance 136
6.3.8 Geometric Yield Partition 137
6.3.9 Geometric Yield 138
6.3.10 General Worst-Case Analysis/Geometric Yield
Analysis 141
6.3.11 Approximate Geometric Yield Analysis 142
6.4 Exercise 143
7. YIELD OPTIMIZATION/DESIGN CENTERING 145
7.1 Statistical-Yield Optimization 145
7.1.1 Acceptance-Truncated Distribution 145
7.1.2 Statistical Yield Gradient 146
7.1.3 Statistical Yield Hessian 148
7.1.4 Solution Approach to Statistical-Yield Optimization 149
7.1.5 Tolerance Assignment 150
7.1.6 Deterministic Design Parameters 150
7.2 Geometric-Yield Optimization 153
7.2.1 Worst-Case-Distance Gradient 153
7.2.2 Solution Approaches to Geometric-Yield Optimization 156
7.2.3 Least-Squares/Trust-Region Solution Approach 157
7.2.4 Min-Max Solution Approach 162
7.2.5 Linear-Programming Solution Approach 162
7.2.6 Tolerance Assignment, Other Optimization Parameters 163
Appendices
A Expectation Values 165
A.1 Expectation Value 165
A.2 Moments 165
A.3 Mean Value 165
A.4 Central Moments 166
A.5 Variance 166
A.6 Covariance 166
A.7 Correlation 166
A.8 Variance/Covariance Matrix 166
A.9 Calculation Formulas 167
A.10 Standardization of Random Variables 168
Contents ix

A.11 Exercises 168

B Statistical Estimation of Expectation Values 169
B.1 Expectation-Value Estimator 169
B.2 Variance Estimator 169
B.3 Estimator Bias 170
B.4 Estimator Variance 170
B.5 Expectation-Value-Estimator Variance 170
B.6 Estimator Calculation Formulas 171
B.7 Exercises 171
C Optimality Conditions of Nonlinear Optimization Problems 173
C.1 Unconstrained Optimization 173
C.2 First-Order Unconstrained Optimality Condition 174
C.3 Second-Order Unconstrained Optimality Condition 175
C.4 Constrained Optimization 176
C.5 First-Order Constrained Optimality Condition 177
C.6 Second-Order Constrained Optimality Condition 179
C.6.1 Lagrange-Factor and Sensitivity to Constraint 180
C.7 Bounding-Box-of-Ellipsoids Property (37) 181
References 183
Index 193
List of Figures

1 Trends in process and design technology. Data have

been taken from the Intel web site and from the
International Technology Roadmap for Semiconductors
(ITRS) web site. 2
2 Mixed-signal system-on-chip ICs. 3
3 A CMOS operational transconductance amplifier (OTA,
left) and a CMOS folded-cascode operational amplifier
(right). 4
4 An OTA-C biquad filter with OTAs from Figure 3 as
building blocks [57]. 5
5 A phase-locked loop (PLL) with digital components,
phase frequency detector (PFD), divider (DIV), and
analog components, charge pump (CP), loop filter (LF),
voltage-controlled oscillator (VCO). 5
6 Scope of analog design. 6
7 Nonlinear analog circuit. Structural model in form of a
circuit netlist, behavioral model in form of a differential
equation. Concerning the output voltage, both models
are equivalent. 8
8 Simplified models of the OTA in Figure 3. Behavioral
model in form of a description language code, Structural
model in form of a circuit netlist. Concerning the OTA
ports, both models are equivalent and approximate to
the transistor netlist in Figure 3. 9
9 Analog synthesis and analysis. 9
10 Design flow of a PLL. As in Figure 9, boxes denote the
structural view, rounded boxes the behavioral view. 11
11 Analog synthesis path of a PLL. 12
xii ANALOG DESIGN CENTERING AND SIZING

12 Analog sizing tasks. 13

13 Tolerance design tasks. 14
14 Elementary RC circuit with performance features time
constant τ and area A as a function of resistor value R
and capacitor value C. 19
15 RC circuit. (a) Performance values of RC circuit after
nominal design. (b) Parameter values of RC circuit after
nominal design. 21
16 Parameter values and parameter tolerances of the RC
circuit after nominal design. 23
17 Probability density function after nominal design of the
RC circuit, truncated by the performance specification. 24
18 (left) Parameter values and truncated probability den-
sity function. after nominal design of the RC circuit.
(right) Parameter values and truncated probability den-
sity function after yield optimization/design centering
of the RC circuit. 25
19 (a) Box tolerance region. (b) Polytope tolerance region.
(c) Ellipsoid tolerance region. (d) Nonlinear tolerance
region. 30
20 Probability density function pdf and cumulative distri-
bution function cdf of a univariate normal distribution. 34
21 Multivariate normal distribution for two parameters
according to (23), (24) or (34), respectively, with mean
values xs,0,1 = xs,0,2 = 0, correlation  = 0.5, and
variances σ1 = 2, σ2 = 1. 35
22 Different sets of level contours β 2 = (xs − xs,0 )T ·
C−1 · (xs − xs,0 ) (24) of a two-dimensional normal
probability density function pdfN . (a) Varying β, con-
stant positive correlation , constant unequal variances
σ1 , σ2 . (b) Varying β, constant zero correlation, con-
stant unequal variances. (c) Varying β, constant zero
correlation, constant equal variances. (d) Varying cor-
relation, constant β, constant unequal variances. 38
23 Sensitivity matrix in the parameter and performance
space. 50
List of Figures xiii

24 (a) Set M > (f ∗ ) of all performance vectors that are

greater than f ∗ , i.e. inferior to f ∗ with regard to multiple-
criteria minimization according to (92) and (93). (b) Set
M < (f ∗ ) of all performance vectors that are less than f ∗ ,
i.e. superior to f ∗ with regard to multiple-criteria min-
imization. 58
25 (a) Pareto front P F (f1 , f2 ) of a feasible performance
region of two performance features. Different Pareto
points addressed through different weight vectors from
reference point fRP , which is determined by the indi-
vidual minima of the performance features. (b) Pareto
front P F (f1 , f2 , f3 ) of three performance features with
boundary B = {P F (f1 , f2 ), P F (f1 , f3 ), P F (f2 , f3 )}. 60
26 (a) Continuous Pareto front of a convex feasible perfor-
∗ is monotone in f ∗ . (b) Continu-
mance region. fI,2 I,1
ous Pareto front of a nonconvex feasible performance
∗ is monotone in f ∗ . (c) Discontinuous
region. fI,2 I,1
Pareto front of a nonconvex feasible performance
∗ is nonmonotone in f ∗ .
region. fI,2 61
I,1

27 (a) Level contours of the weighted l1 -norm. (b) Level

contours of the weighted l2 -norm. (c) Level contours
of the weighted l∞ -norm. 63
28 Input and output of a worst-case analysis. 66
29 (a) Input of a worst-case analysis in the parameter and
performance space: nominal parameter vector, toler-
ance regions. (b) Output of a worst-case analysis in
the parameter and performance space: worst-case para-
meter vectors, worst-case performance values. 67
30 Nested loops within worst-case optimization. 69
31 (a) The volume under a probability density function of
statistical parameters, which has ellipsoid equidensity
contours, corresponds to 100% yield. (b) The perfor-
mance specification defines the acceptance region Af ,
i.e. the region of performance values of circuits that are
in full working order. The yield is the portion of circuits
in full working order. It is determined by the volume
under the probability density function truncated by the
corresponding parameter acceptance region As . 71
xiv ANALOG DESIGN CENTERING AND SIZING

32 Parameter acceptance region As partitioned into para-

meter acceptance region partitions, As,L,1 , As,U,1 , As,L,2 ,
As,U,2 , for four performance-specification features, f1 ≥
fL,1 , f1 ≤ fU,1 , f2 ≥ fL,2 , f2 ≤ fU,2 . As results
from the intersection of the parameter acceptance region
partitions. 74
33 Input and output of a yield analysis. 76
34 Yield optimization/design centering determines a
selected point of Pareto front of performance features. 77
35 Nested loops within yield optimization/design
centering. 78
36 (a) Initial situation of yield optimization/design center-
ing by tuning of design parameters xd that are disjunct
from statistical parameters. (b) After yield optimiza-
tion/design centering by tuning of design parameters
xd that are disjunct from statistical parameters. The
parameter acceptance region As depends on the val-
ues of design parameters xd . The equidensity contours
of a normal probability density function are ellipsoids
according to (24). 79
37 (a) Yield optimization/design centering by tuning of the
mean values of statistical parameters xs,0 . Parameter
acceptance region As is constant. (b) Yield optimiza-
tion/design centering by tuning of the mean values, vari-
ances and correlations of statistical parameters xs,0 , C
(tolerance assignment). Parameter acceptance region
As is constant. Level contours of the normal proba-
bility density function (24) change their shape due to
changes in the covariance matrix C. 80
38 To go beyond 99.9% yield for maximum robustness,
yield optimization/design centering requires specific
measures. 82
39 Classical worst-case analysis. 87
40 Classical worst-case analysis with undefined elements
xr,W L,1 , xr,W U,1 , of worst-case parameter vectors xr,W L ,
xr,W U . 89
41 Normal probability density function of a manufactured
component splits into truncated probability density func-
tions after test according to different quality classes. 90
42 Realistic worst-case analysis. 91
List of Figures xv

43 General worst-case analysis, in this case the solution is

on the border of the tolerance region. 96
44 (a) Comparison between true parameter acceptance
region partitions (gray areas) and approximate para-
meter acceptance region partitions (linen-pattern-filled
areas) of a general worst-case analysis of an upper worst-
case performance value. (b) Lower worst-case perfor-
mance value. (c) Lower and upper worst-case perfor-
mance value. 102
45 Duality principle in minimum norm problems. Shown
are two acceptance regions (gray), the respective nomi-
nal points within acceptance regions and two points on
the border of each acceptance region. Points (a) and (d)
are the worst-case points. 103
46 Statistical yield estimation (Monte-Carlo analysis) con-
sists of generating a sample according to the underly-
ing statistical distribution, simulating of each sample
element and flagging of elements satisfying the perfor-
mance specification. 109
47 Tolerance box Ts,B and normal probability density func-
tion pdfN,0R with zero mean and unity variance. 117
48 Ellipsoid tolerance region Ts,E and equidensity con-
tours of normal probability density function pdfN . 119
49 Single-plane-bounded tolerance region Ts,SP,U and equiden-
sity contours of normal probability density function pdfN . 122
50 Single-plane-bounded tolerance regions Ts,SP,L , Ts,SP,U
for a single performance feature with either an upper
bound (first row) or a lower bound (second row), which
is either satisfied (first column) or violated at the nom-
inal parameter vector (second column). 124
51 Single-plane-bounded tolerance region Ts,SP with cor-
responding worst-case parameter vector xW,SP and worst-
case distance βW for two parameters. βW denotes a
±βW times the covariances tolerance ellipsoid. Toler-
ance box Ts,B determined by ±βW times the covari-
ances with corresponding worst-case parameter vector
xW,B and worst-case distance βW,B . 126
xvi ANALOG DESIGN CENTERING AND SIZING

52 Geometric yield analysis for a lower bound on a per-

formance feature, f > fL , which is satisfied at the
nominal parameter vector. Worst-case parameter vec-
tor xs,W L has the smallest distance from the nominal
statistical parameter vector xs,0 measured according to
the equidensity contours among all parameter vectors
outside of or on the border of the acceptance region
partition As,L . The tangential plane to the tolerance
ellipsoid through xs,W L as well as to the border of As,L
at xs,W L determines a single-plane-bounded tolerance
region As,L . 129
53 Two stationary points xs,A and xs,W of (277), which
satisfy the first-order optimality condition. Only xs,W
satisfies the second-order optimality condition and is
therefore a solution of (270). 134
54 Worst-case distances and approximate yield values from
a geometric yield analysis of the operational amplifier
from Table 14. 139
55 Parameter acceptance region As (gray area) originat-
ing from four performance-specification features, f1 ≥
fL,1 , f1 ≤ fU,1 , f2 ≥ fL,2 , f2 ≤ fU,2 (Figure 32). A
geometric yield analysis leads to four worst-case para-
meter vectors xW L,1 , xW U,1 , xW L,2 , xW U,2 and four
single-plane-bounded tolerance regions. The intersec-
tion of these single-plane-bounded tolerance regions
forms the approximate parameter acceptance region As
(linen-pattern-filled area). 140
56 General worst-case analysis and geometric yield ana-
lysis as inverse mappings exchanging input and output. 142
57 (a) Statistical-yield optimization before having reached
the optimum. Center of gravity xs,0,δ of the probability
density function truncated due to the performance spe-
cification (remaining parts drawn as bold line) differs
from the center of gravity of the original probability
density function. A nonzero yield gradient ∇Y (xs,0 )
results. (b) After statistical-yield optimization having
reached the optimum. Centers of gravity of original and
truncated probability density function are identical. 148
58 Worst-case distances and approximate yield partition
values before and after geometric-yield optimization of
the operational amplifier from Table 15. 158
List of Figures xvii

59 Least-squares/trust-region approach to geometric-yield

optimization according to (366). Quarter-circles repre-
sent trust-regions of a step. Ellipses represent level con-
tours of the least-squares optimization objective (363).
r∗ is the optimal step for the respective trust region
determined by ∆. 160
60 Pareto front of optimal solutions of least-squares/trust-
region approach (366) in dependence of maximum step
length ∆, i.e. Lagrange factor λ (367). A point in
the bend corresponds to a step with a grand progress
towards the target worst-case distances at a small step
length. 161
61 Min-max solution to geometric-yield optimization. 163
C1 Descent directions and steepest descent direction of a
function of two parameters. 174
C2 Different types of definiteness of the second-derivative
of a function of two parameters: positive definite
(upper left), positive semidefinite (upper right), indefi-
nite (lower left), negative definite (lower right). 176
C3 Descent directions and unconstrained directions of a
function of two parameters with one active constraint. 178
List of Tables

1 Analog design terms. 7

2 Some modeling approaches for analog circuits in
behavioral and structural view. 8
3 Characteristics of numerical optimization. 15
4 Sizing of an operational amplifier. 17
5 Selected cumulative distribution function values of the
univariate normal distribution. 33
6 Worst-case analysis (WCA) types and characterization. 85
7 Standard deviation of the yield estimator if the yield
is 85% for different sample sizes evaluated according
to (225). 113
8 Required sample size for an estimation of a yield of 85%
for different confidence intervals and confidence levels
according to (232). 114
9 Tolerance intervals Ts,I and corresponding yield values
YI according to (235). 116
10 Yield values YB for a tolerance box Ts,B with ∀k βak =
−βbk = −βW = −3 in dependence of the number of
parameters nxs and of the correlation. ∀k=l k,l = 0.0
according to (241), ∀k=l k,l = 0.8 according to (237). 118
11 Yield values YE for an ellipsoid tolerance region Ts,E
with βW = 3 in dependence of the number of para-
meters nxs according to (247) and (248). 120
12 Single-plane-bounded tolerance region Ts,SP and cor-
responding yield partition values YSP according to (258)
or (259). 125
xx ANALOG DESIGN CENTERING AND SIZING

13 Exaggerated robustness βW,B represented by a corner

worst-case parameter vector of a classical worst-case
analysis, for different correlation ∀k=l k,l =  among
the parameters, and for different numbers of parameters
nxs . 127
14 Geometric yield analysis of an operational amplifier. 139
15 Geometric-yield optimization of an operational ampli-
fier from Section 6.3.9. 157
Preface

This book represents a compendium of fundamental problem formulations

of analog design centering and sizing. It provides a differentiated knowledge
about the tasks of analog design centering and sizing. In particular the worst-
case problem will be formulated. It stands at the interface between process
technology and design technology. This book wants to point out that and how
both process and design technology are required for its solution. Algorithms
based on the presented material are for instance available in the EDA tool
WiCkeD [88].
The intention is to enable analog and mixed-signal designers to assess CAD
solution methods that are presented to them. On the other side, the intention
is to enable developers of analog CAD tools to formulate and develop solution
approaches for analog design centering and sizing. The structure of the book
is geared towards a combination of a reference book and a textbook. The
presentation goes from general topics to the more specific details, preceding
material is usually a prerequisite for succeeding material. The formulations of
tasks and solution approaches by mathematical means makes the book suitable
as well for students dealing with analog design and design methodology.
The contents is structured as follows:
Chapter 1 sketches the role of analog circuits and analog design in integrated
circuits. An overview of analog sizing tasks and the corresponding terminology
is introduced.
Chapter 2 illustrates analog sizing and yield optimization/design centering
with the simplest example of an RC circuit.
Chapter 3 describes the basic input and output quantities of analog sizing.
Parameters and performance features are defined. Tolerance ranges and statis-
tical distributions of parameters as well as the performance specification are
introduced. The multivariate normal distribution as most important distribu-
tion type is described in the univariate and multivariatecase. In addition, the
xxii ANALOG DESIGN CENTERING AND SIZING

transformation of one type of distribution into another is sketched and illustrated

with examples like that of the generation of a normally distributed sample
element.
Chapter 4 formulates the basic tasks of analog sizing. The first task is a
sensitivity-based analysis of the effects of multiple parameters and multiple
design objectives. The second task is the scaling of parameters and design
objectives, which is crucial for the effectiveness of sizing algorithms. Perfor-
mance optimization is introduced as a multiple-objective approach, which leads
to Pareto optimization, and as a single-objective approach using vector norms.
After that, worst-case analysis and optimization are formulated and character-
ized. In the same way, yield analysis and optimization are treated. It is pointed
out how yield and worst-case refer to each other, how yield optimization/design
centering works beyond 99.9% yield, and how different quantities contribute
to the yield.
Chapter 5 develops three types of worst-case analysis, which are called clas-
sical, realistic and general. Solution approaches for the three types of worst-case
analysis are developed. In the case of classical and realistic worst-case analysis,
analytical solutions are obtained. In a general worst-case analysis, a numeri-
cal optimization problem is formulated as a starting point for the development
of solution algorithms. A general approach is described that relates a yield
requirement to the tolerance input of a worst-case analysis.
Chapter 6 describes two types of yield analysis. The first type is a statistical
estimation based on a Monte-Carlo analysis, the second type is a geometric
approximation that is closely related to worst-case analysis. Advantages and
limitations of both types are treated.
Chapter 7 develops two types of yield optimization/design centering. They
are based on the two types of yield analysis described in Chapter 6 and have
corresponding advantages and limitations. The problem formulations and
approaches for solution algorithms are developed. Derivatives of the objective
functions of yield optimization/design centering will be derived for sensitivity-
based optimization approaches.
It is recommended that the reader is familiar with matrix/vector notations
and the concept of optimality conditions. Appendices A-C summarize statisti-
cal expectation values and their estimation as well as optimality conditions of
nonlinear optimization.
Chapter 1

INTRODUCTION

1.1 Integrated Circuits

Integrated circuits (ICs), so-called chips, have become the brain and nervous
system of all kinds of devices. ICs are involved in everybody’s everyday life
more and more comprehensively. They are parts of domestic appliances, indi-
vidual or public transportation devices like cars and trains, traffic management
systems, communication networks, mobile phones, power plants and supply
networks in high and low voltage, medical devices outside or inside the body,
clothes, computers, and so on.
ICs are parts of systems, but ICs more and more form complete systems with
hardware and software components in themselves. What had been an electronic
cubicle with many components as tall as a human being years ago, today can
be implemented on one IC as large as a thumb nail.
This development has been made possible by an ongoing extraordinary pro-
gress in process technology and design technology of integrated circuits made
of silicon material. The progress in process technology is generally expressed
in the increase in the number of transistors, the fundamental electronic devices,
in one IC over the years. Over the past decades, an exponential growth of
the number of transistors per IC could be observed at a rate that has been
reported as a doubling of the number of transistors either every 12, 18 or 24
months. This exponential growth is called Moore’s Law after a forecast of the
Intel co-founder [85]. This forecast has become a goal of the semiconductor
industry in the meantime.
Process technology contributes to achieving this goal by developing the abil-
ity to realize ever smaller transistor dimensions and the ability to manufacture
ICs

with ever increasing number of transistors

2 ANALOG DESIGN CENTERING AND SIZING

Figure 1. Trends in process and design technology. Data have been taken from the Intel web
site and from the International Technology Roadmap for Semiconductors (ITRS) web site.

at nearly constant percentage of ICs that pass the production test (yield) and
with a reasonable life time (reliability).
Design technology contributes to achieving the exponential growth of IC com-
plexity by developing the ability to design ICs with ever increasing number of
functions by
bottom-up library development methods for complex components and
top-down system design methodologies
for quickly emerging system classes.
As a result it has been possible to produce ICs with an exponentially grow-
ing number of transistors and functions while keeping the increase in the cost
of design, manufacturing and test negligible or moderate. Figure 1 illustrates
the technology and design development. The dashed line shows the increase in
number of transistors on Intel processors, which exhibits a doubling of this num-
ber every 24 months for 35 years. Similar developments have been observed for
other circuit classes like memory ICs or signal processors. The solid line shows
the development of design cost for system-on-chip ICs. If no progress in design
technology and electronic design automation (EDA) had been achieved since
Introduction 3

Figure 2. Mixed-signal system-on-chip ICs.

1990, design cost would have exhibited an exponential growth as indicated by

the dotted line in Figure 1. But thanks to EDA innovations, design cost grew
only slowly or even decreased. The decrease in the mid 90s for instance has
been ascribed to small block reuse methodologies.
Because of the obvious economic advantage of integrated circuits, more and
more system functionality moves into ICs, which become ever more heteroge-
neous [70]. Modern system-on-chip ICs, as sketched in Figure 2, contain for
instance
memory components, processors, application-specific ICs (ASICs),
analog and digital components, high-frequency (RF) components,
hardware and software components (embedded systems),
optical or mechanical components.
But as miniaturization of transistors approaches atomic layers and as system
complexity approaches billions of transistors, the challenge of IC design, man-
ufacturing and test at reasonable cost has grown critically today. In addition,
relative variations in process technology and the impact of technological ef-
fects on system level are increasing and cannot be compensated by circuit
design alone. The result are significant variations in system performance, as
for instance processor operating frequency, that have to be coped with. Statis-
tical methods, which have been a topic of research and application in analog
4 ANALOG DESIGN CENTERING AND SIZING

Figure 3. A CMOS operational transconductance amplifier (OTA, left) and a CMOS folded-
cascode operational amplifier (right).

circuit design for decades, therefore become more and more important for dig-
ital design as well.

1.2 Analog Circuits

Analog components play an important role in integrated circuits. On one
hand, important system functions like clock generation, or signal conversion
between the off-chip analog environment and the on-chip digital signal pro-
cessing, are realized by analog circuits. On the other hand, analog circuits are
difficult to design and reluctant to design automation.
According to EDACafé Weekly (http://www.edacafe.com) on March 21,
2005, analog components use on average 20% of the IC area, as indicated
in Figure 2. At the same time, the analog components require around 40% of
the IC design effort and are responsible for about 50% of the design re-spins.
Analog design automation is therefore urgently needed to improve the design
quality and reduce the design effort, all the more as analog circuits are included
in more and more ICs. EDACafé Weekly reported that 75% of all ICs would
contain analog components by 2006.
Analog components appear in integrated circuits in various complexities.
For a long time, analog design has been addressing analog integrated circuits
as operational amplifiers (Figure 3) or filters (Figure 4). These circuits range
from a few transistors to around hundred(s) of transistors and are designed
based on numerical simulation with SPICE [89] and its advanced successors
for the computer-aided evaluation of a circuit’s performance.
But analog circuits on modern ICs are more complex and SPICE-like sim-
ulation consumes too much time so that more efficient modeling levels have
to be entered. Examples for such complex analog circuits are analog/digital
Introduction 5

Figure 4. An OTA-C biquad filter with OTAs from Figure 3 as building blocks [57].

Figure 5. A phase-locked loop (PLL) with digital components, phase frequency detector (PFD),
divider (DIV), and analog components, charge pump (CP), loop filter (LF), voltage-controlled
oscillator (VCO).

converters (ADCs), digital/analog converters (DACs), or phase-locked loops

(PLLs). These circuits are mixed-signal circuits with both analog and digi-
tal parts with an emphasis on the analog circuit behavior. Popular functions
of PLLs are clock generation, clock recovery or synchronization. Figure 5
shows the structure of a charge-pump PLL that can be used to generate a high-
frequency signal through an oscillator that is synchronized to a high-precision
low-frequency signal in a control loop.
The complexity of these analog and mixed-signal circuits ranges around
thousand(s) of transistors. The filter in Figure 4, which is hierarchically built
from OTA building blocks, can easily be simulated flat on circuit level using
the OTA transistor netlist in Figure 3, whereas the simulation time of the PLL
6 ANALOG DESIGN CENTERING AND SIZING

Figure 6. Scope of analog design.

in Figure 5 flat on circuit level reaches hours or days. The usual design process
that requires frequent numerical simulations becomes infeasible in this case.
A remedy offers the transition from modeling on circuit level to modeling on
architecture level using behavioral models. By modeling the whole PLL on
architecture level or by combining behavioral models of some blocks with
transistor models of other blocks, the PLL simulation time reduces to minutes
and the usual design process remains feasible.

1.3 Analog Design

Analog design usually stands for the design of analog circuits. Its most
important characteristic is the consideration of signals that are continuous in
time and value. Contrary to that, digital design considers values that are discrete
in time and value. Please note that value-continuous and time-continuous sig-
nals can be considered for any kind of circuit class, which leads to an extended
meaning of analog design in the sense of an analog design approach (Figure 6).
Analog design therefore not only refers to the design of analog and mixed-
signal circuits but includes for example the design of micromechanical compo-
nents, the characterization of digital library components on circuit level, or the
characterization of physical effects from interconnects or substrate.
As in digital design, the analog design process leads from a description of
the system requirements to a complete hardware and software realization (i.e.
implementation) in a cascade of individual design steps. Each design step can
be classified in terms of the design levels, the design partitioning, and the design
views it involves (Table 1).
The design partitioning and the design levels reflect the complexity of an
integrated circuit, which is recursively decomposed into a hierarchy of sub-
blocks that are designed individually and then composed to the system in a
divide-and-conquer manner.
Introduction 7

Table 1. Analog design terms.

Design Partitioning Design level Design view

HW/SW System Behavior
Analog/digital Architecture Structure
Functions, blocks Circuit Geometry
Device, process

Typical analog design levels are system level, architecture level, circuit level,
device level and process level. An example of an analog system is a receiver
frontend, whose architecture consists of different types of blocks like mixers,
low-noise amplifiers (LNAs) and ADCs. On circuit level, design is typically
concerned with a transistor netlist and applies compact models of components
like transistors. These compact transistor models are determined on device
level requiring dopant profiles, which in turn are determined on process level,
where the integrated manufacturing process is simulated.
The design partitioning concerns the hierarchical decomposition of the sys-
tem into subblocks. Design partitioning is required to handle the design com-
plexity and refers to important design decisions with respect to the partitioning
between hardware and software, between analog and digital parts, and between
different functions and blocks. The design partitioning yields graphs of design
tasks that represent design aspects like block interdependencies or like critical
paths for the design project management.
During the design process, a refinement of the system and an enrichment
with more and more details concerning the realization occurs. Along with it,
the system consideration traverses the design levels from the system level over
the architecture level down to the circuit level. Moreover, the design process
switches among three different design views, which are the behavioral view, the
structural view and the geometric view (Table 2). The design level rather refers
to the position of the considered system component within the hierarchical
system (de)composition. The design view on the other hand rather refers to
the modeling method used to describe the considered system component. The
behavioral view refers to a modeling with differential equations, hardware and
behavior description languages, transfer functions, or signal flow graphs. It is
oriented towards a modeling of system components with less realization details.
The structural view refers to a modeling with transistor netlists or architecture
schematics. It is oriented towards a modeling of system components with
more realization details. The geometric view refers to the layout modeling.
On the lowest design level, it represents the ultimate realization details for IC
production, the so-called mask plans. During the design process, not only a
8 ANALOG DESIGN CENTERING AND SIZING

Table 2. Some modeling approaches for analog circuits in behavioral and structural view.

Behavior Linear/nonlinear differential equations

Transfer function (e.g. pole/zero form)
Behavior description code (e.g. VHDL, Verilog)
Signal flow graph
Structure Circuit netlist (R, L, C, transistor, source elements)
LTI netlist (adding, scaling, integrating elements)
Architecture schematic

dv
dt + 1
RC v + IS v
C (exp( VT ) − 1) − 1
C i0 =0

Figure 7. Nonlinear analog circuit. Structural model in form of a circuit netlist, behavioral
model in form of a differential equation. Concerning the output voltage, both models are
equivalent.

refinement towards lower design levels takes place, but also a trend from the
behavioral view towards the structural and geometric view.
In analog design, often the structural view is to the fore, as transistor netlists
on circuit level or block netlists are prevalent in the design. But the behavioral
view is immediately included, for instance by the implicit algebraic differential
equations behind the transistor netlist or by program code. Behavioral and
structural models can be equivalent as in the two examples in Figures 7 and 8
but they do not have to be equivalent.
The behavioral model and the structural model in Figure 8 represent more
abstract models of the OTA circuit in Figure 3. While Figure 3 represents the
OTA on circuit level using detailed transistor models, Figure 8 represents the
OTA on architecture level to be used as a component of the filter in Figure 4.
The architecture-level structural and behavioral models are less complex, less
accurate and faster to simulate than the circuit-level models.
The relationship between design levels and design view in the design process
has been illustrated by the so-called Y-chart, in which each branch corresponds
to one of the three design views and in which the design levels correspond to
Introduction 9

Figure 8. Simplified models of the OTA in Figure 3. Behavioral model in form of a description
language code, Structural model in form of a circuit netlist. Concerning the OTA ports, both
models are equivalent and approximate to the transistor netlist in Figure 3.

Figure 9. Analog synthesis and analysis.

concentric circles around the center of the Y [47]. From the design automation
point of view, the change in design view during the design process from the
behavioral view towards the structural and geometric view is of special interest.
The term synthesis as the generic term for an automatic design process has
therefore been assigned to the transition from the behavioral view towards the
structural view, and from the structural view to the geometric view, rather than to
the transitions in design level or design partitioning [84]. Figure 9 illustrates this
concept, where synthesis and optimization denote the transition from behavioral
to structural and geometric view, and where analysis and simulation denote the
inverse transition.
10 ANALOG DESIGN CENTERING AND SIZING

In analog design, we can distinguish three synthesis phases regarding the

transition from the behavioral towards the structural and geometric view. Start-
ing from a behavioral formulation of the requirements on the performance of
a system or system component, structural synthesis serves to find a suitable
structure of the system component.
In the subsequent step, specific for analog design, parameter values for the
given structure, as for instance transistor widths and lengths in CMOS technol-
ogy, have to be determined by parametric synthesis. This process, and likewise
the result, is also called analog sizing. Parametric synthesis can be interpreted
as an additional part of the structural synthesis, which completes the structure
with required attributes, i.e. sizing. After that, layout synthesis leads to the
placement and routing.
This top-down process of synthesis and optimization [55, 24] is always
accompanied by bottom-up steps of analysis and simulation. These are required
for function modeling and evaluation during the synthesis and optimization pro-
cess, or for verification after synthesis.
An exemplary design flow of a PLL is depicted in Figure 10 [128]. The design
process starts on architecture level from a specification of the PLL behavior
concerning the locking behavior, noise, power, stability, and output frequency.
Structural synthesis on architecture level (de)composes the PLL (into)of a
netlist of five blocks. In this step, the design view changes from the behavioral
view to a mixed structural-behavioral view: a netlist is obtained, whose com-
ponents are modeled behaviorally. Additionally, an important partitioning into
digital and analog blocks is carried out. The open boxes for the three analog
blocks charge pump, loop filter and current-controlled oscillator indicate that
the structure has not been completed yet. This is done in the next step, which is
a parametric synthesis step on architecture level, where architectural parameter
values of the three analog blocks are computed. For instance values for the
charge current and jitter of the charge pump, values for the R’s and C’s of a
passive loop filter, and values for the gain, current and jitter of the controlled
oscillator are determined. Only now, the architectural realization of the PLL is
complete.
The parametric synthesis step is done on architecture level using behavioral
models for the PLL blocks for two reasons. First, the circuit-level transistor
netlist of the PLL is much too expensive to simulate as mentioned before, which
makes the parametric synthesis of a PLL on circuit level infeasible. Second,
using behavioral models for the PLL blocks allows to hide more realization
details at this stage of the design process and leads to a top-down specification
propagation. However, there must be a bottom-up modeling of the performance
capabilities of PLL blocks to prevent the parametric synthesis on architecture
level from producing unrealistic requirements on the PLL blocks.
Introduction 11

Figure 10. Design flow of a PLL. As in Figure 9, boxes denote the structural view, rounded
boxes the behavioral view.

Next, the design level switches to the circuit level, where the same process
from the architecture level is repeated for all blocks simultaneously, as indicated
by three parallel arrows in Figure 10. A transistor netlist capable of satisfying
the block specification is determined by structural synthesis. The circuit-level
parameter values, as the transistor widths and lengths, are computed by para-
metric synthesis.
The transition from the architecture level to the circuit level includes a tran-
sition from the behavioral view of the three analog blocks “under design” to the
structural view. Here a transformation of the architectural parameter values to
a circuit-level performance specification of each individual block is required.
Adequate behavioral block models do not only consider the input and output
behavior but prepare the required transformation by including model parameters
that have a physical meaning as performances of a block on circuit level, as for
instance the gain of the VCO.
12 ANALOG DESIGN CENTERING AND SIZING

Figure 11. Analog synthesis path of a PLL.

After the synthesis process, the obtained real block performance values on
circuit level can be used to verify the PLL performance.
If we illustrate the synthesis process just described in a plane with design
level and design view as coordinates, we obtain a synthesis path as shown in
Figure 11.
This figure reflects that both structural and parametric synthesis generally
rely on function evaluations, which happen in a behavioral view. This leads to
periodical switches to the behavioral view within the synthesis flow that can be
illustrated with a “fir”-like path.
A top-down design process as just described requires on each level adequate
models from the lower design levels. These models are computed beforehand
or during the design process. Ideally, the results of the top-down process can
be verified. But as the underlying models may not be accurate enough or
incomplete, for instance for reasons of computational cost, verification may
result in repeated design loops.
The design of the PLL was illustrated starting on the architecture level. On
system level, signal flow graphs, which represent a yet non-electric information
flow, are often applied. Switching to the architecture level then involves the
important and difficult transition to electrical signals.
Introduction 13

Figure 12. Analog sizing tasks.

1.4 Analog Sizing

Among the three synthesis steps concerning structure, sizing and layout, this
book deals with the sizing step, regardless of the design level or hierarchical
design stage in which it happens. Sizing is done in two steps (Figure 12).
In a first nominal design step, no parameter tolerances are considered, and
sizing means optimizing nominal performance values subject to constraints that
have to be satisfied.
In the subsequent tolerance design step, the inevitable variations of the man-
ufacturing process and of the operating conditions are considered.
The variations of the manufacturing process are modeled by global variations
of parameters like oxide thickness and by local variations of parameters like
threshold voltage. Global parameter variations are usually modeled as equal
for all transistors in a circuit, whereas local parameter variations are usually
modeled as independent for each transistor in a circuit.
The variations of the operating conditions are specified by intervals of para-
meters like supply voltage or temperature for which has to be guaranteed that
the system works properly.
The tolerance design task can be subdivided into subtasks as illustrated in
Figure 13.
Worst-case optimization denotes the task of minimizing the worst-case per-
formance deviation in a given tolerance region of parameters. This goal includes
a minimum performance sensitivity with regard to parameter tolerances. Every
optimization problem requires the evaluation of its optimization objective.
14 ANALOG DESIGN CENTERING AND SIZING

Figure 13. Tolerance design tasks.

Worst-case optimization requires the evaluation of worst-case performance val-

ues. This task in turn is denoted as worst-case analysis.
Yield optimization and design centering equally denote the task of maxi-
mizing the so-called yield, which is the percentage of systems that satisfy the
performance specification despite performance variations due to manufacturing
tolerances and operating tolerances.
There are two approaches to yield optimization/design centering, a statistical
approach and a geometric approach.
Statistical-yield optimization is based on a statistical estimation of the yield.
The underlying statistical yield analysis is done by means of a Monte-Carlo
analysis. Please note that the term statistical-yield optimization refers to the
statistical estimation of the yield, but not to the optimization method.
Geometric-yield optimization replaces the statistical estimation of yield by
a yield approximation that is based on geometric considerations. The manufac-
turing variations are described by a probability density function of statistical
parameters. Those parameter vectors that lead to a violation of the perfor-
mance specification are cut away. A geometric yield analysis is now based
on an approximation of the geometry of the performance specification in the
statistical parameter space.
Introduction 15

Table 3. Characteristics of numerical optimization.

Objective dimension: Single-objective optimization Multiple-objective optimization

Objective nature: Deterministic objective Statistical objective
Solution nature: Deterministic optimization Statistical optimization
Search basis: Gradient-free approach Gradient/sensitivity-based approach

The term design centering for yield optimization emanates from such a spa-
tial interpretation of the optimization problem. Section 7.1.2 will show that a
maximum yield is obtained when the center of gravity of the original probability
density function hits the center of gravity of the truncated probability density
function. Yield optimization as finding an equilibrium of the center of gravity of
the manufactured “mass” concerning the performance specifications becomes
design centering in this interpretation. This interpretation suggests that the
terms yield optimization and design centering should be used as synonyms.
As indicated in Figure 13, we emphasize the distinction between the analysis
of an objective and the optimization of this objective. We will see that yield
optimization approaches result immediately from the two ways of either statis-
tically estimating or geometrically approximating the yield. For that reason, the
terms “statistical” and “geometric” will be used to distinguish yield optimiza-
tion/design centering approaches. We will derive the first- and second-order
gradient of the yield based on the results of a Monte-Carlo analysis in Section
7.1. This leads to the formulation of a deterministic optimization method for the
statistically analyzed yield, which could be denoted as deterministic statistical-
yield optimization approach. We will derive the gradients of the worst-case
distances, which are geometric measures of yield, in Section 7.2.1, which leads
to a deterministic geometric-yield optimization approach.
Another important issue about yield optimization/design centering is the
difficulty to maximize yield values beyond 99.9%, which are described by
terms like six-sigma design. We will explain how to deal with these cases in
statistical-yield optimization in Section 4.8.7. Geometric-yield optimization
inherently covers these cases, as illustrated in Section 6.3.9.
Analog sizing is a systematic, iterative process that applies mathematical
methods of numerical optimization. The solution methods can be characterized
in different ways, which are summarized in Table 3.
16 ANALOG DESIGN CENTERING AND SIZING

One way of characterization is to distinguish between single-objective and

multiple-objective optimization problems. Nominal design with many perfor-
mance features usually is amultiple-objective optimization problem. Statistical-
yield optimization with the single objective yield on the other hand is a
single-objective optimization problem. We will describe the two types of opti-
mization problems in Sections 4.5 and 4.6. We will formulate yield optimiza-
tion/design centering both as a single-objective and as a multiple-objective
optimization approach in Section 4.8.5. This allows us to consider nominal
design and tolerance design in a unifying manner.
Other ways of characterizing the optimization problems of nominal and tol-
erance design are according to the nature of the objectives or of the solution
approaches. The objective and the solution approach either can be statistical or
deterministic. Concerning yield optimization/design centering, we will intro-
duce a way to closely couple the statistical objective yield with a deterministic
objective worst-case distance in Sections 6.2.4, 5.5 and 6.3.7. This allows a
unified view on the yield optimization/design centering tasks.
Another way to characterize optimization tasks is according to the usage
of gradients in the solution approach. While statistical optimization methods
often work without gradients, deterministic optimization methods often are
based on gradients and second-order derivatives. In this work, deterministic
solution methods for yield optimization/design centering that use gradients will
be formulated in Chapter 7.
It is important to note that numerical optimization is not only required for
yield optimization/design centering and performance optimization, but for tasks
that are named analysis in Figure 13. Worst-case analysis as well as geometric
yield analysis will turn out to be optimization problems as well. Solution
methods for these tasks will be presented in Chapter 5 and Section 6.3.
Optimization methods applied to sizing generally have to frequently evaluate
the optimization objective. In analog design this is done through numerical
simulation with SPICE-like simulators or through a performance model that has
to be computed a-priori based on numerical simulation. As a single numerical
simulation is very expensive in terms of CPU time, and as many simulations
are required within sizing, the number of required numerical simulations is the
essential measure for the CPU cost of a sizing algorithm.
Table 4 illustrates nominal design and yield optimization/design centering
in a typical sizing process of operational amplifiers like those shown in Figure
3. Five performance features (first column of Table 4) are considered in this
case, which characterize the transfer function (gain, transit frequency), stability
(phase margin), speed (slew rate) and power. For each performance feature,
a lower or upper bound is specified that defines the range of full performance
(second column).
Introduction 17

Table 4. Sizing of an operational amplifier.

1 2 3 4 5
Performance Specification Initial After nominal design After yield optimiza-
feature feature tion/design centering
Gain ≥ 80dB 67dB 100dB 100dB
Transit frequency ≥ 10M Hz 5M Hz 20M Hz 18M Hz
Phase margin ≥ 60◦ 75◦ 68◦ 72◦
Slew rate ≥ 10V /µs 4V /µs 12V /µs 12V /µs
DC power ≤ 50µW 122µW 38µW 39µW
Yield 0% 89% 99.9%

Numerical simulation of an initial sizing of circuit parameters that could for

instance originate from a previous process technology leads to the performance
values in the third column. We can see that all performance features but the
phase margin initially violate the respective performance-feature bounds.
A performance optimization performance optimization that aims at satisfying
all performance-feature bounds with as much safety margin as possible results
in the performance values given in the fourth column. All bounds are satisfied
with a certain safety margin. We cannot decide if one performance safety margin
is better than another one, but an analysis of the yield shows that 89% of all
manufactured circuits will satisfy the performance specification after nominal
design.
Next, yield optimization/design centering is executed and results in the per-
formance values in the last column. Essentially, the safety margin for the transit
frequency has been decreased while the safety margin for the phase margin has
been increased. The effect is an increase of the yield to 99.9% after yield
optimization/design centering.
The main reason for splitting the sizing into the two stages nominal design
and tolerance design is the cost of numerical simulation. Including parameter
tolerances and yield as an optimization objective leads to a significant increase
in the required number of simulations during optimization. It is therefore
advantageous to first optimize the circuits’ safety margins as much as possi-
ble without inclusion of tolerances and yield.
Chapter 2

TOLERANCE DESIGN: EXAMPLE

2.1 RC Circuit
Using an elementary RC circuit, we will illustrate the tasks of nominal design
and tolerance design. Figure 14 shows the circuit netlist and circuit variables.
Two performance features, the time constant τ and the area A, and two
circuit parameters, the resistor value R and the capacitor value C, are consid-
ered. Please note that R and C are normalized, dimensionless quantities. This
example is well suited as we can explicitly calculate the circuit performance in
dependence of the circuit parameters and as we can visualize the design situa-
tion in a two-dimensional performance space and a two-dimensional parameter
space.

τ = R·C
A = R+C

Figure 14. Elementary RC circuit with performance features time constant τ and area A as a
function of resistor value R and capacitor value C.
20 ANALOG DESIGN CENTERING AND SIZING

2.2 Performance Evaluation

The right side of Figure 14 displays the performance functions of the RC
circuit. The performance features, i.e. the time constant τ and the area A, for
the given parameters, i.e. the resistor value R and the capacitor value C, can
be explicitly evaluated in this case. But in general, performance evaluation
requires expensive numerical simulation (Section 3.11).
Please note that the RC circuit is a linear circuit because it consists of lin-
ear circuit elements, as opposed to nonlinear circuit elements like transistors.
Nevertheless the performance feature τ is nonlinear in the parameters.

2.3 Performance-Specification Features

For the RC circuit, a lower and upper bound on the time constant τ and an
upper bound on the area A are specified:

0.5 ≤ τ ≤ 2.0 0.5 ≤ R · C ≤ 2.0

≡ (1)
A ≤ 4.0 R + C ≤ 4.0

We call the inequality resulting from one bound a performance-specification

feature. (1) hence formulates three performance-specification features of the
RC circuit.

2.4 Nominal Design/Performance Optimization

Nominal design naturally tries to keep the nominal performance within the
specified limits with as much safety margin as possible. In our example, the time
constant τ has to be centered between its lower and upper bound, and the safety
margin of the area A with regard to its upper bound has to be maximized. This
is achieved by inspection of the problem or through the optimality conditions
of the following optimization problem:

min A subject to τ = 1.25

(2)
≡ min R + C s.t. R · C = 1.25

The solution results in:

√
R = C  = 5/2 ≈ 1.118
τ  = 1.25 (3)
√
A = 5 ≈ 2.236

Figure 15 illustrates the situation of the RC circuit after nominal design. Figure
15(a) sketches the two-dimensional space spanned by the performance fea-
tures “area” and “time constant,” Figure 15(b) shows the corresponding two-
dimensional space spanned by the parameters R and C.
Tolerance Design: Example 21

Figure 15. RC circuit. (a) Performance values of RC circuit after nominal design. (b) Parameter
values of RC circuit after nominal design.

A performance specification as in (1) constitutes box constraints as illustrated

by the gray box in Figure 15(a). In this example, the box of acceptable perfor-
mance values is not closed as no lower bound for the area is given. Figure 15(a)
shows the nominal performance values τ  and A and their centering between
the lower and upper bound of the time constant τ . From Figure 15(a) it is not
yet clear why the safety margin of the area to its upper bound is not larger.
Figure 15(b) shows the corresponding situation for the parameters. Using
the performance equations in Figure 14, the performance specification can be
formulated as inequalities for the parameters as done in (1). This results in
a bounding curve for each of the three performance-specification features in
Figure 15(b). The correspondence between the bounding curves in Figure 15
is indicated by single, double and triple slashes. The single slash belongs to the
performance-specification feature A ≤ 4.0 ≡ C = 4.0 − R. The double slash
belongs to the performance-specification feature τ ≤ 2.0 ≡ C ≤ 2.0/R, and
the triple slash belongs to τ ≥ 0.5 ≡ C ≥ 0.5/R. We can see that the region
of acceptable parameter values is closed. The nominal parameter values R
and C  are also shown. Here we can graphically reproduce the nominal design
process. Centering the design between the lower and upper bound of the time
constant is achieved by all points on the curve C = 1.25/R that runs between
the double-slashed curve and the triple-slashed curve. From all points on that
curve, R and C  is the one with the largest safety margin to the single-slashed
line representing the upper bound on the area. A larger safety margin for the
upper area bound could only be achieved at the cost of the safety margin for
the lower bound on the time constant.
22 ANALOG DESIGN CENTERING AND SIZING

We can also see that although we have centered the performance values, we
have not centered the parameter values: R and C  are closer to the double-
slashed curve than to the triple-slashed curve. This difference is due to the
nonlinearity of the time constant in the parameters. If we had centered the
parameter values instead of the performance values, using R = C = x, we
would not have centered x2 between 2, i.e. x2 = 0.5 · (0.5
√0.5 and √ √ + 2),√but
we would have centered x between 0.5 and 2, i.e. x = 0.5 · ( 0.5 + 2).
As a result, we would have designed the following parameter and performance
values: √
R = C  = 0.75 · 2 ≈ 1.061
τ  = 1.125 (4)
√
A = 1.5 · 2 ≈ 2.121
This difference between the situation in the performance space and the para-
meter space leads us to yield optimization/design centering.

2.5 Yield Optimization/Design Centering

In real designs, we are additionally facing statistical variations in the para-
meters. As a result, parameters do not have deterministic values, but are random
variables that are statistically distributed according to a certain statistical distri-
bution. Let us assume that in our example the parameters R and C are normally
distributed. Let the standard deviations of R and C be σR = 0.2 and σC = 0.8,
and let the correlation between R and C be  = 0. Then the joint normal
probability density function pdf(R, C) after nominal design is:

 
 2   2
1 −0.5 R−R
+ C−C
·e
σR σC
pdf(R, C) = (5)
2πσR σC
A probability density function is the continuous counterpart to a relative fre-
quency distribution. A relative frequency distribution describes the percentage
of events in each class of parameter values among all events. For a prob-
ability density function this corresponds to the probability of occurrence in
an infinitesimal parameter class (R + dR) · (C + dC), which is given by
pdf(R, C) · (R + dR) · (C + dC). Due to the quadratic form in the exponent
of (5), the probability density function of a normal distribution looks like a
bell-shaped curve. The level curves of our two-dimensional probability den-
sity function are ellipses around the nominal parameter values R , C  . At R ,
C  the probability density function takes its maximum value indicating maxi-
mum probability of occurrence. The “larger” a level ellipse is, the smaller is
the corresponding constant value of the probability density function indicating
smaller probability of occurrence.
Figure 16 illustrates the situation for the RC circuit after nominal design.
To indicate the decreasing values of the probability density function, the level
Tolerance Design: Example 23

Figure 16. Parameter values and parameter tolerances of the RC circuit after nominal design.

ellipses have decreasing line thickness. Had the standard deviations of R and
C been equal, the level curves would have been circles. σC being larger than
σR corresponds to a higher dilation of the level curves in the direction of C than
in the direction of R. The distribution of the resistor values hence is less broad
than that of the capacitor values. Figure 16 shows that parameter vectors exist
that lead to a violation of the performance specification. The volume under
the probability density function over the whole parameter space is 1, which
corresponds to 100% of the manufactured circuits. We can see that a certain
percentage of circuits will violate the performance specification. The percent-
age of manufactured circuits that satisfy the performance specification is called
parametric yield Y . Parametric variations refer to continuous variations within
a die or between dies of the manufacturing process. The resulting performance
variations can be within the acceptable limits or violate them. In addition, the
manufacturing process is confronted with catastrophic variations that result for
instance from spot defects on the wafer causing total malfunction and resulting
in the so-called catastrophic yield loss. The yield is an important factor of the
manufacturing cost: the smaller the yield, the higher the manufacturing cost.
In the rest of the paper, the term yield will refer to the parametric yield of a
circuit.
The estimated yield after nominal design in this example is:
Y  = Y (R , C  ) = 58.76% (6)
Chapter 6 will deal with methods to estimate the parametric yield. Figure 17
extends Figure 16 with a third axis for the value of the probability density
function to a three-dimensional view.
24 ANALOG DESIGN CENTERING AND SIZING

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0
4
1
3
2
2
3 1
4 0

Figure 17. Probability density function after nominal design of the RC circuit, truncated by the
performance specification.

Only that part of the probability density function that corresponds to accept-
able values of R and C is given. We can think of the probability density function
as a bell-shaped “cake” that describes the variation of the actual parameter vec-
tors due to the manufacturing process. A portion of varying parameter vectors
are cut away by the performance specification. An optimal design will lead to a
maximum remaining volume of the “cake,” which is equivalent to a maximum
parametric yield Ymax . The part of the design process that targets at maximum
yield is equally called yield maximization or design centering.
Figures 16 and 17 illustrate that the result of nominal design does not lead
to a maximum yield. The first reason for that is the mentioned nonlinearity of
the performance with respect to the parameters, which leads to skewing equal
safety margins of the performance with respect to its lower and upper bounds
in the parameter space. The second reason is the parameter distribution, which
Tolerance Design: Example 25

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 0
4 4
1 1
3 3
2 2
2 2
3 1 3 1
4 0 4 0

Figure 18. (left) Parameter values and truncated probability density function. after nominal
design of the RC circuit. (right) Parameter values and truncated probability density function
after yield optimization/design centering of the RC circuit.

leads to skewed spreading of the parameter values depending on the parameter

correlation and variance values.
From Figures 16 and 17 it is intuitively clear that a simultaneous decrease
of the nominal value of R and increase of the nominal value of C will increase
the amount of volume of the “cake” that is not cut away by the performance
specification and hence increase the yield. Figure 18 illustrates this. The
increase in yield can be seen in the larger probability density function “cake” left
by the performance specification and in the increased number of level ellipses
that are completely inside the region of acceptable parameter values.
After yield optimization/design centering, the estimated yield in this example
is:
Y  = Y (R , C  ) = Y (0.569, 2.016) = 74.50% (7)
26 ANALOG DESIGN CENTERING AND SIZING

Chapter 7 will deal with methods for yield maximization/design centering.

Please note that the complexity of performance evaluation through numer-
ical simulation leads to a much more challenging yield optimization/design
centering situation in real analog designs.
Chapter 3

PARAMETERS & TOLERANCES,

PERFORMANCE & SPECIFICATION

3.1 Parameters
In the literature, the term parameter usually denotes a variable quantity that
is related to the circuit structure and the circuit behavior. In this book, the term
parameters shall be used in a more specific sense for those quantities that are
input to the numerical simulation. Quantities related to the output of numerical
simulation shall be denoted as performance features. This distinction is owing
to the fact that the mapping of parameter values onto values of performance
features by means of numerical simulation takes the – computationally very
expensive – center stage of a systematic analog design process.

Definition: Parameters are variable quantities of an analog system or an

analog component that are input quantities to numerical simulation. The ele-
ments xk of the parameter vector x = [x1 . . . xk . . . xnx ]T ∈ Rnx denote the
values of the respective parameters, whose names are addressed through the
index k.

We can distinguish three types of parameters:

Design parameters xd = [xd,1 . . . xd,k . . . xd,nxd ]T ∈ Rnxd are subject to the
sizing process. Appropriate design parameter values have to be calculated
such that the performance is “optimal”. In the RC circuit in Section 2.1, the
capacitance and the resistance are design parameters. In CMOS circuits,
mainly transistor geometries, i.e. channel widths and lengths, form the
design parameters.
Statistical parameters xs = [xs,1 . . . xs,k . . . xs,nxs ]T ∈ Rnxs are subject
to statistical variations for instance due to fluctuations in the manufactur-
ing process. The statistical variations are modeled in form of probability
28 ANALOG DESIGN CENTERING AND SIZING

density functions of the statistical parameters. In the RC circuit in Sec-

tion 2.1, the capacitance and the resistance are statistical parameters. In
CMOS circuits, mainly transistor model parameters, like for instance oxide
thickness, threshold voltage or channel length reduction, form the statistical
parameters.
In the RC circuit, design parameters and statistical parameters are identical.
This corresponds to an analog circuit with discrete elements, as against to an
analog integrated circuit. Discrete resistors for instance can be selected with
different nominal values and with different tolerance classes, for instance
5%, 10%. The situation is different for integrated CMOS circuits. They have
transistor widths whose values can be devised individually. However, the
manufacturing process leads to a global variation of all transistors’ widths,
which is equal for all transistors and which is modeled in a statistical distri-
bution of one parameter “width reduction.” Another example of a statistical
parameter is the oxide thickness. Width reduction and oxide thickness are
not to be devised within analog design but within process design. In inte-
grated circuits, we usually have separate sets of design parameters on the
one hand and statistical parameters on the other hand. We will see in the
course of this book that this discrimination makes statistical approaches to
yield optimization/design centering at practicable computational cost very
difficult.

Range parameters xr = [xr,1 . . . xr,k . . . xr,nxr ]T ∈ Rnxr are parameters

that are subject to a range of values they can acquire. Typical examples of
range parameters are so-called operating parameters that model operating
conditions, as for instance supply voltage or temperature. Range parameters
differ from statistical parameters in that they do not have probability data in
form of statistical distributions on top of their given ranges of values. For
the temperature for instance, a range of -40◦ C to 125◦ C can be specified.
Regardless of the distribution of the temperature, the performance has to
be guaranteed for any temperature in this interval. The circuit operating
conditions usually are a part of the circuit specification, but as the operating
quantities are parameters, this part of the specification is categorized as
parameter tolerances within the numerical-simulation-based sizing process.

3.2 Parameter Tolerances

We distinguish two types of parameter tolerances, parameter ranges and
parameter distributions. Correspondingly, we have defined range parameters
and statistical parameters. Typical examples of parameter tolerances and the
essential sources of IC performance variation are the manufacturing fluctuations
and the operating conditions. The inevitable fluctuations in the manufacturing
process result in a whole range of performance values after production and test.
Parameters & Tolerances, Performance & Specification 29

The manufactured ICs are arranged by quality classes, for instance of faster or
slower ICs, and usually sold at different prices. The IC that is actually bought
therefore can have a quite different performance from what is specified as its
nominal performance in the data sheet. In addition, its performance will dynam-
ically vary with altering operating conditions. These dynamic variations during
operation may also be larger or smaller, depending on the quality of the actual
IC. Operating conditions are usually modeled as intervals of range parameters.
Manufacturing fluctuations are modeled as statistical parameter distributions
on the regarded design level. The calculation of the statistical parameter distri-
bution is a difficult task that requires specific optimization algorithms and the
consistent combination of measurements that belong to different design levels
and design views [40, 29, 109, 87, 98].

3.3 Range-Parameter Tolerances

Tolerances of range parameters are described by the region Tr ⊂ Rnxr of
values that the range parameters can adopt. The tolerance region Tr of range
parameters can be defined in different ways. Typical tolerance regions are
boxes, polytopes, ellipsoids or a general nonlinear region:
box region

Tr,B = {xr | xr,L ≤ xr ≤ xr,U }

(8)
xr,L , xr,U ∈ Rnxr , −∞ ≤ xr,L,i < xr,U,i ≤ +∞

polytope region

Tr,P = {xr | Apoly · xr ≤ bpoly }

(9)
Apoly ∈ Rnpoly ×nxr , bpoly ∈ Rnpoly

ellipsoid region




Tr,E = {xr
(xr − xr,0 )T · Aellips · (xr − xr,0 ) ≤ b2ellips
(10)
Aellips ∈ Rnxr ×nxr , Aellips symmetric, positive definite

nonlinear region

Tr,N = {xr | ϕnonlin (xr ) ≥ 0}

(11)
ϕnonlin , 0 ∈ Rnnonlin
30 ANALOG DESIGN CENTERING AND SIZING

Figure 19. (a) Box tolerance region. (b) Polytope tolerance region. (c) Ellipsoid tolerance
region. (d) Nonlinear tolerance region.

A vector inequality of the form x ≤ y assumes that x, y ∈ Rnx and is defined

as
x≤y ⇔ ∀ xµ ≤ yµ (12)
µ

Figure 19 illustrates the four types of tolerance regions. The box region accord-
ing to (8), illustrated in Figure 19(a), is the prevalent type of range-parameter
tolerances. It results if minimum and maximum values for each range para-
meter are specified. The dashed line in Figure 19(a) for instance corresponds
to the upper bound of parameter xr,1 ≤ xr,U,1 . In many cases, the tolerance
region Tr will be closed, that means a finite range of values will be given for
each range parameter. This reflects practical design situations where neither
process parameters nor operating parameters should attain excessive values,
be it for physical limitations or limited resources. The oxide thickness as well
as the channel length for example are process quantities that cannot go below
certain values. On the other hand, the channel length will be bounded above
due to the limited chip area. For the operating range in turn, reasonable finite
ranges have to be negotiated between manufacturer and customer. The size of
this range is related to the product price: an IC that works between −40◦ C and
125◦ C will be more expensive than the same IC designed to work between 0◦ C
and 70◦ C.
If the range parameters’ bounds are determined in a linear dependence of
each other, the tolerance region will have the shape of a polytope according
to (9), as illustrated in Figure 19(b). (9) describes a set of npoly inequalities
of the form aTµ· · xr ≤ bpoly,µ , where aTµ· = [aµ,1 aµ,2 aµ,3 . . .]T is the µ-th
row of the matrix Apoly . Every inequality corresponds to a bounding plane
as indicated by the dashed line in Figure 19(b), determining a halfspace of
acceptable parameter vectors. The tolerance region Tr is the intersection of
Parameters & Tolerances, Performance & Specification 31

all these halfspaces. Inequalities may be redundant, leaving them out would
not alter the tolerance region Tr . The bounding line at the top of Figure 19(b)
corresponds to such a redundant inequality.
(10) defines a tolerance region that is shaped as an ellipsoid, as illustrated by
the ellipse in Figure 19(c). It will be detailed in the following Section 3.4 that
the level curves of a Gaussian probability density function are ellipsoids. If the
shape of a parameter region corresponds to the shape of the level contours of a
probability density function, we are able to complement the parameter region
with probability values for the occurrence. In this way, we allow a transition
between range parameters and statistical parameters.
(11), illustrated by Figure 19(d), finally provides the most general descrip-
tion of a nonlinear tolerance region described by single nonlinear inequalities,
ϕnonlin,µ (xr ) ≥ 0, as nonlinear bounding surfaces, as indicated by the dashed
line in Figure 19(d).

3.4 Statistical Parameter Distribution

The manufacturing variations are modeled through a joint, multidimensional,
continuous distribution of statistical parameters.
In the discrete case, distributions are characterized by dividing the complete
value range into an ordered sequence of intervals and counting the number of
events in each interval. By normalizing these numbers to the total number of
events, we obtain the relative frequency function. By accumulating the relative
frequencies of all parameter intervals up to a maximal parameter value, we
obtain the cumulative frequency function.
In the continuous case, the relative frequency function becomes a probability
density function pdf, and the cumulative frequency function becomes a cumu-
lative distribution function cdf. pdf and cdf are obtained from each other in the
following way:

 xs,1  xs,nxs
cdf(xs ) = ... pdf(t) · dt , dt = dt1 · dt2 · . . . · dtnxs (13)
−∞ −∞

∂ nxs
pdf(xs ) = cdf(xs ) (14)
∂xs,1 · ∂xs,2 · . . . · ∂xs,nxs

In (13), the accumulation of relative frequencies has become an integration over

the probability density function. The probability density function vice versa
results from the partial derivatives of the cumulative distribution function with
respect to all parameters.
For an interpretation of the cumulative distribution function values as relative
portions of occurrence, we require a normalization of the probability density
32 ANALOG DESIGN CENTERING AND SIZING

function such that:

 +∞  +∞
lim cdf(xs ) = ... pdf(t) · dt = 1 (15)
xs→∞ −∞ −∞

An exemplary continuous distribution is the normal distribution, which is also

called Gaussian distribution.

3.5 Univariate Normal Distribution

The cumulative distribution function and the probability density function of
a univariate normal distribution are:
x 2
s −xs,0
1 −1
pdfN (xs ) = √ e 2 σ
, σ>0 (16)
2π · σ
 xs  xs −xs,0
σ 1 1 2
cdfN (xs ) = pdfN (t)·dt = √ e− 2 t ·dt (17)
−∞ −∞ 2π
  
pdfN N (t)
 xs −xs,0 
1 σ 1 1 2 1 xs − xs,0
= + √ e− 2 t · dt = + φ0 (18)
2 0 2π 2 σ
 xs −xs,0
√ 
1 1 2 2·σ
−t2 1 1 xs − xs,0
= + √ e · dt = + erf √ (19)
2 2 π 0 2 2 2·σ
In (16), xs,0 denotes the mean value (expected value, expectation value) of
parameter xs and σ denotes the standard deviation of xs . Background material
about expectation values can be found in Appendix A. A random variable1 xs
that originates from a normal distribution with mean value xs,0 and standard
deviation σ is also denoted as:

xs ∼ N (xs,0 , σ 2 ) (20)

1 Speaking of a random variable x, we immediately denote the VALUE x of a random variable X. This
value x refers e.g. to
a random number generated according to the statistical distribution of the random variable,
the argument in the probability density pdf(x),
 x+dx
the infinitesimal range x + dx having a probability of occurrence prob(X ≤ x) = x pdf(t)dt =
cdf(x + dx) − cdf(x).
x
The probability that a random variable X has a value less or equal x is prob(X ≤ x) = −∞ pdf(t)dt =
cdf(x). Please note that the probability that a random variable X has a value of x is prob(X = x) =
 x
x pdf(t)dt = 0!
Parameters & Tolerances, Performance & Specification 33

Table 5. Selected cumulative distribution function values of the univariate normal distribution.

xs − xs,0 −3σ −2σ −σ 0 +σ +2σ +3σ +4σ

cdf(xs − xs,0 ) 0.1% 2.2% 15.8% 50% 84.1% 97.7% 99.8% 99.99%

σ 2 denotes the variance of the univariate normal distribution.

In (17), the cumulative distribution function is formulated as an integral over
the probability density function according to (13). By variable substitution, the
integrand is simplified to the standard normal distribution with the mean value
0 and the standard deviation 1:
xs − xs,0
∼ N (0, 1) (21)
σ
The standard normal probability density function pdfN N in (17) can also be
obtained by setting xs,0 = 0 and σ = 1 in (16). The cumulative distribution
function of the variable xs taken from a normal distribution is hence obtained by
x −x
the standard cumulative distribution function value at s σ s,0 , which denotes
the difference of xs to the mean value as a multiple of the standard deviation.
In (18) and (19), the cumulative distribution function is formulated using the
function φ0 and the error function erf, which are available in statistical tables
and program libraries.
Figure 20 illustrates the probability density function pdfN and the cumulative
distribution function cdfN of a univariate normal distribution.
pdfN shows the typical bell shape of the normal distribution. As pdfN is
symmetrical around xs,0 , cdfN (xs,0 ) = 0.5. The hatched area under pdfN
from −∞ till xs,0 + σ is determined by cdfN (xs,0 + σ) according to (13), as
indicated by the curved arrow in Figure 20. Vice versa, according to (14), the
slope at cdfN (xs,0 + σ) is determined by pdfN (xs,0 + σ).
Table 5 shows selected cumulative distribution function values of the uni-
variate normal distribution. It can be seen that there is a unique relation between
tolerance intervals and yield values in the univariate case:
Y = prob{xs ∈ Ts } = prob{xs ∈ [xs,L , xs,U ]}
= cdf(xs,U ) − cdf(xs,L ) (22)
In this case, yield optimization/design centering can equivalently be formulated
as centering the nominal parameter value between the interval bounds or as
maximizing the interval bounds around the nominal parameter value.
34 ANALOG DESIGN CENTERING AND SIZING

Figure 20. Probability density function pdf and cumulative distribution function cdf of a uni-
variate normal distribution.

The relationship between tolerance ranges and yield values becomes more
complicated for multivariate distributions.

3.6 Multivariate Normal Distribution

The probability density function of a multivariate normal distribution is given
by:

1

pdfN (xs ) = √ nxs √ · exp −0.5 · β 2 (xs ) (23)
2π · det C
β 2 (xs ) = (xs − xs,0 )T · C−1 · (xs − xs,0 ) (24)
β ≥ 0
Parameters & Tolerances, Performance & Specification 35

0.16
0.14 f(x,y)
0.15
0.12 0.1
0.05
0.1
0.08
0.06
0.04
0.02
0

-4
-3
-2
-1
0
1
2 3 4
3 1 2
-1 0
4 -4 -3 -2

Figure 21. Multivariate normal distribution for two parameters according to (23), (24) or (34),
respectively, with mean values xs,0,1 = xs,0,2 = 0, correlation  = 0.5, and variances σ1 = 2,
σ2 = 1.

In (24), the vector xs,0 = [xs,0,1 . . . xs,0,k . . . xs,0,nx ]T denotes the vector of
mean values xs,0,k , i = 1, . . . , nxs , of the multivariate normal distribution of
the parameters xs,k , k = 1, . . . , nxs . Background material about expectation
values can be found in Appendix A. Figure 21 illustrates the probability density
function of a multivariate normal distribution for two parameters.
It shows the typical bell shape of the probability density function of two para-
meters and the quadratic form of the equidensity contours determined by (24).
The matrix C denotes the variance/covariance matrix, or covariance matrix,
of the parameter vector xs .
A random vector xs (see Note 1 in Section 3.5) that originates from a normal
distribution with mean value xs,0 and covariance matrix C is also denoted as:

xs ∼ N (xs,0 , C) (25)

The covariance matrix C is formed by the standard deviations σk , k =

1, . . . , nxs of the individual parameters xs,k , k = 1, . . . , nxs , and by the
mutual correlations k,l , k, l = 1, . . . , nxs , k = l, between parameters xs,k
and xs,l in the following way:

C = Σ·R·Σ (26)
36 ANALOG DESIGN CENTERING AND SIZING
⎡ ⎤
σ1 0 ... 0
⎢ .. ⎥
⎢ 0 σ2 . ⎥
Σ = ⎢ . ⎥ (27)
⎣ .. ..
. 0 ⎦
0 ... 0 σnxs
⎡ ⎤
1 1,2 ... 1,nxs
⎢ .. ⎥
⎢  1 . ⎥
R = ⎢ 2,1
.. ⎥ (28)
⎣ .
..
. nxs −1,nxs ⎦
nxs ,1 ... nxs ,nxs −1 1
⎡ ⎤
σ12 σ1 1,2 σ2 ... σ1 1,nsx σnxs
⎢ .. ⎥
⎢ σ  σ1 σ22 . ⎥
C = ⎢ 2 2,1
.. ⎥ (29)
⎣ .
..
. σnxs −1 nxs −1,nxs σnxs ⎦
nxs ,1 ... nxs ,nxs −1 σn2 xs
σk2 is the variance of parameter xs,k . σk k,l σl is the covariance of parameters
xs,k , xs,l . For the variance, we have:
σk > 0 (30)
For the correlation, we have:
k,l = l,k (31)
−1 ≤ k,l ≤ +1 (32)
From (26)–(32) follows that the covariance matrix C is symmetric and pos-
itive semidefinite.
If k,l = 0, this is denoted as the two parameters xs,k , xs,l being uncorrelated.
In the case of a multivariate normal distribution, “uncorrelated” is equivalent
to “statistically independent.” In general, “statistically independent” implies
“uncorrelated.”
If two parameters xs,k , xs,l are perfectly correlated, then |k,l | = 1. Perfect
correlation refers to a functional relationship between the two parameters. For
each value of one parameter a unique value of the other parameter corresponds,
i.e. the statistical uncertainty between the two parameters disappears. In this
case, the covariance matrix has a zero eigenvalue and the statistical distribution
is singular, i.e. (23) cannot be written. In the following, we will assume that
|k,l | < 1 (33)
This means that C is positive definite and that the multivariate normal distribu-
tion is nonsingular. Asingular distribution can be transformed into a nonsingular
one [5].
Parameters & Tolerances, Performance & Specification 37

For two parameters, (23) and (24) have the following form:

 


xs,1 xs,0,1 σ12 σ1  σ2
xs = ∼ N xs,0 = , C=
xs,2 xs,0,2 σ1  σ2 σ22



1 1 (xs,1 − xs,0,1 )2
pdfN (xs,1 , xs,2 ) =  · exp − ·
2πσ1 σ2 1 − 2 2(1 − 2 ) σ12

xs,1 − xs,0,1 xs,2 − xs,0,2 (xs,2 − xs,0,2 )2
−2 + (34)
σ1 σ2 σ22

If the two parameters are uncorrelated, i.e.  = 0, (34) becomes:



1 1 (xs,1 − xs,0,1 )2 (xs,2 − xs,0,2 )2
pdfN (xs,1 , xs,2 ) = · exp − + (35)
2πσ1 σ2 2 σ12 σ22
   
1 1 (xs,1 − xs,0,1 )2 1 1 (xs,2 − xs,0,2 )2
= √ exp − · √ exp −
2πσ1 2 σ12 2πσ2 2 σ22

= pdfN (xs,1 ) · pdfN (xs,2 )

If the two parameters are uncorrelated, i.e.  = 0, and if their mean values
are xs,0,1 = xs,0,2 = 0, and if their variance values are σ1 = σ2 = 1, (34)
becomes:
   
1 1 2 2 1 1 T
pdfN (xs,1 , xs,2 ) = · exp − xs,1 + xs,2 = √ nxs · exp − xs xs (36)
2π 2 2π 2

The level curves of the probability density function of a normal distribution, i.e.
the equidensity curves, are determined by the quadratic form (24). Therefore,
the level curves of pdfN are ellipses for nxs = 2, ellipsoids for nxs = 3 (Figure
21) or hyper-ellipsoids for nxs ≥ 3. In the rest of this book, three-dimensional
terms like “ellipsoid” will be used regardless of the dimension.
Figure 22 illustrates different shapes of the equidensity curves of a normal
probability density function in dependence of variances and correlations. Each
equidensity curve is determined by the respective value of β in (24). For
β = 0, the maximum value of pdfN , given by √2π nxs1·√det C , is obtained at
the mean value xs,0 . Increasing values β > 0 determine concentric ellipsoids
centered around the mean value xs,0 with increasing value of β and decreasing
corresponding value of pdfN .
Figure 22(a) illustrates the situation if the two parameters are positively cor-
related. The major axes of the ellipsoids then have a tilt in the parameter space.
Note that the equidensity ellipsoids denote a tolerance region corresponding
to (10). If the two parameters are uncorrelated, then the major axes of the
ellipsoids correspond to the coordinate axes of the parameters, as illustrated
in Figure 22(b). If uncorrelated parameters have equal variances, then the
ellipsoids become spheres, as illustrated in Figure 22(c).
38 ANALOG DESIGN CENTERING AND SIZING

Figure 22. Different sets of level contours β 2 = (xs − xs,0 )T · C−1 · (xs − xs,0 ) (24) of
a two-dimensional normal probability density function pdfN . (a) Varying β, constant positive
correlation , constant unequal variances σ1 , σ2 . (b) Varying β, constant zero correlation,
constant unequal variances. (c) Varying β, constant zero correlation, constant equal variances.
(d) Varying correlation, constant β, constant unequal variances.
Parameters & Tolerances, Performance & Specification 39

Figure 22(d) illustrates a set of equidensity contours for a constant value

of β. A positive correlation between two parameters leads to a contraction
of the equidensity ellipsoids towards the axis of parameter alterations in the
same direction. A negative correlation leads to a contraction of the equidensity
ellipsoids towards the axis of parameter alterations in the opposite direction.
The larger the absolute value of correlation, the narrower the shape of the
ellipsoid. For a correlation of ±1, this part of the ellipsoid degenerates to the
corresponding line of parameter alteration. In addition, for a constant value of
β, any equidensity contour for an arbitrary value of the correlation k,l between
two parameters lies in a bounding box around the mean value xs,0 that has a size
of β · 2σk in direction of parameter xs,k and β · 2σl in direction of parameter
xs,l . This bounding-box-of-ellipsoids property for two statistical parameters
xs,k , xs,l can be formulated by:

{xs | (xs − xs,0 )T C−1 (xs − xs,0 ) ≤ β 2 }
|k,l | ≤ 1


= xs
|xs,k/l − xs,0, k/l | ≤ β σk/l (37)

While the left set in (37) describes an ellipsoid tolerance region according
to (10), the right set in (37) describes a box tolerance region according to
(8). The bounding-box-of-ellipsoids property (37) says that the union of all
ellipsoid tolerance regions that are obtained by varying the correlation value of
two parameters for a constant β yields exactly a box tolerance region ±βσk/l
around the mean value. A motivation of (37) is given in Appendix C.7.
The cumulative distribution function cdfN (xs ) of the multivariate normal
distribution can be formulated according to (13). A more specific form analo-
gous to (18) and (19) of the univariate normal distribution cannot be obtained.
Only if the correlations are zero, then the probability density function of the
multivariate normal distribution can be formulated as a product of the univari-
ate probability density functions of the individual parameters (see (35)) and we
obtain:
nxs 
 xs,k 
nxs
R = I ⇒ cdfN (xs ) = pdfN (t)dt = cdfN (xs,k ) (38)
i=1 −∞ k=1

I denotes the identity matrix.

3.7 Transformation of Statistical Distributions

Due to the quadratic function (24) in the probability density function (23),
the formulation of tolerance design tasks based on a multivariate normal dis-
tribution of parameters can be very thorough and rich in technical interpretation.
40 ANALOG DESIGN CENTERING AND SIZING

Therefore, we will assume that the statistical parameters are normally distributed.
But parameters are not normally distributed in general. For instance para-
meters like oxide thickness that cannot attain values below zero naturally have a
skew distribution unlike the symmetric normal distribution. These parameters
have to be transformed into statistical parameters whose distribution is normal.
The resulting transformed parameters of course have no longer their original
physical meaning. This is in fact not necessary for the problem formulation
of yield optimization/design centering. The reader is asked to keep in mind
that the normally distributed statistical parameters that are assumed in the rest
of this book may originate from physical parameter that are distributed with
another distribution.
Given two random vector variables, y ∈ Rny and z ∈ Rnz with ny = nz ,
which can be mapped onto each other with a bijective function2 , i.e. z(y) and
y(z), the probability density function pdfz of random variable z is obtained
from the probability density function pdfy of random variable y by:




∂y


pdfz (z) = pdfy (y(z)) ·

det (39)
∂zT





∂y

pdfz (z) = pdfy (y(z)) ·




∂z

The idea behind the transformation is to keep the probability of an infinitesimal

parameter interval equal while transforming it to another infinitesimal interval:

pdfz (z) · dz = pdfy (y) · dy (40)

This corresponds to a substitution of variable y in the integration over pdfy to

calculate cdfy (13):
 y1  yny
cdfy (y) = ... pdfy (y)dy
−∞ −∞
 


z1 (y) znz (y)
∂y


= ...

pdfy (y(z))
det dz
−∞ −∞ ∂zT

 z1  znz
= ... pdfz (z)dz = cdfz (z) (41)
−∞ −∞

Exercise. Prove (17) using (41).

2 z(y) is meant to be read as “z evaluated at y.” This implies that z is a function of y, i.e. z = ϕ(y). The

function ϕ is not denoted separately.

Parameters & Tolerances, Performance & Specification 41

Example 1. This example shows how a transformation function from one

random variable to another random variable can be derived if the two probability
density functions are given.
Let us consider the case that the univariate random variable z is originating
from a uniform distribution, z ∼ U(0, 1) with the probability density function
pdfU :

1, 0 < z < 1
pdfU (z) = (42)
0, else
We are looking for the function that maps the uniformly distributed variable z
into a random variable y that is distributed according to pdfy . Towards this, we
insert (42) in (40), i.e. 1 · dz = pdfy (y) · dy, and obtain:

y = cdf−1
y (z) (43)

(43) says that a uniformly distributed random variable z is transformed into

a random variable y that is distributed according to pdfy through the inverse
cumulative distribution function of y. If pdfy is given, this is done by
y
solving −∞ pdfy (t)dt to compute cdfy (y) and then

computing the inverse cdf−1

y .

Example 2. This example shows how the probability density function pdfz
of random variable z can be derived if the probability density function pdfy of
random variable y and the transformation function y → z are given.
Given as an example are a random variable y originating from a normal
distribution (44) and an exponential function (45) that maps y onto z:
 2
1 −1
y−y0
pdfy (y) = √ e 2 σ
(44)
2π · σ
z = ey (45)

We are looking for the probability density function pdfz (z). As the function
z = ey is bijective w.r.t. R → R+ , (39) is applied:






∂y

∂y
1
pdfz (z) = pdfy (y(z)) ·



with y = lnz and



= yields
∂z ∂z z
 2
1 −1
lnz−y0

pdfz (z) = √ e 2 σ
(46)
2π · σ · z
(46) denotes the pdf of a lognormal distribution. As can be seen from the
calculations above, the logarithm y = lnx of a lognormally distributed random
42 ANALOG DESIGN CENTERING AND SIZING

variable z is normally distributed. As z > 0, the lognormal distribution may

be a candidate for modeling the distribution of a random variable that does
not attain values below zero. Based on the transformation (45), a normally
distributed random variable can be obtained.

Example 3. This example is of the same type as Example 2. Given are a

random variable y originating from a standard normal distribution (47) and a
quadratic function (48) that maps y onto z:

1 1 2
pdfy (y) = √ e− 2 y (47)
2π
z = y2 (48)

We are looking for the probability density function pdfz (z). The function
z = y 2 is not bijective, but surjective w.r.t. R → R+ . In this case we can add
the probabilities of all infinitesimal intervals dy that contribute
√ to the probability
√
of an infinitesimal interval dz. Two values, y (1) = + z and y (2) = − z, lead
to the same value z and hence two infinitesimal intervals in y contribute to the
probability of dz:
   
pdfz (z) · dz = pdfy y (1) · dy (1) + pdfy y (2) · dy (2) (49)

This leads us to:

 
∂y (1)
 
∂y (2)

(1)

(2)


pdfz (z) = pdfy y (z)

+ pdfy y (z)



∂z

∂z






1 − 12 z

1 −1

1 −1

= √ e ·

z 2

+

− z 2


2π 2 2
1 1 1
= √ · z − 2 · e− 2 z (50)
2π

(50) denotes the probability density function of a χ2 (chi-square)-distribution

with one degree of freedom. The χ2 distribution can be applied to describe the
probability of ellipsoidal tolerance regions determined by β (24).
The transformation of statistical distributions and of the corresponding ran-
dom variables is also important for statistical estimations, for instance by a
Monte-Carlo analysis. Towards that, a statistical sample, i.e. a finite set of
sample elements, is generated, which is originating from a certain statistical
distribution. In the next section, the generation of sample elements from a
normal distribution will be described.
Parameters & Tolerances, Performance & Specification 43

3.8 Generation of Normally Distributed Sample Elements

The generation of normally distributed sample elements3 xs ∼ N (xs,0 , C)
is done in three steps:
Create a vector z with vector elements that are independently uniformly
distributed, i.e. zk ∼ U (0, 1), k = 1, . . . , nxs . This step is based on
methods to create pseudo-random sequences of real numbers between 0
and 1. This is implemented for instance by a linear feedback shift register,
where the XOR operation of several bits is fed back to the input. Another
approach is a multiple recursive generator zk = wn /ν, wn := (a1 wn−1 +
. . . + aµ wn−µ ) mod ν, of order µ and a prime modulus ν. Programming
libraries usually contain functions for the generation of uniformly distributed
pseudo-random numbers.
Map z onto a random variable y that is distributed under a standardized nor-
mal distribution, i.e. y ∼ N (0, I), y ∈ Rnxs . This is done as in Example
1 of the previous section. According to (43), the value zk obtained from
a uniform distribution is mapped onto a value yk from a univariate stan-
−1 −1
dard normal distribution by yk = cdfN (zk ). cdfN is evaluated through
functions like ϕ0 (18) or erf (19), which are contained in programming
libraries. The uniformly distributed values zk , k = 1, . . . , nxs , are statisti-
cally pseudo-independent, and the values yk , k = 1, . . . , nxs , are collected
in a random vector y of normally distributed, pseudo-uncorrelated vector
elements.
Map y into a random variable xs that is distributed under a normal distri-
bution, i.e. xs ∼ N (xs,0 , C). In this third step, y is transformed into a
random variable xs ∼ N (xs,0 , C) by a linear transformation:
xs = A · y + b (51)
The unknown coefficients A and b can be derived by the requirements that
the expectation value of xs is xs,0 , and that the variance of xs is C:
E {xs } = xs,0 (52)
V {xs } = C (53)
Inserting (51) in (52) and (53) and applying (A.12) and (A.13) we obtain:
E {xs } = A · E {y} + b = b (54)
V {xs } = A · V {y} · AT = A · AT (55)

3A sample element is represented by a vector of parameter values x. It refers to the stochastic event of an
infinitesimal range of values dx around x.
44 ANALOG DESIGN CENTERING AND SIZING

Comparison of (52) and (53) with (54) and (55) finally yields

y ∼ N (0, I) → xs ∼ N (xs,0 , C), C = A · AT (56)

xs = A · y + xs,0 (57)
y = A−1 · (xs − xs,0 ) (58)

Transformation formulas (57) and (58) can be verified by showing that

they transform the probability density function of the multivariate standard
normal distribution (36) into the probability density function of the general
multivariate normal distribution (23)

Exercise. Apply (39) to verify transformation formulas (57) and (58).

In the transformation (57) and (58), no specific assumption about the form
of the matrix A is included. Therefore, not only a Cholesky decomposition,
which yields a triangular matrix, can be applied, but also an eigenvalue
decomposition. As C is positive definite, we obtain:
1 1
C = VΛVT = VΛ 2 Λ 2 VT , (59)
1
where V is the matrix of eigenvectors and where Λ 2 is the diagonal matrix
of the roots of the eigenvalues. A comparison with (56) leads to:
1
A = VΛ 2 (60)

An eigenvalue decomposition is to be preferred if the number of statistical

parameters is not too large because of the numerical ill-conditioning of C
due to highly correlated parameters.

3.9 Global and Local Parameter Tolerances

The modeling of the statistical variations of an integrated circuit manufac-
turing process is a complex task. On circuit and architecture level, two types
of parameter tolerances are assumed, global and local tolerances.
Global parameter tolerances are due to chip-to-chip (“inter-chip”) and wafer-
to-wafer fluctuations of the manufacturing process. They are modeled as para-
meter variations that affect all transistors of a circuit in the same way, hence
the notation global. Typical examples of statistical parameters with global
tolerances are oxide thickness or channel length reduction.
Besides global parameter tolerances, there are local parameter tolerances
that are due to variations within a chip (“intra-chip”) and that affect transistors
individually.
Parameters & Tolerances, Performance & Specification 45

Local variations on circuit level are usually assumed as statistically inde-

pendent variations of individual transistors. Local variations decrease with
increasing gate area and increasing distance between transistors [95]. They
lead to different behaviors of transistors in a circuit, a so-called mismatch. But
analog circuits are formed of transistor pairs like current mirrors requiring an
identical behavior of the transistors in a pair and are therefore very sensitive to
mismatch. Local variations have always been an issue in analog design. But
as the device dimensions become smaller and smaller, the significance of local
variations compared to global variations grows critical for digital circuits as
well. The statistical variation of critical paths’ delays for instance are strongly
depending on the local and global process variations and their effect on the
correlation of gate delays.
A typical example of statistical parameters with local tolerances are the tran-
sistor threshold voltages. Local variations lead to an increase in the number of
statistical parameters that corresponds to the number of transistors in a circuit.
In a circuit with 100 transistors for instance there will be one global thresh-
old voltage Vth,glob that affects all transistors equally and 100 local threshold
voltages Vth,loc,i , i = 1, . . . , 100 that affect each transistor individually and
independently from the others.
According to the assumptions that global statistical parameters are not corre-
lated with local statistical parameters and that local statistical parameters are not
correlated with each other, the covariance matrix has the following structure:
   
xglob Cglob 0
xs = , C= , Σloc = diag(σloc,1 . . . σloc,nloc ) (61)
xloc 0 Σloc

Note that the actual transistor parameter at the interface to the circuit simulator
may be a sum of a deterministic design parameter component xd,k , of a global
statistical parameter component xs,glob,l , and of a local statistical parameter
component xs,loc,m .

3.10 Performance Features

Corresponding to the definition of parameters, performance features shall be
used specifically for those quantities that are output of the numerical simulation.

Definition: Performance Features are variable quantities of an analog

system or an analog component that are output quantities of numerical simula-
tion. The elements fi of the performance vector f = [f1 . . . fi . . . fnf ]T ∈ Rnf
denote the values of the respective performance features, whose names are
addressed through the index i.

Examples of performance features of an operational amplifier are gain, phase

margin, slew rate, power supply rejection ratio.
46 ANALOG DESIGN CENTERING AND SIZING

Performance features are determined by the parameters. For a certain para-

meter vector, a certain vector of performance features is obtained:

x → f (62)

As mentioned in Note 2 in Section 3.7, we do not explicitly denote the function

that maps a parameter vector onto a performance vector: ϕ : x → f . Instead,
f (x) is meant to denote “f evaluated at x,” which implies that such a function
exists.
The performance features f for given parameters x are usually evaluated by
means of numerical simulation.

3.11 Numerical Simulation

It is important to note that in general the circuit behavior is not formulated
explicitly as in the RC circuit in Chapter 2, but is described by nonlinear dif-
ferential algebraic equations (DAEs),
 
ϕ t, x, un (x, t), iZ (x, t), q̇(un , iZ , x, t), φ̇(un , iZ , x, t), = 0, (63)

where t denotes the time, x the parameters, un the node voltages, iZ the currents
in Z-branches, q the capacitor charges, φ the inductor fluxes, and ˙ the derivative
with respect to time. Performance evaluation in general is a map of a certain
parameter vector x onto the resulting performance feature vector f , which is
carried out in two stages. First, numerical simulation [97, 75, 121, 101, 89]
by means of numerical integration methods maps a parameter vector onto the
node voltages and Z-branch currents:

x → un , iZ (64)

In case of a DC simulation, the resulting node voltages and Z-branch currents

are constant values. In case of an AC simulation, they are a function of fre-
quency, and in case of a transient (TR) simulation, they are a function of time.
These functions are given in the original sense of a relation as a finite set of
(x, [uTn iTZ ]T )-elements. In a second post-processing step, the node voltages
and Z-branch currents are mapped onto the actual performance features by
problem-specific operations:

un , iZ → f (65)

Depending on the type of simulation, we obtain corresponding DC performance

features like for instance DC power consumption or transistor operating volt-
ages, AC performance feature like gain or phase margin, and TR performance
features like slew rate.
Parameters & Tolerances, Performance & Specification 47

Numerical simulation is computationally extremely expensive, a single sim-

ulation can even take days. Contrary to our simple illustrative example in
Chapter 2, general analog performance functions

can only be established element-wise (x, f ) for selected parameter vectors,

and the calculation of each function element costs minutes to hours CPU
time.

Algorithms for analog optimization must therefore very carefully spend numer-
ical simulations, just as oil exploration has to carefully spend the expensive test
drillings.
In order to avoid the high computational cost of numerical simulation, per-
formance models or macromodels for a cheaper function evaluation would be
welcomed. Performance models can be obtained through numerical methods
called response surface modeling [31, 39, 3, 76, 37, 124, 16, 125, 127, 43, 64,
78, 17, 4, 126, 14, 29] or through analytical methods like symbolic analysis.
Macromodels can be obtained by simplified DAEs [22], by symbolic analysis
[108, 112, 113, 30, 107, 106, 32, 117, 67, 120, 119, 44, 18, 114, 15, 51, 50, 53,
52], by equivalent circuits [49, 20], or by behavioral models [128, 54, 65, 71,
72, 82]. However, the generation of such models requires a large number of
numerical simulations. Only if the resulting models are applied often enough
the setup cost are worthwhile. We are not aware of investigations of the break-
even where pre-optimization modeling undercuts the cost of optimization on
direct numerical simulation.

3.12 Performance-Specification Features

The performance features are subject to constraints that have to be satisfied.
These constraints are called performance-specification features and are derived
from the system performance that has been contracted with the customer. The
verification or validation of a circuit with regard to the performance specification
may be required at different stages of the design and test process, for instance
after structural synthesis prior to the layout design phase, or after production
test prior to the delivery.

Definition. A performance-specification feature is an inequality of the form:

ϕP SF,µ (f ) ≥ 0 (66)

The performance specification consists of performance-specification features,

which determine the acceptable performance region:

ϕP SF,µ (f ) ≥ 0, µ = 1, . . . , nP SF (67)
48 ANALOG DESIGN CENTERING AND SIZING

If the values of the performance features satisfy all performance-specification

features, i.e. are inside the acceptable performance region, then a circuit is
classified as “in full working order.”

The prevalent form of performance-specification features are upper boundary

values fU or lower boundary values fL that the performance may not exceed or
fall below. In this case, a performance-specification feature has the following
form,

fi − fL,i ≥ 0 or fU,i − fi ≥ 0 (68)

and the performance specification is:

fi − fL,i ≥ 0 ∧ fU,i − fi ≥ 0, i = 1, . . . , nf (69)

The set of all boundary values determines a performance acceptance region Af :

Af = {f | fL ≤ f ≤ fU }
= {f | fi ≥ fL,i ∧ fi ≤ fU,i , i = 1, . . . , nf } (70)

The vector inequality is defined as in (12). For each performance feature fi ,

either a lower bound fU,i or an upper bound fL,i or both a lower and an upper
bound may be specified. If a bound is not specified, the value of the corre-
sponding bound in (70) would be fL,i → −∞ or fu,i → ∞.
(70) has the same form as (8) and the acceptable performance region is a
box region as illustrated in Figure 19(a). Performance-specification features
in general may acquire the forms given in (8)–(11) and the corresponding box,
polytope, ellipsoid or general nonlinear geometries, or combinations of these
forms.
Chapter 4

ANALOG SIZING TASKS

4.1 Sensitivity-Based Analysis

Continuous optimization of analog circuits assumes a certain “smoothness”
of the optimization objectives in the parameters, i.e. continuously differentiable
objective functions with a certain degree of continuous differentiability [56].
For such functions, the concept of sensitivity can be applied. Basically, the
sensitivity is described by the first-order derivatives of objectives, for instance
the performance features f , with respect to parameters x:


∂f


S = = ∇f (xT ) = J, S ∈ Rnf ×nx
∂xT
x
⎡ ∂f ⎤

1 ∂f1
. . . ∂x∂f1

∂x1 ∂x2

⎢ ∂f2 nx
⎥

⎢ ∂f2 ∂f2 ⎥

. . . ∂xn ⎥

⎢ ∂x ∂x2
S = ⎢ .1 . .. .. ⎥
x


⎢ .. .. . . ⎥

⎣ ⎦

∂fnf ∂fnf ∂fnf

∂x ∂x . . . ∂x


1 2 nx x
⎡ ⎤
∇f1 (x1 ) ∇f1 (x2 ) ... ∇f1 (xnx )
⎢ ∇f (x ) ∇f2 (x2 ) . . . ∇f2 (xnx ) ⎥
⎢ 2 1 ⎥
= ⎢ .. .. .. .. ⎥
⎣ . . . . ⎦
∇fnf (x1 ) ∇fnf (x2 ) . . . ∇fnf (xnx )
⎡ ⎤
∇f1 (x)T
⎢ ∇f2 (x)T ⎥  
⎢ ⎥
= ⎢ .. ⎥ = ∇f (x1 ) ∇f (x2 ) . . . ∇f (xnx ) (71)
⎣ . ⎦
∇fnf (x)T
50 ANALOG DESIGN CENTERING AND SIZING

Figure 23. Sensitivity matrix in the parameter and performance space.

The matrix in (71) is called sensitivity matrix S, gradient ∇f (xT ), or Jacobian

matrix J. The i, kth element of S denotes the partial derivative

of the perfor-
∂fi

mance feature fi with respect to the parameter xk : ∂xk
= ∇fi (xk ).1 The ith
x
row ∇fi (x)T denotes the gradient of performance feature fi with respect to the
parameters x, the kth column ∇f (xk ) denotes the gradient of the performance
f with respect to parameter xk .
The sensitivity matrix represents the first-order change of the performance
feature values ∆f in dependence of a change in the parameters ∆x:


nx
∆f = S · ∆x = ∇f (xk ) · ∆xk (72)
k=1

The sensitivity matrix allows to separate the effects of a single-parameter

alteration or a multiple-parameter alteration on a single or multiple perfor-
mance features and is therefore of eminent significance for the analysis and
optimization of analog circuits. Figure 23 illustrates the sensitivity matrix for
an example with two parameters and two performance features.
On the left side, the directions of steepest ascent for each of the two perfor-
mance features are shown. On the right side, the effects of alterations in either
of the two parameters on the performance feature values are shown.
In the following, some aspects of a sensitivity-based analysis are formulated.

1 Please note that the Nabla symbol does not denote the Nabla operator here. In the abbreviated notation of
the partial derivative with the Nabla sign the parameters in brackets denote the parameters with regard to
which the partial derivative is calculated. Other parameters, which are held constant at certain values, are
not mentioned explicitly in this notation, but they determine the result value as well.
Analog Sizing Tasks 51

4.1.1 Similarity of Performance Features

If the two gradients ∇fi (x) and ∇fj (x) are orthogonal to each other, i.e.
linearly independent, an altered parameter vector in the direction of one of the
gradients will only affect one performance feature, while the other one remains
unchanged. Linear independence of the two gradients ∇fi (x) and ∇fj (x)
hence means that the two corresponding performance features can be tuned
independently from each other. We could call them minimum similar.
If on the other hand, the two gradients ∇fi (x) and ∇fj (x) are parallel to each
other, i.e. linearly dependent, changes in the two corresponding performance
features are inseparable from each other to the first order. We could call the
two performance features maximum similar.
Similarity of two performance features fi , fj is a measure of the similarity of
their reaction on alterations in the parameter vector. It can be described through
the angle φ(fi , fj ) between their gradients with respect to the parameters:
∇fi (x)T · ∇fj (x)
cos φ(fi , fj ) = , −1 ≤ cos φ(fi , fj ) ≤ +1 (73)
∇fi (x) · ∇fj (x)
The smaller the angle between the gradients of the two performance features
with respect to all parameters is, the more similar is the effect of parameter
alteration on them. A similarity of 0 corresponds to linear independence and
independent adjustment through parameter alteration. A similarity of ±1 corre-
sponds to linear dependence and inseparable reaction on parameter alterations.
The concept of similarity corresponds to the statistical concept of correlation.

4.1.2 Similarity of Parameters

The similarity of two parameters xk , xl is described by the cosine of the
angle between the gradients of the performance vector with respect to these
two parameters:
∇f (xk )T · ∇f (xl )
cos φ(xk , xl ) = , −1 ≤ cos φ(fi , fj ) ≤ +1 (74)
∇f (xk ) · ∇f (xl )
The smaller the angle between these gradients is, the more similar are the two
corresponding parameters and the more similar is their effect on all performance
features.

4.1.3 Significance of Parameters

The significance of a parameter xk corresponds to the amount of change in
the values of all performance features, ∆f (∆xk ), that it effects:
∆f (∆xk ) = S · ∆xk · ek = ∇f (xk ) · ∆xk (75)
ek is a vector with a 1 in the kth position and zeros in the other positions.
52 ANALOG DESIGN CENTERING AND SIZING

If we assume the same alteration ∆x for all parameters, then a ranking of the
parameters according to their significance is obtained by comparing the lengths
of the performance gradients with respect to the individual parameters:
xk more significant than xl ⇔ ∇f (xk ) > ∇f (xl ) (76)

4.1.4 Adjustability of Performance Features

Analogously, the adjustability of a performance feature fi is based on the
amount of change in its value that it shows for an alteration of parameter values:
∆fi (∆x) = ∇fi (x)T · ∆x (77)
By inspection of (77), we can see that the maximum amount of change in the
value of fi is obtained, if the parameters are altered according to the gradient
of the performance feature:
max ∆fi (∆x) → ∆xmax = ∇fi (x) , ∆fi,max = ∇fi (x)22 (78)
∆x

Aranking of the performance features according to their adjustability is obtained

by comparing the lengths of the gradients of the individual performance
features:
fi more adjustable than fj ⇔ ∇fi (x) > ∇fj (x) (79)

4.1.5 Multiple-Objective Behavior

The maximum increase ∆fi,max in the value of performance feature fi is
obtained if the parameters are changed according to the corresponding gradient
of this performance. The resulting change in all other performance features
results from inserting the corresponding gradient of this performance, ∇fi (x),
in (72):
∆f |∆fi,max = S · ∇fi (x) (80)
In this way, we can analyze how performance features change their values if
one them is optimized. This is a first step towards optimization of multiple
objectives.

4.1.6 Exercise
Given is the sensitivity matrix for two performance features f1 , f2 and two
parameters x1 , x2 at a certain point x0 (Figure 23):
 
1 2
S= (81)
−1 1
Analyze the design situation regarding similarity, significance, adjustability and
multiple-objective behavior.
Analog Sizing Tasks 53

4.1.7 Sensitivity-Based Analysis of Tolerance Objectives

Please note that the given formulas for analyzing multiple-performance/
multiple-parameter design tasks can be applied for any type of objective. In
nominal design, the nominal performance feature values are the objectives.
Tolerance design’s objectives are worst-case performance feature values and
worst-case distances.

4.2 Performance-Sensitivity Computation

A sensitivity-based analysis requires the availability of the gradients of the
respective objectives.
The performance-to-parameter sensitivities ∇f (xT ) are computed in two
ways, depending on the capabilities of the numerical simulator.

4.2.1 Simulator-Internal Computation

Some simulators provide the sensitivities ∇un (xT ), ∇iZ (xT ), from which
∇f (xT ) can be computed by applying the chain rule of differentiation:

∇f (xT ) = ∇f (uTn ) · ∇un (xT ) + ∇f (iTZ ) · ∇iZ (xT ) (82)

The formulas to compute ∇un (xT ) and ∇iZ (xT ) require the corresponding
parts in the transistor models. This increases the transistor modeling effort,
as not only the transistor equations but also their first derivatives have to be
provided.
The formulas to compute f (uTn ) and ∇f (iTZ ) are part of the postprocessing
operations that have to include the corresponding first derivatives.
The main effort is the a-priori calculation of the formulas for the first deriva-
tives. This is done once for each transistor class of the underlying production
technology and once for the considered circuit class.
The simulation overhead for sensitivity computation remains relatively small.
It amounts to about 10% of the CPU time of one simulation for the computation
of the sensitivity of all performance features with respect to one parameter.

4.2.2 Finite-Difference Approximation

Many simulators focus on providing an interface to apply user-defined tran-
sistor and device models. User-defined device models mostly do not include
the derivatives with respect to parameters. In this frequent case, a simulator-
internal sensitivity computation is not possible.
Instead, a sensitivity computation can be done by a finite-difference approxi-
mation of the gradient, which can be interpreted geometrically as a secant to the
performance function at the current parameter vector. Two simulations are per-
formed, one at the current parameter vector x, one at a parameter vector where
component xk is altered by a certain amount ∆xk . The gradient with respect to
54 ANALOG DESIGN CENTERING AND SIZING

parameter xk is approximately the quotient of the performance difference and

the parameter difference:
f (x + ∆xk · ek ) − f (x)
∇f (xk ) ≈ (83)
∆xk
The finite-difference approximation has no a-priori overhead, as no analyti-
cal first-order derivatives are calculated. But the simulation overhead is one
simulation, i.e. 100% of the CPU time of one simulation, for computing the
sensitivity of all performance features with respect to one parameter.
The choice of ∆xk is critical. It has to be large enough to surmount numerical
noise and small enough to compare to the gradient.

4.3 Scaling Parameters and Performance Features

In a multiple-parameter/multiple-objective optimization problem like ana-
log circuit design, an appropriate scaling of the parameters and performance
features is crucial for the effectiveness of the solution algorithms. Missing or
improper scaling leads to a deterioration in the numerical problem condition
and probable failure of the solution methods. Scaling means a transformation
of parameters, performance features or other design variables:

x → x , f → f  (84)

4.3.1 Scaling to a Reference Point

Huge differences in orders of magnitudes of physical variables like capacitors
in 10−12 F and frequencies in 106 Hz have to be eliminated. This suggests to
scale a variable with respect to a reference point:
xk fi
xk = , k = 1, . . . , nx , fi = , i = 1, . . . , nf (85)
xRP,i fRP,i

Examples for reference points are lower bounds, upper bounds or initial values:

xRP ∈ {xL , xU , x0 , . . .} , fRP ∈ {fL , fU , f0 , . . .} (86)

4.3.2 Scaling to the Covered Range of Values

Another approach is to scale the design variables according to the range of
values they cover during the design [56]:
xk − xRP,L,k
xk = , i = 1, . . . , nx
xRP,U,k − xRP,L,k
fi − fRP,L,i
fi = , i = 1, . . . , nf (87)
fRP,U,i − fRP,L,i
Analog Sizing Tasks 55

The scaled variables according to (87) will be in the interval between 0 and
1: xk , fi ∈ [0, 1]. The values xRP,L,k , xRP,U,k , fRP,L,i , fRP,U,i have to be
chosen very carefully. The performance specification (70) or tolerances range
of parameters (8) may be helpful at this point. If only an upper or a lower
bound of a variable is available, the proper choice of the opposite bound is very
difficult. Initial guesses are used in these cases, which have to be tightened or
relaxed according to the progress of the optimization process.

4.3.3 Scaling by Affine Transformation

A more general approach is to not only scale variables separately, but to apply
affine transformations. An example for such a transformation is to scale the
statistical parameters according to (56)–(58):

xs = A−1 · (xs − xs,0 ) , C = A · AT (88)

Statistical parameters scaled according to (88) are independent from each other
and have most probably values between −3 and 3. The resulting covariance
matrix is the unity matrix and has the best condition number, i.e. 1.

4.3.4 Scaling by Equalization of Sensitivities

An extended approach to the scaling of variables is the inclusion of sensitiv-
ity. The linear model (72) can be interpreted as a first-order approach to map
a designated change in the performance feature values ∆f onto the required
parameter changes ∆x:

∆f , S · ∆x = ∆f → ∆x (89)

The solution of (89) requires the solution of a system of linear equations that
may be rectangular and rank-deficient. The computational accuracy of the
solution is determined by the condition number of the system matrix, cond(S) =
S−1 ·S. Gradients that differ very much cause an ill-conditioned sensitivity
matrix. An improvement is to equalize the sensitivities, which in turn leads to
a corresponding scaling of performance features and parameters.
There are two approaches to equalize the sensitivity matrix S.
The first approach is to equalize the rows of S to an equal length of 1:

∆fi ∇fi (x)T

= · ∆x (90)
∇fi (x) ∇fi (x)
     

∆fi = ∇fi (x)T · ∆x, ∇fi (x) = 1

(90) shows that normalization of row i of the sensitivity matrix to length 1

corresponds to scaling the performance feature fi by the norm of row i.
56 ANALOG DESIGN CENTERING AND SIZING

The second approach is to equalize the columns of S to an equal length of 1:

 ∇f (xk )
∆f = · ∆xk · ∇f (xk ) (91)
∇f (xk )   
k   

= ∇f (xk ) · ∆xk
k

(91) shows that normalization of column k of the sensitivity matrix to length 1

corresponds to scaling parameter xk by the reciprocal norm of the column k.
An equalization of the sensitivity matrix is an additional measure to the initial
scaling of performance features and parameters. The resulting improvement in
the numerical condition can be interpreted as follows: Along with the equaliza-
tion of the sensitivity values goes an equalization of the visibility of parameters
and performance features within an optimization process. Optimization direc-
tions will be considered that otherwise would have been obscured by dominating
numerical entries in the sensitivity matrix.
The formulas in this book are given for unscaled variables for a better
illustration of the technical tasks. Scaled variables have to be used in an
implementation.

4.4 Nominal Design

Nominal design naturally optimizes the performance features without consid-
ering parameter tolerances. In the following, the multiple-objective approach
and the single-objective approach to performance optimization will be
described.
Note that the description holds as well for tolerance design objectives. For
that reason, the description is not put into subsections of this section, but into
own sections.

4.5 Multiple-Objective Optimization

While a unique solution has been calculated explicitly for the RC circuit in
Chapter 2, nominal design in general involves a multiple-objective optimization
problem or multiple-criteria optimization problem respectively [35, 34, 41, 73,
66, 86], which can be formulated as:

min f (xd ) subject to c(xd ) ≥ 0 (92)

xd
In (92), the vector inequality is defined as in (12). (92) is formulated as a
minimization task without loss of generality. Performance features that have
to be maximized are included by max f = − min −f .
Analog Sizing Tasks 57

The constraints c(xd ) ≥ 0 basically describe technological and structural

requirements concerning DC properties of transistors that have to be fulfilled
for a proper function and robustness [58]. Usually these constraints determine
the achievable performance feature values of a given circuit structure.
Nominal design according to (92) is concerned with tuning the design para-
meters xd in order to obtain an optimal performance. Doing so, it assumes that
the statistical parameters xd and the range parameters xr have certain values.
These values can be selected such that they represent worst-case conditions in
statistical and range parameters.
If in addition, a performance specification according to (70) is given, then
the multiple-objective minimization problem is formulated as:

fII,L ≤ fII (xd ) ≤ fII,U
min fI (xd ) subject to (93)
xd c(xd ) ≥ 0

fI , fII denote subsets of performance features. This corresponds to the practical

situation, where a subset of performance features are optimization objectives
while others are kept as constraints concerning their performance-specification
features. This happens for instance if two or three performance features shall
be compared visually or if computational costs shall be kept low.
The simultaneous minimization of several performance features as in (92)
and (93) requires a compromise, i.e. a “trade-off,” between competing opti-
mization objectives. Typical examples of such trade-off situations are speed
vs. power or gain vs. bandwidth. The optimality of one performance feature in
a multiple-objective optimization problem can only be evaluated in connection
with other performance features. This leads to the concept of Pareto optimality.

Definition. A performance feature is called Pareto optimal if it can only be

improved at the prize of deteriorating another performance feature.

4.5.1 Smaller or Greater Performance Vectors

Pareto optimality is based on a relation of “less than” and “greater than”
between performance vectors:

f > f ∗ ⇔ f ≥ f ∗ ∧ f = f ∗ ⇔ ∀i f i ≥ f ∗i ∧ ∃i f
i = fi∗ (94)

f < f ∗ ⇔ f ≤ f ∗ ∧ f = f ∗ ⇔ ∀i f i ≤ f ∗i ∧ ∃i f
i = fi∗ (95)

(94) and (95) define that a vector is less (greater) than another vector if all its
components are less (greater) or equal and if it differs from the other vector.
58 ANALOG DESIGN CENTERING AND SIZING

Figure 24. (a) Set M > (f ∗ ) of all performance vectors that are greater than f ∗ , i.e. inferior to
f ∗ with regard to multiple-criteria minimization according to (92) and (93). (b) Set M < (f ∗ ) of
all performance vectors that are less than f ∗ , i.e. superior to f ∗ with regard to multiple-criteria
minimization.

This equivalently means that a vector is less (greater) than another vector if at
least one of its components is less (greater) than the corresponding component
of the other vector. In Figure 24(a), the shaded area indicates all performance
vectors that are greater than a given performance vector f ∗ for a two-dimensional
performance space.
These form the set M > (f ∗ ):

M > (f ∗ ) = { f | f > f ∗ } (96)

The solid line at the border of the shaded region is meant as a part of M > (f ∗ ).
The small kink of the solid line at f ∗ indicates that f ∗ is not part of M > (f ∗ )
according to the definition in (94). The solid line represents those performance
vectors for which one performance feature is equal to the corresponding com-
ponent of f ∗ and the other one is greater. In the shaded area in Figure 24(a),
both performance features have values greater than the corresponding star val-
ues. With respect to a multiple-objective minimization problem, performance
vectors in M > (f ∗ ) are inferior to f ∗ . Figure 24(b) illustrates the analogous
situation for all performance vectors that are less than, i.e. superior to, a given
performance vector f ∗ for a two-dimensional performance space. These form
the set M < (f ∗ ):
M < (f ∗ ) = { f | f < f ∗ } (97)
From visual inspection of Figure 24(b), we can see that the border of the set of
superior performance vectors M < (f ∗ ) to f ∗ corresponds to the level contour of
Analog Sizing Tasks 59

the l∞ -norm of a vector, which is defined as the max operation on the absolute
values of its components. The set of superior performance vectors M < (f ∗ )
to f ∗ can therefore be formulated based on the max-operation or the l∞ -norm,
if the performance vectors are assumed to be scaled such that they only have
n
positive values, f ∈ R+f :




nf < ∗
∗
f ∈ R+ : M (f ) = f
max (fi /fi ) ≤ 1 ∧ f = f ∗
(98)
i

= {f |  [ . . . fi /fi∗ . . . ] ∞ ≤ 1 ∧ f = f ∗ } (99)
Based on (97)–(99), Pareto optimality is defined by excluding the existence of
any superior performance vector to a Pareto-optimal performance vector:
f ∗ is Pareto optimal ⇔ M < (f ∗ ) = { } (100)

4.5.2 Pareto Point

From (92) and (98)–(100) follows the formulation of a Pareto-optimal point
(Pareto-efficient point, Pareto point) fI∗ as the performance vector with the
minimum value of its maximum component among all feasible performance
vectors:
f ∗ (w) ← min max wi ·(fi (xd ) − fRP,i ) s.t. c(xd ) ≥ 0 (101)
xd
i  
f (xd )−fRP ∞,w

wi ≥ 0 , wi = 1 , fi − fRP,i ≥ 0
i

(101) shows that the min-max operation, or the l∞ -norm respectively, is an

inherent part of the definition of a Pareto-optimal point.
If the reference point fRP = [ fRP,1 . . . fRP,nf ]T is chosen such that the
weights wi refer to shifted performance feature values fi −fRP,i that are always
positive, a Pareto point f ∗ can be addressed from fRP through a unique weight
vector w = [ w1 . . . wnf ]T .
The individual minima of the performance features can be used to define the
reference point:
fRP,i ≡ min fi (xd ) s.t. c(xd ) ≥ 0 (102)
xd

4.5.3 Pareto Front

In a multiple-objective optimization problem exists a set of Pareto points,
which is called Pareto-optimal front or simply Pareto front P F . Figure 25(a)
illustrates a Pareto front for two performance features. The gray region denotes
60 ANALOG DESIGN CENTERING AND SIZING

Figure 25. (a) Pareto front P F (f1 , f2 ) of a feasible performance region of two performance
features. Different Pareto points addressed through different weight vectors from reference point
fRP , which is determined by the individual minima of the performance features. (b) Pareto front
P F (f1 , f2 , f3 ) of three performance features with boundary B = {P F (f1 , f2 ), P F (f1 , f3 ),
P F (f2 , f3 )}.

the set of all performance vectors that are achievable under consideration of the
constraints of the optimization problem. The solid line at the left lower border
of the region of achievable performance values represents the set of Pareto
points, i.e. the Pareto front. Each point on the front is characterized in that
no point superior to it exists, i.e. M < (f ∗ ) is empty. Three Pareto points are
illustrated and the corresponding empty sets M < (f ∗ ) are indicated. We can as
well see that each Pareto point represents the minimum weighted l∞ -norm with
respect to the reference point fRP with components according to (102) that is
achievable for a given weight vector w. Depending on the weight vector w, a
specific Pareto point on a ray from the reference point fRP in the direction w
is determined. The Pareto front P F and its boundary B can be defined as:
P F (f ) = {f ∗ (w) | wi ≥ 0 , i wi = 1 }
 (103)
B(P F (f1 , . . . , fnf )) = i P F (f1 , . . . , fi−1 , fi+1 . . . fnf )
A Pareto front for three performance features and its boundary is shown in
Figure 25(b). Starting from the individual minima, the Pareto front can be
formulated hierarchically by adding single performance features according
to (103).
Pareto fronts may exhibit different shapes and may be discontinuous. Figure
26 shows three example shapes of Pareto fronts in relation to the convexity
Analog Sizing Tasks 61

∗
Figure 26. (a) Continuous Pareto front of a convex feasible performance region. fI,2 is mono-
∗ ∗
tone in fI,1 . (b) Continuous Pareto front of a nonconvex feasible performance region. fI,2 is
∗
monotone in fI,1 . (c) Discontinuous Pareto front of a nonconvex feasible performance region.
∗ ∗
fI,2 is nonmonotone in fI,1 .

of the feasible performance region and the monotonicity of the explicit Pareto
front function of one performance feature.

4.5.4 Pareto Optimization

The task of computing the Pareto front is called Pareto optimization. Usually
the Pareto front is discretized by a number of Pareto points. The art of Pareto op-
timization is an efficient method to compute Pareto points that are evenly spread
on the Pareto front. The computation of a Pareto point involves the solution
of a single-objective optimization problem. Evolutionary algorithms as well
as the deterministic Normal-Boundary-Intersection [34] and Goal-Attainment
[48] approaches are suitable for Pareto optimization. A weighted sum approach
with level contours (Figure 27(a)) that will end as tangents to the Pareto front
is suitable for convex feasible performance regions only.

4.6 Single-Objective Optimization

We can distinguish two cases of single-objective optimization in analog
design:

Optimization of a selected single design target. E.g.:

– As mentioned in the previous section, the individual minima of the

performance features are very important in multiple-objective optimiza-
tion. The individual minima are determined by optimizing a single per-
formance feature while the other performance features are constraints
or ignored.
62 ANALOG DESIGN CENTERING AND SIZING

– The yield, i.e. the probability of satisfying the performance specifica-

tion under given statistical parameter distributions and range-parameter
tolerances, is an important single objective that has to be maximized.
Optimization of a dedicated scalar function of several design targets. E.g.:
– Multiple-objective optimization problems as in (92) and (93) require the
solution of single-objective optimization problems as in (101) in order
to calculate the Pareto front.
– A set of performance features shall be optimized to meet the given per-
formance specification with as much safety margin as possible. This is
referred to as performance centering. Many kinds of scalar functions
of several objectives are possible. However they introduce additional
nonlinearities, as for instance a sum of exponential functions, and may
deteriorate the numerical problem condition. The combination of multi-
ple objectives into one scalar objective has to be defined very carefully.
A single-objective solution [19, 36, 91, 80, 45, 56, 61, 79] is approached with
deterministic [115, 37, 81, 77, 28, 116, 83, 42, 105, 104, 16, 103, 27, 64, 62, 9,
92, 13, 12, 25, 23, 59, 7, 40, 74, 60, 2, 21] or statistical methods [33, 99, 26, 3,
46, 100, 110, 118, 96, 90, 93, 94, 102, 53, 111, 68, 69].
A practical problem formulation of single-objective analog optimization is:

fL ≤ f (xd ) ≤ fU
min fI (xd ) − fI,target  subject to (104)
xd c(xd ) ≥ 0

As mentioned, the constraints c(xd ) ≥ 0 basically describe technological and

structural requirements concerning DC properties of transistors that have to
be fulfilled for a proper function and robustness [58]. The constraints fL ≤
f (xd ) ≤ fU correspond to the performance specification. And a subset of
performance features, fI , are selected as part of the optimization objective,
which is a vector norm . of their distances to target values, fI,target .
Single-objective analog optimization could of course be formulated as a
minimum-norm problem without performance target values. But this would
make the solution unnecessarily difficult. Because of the physical nature of the
objectives, they cannot reach arbitrary values. Even achieving the individual
minima of each performance feature simultaneously for all performance fea-
tures is an utopian goal. Performance target values represent realistic goals for
optimization that consider the practical conditions. They can be determined
using input data and circuit knowledge and should be used in the sense of an
appropriate scaling of performance features as described in Section 4.3 to ease
the solution.
A concrete problem formulation now has to select the vector norm to be
applied and find appropriate performance target values.
Analog Sizing Tasks 63

Figure 27. (a) Level contours of the weighted l1 -norm. (b) Level contours of the weighted
l2 -norm. (c) Level contours of the weighted l∞ -norm.

4.6.1 Vector Norms

Several vector norms . can be applied in (104):
weighted l1 -norm

h1,w = |w|T · |h| = |wi | · |hi | (105)
i

weighted l2 -norm
√ !
W=diag(wi )
h2,w = hT ·W·h = wi2 · h2i (106)
i

weighted l∞ -norm
h∞,w = max |wi | · |hi | (107)
i

Figure 27 illustrates the level contours of these three weighted norms.

The corners in their level contours illustrate that the l1 -norm and l∞ -norm are
differentiable only with restrictions. The l2 -norm being differentiable makes it
a favored norm for gradient-based numerical optimization.
Like the selected norm in (104), the selection of the performance target values
fI,target,i , i = 1, . . . , nf I is very important.

4.6.2 Performance Targets

In the case of performance centering, the performance target values are
defined using the performance-feature bounds. If a lower bound and an upper
bound are specified for a performance feature at the same time, the target value
64 ANALOG DESIGN CENTERING AND SIZING

is the center of the resulting interval:

1
∞ < fI,L,i ≤ fI,i ≤ fI,U,i < ∞ : (fI,L,i + fI,U,i ) (108)
fI,target,i =
2
This target value provides a maximum performance margin with regard to the
respective performance-specification features.
If only an upper or only a lower bound is specified for a performance fea-
ture, a target value of this performance feature is hard to determine. It has to
be determined based on experience and utilizing the sensitivity matrix. The
performance target vector fI,target should be updated during the optimization
process.

4.7 Worst-Case Analysis and Optimization

In nominal design, the statistical and range parameters are considered at
selected specific values. On the other hand, tolerance design considers a whole
tolerance range of their values.
The task of worst-case analysis is to compute the worst-case performance
that appears if the statistical parameters and the range parameters can take any
value within their given tolerance regions.
Worst-case optimization hence is the task of minimizing the worst-case
deviation of the performance from its nominal behavior.
Intuitively, the deviation of the worst-case from the nominal performance
is proportional to the performance sensitivity. Then, worst-case optimization
means a minimization of the performance sensitivity while taking care that the
nominal performance does not drift away.

4.7.1 Worst-Case Analysis

Given a tolerance region of the statistical parameters Ts , a tolerance region of
the range parameters Tr , and a nominal parameter vector for design parameters,
statistical parameters and range parameters x = [ xTd xTs,0 xTr ]T , the worst-
case analysis is formulated as:
fi ≥ fL,i : min fi (xs , xr ) s.t. xs ∈ Ts (YL ) , xr ∈ Tr (109)
xs ,xr

fi ≤ fU,i : max fi (xs , xr ) s.t. xs ∈ Ts (YL ) , xr ∈ Tr (110)

xs ,xr

i = 1, . . . , nf
If a lower bound fL,i is specified or required for a performance feature fi , the
worst-case represents the maximal deviation of the performance feature from
its nominal value in negative direction. This value is obtained by computing
the minimum value of the performance feature that occurs over the statistical
and range parameters xs , xr within their given tolerance regions Ts , Tr .
Analog Sizing Tasks 65

Vice versa, if an upper performance-feature bound fU,i is specified or

required, the worst-case will be the maximal deviation of the performance
feature from its nominal value in positive direction.
Even if no performance-feature bounds are given, it is often known if a worst-
case of a performance feature happens in the direction of larger or of smaller
values. The worst-case power for instance will be greater than the nominal
power consumption, and a worst-case clock frequency will be less than the
nominal clock frequency.
We may not be interested in a worst-case value for each performance feature.
And if we are interested in the worst-case, then either a lower worst-case value
or an upper worst-case value, or both a lower and upper worst-case value may
be required, like the earliest and latest arrival times of a signal.
It follows that the number of worst-case performance values nW C that we
have to compute is
1 ≤ nW C ≤ 2 · n f (111)
The tolerance region of the range parameters Tr is given as an input and can
have the forms in (8)-(11).
The tolerance region of the statistical parameters Ts depends on the chosen
type of tolerance class and a required minimum yield YL . As the relation
between yield and a tolerance region is ambiguous in the multivariate case, an
a-priori determination of a specific tolerance region Ts for a given yield value
YL has to be done heuristically (Section 6.2).
Note that the worst-case analysis (109), (110) in general requires the solution
of an optimization problem.
The solution of the worst-case analysis problem (109), (110) yields worst-
case parameter vectors xW L,i and xW U,i as well as the corresponding worst-case
performance feature values fW L,i and fW U,i .
The solution of (109) or (110) therefore implements the following mapping:

Worst-case parameter vectors

xW L,i = [ xTd xTs,W L,i xTr,W L,i ]T
xW U,i = [ xTd xTs,W U,i xTr,W U,i ]T
Tolerance region Ts (YL ) i = 1, . . . , nf
Tolerance region Tr Worst-case performance feature values
→
Nominal parameter vector fW L,i = fi (xW L,i )
x = [ xTd xTs,0 xTr ]T fW L,i = f (xW L,i )
fW U,i = fi (xW U,i )
fW U,i = f (xW U,i )
i = 1, . . . , nf
(112)
66 ANALOG DESIGN CENTERING AND SIZING

Figure 28. Input and output of a worst-case analysis.

In each worst-case parameter vector xs,W L,i , xs,W U,i , where the performance
feature fi has its worst-case value fW L,i , fW U,i , the values of the other perfor-
mance features add to the worst-case performance vectors f (xW L,i ), f (xW U,i ).
A worst-case performance vector f (xW L,i ), f (xW U,i ) hence describes the situ-
ation of all performance features, if performance feature fi is at its individual
lower or upper worst-case. Figure 28 shows a graphical overview of the input
and output of this mapping.
Figure 29 illustrates the input and output of a worst-case analysis in the
spaces of design, statistical and range parameters and of performance fea-
tures. For graphical reasons, each of these spaces is two-dimensional in this
example. Figure 29(a) shows the input situation with nominal values for design
parameters xd , statistical parameters xs,0 , and range parameters xr , and with
tolerance regions of the range parameters Tr and of the statistical parameters Ts .
The tolerance regions have the typical shape of a box for the range parameters
according to (8) and of an ellipsoid for the statistical parameters according to
(10), (23) and (24). The tolerance region of statistical parameters has been
determined to meet a certain yield requirement YL according to Section 6.2.
Note that the design parameters do not change their values during worst-case
analysis. A worst-case analysis computes lower and upper worst-case values
for each of the two performance features. The results are illustrated in Figure
29(b) and consist of:
Analog Sizing Tasks 67

Figure 29. (a) Input of a worst-case analysis in the parameter and performance space: nominal
parameter vector, tolerance regions. (b) Output of a worst-case analysis in the parameter and
performance space: worst-case parameter vectors, worst-case performance values.
68 ANALOG DESIGN CENTERING AND SIZING

four worst-case performance values fW L,1 , fW U,1 , fW L,2 and fW U,2 ,

four worst-case parameter vectors
xW L,i = [ xTd xTs,W L,i xTr,W L,i ]T , xW U,i = [ xTd xTs,W U,i xTr,W U,i ]T ,
i = 1, 2, which are given separately for the statistical parameters xs,W L,1 ,
xs,W U,1 , xs,W L,2 and xs,W U,2 , and the range parameters xr,W L,1 , xr,W U,1 ,
xr,W L,2 and xr,W U,2 , and
the overall performance vectors at these parameter vectors fW L,1 , fW U,1 ,
fW L,2 and fW U,2 .
Figure 29 illustrates the important characteristic of worst-case analysis that
for each lower or upper direction of each performance feature, an individual
worst-case parameter vector exists.
It may happen that worst-case parameter vectors are equal, illustrated by
xr,W L,2 and xr,W U,1 , or close to each other, illustrated by xs,W U,1 , xs,W L,2 in
Figure 29(b). An analysis of a clustering of worst-case parameter vectors for
certain classes of performance features or certain classes of analog circuits may
lead to a reduction in the required number of worst-case parameter vectors.
It may also happen that a worst-case parameter vector is not on the border
of a tolerance region but inside the tolerance region, as illustrated by xs,W L,1
in Figure 29(b). If a performance function is unimodal over the parameters,
than it has exactly one maximum or minimum. If this maximum or minimum is
inside the parameter tolerance region, the corresponding upper or lower worst-
case performance will be taken at the corresponding parameter values inside
the tolerance region. If this maximum or minimum is outside of the parameter
tolerance region, the maximum or minimum performance value will be on the
border of the tolerance region.
Worst-case parameter vectors are very popular in analog design in order
to check if a circuit satisfies the performance specification under tolerance
conditions. (109) or (110) and the explanations above show that the computation
of worst-case parameter vectors requires the cooperation of process technology
and circuit design. The availability of statistical parameter distributions alone is
not sufficient to compute worst-case parameter vectors. In addition, a concrete
circuit and concrete performance features have to be provided by analog design.

4.7.2 Worst-Case Optimization

The goal of worst-case optimization is the minimization of the worst-case
performance deviations from the nominal values, which have been computed
by a worst-case analysis:
⎧ ⎫ ⎧
⎨ |fi (xd ) − fW L,i (xd )| ⎪
⎪ ⎬ ⎨ fL,i ≤ fi /fW L/U,i (xd ) ≤ fU,i
⎪
min |fW U,i (xd ) − fi (xd )| s.t. i = 1, . . . , nf (113)
xd ⎪⎩ i = 1, . . . , n ⎪
⎭ ⎪
⎩ c(x ) ≥ 0
f d
Analog Sizing Tasks 69

Figure 30. Nested loops within worst-case optimization.

During computation of the worst-case performance values over the statistical

and range parameters, the design parameters have been kept at their nominal
values.
Worst-case optimization utilizes that the nominal performance values and
the worst-case performance values depend on the design parameters as well.
By sizing the design parameter values xd , the amount of lower and upper worst-
case deviations, |fi (xd ) − fW L,i (xd )| and |fW U,i (xd ) − fi (xd )|, is minimized.
At the same time, neither the worst-case nor the nominal performance values
should violate the performance specification. This constraint can be relaxed to
include only nominal performance values.
Worst-case optimization (113) is a multiple-objective optimization problem
that has to be transformed into a suitable single-objective optimization problem.
It targets worst-case performance values as optimization objectives, which in
turn have been computed by means of an optimization problem. This prob-
lem partitioning necessitates a repeated worst-case analysis according to (109)
or (110) within the iterative solution of the worst-case optimization process
according to (113). Therefore worst-case optimization applies worst-case ana-
lysis “in a loop.” It is thus a two-layered optimization process. Worst-case
analysis in turn applies sensitivity computation for gradient calculation as well
“in a loop.” And if sensitivities are computed by finite differences, a sensitivity
analysis in turn applies numerical simulation “in a loop.” Figure 30 illustrates
this situation of four nested loops.
Simplifications of this general approach result from the implementation of the
loops in which worst-case optimization calls worst-case analysis and in which
worst-case analysis calls sensitivity analysis. Simplifications are for instance:

Call a worst-case analysis once for each iteration step of worst-case

optimization.
70 ANALOG DESIGN CENTERING AND SIZING

Call a worst-case analysis in every µth iteration step of worst-case

optimization.
These are combined with:
Call a sensitivity analysis once for each iteration step of a worst-case
analysis.

Call a sensitivity analysis once in a worst-case analysis.

A comparison of the worst-case optimization problem according to (113) and
the nominal design problem according to (104) shows the similarity between
these two design tasks. In both tasks, certain performance values (nominal
and/or worst-case) are tuned within a given performance specification. As
nominal design in practice will consider at least an approximation of the worst-
case behavior based on a linear performance model, for instance based on
sensitivities, the transition between nominal design and worst-case optimization
is smooth. In order to save design time and computational cost, nominal design
can already include assumptions about the worst-case behavior.

4.8 Yield Analysis, Yield Optimization/Design Centering

The yield is the percentage of manufactured circuits that satisfy the per-
formance specification in face of statistical parameter variations and range-
parameter tolerances.
Yield analysis denotes the estimation of the yield, yield optimization/design
centering denotes the maximization of yield by tuning of design parameters.

4.8.1 Yield
The yield Y can be defined as the probability that a manufactured circuit
satisfies the performance specification under all operating conditions:

Y = prob{ ∀
xr ∈ Tr
fL ≤ f (xd , xs , xr ) ≤ fU } (114)

Depending on the perspective, the yield considered as a probability will be

between 0 and 1, and considered as a percentage between 0% and 100%:

0 ≤ Y ≤ 1 ≡ 0% ≤ Y ≤ 100% (115)

On the one hand, the yield is determined by the probability density function of
the statistical parameters (13), (14). If no performance specification was given,
then the yield would be 100% according to (15). Figure 31(a) shows a normal
probability density function with its ellipsoidal equidensity contours, which is
unaffected by any performance specification.
Analog Sizing Tasks 71

0.16
0.14 f(x,y)
0.15
0.12 0.1
0.05
0.1
0.08
0.06
0.04
0.02
0

-4
-3
-2
-1
0
1
2 3 4
3 1 2
-1 0
4 -4 -3 -2

Figure 31. (a) The volume under a probability density function of statistical parameters, which
has ellipsoid equidensity contours, corresponds to 100% yield. (b) The performance specification
defines the acceptance region Af , i.e. the region of performance values of circuits that are in
full working order. The yield is the portion of circuits in full working order. It is determined
by the volume under the probability density function truncated by the corresponding parameter
acceptance region As .

The volume under this probability density function is 1, which refers to 100%
yield. On the other hand, the yield is determined by the performance specifica-
tion (70), (67) and the range parameters’ tolerance region (8)–(11). These lead
to a yield loss because a certain percentage of the statistically varying para-
meters will violate the performance specification for some operating condition.
Figure 31(b) illustrates how the ellipsoidal equidensity contours of the proba-
bility density function of statistical parameters are truncated by the parameter
acceptance region As , which corresponds to the performance acceptance region
Af defined by the performance specification. Figure 17 earlier illustrated the
72 ANALOG DESIGN CENTERING AND SIZING

remaining part of a normal probability density function if all parameter values

that violate a performance specification are left out. The volume under this
truncated probability density function refers to the probability that a circuit will
satisfy the performance specification and is between 0 and 1, which refers to a
yield value between 0% and 100%.
The yield Y is defined by integrating the probability density function over
the acceptance region imposed by the performance specification. This can be
done either in the statistical parameter space,
 
Y = ... pdf(xs ) · dxs , dxs = dxs,1 · dxs,2 · . . . · dxs,ns (116)
xs ∈As

or in the performance space:

 
Y = ... pdff (f ) · df , df = df1 · df2 · . . . · dfnf (117)
f ∈Af

4.8.2 Acceptance Region Partitions

(116) and (117) illustrate that the yield formulation requires the formu-
lation of both the probability density function and the acceptance region in
either the performance space or the statistical parameter space. The acceptance
region however is defined in the performance space, and the probability den-
sity function is defined in the statistical parameter space. Yield formulation
(117) therefore requires an approximation of the probability density function
of the performance features pdff , whereas yield formulation (116) requires an
approximation of the acceptance region in the statistical parameter space As .
An approximation of pdff could be based on an expansion of the probability
density function in dependence of statistical moments and an estimation of the
statistical moments of the performance features.
An approximation of the acceptance region in the statistical parameter space
starts from the formulation of the performance acceptance region Af . Af
is partitioned into individual performance acceptance region partitions Af,L,i ,
Af,U,i , which refer to the individual performance-specification features:

Af = {f | fL ≤ f ≤ fU } (118)
Af,L,i = {f | fi ≥ fL,i } (119)
Af,U,i = {f | fi ≤ fU,i } (120)
)
Af = Af,L,i ∩ Af,U,i (121)
i=1,...,nf
Analog Sizing Tasks 73

Correspondingly, the parameter acceptance region As , and individual parameter

acceptance region partitions As,L,i , As,U,i are defined:
*
+


As =


xs


xr ∈ Tr L
∀ f ≤ f (xs , xr ) ≤ fU (122)

*
+


As,L,i =


xs
∀ f (x , x ) ≥ fL,i

xr ∈ Tr i s r
(123)

*
+


As,U,i =


xs
∀ f (x , x ) ≤ fU,i

xr ∈ Tr i s r
(124)

)
As = As,L,i ∩ As,U,i (125)
i=1,...,nf

The definitions of the parameter acceptance region and its partitions include
that the respective performance-specification features have to be satisfied for all
range-parameter vectors within their tolerance region. This more complicated
formulation is due to the fact that the tolerance region of range parameters has
the form of a performance specification but is an input to circuit simulation.
Figure 32 illustrates the parameter acceptance region and its partitioning
according to performance-specification features. This partitioning can be of
use for an approximation of As in (116). In Figure 32, the complementary
“non-acceptance” region partitions are shown as well. The performance non-
acceptance region Āf and its partitions are defined as:
*
+


Āf =


f


i i
∃ f < fL,i ∨ fi < fU,i (126)

Āf,L,i = { f | fi < fL,i } (127)

Āf,U,i = { f | fi > fU,i } (128)
,
Āf = Āf,L,i ∪ Āf,U,i (129)
i=1,...,nf

The parameter non-acceptance region Ās and its partitions are defined as:
*
+


Ās =


xs
∃ ∃ fi (xs , xr ) < fL,i ∨

xr ∈ Tr i ∨fi (xs , xr ) > fU,i
(130)

*
+


Ās,L,i =


xs
∃ f (x , x ) < fL,i

xr ∈ Tr i s r
(131)
74 ANALOG DESIGN CENTERING AND SIZING

Figure 32. Parameter acceptance region As partitioned into parameter acceptance region par-
titions, As,L,1 , As,U,1 , As,L,2 , As,U,2 , for four performance-specification features, f1 ≥ fL,1 ,
f1 ≤ fU,1 , f2 ≥ fL,2 , f2 ≤ fU,2 . As results from the intersection of the parameter acceptance
region partitions.
Analog Sizing Tasks 75
*
+


Ās,U,i =


xs
∃ f (x , x ) > fU,i

xr ∈ Tr i s r
(132)

,
Ās = Ās,L,i ∪ Ās,U,i (133)
i=1,...,nf

4.8.3 Yield Partitions

By introducing the acceptance functions δ(xs ), δL,i (xs ) and δU,i (xs ),

  
1 , xs ∈ As 1 , f ∈ Af
δ(xs ) = = (134)
0 , xs ∈ Ās 0 , f ∈ Āf

1 , xs ∈ As,L,i
δL,i (xs ) = (135)
0 , xs ∈ Ās,L,i

1 , xs ∈ As,U,i
δU,i (xs ) = (136)
0 , xs ∈ Ās,U,i

the yield can be formulated as the expectation value of the acceptance function
according to Appendix A:
 +∞  +∞
Y = ... δ(xs ) · pdf(xs ) · dxs = E {δ(xs )} (137)
−∞ −∞

YL,i = E {δL,i (xs )} (138)

YU,i = E {δU,i (xs )} (139)

Y represents the overall yield, YL,i and YU,i represent the yield partitions
of the respective performance-specification features. An estimation of the
yield according to (137)–(139) is based on statistical estimators according to
Appendix B.
The yield partitions allow a ranking of the individual performance-
specification features concerning their impact on the circuit robustness. Figure
32 illustrates that each added performance-specification feature usually leads
to another yield loss. The smallest specification-feature yield value hence is an
upper bound for the overall yield Y .

4.8.4 Yield Analysis

A yield analysis according to (137)–(139) calculates the yield and yield par-
tition values for given performance-specification features, a given tolerance
76 ANALOG DESIGN CENTERING AND SIZING

Figure 33. Input and output of a yield analysis.

region of range parameters and a given nominal design:

Performance-specification features
fL,i , fU,i , i = 1, . . . , nf Yield Y
Tolerance region Tr → Yield Partitions
Nominal design YL,i , YU,i , i = 1, . . . , nf
x = [ xTd xTs,0 xTr ]T
(140)
This mapping is illustrated in Figure 33.
The yield analysis can be related to a worst-case analysis, which was
illustrated in Figure 28. The worst-case performance values from a worst-case
analysis become upper or lower performance-feature bounds as input of a yield
analysis. The yield value that is obtained as an output of yield analysis becomes
an input yield requirement of the worst-case analysis where it is transformed to
a tolerance region of statistical parameters according to Section 6.2.

4.8.5 Yield Optimization/Design Centering

The goal of yield optimization/design centering is the maximization of the
yield values computed by a yield analysis. According to (137)–(139) yield
optimization/design centering can be formulated either as a single-objective
optimization problem taking the yield,

max Y (xd ) s.t. c(xd ) ≥ 0 (141)

xd
Analog Sizing Tasks 77

Figure 34. Yield optimization/design centering determines a selected point of Pareto front of
performance features.

or as a multiple-objective optimization problem taking the yield partitions:

⎧ ⎫
⎪
⎨ YL,i (xd ) ⎪
⎬
max Y (x
U,i d ) s.t. c(xd ) ≥ 0 (142)
xd ⎪⎩ i = 1, . . . , n ⎪⎭
f

Interestingly, nominal design (93) and worst-case optimization (113)

inherently are multiple-objective optimization problems, whereas yield opti-
mization/design centering (141) inherently is a single-objective optimization
problem. Apparently, it is difficult to combine various performance features
into a single optimization objective. The solution of this task requires a mathe-
matical approach like vector norms to combine multiple performance features
into a single objective. The combination of parameter tolerances and the perfor-
mance specification leads to an inherent single-objective analog design target,
i.e. the yield. The solution of the yield optimization/design centering problem
refers to a restricted part of the Pareto front of performance features, or even
a single point on it, which is determined by the performance specification and
the performance tolerances. Figure 34 illustrates an example of a Pareto front
with three Pareto points and corresponding tolerance ellipsoids.
The performance tolerances have been assumed as equal for the three points,
which is not the general case, but sufficient for illustration. As can be seen, the
black point has maximum yield.
78 ANALOG DESIGN CENTERING AND SIZING

Figure 35. Nested loops within yield optimization/design centering.

Section 6.1.3 explains why a yield estimation is computationally expensive.

Therefore, it may be advantageous to formulate yield optimization/design cen-
tering as a multiple-objective optimization problem (142). The partitioning
into yield partition values may lead to an easier yield computation that pays off
in terms of computational efficiency despite the approximation of the overall
yield target.
An overview of the nested loops of yield optimization/design centering,
which calls yield analysis in a loop, which in turn calls numerical simulation
in a loop, is given in Figure 35. Similar to worst-case optimization, simplifi-
cations of this general approach result from the implementation of the loops
in which yield optimization/design centering calls yield analysis and in which
yield analysis calls simulation. Simplifications are for instance:
Call a yield analysis once for each iteration step of yield optimization/design
centering.

Call a yield analysis in every kth iteration step of yield optimization/design

centering.
These can be combined with the accuracy of the yield estimation in a yield
analysis.
Figure 36 illustrates how yield optimization/design centering changes the
situation in the statistical parameter space.
It shows the parameter acceptance region from Figure 32 plus level contours
of a normal probability density function (23). Each level contour represents a
constant value of the probability density function, which decreases with the size
of the level contour. The parameter acceptance region depends on the values of
design parameters that are disjunct from the statistical parameters. After yield
optimization/design centering, the shape of the parameter acceptance region
has changed in such a way that the volume of the truncated probability density
function is at its maximum. This could look for instance as in Figure 36(b).
Analog Sizing Tasks 79

Figure 36. (a) Initial situation of yield optimization/design centering by tuning of design para-
meters xd that are disjunct from statistical parameters. (b) After yield optimization/design
centering by tuning of design parameters xd that are disjunct from statistical parameters. The
parameter acceptance region As depends on the values of design parameters xd . The equidensity
contours of a normal probability density function are ellipsoids according to (24).

Note that maximum yield does not equally mean a maximum tolerance region
inside the acceptance region.
The picture looks different if the design parameter space and the statistical
parameter space are identical. In that case, the parameter acceptance region
As will be constant. Yield optimization/design centering then happens through
tuning of the statistical parameter distribution, which basically concerns the
mean values, variances, correlations or higher-order moments. The first choice
of yield optimization/design centering in this case is to tune the mean value
xs,0 :

max Y (xs,0 ) s.t. c(xs,0 ) ≥ 0 (143)

xs,0

Figure 37(a) illustrates how yield optimization/design centering changes the

situation in the statistical parameter space, if the nominal values of statistical
parameters xs,0 are tuned. In Figure 36, the maximum yield had been achieved
by appropriate changes in the acceptance region, now it is achieved by appro-
priate changes in the probability density function. Note that a combination
of tuning of nominal values of design and of statistical parameters may occur
as well.
80 ANALOG DESIGN CENTERING AND SIZING

Figure 37. (a) Yield optimization/design centering by tuning of the mean values of statistical
parameters xs,0 . Parameter acceptance region As is constant. (b) Yield optimization/design
centering by tuning of the mean values, variances and correlations of statistical parameters xs,0 ,
C (tolerance assignment). Parameter acceptance region As is constant. Level contours of the
normal probability density function (24) change their shape due to changes in the covariance
matrix C.

4.8.6 Tolerance Assignment

Figure 37(b) illustrates how yield optimization/design centering changes
the situation in the statistical parameter space, if the nominal values, vari-
ances and correlations of statistical parameters are tuned. This type of yield
optimization/design centering is called tolerance assignment:

c(xs,0 ) ≥ 0
max Y (xs,0 , C) s.t. (144)
xs,0 , σk , k,l det C = const ≡ 0
k, l = 1, . . . , nxs , k = l

Without the additional constraint concerning the covariance matrix C in (144),

the trivial solution C∗ = 0 would be obtained. det C describes the volume
of a parallelepiped spanned by the columns or rows of C. The volume det C
corresponds to the volume of a reference tolerance region, which is kept constant
during the optimization to avoid the trivial solution.
Analog Sizing Tasks 81

From the definition of the parameter acceptance region and its partitions
in (122)–(124) follows that the yield depends on the range-parameter bounds
xr,L,k , xr,U,k , k = 1, . . . , nxr as well.
Therefore, a yield sensitivity and yield improvement can be formulated with
regard to the range-parameter bounds as another type of tolerance assignment.
Tolerance assignment plays a role in process tuning.
Tolerance assignment is also applied for the selection of discrete parameters
with different tolerance intervals, like for instance resistors with ±1% tolerances
or ±10% tolerances. Here the goal is to select the largest possible tolerance
intervals without affecting the yield in order to save production costs.
Note that the yield also depends on the specified performance-feature bounds
fL,i , fU,i , i = 1, . . . , nf , which can become a subject of yield optimiza-
tion/design centering. Here the goal is to select the best possible performance
specification that can be guaranteed with a certain yield.

4.8.7 Beyond 99.9% Yield

Figure 38 illustrates that the yield represents a very weak optimum for values
above 99.9%, where further yield improvements are hardly achievable because
the termination criteria of the optimization process will become active. If the
design problem allows a yield above 99.9%, this property therefore results
in a premature termination of the yield optimization/design centering process.
In this case, the design cannot achieve its full robustness in terms of a parts-
permillion yield loss and an increased robustness with respect to tightening of
the performance specification and worsening of the parameter tolerances, as
indicated in Figure 38.
To overcome the problem of premature termination, the parameter tolerances
have to be controlled during the yield optimization process. Once a large enough
yield value is reached, the tolerances are inflated. This scales the yield value
down below 99% and avoids entering the region where the yield runs into
saturation. In this way, the yield optimization/design centering process can
continue to increase the scaled yield, and the true yield can go beyond 99.9%.
This process is continued until no further improvement in the scaled yield is
possible:

(1) maxxd Y (xd ) s.t. c(xd ) ≥ 0

(2) if Y (x∗d ) > 99%
(3) C := a · C with a > 1 (145)
(4) goto (1)
(5) endif

In the case of tolerance assignment, a premature termination of the optimization

82 ANALOG DESIGN CENTERING AND SIZING

Figure 38. To go beyond 99.9% yield for maximum robustness, yield optimization/design
centering requires specific measures.

process can be avoided by exchanging objective and constraint in (144):


c(xs,0 ) ≥ 0
max det C s.t. (146)
xs,0 , σk , k,l Y (xs,0 , C) = const ≡ 50%
k, l = 1, . . . , nxs , k =
 l

In (146), the volume of the parallelepiped determined by the rows or columns

of the covariance matrix is maximized while maintaining the yield at a constant
value of for instance 50%. A yield value of 50% is advantageous because it is
most sensitive to changes. This is illustrated in Figure 20, where the cumulative
distribution function has its maximum slope if its value is 0.5. Solving problem
(146) can be interpreted as inflating the set of parameter tolerance bodies as
much as possible and adjusting their center without going below a certain yield
Analog Sizing Tasks 83

value. It is intuitively clear that this will find a solution that can go beyond a
yield of 99.9%.
A corresponding formulation of (146) for yield optimization/design center-
ing, which only adjusts the nominal parameter vector and leaves the second-
order moments unchanged, is:

c(xs,0 ) ≥ 0
max det(a · C) s.t. (147)
xs,0 , a Y (xs,0 , a · C) = const ≡ 50%

The solution of (147) will lead to a picture as in the lower right half of
Figure 38.
A yield optimization/design centering is only complete if tries to go beyond
yield values of 99.9% as described above. The geometric approach to yield opti-
mization/design centering, described in Section 7.2, does not have the property
of a weak optimum illustrated in Figure 38. It therefore leads to an optimum
as illustrated on the lower right side of this figure without further ado.
Chapter 5

WORST-CASE ANALYSIS

In the following, three main types of worst-case analysis will be described.

They differ in the type of tolerance region and in the modeling of the perfor-
mance function and are suitable for different practical tasks. Table 6 gives an
overview of the three approaches to worst-case analysis.
The classical worst-case analysis comes from a box-type tolerance region and
a linear or linearized performance function. A box-type tolerance region results
if tolerance intervals are defined for each individual parameter independently
from the other parameters (8). Therefore the classical worst-case analysis is
suitable for range parameters, for which the tolerance region is defined exactly
in this way. The classical worst-case analysis is also suitable for uniformly
distributed parameters, for which an interval of parameter values with equal
probability density is defined, or if the type of distribution is not known.
The realistic worst-case analysis starts from an ellipsoid tolerance region
and a linear or linearized performance function. An ellipsoid tolerance region
results if the parameters are normally distributed (10), (23), (24). The realistic

Table 6. Worst-case analysis (WCA) types and characterization.

WCA type Tolerance region Performance function Suitable for

Classical Box Linear Uniform or unknown distribution
Discrete or range parameters
Realistic Ellipsoid Linear Normal distribution
IC transistor parameters
General Ellipsoid Nonlinear Normal distribution
IC transistor parameters
86 ANALOG DESIGN CENTERING AND SIZING

worst-case analysis is therefore suitable for normally distributed parameters

like transistor model parameters of integrated circuits.
We will illustrate that worst-case parameter vectors obtained by the classical
worst-case analysis may correspond to an exaggerated robustness if applied to
integrated circuits. This is mainly due to the missing consideration of the actual
correlations. The realistic worst-case analysis considers the actual distribution
of integrated circuits’ parameters and achieves that the worst-case parameters
represent more realistic yield values, that is the reason for its name.
Both classical and realistic worst-case analysis assume a linear function of the
performance features in the parameters. Either the performance is linear in the
parameters, or linearized performance models are used, for instance based on
a sensitivity computation. A linearized model is generally not sufficient. This
leads to the formulation of the general worst-case analysis, which starts from
an ellipsoid tolerance region and a nonlinear performance in the parameters.

5.1 Classical Worst-Case Analysis

The classical worst-case analysis determines the worst-case value of one
performance feature f for a given box tolerance region of range parameters xr ,

xr,L ≤ xr ≤ xr,U (148)

and for a given linear performance model,

f¯(xr ) = f0 + ∇f (xr,0 )T · (xr − xr,0 ) (149)

based on (109) and (110):

f ≥ fL : min +∇f (xr,0 )T · (xr − xr,0 ) s.t. xr ≥ xr,L , xr ≤ xr,U (150)

xr

f ≤ fU : min −∇f (xr,0 )T · (xr − xr,0 ) s.t. xr ≥ xr,L , xr ≤ xr,U (151)

xr

In (150) and (151) the index i denoting the ith performance feature has been
left out. (150) and (151) can be itemized concerning any performance feature
and any type or subset of parameters.
Figure 39 illustrates the classical worst-case analysis problem in a two-
dimensional parameter space for one performance feature. The gray area is
the box tolerance region defined by a lower and upper bound of each para-
meter, xr,L,1 , xr,U,1 , xr,L,2 , xr,U,2 . The dotted lines are the level contours
according to the gradient ∇f (xr,0 ) of a linear performance model, which are
equidistant planes. Each parameter vector on such a plane corresponds to a cer-
tain performance value. The plane through the nominal point xr,0 corresponds
to the nominal performance value f0 . As the gradient points in the direction of
steepest ascent, the upper and lower worst-case parameter vectors xr,W U and
Worst-Case Analysis 87

Figure 39. Classical worst-case analysis.

xr,W L can readily be marked in. They correspond to the level contours of f
that touch the tolerance region furthest away from the nominal level contour.
This will usually happen in a corner of the tolerance box. These level contours
through the worst-case parameter vectors represent the worst-case performance
values fW U and fW L . The problem formulations (150) and (151) describe a
special case of a linear programming problem that can be solved analytically.

Lower Worst-Case Performance. For an analytical solution, we first write

the Lagrangian function of (150) according to Appendix C:

L(xr , λL , λU ) = ∇f (xr,0 )T · (xr − xr,0 ) − λTL · (xr − xr,L ) − λTU · (xr,U − xr )

(152)
Next, we write the first-order optimality condition of (152), which describes a
solution of (150):

∇f (xr,0 ) − λL + λU = 0 (153)
λL,k · (xr,k − xr,L,k ) = 0 , k = 1, . . . , nxr (154)
λU,k · (xr,U,k − xr,k ) = 0 , k = 1, . . . , nxr (155)
88 ANALOG DESIGN CENTERING AND SIZING

(153) results from the condition that ∇L(xr ) = 0 must hold for a stationary
point of the optimization problem. (154) and (155) represent the complemen-
tarity condition of the optimization problem.
As either only the lower bound xr,L,k or the upper bound xr,U,k of a parameter
xr,k can be active but not both of them, we obtain from (153) and the property
that a Lagrange factor at the solution xr,W L is positive:

Either: λW L,k = +∇f (xr,0,k ) > 0 (156)

or: λW U,k = −∇f (xr,0,k ) > 0 (157)

It depends on the sign of the gradient if (156) or (157) holds. Inserting either
(156) in (154) or (157) in (155) leads to the formula of an element of a worst-case
parameter vector for a worst-case lower performance value xr,W L,k . Analo-
gously an element of a worst-case parameter vector for a worst-case upper
performance value xr,W U,k can be derived.

5.1.1 Classical Worst-Case Parameter Vectors

xr,W L/U = [ . . . xr,W L/U,k . . . ]T
⎧
⎪
⎨ xr,L,k , ∇f (xr,0,k ) > 0
xr,W L,k = xr,U,k , ∇f (xr,0,k ) < 0 (158)
⎪
⎩ undefined, ∇f (xr,0,k ) = 0

⎧
⎪
⎨ xr,U,k , ∇f (xr,0,k ) > 0
xr,W U,k = xr,L,k , ∇f (xr,0,k ) < 0 (159)
⎪
⎩ undefined, ∇f (xr,0,k ) = 0

Figure 40 illustrates the case when an element of the performance gradient is

zero. The gradient with regard to parameter xr,1 is zero. As the performance is
insensitive with regard to xr,1 , the level lines are parallel to the xr,1 coordinate
axis. According to (158) and (159), any value of parameter xr,1 is a solution of
problems (150) and (151). The worst-case values xr,W L,2 , xr,W U,2 of parameter
xr,2 are determined, but the worst-case values xr,W L,1 , xr,W U,1 of parameter
xr,1 may therefore take any value in its defined tolerance interval, xr,L,1 ≤
xr,W L/U,1 ≤ xr,U,1 , as indicated as pieces of thick lines.
In practice, the independence of a performance feature value from a para-
meter, which results from a zero gradient, could be translated into a worst-case
parameter that stays just at its nominal value.
Worst-Case Analysis 89

Figure 40. Classical worst-case analysis with undefined elements xr,W L,1 , xr,W U,1 , of worst-
case parameter vectors xr,W L , xr,W U .

5.1.2 Classical Worst-Case Performance Values

The worst-case performance values result from insertion of the worst-case
parameter vectors into the linear performance function (149):

fW L/U = f0 + ∇f (xr,0 )T · (xr,W L/U − xr,0 ) (160)

The index L/U means that the corresponding formula holds once for a lower
performance-feature bound and once for an upper performance-feature bound.

5.1.3 Discrete Parameters

High-volume integrated circuit production will lead to a normal distribution
of the parameters. If however discrete components from different productions
are combined on a board, the distributions of these components will be quite
different from each other. Figure 41 illustrates how the production test leads to
truncated distributions for a component with different quality classes.
F refers to the clock frequency of a component. The test leads to classes
ranging from very fast (F +++ ) to slow (F 0 ), which are sold for prices ranging
from high to low. The distributions for the different price categories will be
truncated distributions.
If boards with discrete electronic components like resistors with selectable
tolerance classes are composed, it becomes clear from Figure 41 that a ±10%
resistor will most probably have a value very close to its ±10% value. Other-
wise, this resistor would have been assigned to a narrower tolerance class and
been sold at a higher price.
90 ANALOG DESIGN CENTERING AND SIZING

Figure 41. Normal probability density function of a manufactured component splits into trun-
cated probability density functions after test according to different quality classes.

If a discrete system is made of such components, it is difficult to assume the

type of distributions for the components. A classical worst-case analysis is an
adequate choice for analyzing the impact of component tolerances in such a
case.

5.1.4 Corner Worst Case

Classical worst-case parameter vectors applied in integrated circuit design
are also referred to as corner worst case. For a specific performance feature, a
set of relevant statistical parameters is selected. Each parameter is altered by a
multiple of its standard deviation in either the positive or the negative direction
of deteriorating performance. For the gate delay for instance, the resulting
corner worst-case parameter vectors are the slow and fast worst-case parameter
vectors. In order to consider more performance features, additional parameters
have to be considered and more corner worst case parameter vectors result, like
slow-slow, slow-fast, fast-fast.

5.2 Realistic Worst-Case Analysis

The realistic worst-case analysis assumes a joint distribution of parameters
that are either normally distributed, or have been transformed into parameters
that are normally distributed. The worst-case value of one performance feature
f is then determined for a given ellipsoid tolerance region of parameters xs ,

(xs − xs,0 )T · C−1 · (xs − xs,0 ) ≤ βW

2
(161)
Worst-Case Analysis 91

Figure 42. Realistic worst-case analysis.

and for a given linear performance model,

f¯(xs ) = f0 + ∇f (xs,0 )T · (xs − xs,0 ) (162)
based on (109) and (110):
f ≥ fL : min +∇f (xs,0 )T · (xs − xs,0 )
xs

s.t. (xs − xs,0 )T · C−1 · (xs − xs,0 ) ≤ βW

2
(163)

f ≤ fU : min −∇f (xs,0 )T · (xs − xs,0 )

xs

s.t. (xs − xs,0 )T · C−1 · (xs − xs,0 ) ≤ βW

2
(164)
In (163) and (164) the index i denoting the ith performance feature has been
left out. (163) and (164) can be itemized concerning any performance feature
and any type or subset of parameters.
Figure 42 illustrates the realistic worst-case analysis problem in a two-
dimensional parameter space for one performance feature. The gray area is
the ellipsoid tolerance region defined by βW , which is determined such that
the worst-case represents a given yield requirement. Section 5.3 describes how
92 ANALOG DESIGN CENTERING AND SIZING

this is done. The dotted lines are the level contours according to the gradient
∇f (xs,0 ) of a linear performance model, which are equidistant planes. Each
parameter vector on such a plane corresponds to a certain performance value.
The plane through the nominal point xs,0 corresponds to the nominal perfor-
mance value f0 .
As the gradient points in the direction of steepest ascent, the upper and lower
worst-case parameter vectors xs,W U and xs,W L can readily be marked in Figure
42. They correspond to the level contours of f that touch the tolerance region
furthest away from the nominal level contour. This happens somewhere on the
border of the ellipsoid. The level contours through the worst-case parameter
vectors represent the worst-case performance values fW U and fW L .
The problem formulations (163) and (164) describe a special programming
problem with a linear objective function and a quadratic inequality constraint.

Lower Worst-Case Performance. An analytical solution can be derived

based on the Lagrangian function of (163):
L(xs , λ) = ∇f (xs,0 )T · (xs − xs,0 )

2 
−λ · βW − (xs − xs,0 )T · C−1 · (xs − xs,0 ) (165)
The first-order optimality condition (Appendix C) of (165) describes a stationary
point xs,W L , λW L of the Lagrangian function (165) and a solution of (163):
∇f (xs,0 ) + 2 · λW L · C−1 · (xs,W L − xs,0 ) = 0 (166)

(xs,W L − xs,0 )T · C−1 · (xs,W L − xs,0 ) = βW

2
(167)
λW L > 0 (168)
(166) results from the condition that ∇L(xs ) = 0 must hold for a stationary
point of the optimization problem. (167) and (168) consider that the inequality
constraint has to be active in the solution because the objective function is linear.
The second-order optimality condition is satisfied because ∇2 L(xs ) = 2 ·
λ · C−1 is positive definite, as C−1 is positive definite and as λW L > 0.
From (166) we obtain:
1
xs,W L − xs,0 = − · C · ∇f (xs,0 ) (169)
2λW L
Inserting (169) in (167) results in
1
· ∇f (xs,0 )T · C · ∇f (xs,0 ) = βW
2
(170)
4λ2W L
Solving (170) for λW L and inserting the resulting equation for λW L in (169)
leads to the analytical formulation of the worst-case parameter vector for a
Worst-Case Analysis 93

worst-case lower performance value xs,W L . Analogously a worst-case para-

meter vector for a worst-case upper performance value xs,W U can be derived.

5.2.1 Realistic Worst-Case Parameter Vectors

βW
xs,W L − xs,0 = − - · C · ∇f (xs,0 ) (171)
∇f (xs,0 )T · C · ∇f (xs,0 )
βW
= − · C · ∇f (xs,0 ) (172)
σf¯
βW
xs,W U − xs,0 = + - · C · ∇f (xs,0 ) (173)
∇f (xs,0 )T· C · ∇f (xs,0 )
βW
= + · C · ∇f (xs,0 ) (174)
σf¯

As the performance feature f¯(xs ) is linear in the normally distributed statistical

parameters according to (162), its variance σf2¯ has been calculated using (A.13)
and inserted in (172) and (174):

σf2¯ = ∇f (xs,0 )T · C · ∇f (xs,0 ) (175)

Exercise. Apply (162) and (A.12) to prove (175).

5.2.2 Realistic Worst-Case Performance Values

The worst-case performance values result from insertion of the worst-case
parameter vectors into the linear performance function (162), fW L/U = f0 +
∇f (xs,0 )T · (xs,W L/U − xs,0 ):
.
fW L = f0 − βW · ∇f (xs,0 )T · C · ∇f (xs,0 ) (176)

= f0 − βW · σf¯ (177)
.
fW U = f0 + βW · ∇f (xs,0 )T · C · ∇f (xs,0 ) (178)

= f0 + βW · σf¯ (179)

5.3 Yield/Worst-Case Distance – Linear Performance

Feature
Assuming the performance feature fi to be a linear function of the parameters
(162), it is normally distributed. This is the result from the linear transformation
of a normal distribution applying (A.13) in the same way as is done in step 3
94 ANALOG DESIGN CENTERING AND SIZING

in Section 3.8.  
f¯i ∼ N f0,i , σf2¯i (180)
The mean value of the probability density function is the nominal performance
value in (162), the variance is given by (175).
From (177), (179) and (180), and applying (17), a unique relationship between
the yield of a performance feature and the value of βW , which characterizes the
ellipsoid tolerance region, can be established:

2
f¯i −f0,i
1 − 12
pdff¯i (f¯i ) = √
σf¯
e i (181)
2π · σf¯i

 fW U,i  fW U,i −f0,i

σf¯ 1 1 2
YU,i = pdff¯i (f¯i ) · df¯i = i
√ e− 2 t · dt
−∞ −∞ 2π
 βW
1 1 2
= √ e− 2 t · dt (182)
−∞ 2π
= YL,i (183)
 fW U,i  βW
1 1 2
Yi = pdff¯i (f¯i ) · df¯i = √ e− 2 t · dt (184)
fW L,i −βW 2π
YL,i , YU,i are the yield partition values regarding the lower and upper worst-
case performance values fW L,i , fW U,i of the considered performance feature
fi . Yi is the yield value regarding performance feature fi with both its lower
and upper worst-case value.
These equations open up the possibility of a technical interpretation of βW
as the measure that relates an ellipsoid tolerance region to a required yield.
We call βW the worst-case distance between the nominal value and the worst-
case value of a performance feature. According to (177) and (179) it is measured
in the unit “performance variance”: the worst-case performance value is βW
times of the performance variance away from the nominal performance value.
βW = 3 therefore refers to a three-sigma safety margin (three-sigma design)
of a performance feature, and βW = 6 refers to a six-sigma safety margin
(six-sigma design).
As a multiple of the performance variance it can immediately be translated
into a yield value for one performance feature according to (182)-(184) and
Table 5. Vice versa, a yield value can be translated into the size of an ellip-
soid tolerance region in a multivariate parameter space of a realistic worst-
case analysis according to (167). Section 5.5 will show that the meaning and
interpretation of the worst-case distance will be valid for nonlinear performance
functions as well.
Worst-Case Analysis 95

5.4 General Worst-Case Analysis

The general worst-case analysis assumes a joint distribution of parameters,
which is either normal or has been transformed into a normal distribution. The
worst-case value of one performance feature f is then determined for a given
ellipsoid tolerance region of parameters xs ,
(xs − xs,0 )T · C−1 · (xs − xs,0 ) ≤ βW
2
(185)
and for a general nonlinear performance feature in the parameters f (xs ), based
on (109) and (110):
f ≥ fL : min +f (x) s.t. (xs − xs,0 )T · C−1 · (xs − xs,0 ) ≤ βW
2
(186)
xs

f ≤ fU : min −f (x) s.t. (xs − xs,0 )T · C−1 · (xs − xs,0 ) ≤ βW

2
(187)
xs

In (186) and (187) the index i denoting the ith performance feature has been
left out. (186) and (187) can be itemized concerning any performance feature
and any type or subset of parameters.
Figure 43 illustrates the general worst-case analysis problem in a two-
dimensional parameter space for one performance feature. The gray area is
the ellipsoid tolerance region defined by βW , which is determined such that
the worst-case represents a given yield requirement. This is done according
to (182)–(184). For a three-sigma design for instance, βW = 3 would be
chosen. Section 5.5 motivates why (182)–(184) are well-suited in the general
nonlinear case.
The dotted lines are the level contours of the nonlinear performance feature.
Each parameter vector on such a level contour corresponds to a certain perfor-
mance value. The contour line through the nominal point xs,0 corresponds to
the nominal performance value f0 . In this example, the performance is uni-
modal in the parameters and its minimum value is outside of the parameter
tolerance ellipsoid.
The upper and lower worst-case parameter vectors xs,W U and xs,W L can
readily be marked in. They correspond to the level contours of f that touch the
tolerance region furthest away from the nominal level contour. This happens
somewhere on the border of the ellipsoid in this case. These level contours
through the worst-case parameter vectors represent the worst-case performance
values fW U and fW L .
The property that performance level contour and tolerance region border
touch each other in the worst-case parameter vector corresponds to a plane
tangential both to the performance level contour and the tolerance region border.
This tangent can be interpreted by a linearized performance model f¯ in the
worst-case parameter vector. The linearization is based on the gradient of the
performance in the worst-case parameter vector.
96 ANALOG DESIGN CENTERING AND SIZING

Figure 43. General worst-case analysis, in this case the solution is on the border of the tolerance
region.

As the performance feature is nonlinear, we have a gradient ∇f (xs,W U ) at the

upper worst-case parameter vector xs,W U and a differing gradient ∇f (xs,W L )
at the lower worst-case parameter vector xs,W L in this example. The lineariza-
tion at the worst-case parameter vector depends on the respective gradient,
worst-case performance value and worst-case parameter vector:
f¯(W L/U ) (xs ) = fW L/U + ∇f (xs,W L/U )T · (xs − xs,W L/U ) (188)
The level contour of such a linearized performance model through the respec-
tive worst-case parameter vector represents the corresponding worst-case per-
formance value, f¯(W L/U ) = fW L/U .
In a realistic worst-case analysis, the performance function is linear in the
parameters as illustrated in Figure 42. The solution of problems (163) and
(164) is unique, and the worst-case parameter vectors will be on the border of
the ellipsoid tolerance region.
In a general worst-case analysis, the solutions of problems (186) and (187)
are not generally unique and the worst-case parameter vectors will be either on
the border or inside of the ellipsoid tolerance region. It has been observed that
usually unique solutions to problems (186) and (187) appear. An important case
Worst-Case Analysis 97

of multiple solutions are mismatch-sensitive performance functions with regard

to local parameter changes, which are characterized by semidefinite second
derivatives. The nominal performance is on a ridge and stays more or less
constant in the direction of equal changes in two parameters and deteriorates in
other directions. In such a case, very regularly concerning two local parameter
variations two worst-case parameter vectors in “opposite” directions from the
nominal parameter vector exist. Solution algorithms for problems (186) and
(187) have to take care of such situations.
Problems (186) and (187) describe a special case of nonlinear programming
with a nonlinear objective function and one quadratic inequality constraint. An
analytical solution of problems (186) and (187) cannot be formulated. The
solution is done by numerical optimization, for instance with a deterministic
approach based on Sequential Quadratic Programming.

Lower Worst-Case Performance, Worst-Case Parameter Vector on the

Border of the Tolerance Region. In the following, we will treat the compu-
tation of a lower worst-case performance value (186). The Lagrangian function
of (186) is:

2 
L(xs , λ) = f (xs ) − λ · βW − (xs − xs,0 )T · C−1 · (xs − xs,0 ) (189)

The first-order optimality condition (Appendix C) of (189) describes a stationary

point xs,W L , λW L of the Lagrangian function (189) and a solution of (186):

∇f (xs,W L ) + 2 · λW L · C−1 · (xs,W L − xs,0 ) = 0 (190)

2 
λW L · βW − (xs,W L − xs,0 )T · C−1 · (xs,W L − xs,0 ) = 0 (191)

(190) results from the condition that ∇L(xs ) = 0 must hold for a station-
ary point of the optimization problem. (191) represents the complementarity
condition of the optimization problem.
The second-order optimality condition holds because ∇2 L(xs ) = 2 · λ · C−1
is positive definite, as C−1 is positive definite and as λW L ≥ 0.
We assume that the solution is on the border of the ellipsoid tolerance region.
This happens for instance if the performance function f (xs ) is unimodal and
if its maximum or minimum is outside of the parameters tolerance region. The
constraint in (189) is therefore active:

λW L > 0 (192)
(xs,W L − xs,0 )T · C−1 · (xs,W L − xs,0 ) = βW
2
(193)

Analogous equations for a worst-case parameter vector concerning a worst-case

upper performance value xs,W U can be derived.
98 ANALOG DESIGN CENTERING AND SIZING

(190), (192) and (193) have the same form as the first-order optimality con-
dition of the realistic worst-case analysis (166)–(168). The general worst-case
parameter vectors and general worst-case performance values therefore also
have the same form as those of the realistic worst-case analysis. The only dif-
ference is in the gradient of the performance. One overall performance gradient
appears in the realistic worst-case analysis, whereas an individual performance
gradient at each worst-case parameter vector appears in the general worst-case
analysis.

5.4.1 General Worst-Case Parameter Vectors

βW
xs,W L − xs,0 = − - · C · ∇f (xs,W L ) (194)
∇f (xs,W L )T · C · ∇f (xs,W L )
βW
= − · C · ∇f (xs,W L ) (195)
σf¯(W L)

βW
xs,W U − xs,0 = + - · C · ∇f (xs,W U ) (196)
∇f (xs,W U )T · C · ∇f (xs,W U )
βW
= + · C · ∇f (xs,W U ) (197)
σf¯(W U )

In (195) and (197) the variance σf2¯(W L/U ) of a performance function f¯(W L/U )
that is linear (188) in normally distributed parameters (25) has been inserted:
σf2¯(W L/U ) = ∇f (xs,W L/U )T · C · ∇f (xs,W L/U ) (198)

5.4.2 General Worst-Case Performance Values

The worst-case performance values fW L and fW U result from the numerical
solution of the optimization problems (186) and (187).
A formulation of the worst-case performance values that corresponds to the
realistic worst-case analysis (176)–(179) can be derived by inserting the nominal
parameter vector xs,0 into the linear performance functions (188):
(W L/U )
f¯(W L/U ) (xs,0 ) = f¯0 = fW L/U + ∇f (xW L/U )T · (xs,0 − xW L/U )
(W L/U )
⇔ fW L/U = f¯0 + ∇f (xW L/U )T · (xW L/U − xs,0 ) (199)
Inserting (194)–(197) into (199) leads to:
.
(W L)
fW L = f¯0 − βW · ∇f (xs,W L )T · C · ∇f (xs,W L ) (200)
(W L)
= f¯0 − βW · σf¯(W L) (201)
Worst-Case Analysis 99
.
(W U )
fW U = f¯0 + βW · ∇f (xs,W U )T · C · ∇f (xs,W U ) (202)
(W U )
= f¯0 + βW · σf¯(W U ) (203)
(200)–(203) are equal to the corresponding equations (176)–(179) of the realis-
tic worst-case analysis concerning the interpretation of the worst-case distance
βW according to Section 5.3. They differ in the definitions of the nominal per-
formance value and the variance. In (200)–(203), the linearized models of the
considered performance feature differ at the worst-case parameter vectors for
the lower and upper worst-case performance value. This results in two different
variance values σf¯(W L/U ) and in two nominal values of the performance feature
(W L/U )
f¯
0 that differ from the true nominal value f0 .

5.5 Yield/Worst-Case Distance – Nonlinear Performance

Feature
With regard to the linear models (188) in the worst-case parameter vectors,
these linearized performance features are normally distributed according to
Section 3.8 and Appendix A.9:

(W L/U ) (W L/U )
f¯i ∼ N f¯0,i , σ 2¯(W L/U ) (204)
fi

The mean value of the probability density function is the value of the linearized
performance function according to (199) in the nominal parameter vector, the
variance is given by (198).
From (201), (203) and (204), and applying (17), a unique relationship
between the yield of the performance feature linearized at the worst-case para-
meter vector and the value of βW , which characterizes the ellipsoid tolerance
region can be established:
⎛ ⎞2
(W L/U ) ¯(W L/U )
f¯ −f
1⎝ i 0,i ⎠
−2
(W L/U ) 1 σ (W L/U )
f¯
pdff¯(W L/U ) (f¯i )= √ e i (205)
i 2π · σf¯(W L/U )
i

 fW U,i
(W U ) (W U )
ȲU,i = pdff¯(W U ) (f¯i ) · df¯i
−∞ i

(W U )
fW U,i −f¯
 0,i
σ (W U ) 1 1 2
= f¯
i √ e− 2 t · dt
−∞ 2π
 βW
1 1 2
= √ e− 2 t · dt (206)
−∞ 2π
100 ANALOG DESIGN CENTERING AND SIZING

= ȲL,i (207)

ȲL,i and ȲU,i are approximated yield partition values regarding the lower and
upper worst-case performance value fW L/U,i of performance feature fi .
This equation opens up the possibility of a technical interpretation of βW as
the measure that relates an ellipsoid tolerance region to an approximated yield.
We call βW the worst-case distance between the nominal value and the worst-
case value of a linearized performance feature according to (188). While the
worst-case value of the real performance feature and the linearized performance
feature are equal, the nominal value of the linearized performance feature differs
from the real nominal value according to (199). According to (201) and (203),
the worst-case distance is measured in the unit “variance of the linearized per-
formance”: the worst-case performance value is βW times the variance away
from the nominal linearized performance value. This variance is determined
by (198).
βW = 3 therefore refers to a three-sigma safety margin (three-sigma design)
of a performance feature, and βW = 6 refers to a six-sigma safety margin
(six-sigma design).
As a multiple of a performance variance, it can immediately be translated into
an approximated yield value for one performance feature according to (206) and
(207) and Table 5. Vice versa, a yield value can be translated approximately into
a value of βW , which determines an ellipsoid tolerance region in a multivariate
parameter space according to (191).

5.5.1 Yield Approximation Accuracy

An approximation error between the true yield partition value and the yield
partition value according to (206), (207) results from approximating a nonlinear
(W L/U )
performance function fi (xs ) by a linear model f¯i (xs ).
This approximation error will be discussed in the following by investigating
the parameter space. It will be assumed that the worst-case parameter vectors
are on the border of the ellipsoid tolerance region.
According to the definition (116), (117), the yield can be formulated equiv-
alently in the performance space and in the parameter space. With the formu-
lations of acceptance regions in the performance space and in the parameter
space in (118)–(125), the approximate yield partition ȲL,i with regard to a
lower performance-feature bound fW L,i and the approximate yield partition
ȲU,i with regard to an upper worst-case performance-feature value fW U,i can
be formulated as:
 
ȲU,i = ... pdfN (xs ) · dxs (208)
(W U )
f¯i (xs )≤fW U,i
Worst-Case Analysis 101
 
= ... pdfN (xs ) · dxs (209)
∇fi (xs,W U,i )T ·(xs −xs,W U,i )≤0
 
ȲL,i = ... pdfN (xs ) · dxs (210)
(W L)
f¯i ≥fW L,i
 
= ... pdfN (xs ) · dxs (211)
∇fi (xs,W L,i )T ·(xs −xs,W L,i )≥0

(188) has been applied to get from the yield definition in the performance space
(208), (210) to the yield definition in the parameter space (209), (211).
pdfN is the normal probability density function according to (23).
On the other hand, the true yield partition values, YL,i with regard to a lower
worst-case performance-feature value fW L,i , and YU,i with regard to an upper
worst-case performance-feature value fW U,i , can be formulated as:
 
YU,i = ... pdfN (xs ) · dxs (212)
fi (xs )≤fW U,i
 
YL,i = ... pdfN (xs ) · dxs (213)
fi (xs )≥fW L,i

(209), (211), (212) and (213) show that the error in the yield approximation
results from the approximation of the parameter acceptance region concerning
a performance-specification feature.
Figure 44 illustrates the situation for the example in Figure 43. In Figure
44(a), the upper worst-case has been picked. The ellipsoid tolerance region
corresponds to the given value of βW . The real level contour of all parameter
vectors that lead to the worst-case performance value fW U is shown by as a
dotted curve. It represents the border of the acceptance region that determines
the true yield partition value YU . This acceptance region is shaded in gray.
In addition, the level contour of all parameter vectors that lead to fW U
according to the linear performance model f¯(W U ) is shown as a solid line. It is
tangential to the true level contour in the worst-case parameter vector xs,W U ,
because f¯(W U ) has been specifically established in xs,W U by the general worst-
case analysis. It represents the border of the approximate acceptance region
that determines the approximate yield partition value ȲU . This region is filled
with a linen pattern.
The integration of the probability density function over the difference region
between the linen-pattern-filled area of the approximate parameter acceptance
region partition and the gray area of the true parameter acceptance region par-
tition determines the error of the yield approximation. In Figure 44(a), the
true yield will be overestimated. In Figure 44(b), the lower worst-case has
102 ANALOG DESIGN CENTERING AND SIZING

Figure 44. (a) Comparison between true parameter acceptance region partitions (gray areas)
and approximate parameter acceptance region partitions (linen-pattern-filled areas) of a general
worst-case analysis of an upper worst-case performance value. (b) Lower worst-case perfor-
mance value. (c) Lower and upper worst-case performance value.
Worst-Case Analysis 103

Figure 45. Duality principle in minimum norm problems. Shown are two acceptance regions
(gray), the respective nominal points within acceptance regions and two points on the border of
each acceptance region. Points (a) and (d) are the worst-case points.

been picked. Here, the true yield will be underestimated. Figure 44(c) shows
the overall approximate acceptance region if the lower and upper worst-case
parameter vectors and their linearizations are combined.
By inspection, the following statements concerning the approximation error
result:
Considering the decreasing values of the probability density function orthog-
onal to its ellipsoid level contours, the yield approximation error becomes
more critical in regions that correspond to ellipsoids closer to the center.
The chosen linearization is very well adapted to this property. It is exact in
the worst-case parameter vector, where the value of the probability density
function is maximum among the difference region. It loses precision in
modeling the exact level contour proportionally to the decrease in the corre-
sponding probability density function values. Apparently the approximation
is very accurate therefore.
Practical experience shows that the approximation error usually is about
1%–3% absolute yield error.
The duality principle in minimum norm problems says that the minimum
distance of a point to a convex set is equivalent to the maximum distance of
the point to all planes that separate the point from the convex set. Figure 45
illustrates the duality principle. In this example, we have assumed without
loss of generality that the variances are equal and that the correlations are
zero. Then the level contours are spheres.
On the left side of Figure 45, point (a) is the worst-case point for the given
nominal point inside the gray acceptance region. The duality principle in
this case applies for the distance of the nominal point to the complement of
104 ANALOG DESIGN CENTERING AND SIZING

the acceptance region, which is convex. It says that point (a) provides the
linearization among all possible points on the border of the acceptance region
from which the nominal point has maximum distance. This is illustrated by
point (b), which leads to a linearization with a smaller distance. We can
also see that any tangent on the acceptance region’s border will lead to
an underestimation of the true yield. Therefore the worst-case distance
obtained due to point (a) is a greatest lower bound among all tangential
approximations of the acceptance region.
On the right side Figure 45, the acceptance region itself is convex now.
We can see that point (c) leads to a larger distance between the nominal
point and the tangential approximation of the acceptance region’s border
than the worst-case point (d). Now the duality principle says that the worst-
case distance will be the smallest value among all possible distances of the
nominal point to tangents of the border of the acceptance region. At the
same time, it can be seen that the yield value will be overestimated by such
a tangent. Therefore the worst-case distance obtained due to point (d) is a
least upper bound among all tangential approximations of the acceptance
region.
In summary, the linearization at a worst-case parameter vector provides
the best yield approximation among all tangential planes of an acceptance
region.

Figure 38 illustrates that the yield approximation according to (206) and

(207) and Table 5 will be very sensitive with regard to the worst-case
distance value for yield values around 50%. With increasing yield value
this sensitivity decreases, which equally means a smaller error in the yield
approximation through worst-case distances.
For yield values above 99.9% the approximation error becomes negligible.
(184) can be applied to obtain an approximate yield Ȳi of a performance feature
fi if an upper and a lower performance-specification feature are simultaneously
regarded. The resulting value is not equivalent to the yield value that results
from the approximate acceptance region as illustrated in Figure 44(c), as long
as the performance gradients in the two worst-case parameter vectors are not
parallel. Nevertheless, Ȳi according to (184) is a valuable approximation in the
general worst-case analysis as well.
The yield approximation as described in this section assumes that worst-case
parameter vectors are on the border of the acceptance region, which usually
happens as practical experience shows.
If a performance function is unimodal and the worst-case parameter vector is
inside the tolerance region, then the worst-case performance value represents the
individual optimum for the considered parameter space. Hence, no performance
Worst-Case Analysis 105

value beyond the worst-case value is achievable in the considered parameter

space and the corresponding yield is 100%.

5.5.2 Realistic Worst-Case Analysis as Special Case

As a linear performance function is a special case of a general nonlinear
function, the realistic worst-case analysis is a special case of the general worst-
case analysis.
A numerical optimization algorithm can identify if the performance function
is linear during the optimization process. It then automatically terminates the
optimization process and determines the worst-case performance values and
worst-case parameter vectors.
An iterative deterministic solution algorithm for a general worst-case analysis
is starting from a sensitivity analysis of the performance features with regard
to the parameters. In the first iteration step, a general worst-case analysis then
behaves like a realistic worst-case analysis. The realistic worst-case analysis
can therefore be interpreted as the first step of a general worst-case analysis.

5.6 Exercise
Given is a single performance function of two parameters:
f = xs,1 · xs,2 (214)
f could be for instance the time constant of the RC circuit in Chapter 2. The
nominal parameter vector is:
xs,0 = [ 1 1 ]T (215)
The parameters are normally distributed with the covariance matrix:
 
0.22 0
C= (216)
0 0.82
Two performance-specification features are given:
f ≥ fL ≡ 0.5 (217)
f ≤ fU ≡ 2.0 (218)
Perform a classical worst-case analysis in the three-sigma tolerance box of
parameters based on a linear performance model established at the nominal
parameter vector. Calculate the worst-case parameter vectors and corre-
sponding worst-case performance values. Compare the worst-case perfor-
mance values from the linearized performance model with the “simulated”
values at the worst-case parameter vectors. Check if the performance spe-
cification is satisfied in the worst-case.
106 ANALOG DESIGN CENTERING AND SIZING

Do the same as a realistic worst-case analysis for an ellipsoid tolerance

region with βW = 3.
Do the same as a general worst-case analysis. Apply the optimality con-
ditions (Appendix C) to calculate a solution. Check if the solution can be
inside the ellipsoid tolerance region (in that case λW L/U = 0 would hold).
Chapter 6

YIELD ANALYSIS

In this chapter, the yield analysis problem formulated in Section 4.8 will
be further developed. The integral in the yield definitions (116) and (117), as
well as in (137), (138) and (139) cannot be solved analytically, but have to be
solved numerically. Two approaches to solve this task will be described in the
following.
The first approach is based on statistical estimation by sampling according
to the parameter distribution. It leads to the so-called Monte-Carlo analysis,
which is a statistical technique for the numerical computation of integrals. It
will be explained that the accuracy of the statistical estimation does not depend
on the number of parameters and performance features, but on the size of the
sample. An acceptable accuracy requires a large number of simulations.
The second approach is based on the partitioning of the performance speci-
fication into the individual performance-specification features and a geometrical
approximation of the integration problem. The resulting problem has the same
form as the general-worst case analysis described in Section 5.4. This approach
is in practice more efficient than a Monte-Carlo technique, but its complexity
grows proportionally to the number of parameters.

6.1 Statistical Yield Analysis

According to (137), the yield is defined as the expectation value of the accep-
tance function (134). This acceptance function takes the value of one (134), if
a circuit represented by a vector of parameter values satisfies the performance
specification (122). According to (138), (139), the yield partition with respect to
a lower or upper performance-specification feature is defined as the expectation
value of the acceptance function (135), (136). This acceptance function takes
the value of one (135), (136), if the parameter vector satisfies the corresponding
performance-specification feature (123), (124).
108 ANALOG DESIGN CENTERING AND SIZING

Based on (B.1) in Section B, an estimator for this expectation value and

hence for the yield can be formulated as follows:
n
MC
1 nok
Y/ / {δ(x)} =
= E δ(x(µ)
s )= = (219)
nM C nM C
µ=1

{accepted sample elements}

=
sample size
n
  1 MC
nok,L/U,i
Y/L/U,i = E
/ δL/U,i (xs ) = δL/U,i (x(µ)
s )= (220)
nM C nM C
µ=1

s ∼ D (pdf(xs )) , i = 1, . . . , nM C
x(µ)

In (219), the yield with regard to the complete performance specification is

estimated. In (220), the yield partition with regard to an individual performance-
specification feature, which is either a lower or upper bound on a performance
feature, is estimated.
The basis of the statistical yield estimation is a sample consisting of nM C
(µ)
sample elements xs , µ = 1, . . . , nM C , which have been generated according
to Section 3.8. As stated in Appendix B, the sample elements are assumed to be
independently and identically distributed according to the given distribution.
This is not strictly true: the computational random number generator produces
a deterministic sequence of pseudo-random numbers for a given initial value.
The generation of a sample can be interpreted as a simulation of the
statistically varying production process on a higher design level, where the
performance vector f is determined in dependence of parameters xs . Each
(µ)
sample element xs is evaluated by numerical simulation. The obtained per-
(µ)
formance vector f (xs ) is checked with regard to the performance specification
and the corresponding value of the acceptance function in (219), (220) can be
determined.
Figure 46 illustrates this process for two statistical parameters. On the left
side, level contours of the probability density function of normally distributed
parameters are given. On the right side, a performance specification with four
performance-specification features is given, which lead to the performance
acceptance region shaded in gray. The corresponding parameter acceptance
region is also indicated by a gray area on the left side. The parameter accep-
tance region though is generally not given in an analytical form.
A sample according to the given distribution yields a cloud of parameter
vectors that corresponds to the probability density function. After simulation
of each parameter vector, a corresponding cloud of performance vectors as on
the right side of Figure 46 is available. Each performance vector that satisfies
Yield Analysis 109

Figure 46. Statistical yield estimation (Monte-Carlo analysis) consists of generating a sam-
ple according to the underlying statistical distribution, simulating of each sample element and
flagging of elements satisfying the performance specification.

the performance specification is filled black. Likewise each parameter vector

can now also be marked according to the corresponding performance vector.
According to (219), the percentage of parameter vectors that are represented
in Figure 46 by circles filled black represents the yield estimate.
This statistical estimation is also called Monte-Carlo analysis.

6.1.1 Monte-Carlo Analysis

A Monte-Carlo analysis is a statistical method for the numerical computation
of an integral I, which applies the formulation of expectation values (A.1):
 
I = ... h(x) · dx
h1 (x)≥0
 ∞  ∞
h(x)
= ... hA (x) · · pdfM C (x) · dx
−∞ −∞ pdfM C (x)
 
h(x)
= E hA (x) · (221)
pdfM C (x) pdfM C (x)


1, h1 (x) ≥ 0
hA (x) =
0, h1 (x) < 0
110 ANALOG DESIGN CENTERING AND SIZING

In (221), the integrand has been extended by a probability density function

pdfM C that is used to scan the integration region. In addition, the integration
is done over the complete parameter space, and an indicator function hA is
introduced in the integrand to include the integration region.
Yield analysis is a special case of Monte-Carlo analysis, where the probability
density function is predetermined by the manufacturing tolerances, and where
the indicator function is determined by the performance specification and the
corresponding acceptance function.
The Monte-Carlo analysis performs a statistical estimation of (221) based
on (B.1):

 
h(x)
I/ = /
E hA (x) ·
pdfM C (x)
pdfM C (x)
n
1 MC
h(x(µ) )
= hA (x(µ) ) (222)
nM C
µ=1
pdfM C (x(µ) )

x(µ) ∼ D (pdfM C (x)) , µ = 1, . . . , nM C

The choice of the probability density function is important for the estimation.
But even if it is predetermined as in a yield analysis, another distribution may
be better for the yield estimation with a Monte Carlo analysis. This is called
importance sampling.

6.1.2 Importance Sampling

From Figure 46, we would intuitively expect that a good yield estimation
scans more or less the whole parameter acceptance region. This requires a
sufficient spread of the sample elements. If the sampling distribution leads
to an equal number of sample elements inside and outside of the parameter
acceptance region, we would have a yield of 50% and the required even spread
of sample elements.
If an estimation, in this case of the yield, is done using a sampling distri-
bution represented by pdfM C that is different from the parameter distribution
represented by pdf, this is called importance sampling:

 ∞  ∞
Y = E {δ(xs )} = δ(xs ) · pdf(xs ) · dxs
...
pdf(xs ) −∞ −∞

 ∞  ∞
pdf(xs )
= ... δ(xs ) · · pdfM C (xs ) · dxs
−∞ −∞ pdfM C (xs )
Yield Analysis 111

 
pdf(xs )
= E δ(xs ) · (223)
pdfM C (xs ) pdfM C (xs )

 
pdf(xs )
Y/ = /
E δ(xs ) ·
pdfM C (xs )
pdfM C (xs )
n
MC (µ)
1 pdf(xs )
= δ(x(µ)
s ) (µ)
(224)
nM C pdfM C (xs )
µ=1

x(µ)
s ∼ D (pdfM C (xs )) , i = 1, . . . , nM C

The change of the original distribution, for which the expectation value of a
function is to be computed, to another sampling distribution leads to a weighting
of the function with the ratio of the two distributions.
Importance sampling allows to use arbitrary sampling distributions for the
yield estimation in order to improve the quality of the yield estimation.
A measure of the yield estimation quality is the variance of the yield estima-
tion value.

6.1.3 Yield Estimation Accuracy

The variance of the yield estimator can be derived using Appendices A and
B:
/
σ 2
= V{Y} = / {δ(xs )}}
V{E
Y
(B.14) 1
= · V {δ(xs )}
nM C
(A.14) 1

= · E{δ 2 (xs )} − E2 {δ(xs )}
nM C
δ 2 =δ 1
= · (Y − Y 2 )
nM C
Y · (1 − Y )
= (225)
nM C
(225) is the yield estimator variance, if the yield Y is known. As the yield
is usually estimated, Y/ , the formula for an estimator of the yield estimator
variance is:
/ Y}
/2
= V{
σ / = / E
V{ / {δ(xs )}}
Y
112 ANALOG DESIGN CENTERING AND SIZING

(B.15) 1 / {δ(xs )}
= ·V
nM C
1 nM C  
(B.18)
= · · E{δ / 2 {δ(xs )}
/ 2 (xs )} − E
nM C nM C − 1
δ 2 =δ 1
= · (Y/ − Y/ 2 )
nM C − 1

Y/ · (1 − Y/ )
= (226)
nM C − 1
(225) and (226) show that the accuracy of a statistical yield estimation by
Monte-Carlo analysis primarily depends on the size of the sample, but also on
the yield value itself.
Obviously, the variance of the yield estimator σ 2
is quadratical in the yield
Y
value. From the first and second-order optimality conditions (Appendix C),
∂σ 2
1
Y
= (1 − 2Y ) ≡ 0 (227)
∂Y nM C
∂ 2 σ 2
2
Y
= − < 0 (228)
∂Y 2 nM C
follows that the variance of the yield estimator has a maximum for a yield of
50%:
arg max σY
= 50% (229)
Y

The variance of the statistical yield estimation corresponds to the sensitivity

of the yield with respect to a shift in the integration bound as illustrated in
Figure 20. The hyperbolic-tangent-like shape of the yield function is typical.
Figure 20 shows a maximum sensitivity of the yield with regard to a shift in
the integration bound if the yield is 50%. For smaller or larger yield values this
sensitivity decreases down to zero for a yield of 0% or 100%.
More important is the dependence of the yield estimator variance on the
number of sample elements. Table 7 evaluates (225) for an assumed yield of
85% for different sample sizes. A yield of 85% is a typical value in practice
and leads to smaller yield estimator variances than a yield of 50% would do.
From Table 7 follows for instance that the three-sigma interval of the esti-
mated yield value is [ 70% . . . 100% ] for a sample size of 50. For a sample
size of 1000, the three-sigma interval of the estimated yield value is [ 81.7% . . .
88.3% ], which is still quite large.
Let us have a look at the number of sample elements that is required to
estimate a yield of 85% with a certain confidence level. For that we will assume
Yield Analysis 113

Table 7. Standard deviation of the yield estimator if the yield is 85% for different sample sizes
evaluated according to (225).

nM C 10 50 100 500 1000

σY
11.3% 5.0% 3.6% 1.6% 1.1%

that the yield estimate is normally distributed. This is an approximation: as

the sample elements are independently and identically distributed and as every
sample element has the same probability Y of being in full working order,
the number of sample elements that are in full working order originally is
binomially distributed. But for an increasing sample size nM C → ∞, the
binomial distribution asymptotically approaches a normal distribution. The
binomial distribution can be approximated under easy conditions. For instance,
the number of sample elements satisfying the performance specification and the
number of sample elements missing the performance specification should each
be larger than 4, and the number of total sample elements should be at least 10.
Considering these constraints concerning the sample, we can assume that the
estimated yield value is normally distributed with the variance (225).
Based on the normal distribution of the yield estimate with variance (225),
we can compute the sample size nM C that is required for a yield to be within
an interval of ±∆Y around the estimated value Y/ with a confidence level of
γ[%] in the following way.
The confidence level γ denotes the probability that the estimated value is
within the given interval. This probability corresponds to an interval that is
described by a multiple kγ of the underlying yield variance σ /Y
. Using (16)–
(19), kγ can be determined as the value that satisfies:

/Y
) − cdf(Y/ − kγ · σ
γ = cdf(Y/ + kγ · σ /Y
) → kγ (230)

kγ is determined without knowing the yield estimate and yield estimate variance
through the corresponding normalized univariate normal distribution.
For instance, a confidence level of γ = 90% denotes a probability of 90% for
the yield estimate to be within an interval of ±1.645/ σY
around the estimated
value Y/ . A confidence level of γ = 95% denotes a probability of 95% for
the yield estimate to be within an interval of ±1.960/ σY
around the estimated
value Y/ . γ = 99% corresponds to ±2.576/ σY
, and γ = 99.9% corresponds to
±3.291/ σY
.
114 ANALOG DESIGN CENTERING AND SIZING

Table 8. Required sample size for an estimation of a yield of 85% for different confidence
intervals and confidence levels according to (232).

Confidence level 90% 95% 99% 99.9%

Y
→
Confidence kγ σ ±1.645σY
±1.960σY
±2.576σY
±3.291σY

interval ↓ Y ± ∆Y
85% ± 10% 35 49 85 139
85% ± 5% 139 196 339 553
85% ± 1% 3,451 4,899 8,461 13,810

The confidence interval is now given by two equal forms:

∆Y = kγ · σ
/Y
(231)

Inserting the obtained value kγ and (231) in (225), the required sample size
nM C for the given confidence level γ → kγ and interval Y/ ± ∆Y can be
computed:

Y · (1 − Y ) · kγ2
nM C ≈ (232)
∆Y 2

For a yield of 85%, Table 8 evaluates (232) for three different confidence in-
tervals, 75% . . . 95%, 80% . . . 90%, and 84% . . . 86%, and for four different
confidence levels, 90%, 95%, 99%, and 99.9%.
Table 8 shows for instance that a Monte Carlo analysis with 8,461 sample
elements is required if we want have a 99% probability that the yield is within
±1% around its estimated value of 85%. It can be seen that thousands of sample
elements are required for a sufficient accuracy of the statistical yield estimation.
According to (225), to increase the accuracy by a factor of F , the sample size
has to be increased by a factor of F 2 . This is illustrated in Table 8, where
the increase in accuracy by a factor of 10 from ∆Y = 10% to ∆Y = 1%
corresponds to an increase in the sample size by a factor of 100.
Each sample element has to be evaluated by simulation. Simulation is com-
putationally very expensive and exceeds by far the computational cost of the
remaining operations. The computational cost of a Monte Carlo analysis is
therefore mainly determined by the number of simulations, i.e. the sample
size.
Yield Analysis 115

Overall we have for a Monte-Carlo analysis:

√
Accuracy ∼ nM C
Complexity ∼ nM C

Interestingly, the accuracy does not depend on any other quantity like for
instance the nonlinearity of the performance, and the complexity does not
depend on any other quantity like for instance the number of parameters.

6.2 Tolerance Classes

An alternative to a statistical yield analysis is a geometric approach. A certain
type of tolerance region of the statistical parameters is assumed and a maximum
size of this tolerance region type within the given parameter acceptance region
is computed. The yield is derived from the size of the tolerance region. The
specific relation between tolerance region Ts and yield Y is denoted as tolerance
class. In the following, some types of tolerance classes will be treated.
We will assume that the statistical parameters are normally distributed or
have been transformed into normally distributed parameters.

6.2.1 Tolerance Interval

If there is only one statistical parameter xs , which is normally distributed
with mean value xs,0 and variance σ 2 , xs ∼ N (xs,0 , σ 2 ), the definition of a
tolerance class is straightforward:

Ts,I = {xs | a ≤ xs ≤ b} (233)

 b
↔ YI = pdfN (xs ) · dxs
a

= cdfN (b) − cdfN (a) (234)

 βb
1 1 2
= √ e− 2 t · dt (235)
βa 2π
a − xs,0 b − xs,0
βa = , βb = , t ∼ N (0, 1)
σ σ
pdfN denotes the probability density function of the univariate normal distribu-
tion as defined in (16). cdfN denotes the corresponding cumulative distribution
function, which is defined in (17)–(19).
Based on (41), the probability density function is transformed into a proba-
bility density function with a mean value of zero and a variance of one (235).
In this way, the integration bounds a and b have been transformed into their
116 ANALOG DESIGN CENTERING AND SIZING

Table 9. Tolerance intervals Ts,I and corresponding yield values YI according to (235).

Ts,I = [βa σ, βb σ] YI
[ −∞ , −1σ ] 15.9%
Ts,I = [βa σ, βb σ] YI
[ −∞ , 0 ] 50.0%
[ −1σ , 1σ ] 68.3% [ −∞ , 1σ ] 84.1%
[ −2σ , 2σ ] 95.5% [ −∞ , 2σ ] 97.7%
[ −3σ , 3σ ] 99.7% [ −∞ , 3σ ] 99.9%

distances to the nominal value as multiples of the standard deviation. Note

that the resulting integration bounds βa , βb represent the univariate case of a
worst-case distance defined in (24), (161) or (185).
Similar to Table 5 and Figure 20, values of tolerance intervals Ts,I and
corresponding yield values YI can be taken from statistical tables. Table 9
shows some examples. We can see the familiar probabilities of a univariate
normal distribution. In the left column, symmetric intervals around the mean
value are presented. In the right column, parameter values bounded from the
right are shown. We can see that worst-case distances may be supplemented
with a negative sign in order to indicate that the mean value is outside the
tolerance interval.

6.2.2 Tolerance Box

If more than one parameter is there, the concept of intervals can be extended to
the multivariate case of a normally distributed parameter vector xs ∼ N (xs,0 , C).
This results in tolerance boxes as defined in (8). Tolerance boxes are also the
given tolerance regions in the classical worst-case analysis described in Section
5.1. Analogously to (234) and (235), the tolerance box Ts,B and corresponding
yield YB are defined:

Ts,B = {xs | a ≤ xs ≤ b} (236)

 b1  bnxs
↔ YB = ... pdfN (xs ) · dxs
a1 anxs
 
βb1 βbnxs
1
−1

= ... √ nxs exp −0.5 · t · R
T
· t · dt (237)
βa1 βanxs 2π

t = Σ−1 · (xs − xs,0 ) , t ∼ N (0, R) (238)

Yield Analysis 117

Figure 47. Tolerance box Ts,B and normal probability density function pdfN,0R with zero
mean and unity variance.

ak − xs,0,k bk − xs,0,k
βak = , βbk = (239)
σk σk
pdfN denotes the probability density function of the multivariate normal distri-
bution as defined in (23)–(29). Using the variable transformation (238) and the
decomposition (26) of the covariance matrix, we obtain parameters t that are
normally distributed each with a mean value of 0 and a variance of 1, and that
are mutually correlated with the correlation matrix R. The probability density
function of the resulting normal distribution N (0, R) is denoted as pdfN,0R .
The variable transformation (238) also transforms the integration bounds ak
and bk into their distances to their nominal values as multiples of their standard
deviations, βak and βbk .
(237) represents an integral over the probability density function pdfN,0R in
a box Ts,B . Figure 47 illustrates level contours of pdfN,0R and Ts,B for two
parameters. In this example, βak = −βbk = −βW has been chosen. The
resulting box is symmetrical around the origin.
(237) has to be solved numerically. Only if the parameters are uncorrelated,
(237) can be evaluated by using the cumulative distribution function values of
the univariate normal distribution:
nxs  βb
 k 1 1 2
R=I : YB = √ e− 2 t · dt (240)
k=1 βak
2π
118 ANALOG DESIGN CENTERING AND SIZING

Table 10. Yield values YB for a tolerance box Ts,B with ∀k βak = −βbk = −βW = −3 in
dependence of the number of parameters nxs and of the correlation. ∀k=l k,l = 0.0 according
to (241), ∀k=l k,l = 0.8 according to (237).

nxs YB |k,l =0.0 YB |k,l =0.8

3 99.1% 99.3%
4 98.8% 99.2%
5 98.5% 99.1%
6 98.2% 99%
7 97.9% 98.9%
8 97.6% 98.85%
9 97.3% 98.8%
10 97.0% 98.7%

⎫
R=I ⎪
⎬  βb nxs
1 1 2
∀k=l βak = βal = βa : YB = √ e− 2 t · dt (241)
⎪
⎭ βa 2π
∀k=l βbk = βbl = βb
Table 10 shows the yield value YB corresponding to a tolerance box Ts,B with
∀k βak = −βbk = −βW = −3. YB is given for different numbers of
parameters nxs and two different correlation values. For  = 0.0, (241) is
applied to compute YB , for k,l = 0.8, (237) is applied. The given correlation
value holds for all parameter pairs.
It can be seen that the yield YB that corresponds to a tolerance box depends
on the dimension of the parameter space and on the correlations between para-
meters. This dependence is more pronounced for smaller tolerance boxes than
the one with βW = 3 in Table 10.
From Figure 22 follows that a tolerance box represents the worst-case if
correlations are unknown. Then, a tolerance box and corresponding yield value
according to (237) seems to be appropriate. Usually, at least a qualitative
knowledge about parameters being rather uncorrelated or strongly correlated is
available. In that case, we are interested to use a tolerance class that does not
depend on the number of parameters and does not depend on the correlation
values. As the parameters or transformed parameters can be assumed to be
normally distributed, it is promising to apply the tolerance region that results
from the level contours of the probability density function, i.e. ellipsoids.

6.2.3 Tolerance Ellipsoid

The ellipsoid tolerance class reproduces the equidensity contours of normally
distributed parameters xs ∼ N (xs,0 , C), which are ellipsoids determined by
Yield Analysis 119

Figure 48. Ellipsoid tolerance region Ts,E and equidensity contours of normal probability
density function pdfN .

the quadratic form (24). Defining a tolerance class according to such an ellip-
soid, the ellipsoid tolerance region Ts,E is:

Ts,E = {xs | (xs − xs,0 )T · C−1 · (xs − xs,0 ) ≤ βW

2
} (242)

Figure 48 illustrates the ellipsoid tolerance region Ts,E for two parameters.
Ellipsoid tolerance regions are also the given tolerance regions in the realistic
worst-case analysis described in Section 5.2 and the general worst-case analysis
describe in Section 5.4.
In order to get to the corresponding yield YE , we apply the variable trans-
formation (58),

t = A−1 · (xs − xs,0 ), (243)

and the ellipsoid tolerance region becomes:

*
n +
 
xs


Ts,E = t | tT · t ≤ βW 2
= t
t2k ≤ βW
2
(244)


k=1

The transformed variable t is normally distributed with mean values of zero,

variances of one, and correlations of zero:

t ∼ N (0, I) ⇔ tk ∼ N (0, 1) , k = 1, . . . , nxs (245)

As the random variables tk are independently and identically normally dis-

tributed, β 2 = tT · t = nk=1
xs 2
tk is χ2 (chi-square)-distributed with nxs degrees
120 ANALOG DESIGN CENTERING AND SIZING

Table 11. Yield values YE for an ellipsoid tolerance region Ts,E with βW = 3 in dependence
of the number of parameters nxs according to (247) and (248).

nxs YE
2 98.9%
3 97.1%
4 93.9%
5 89.1%
6 82.6%
7 74.7%
8 65.8%
9 56.3%
10 46.8%
.. ..
. .
15 12.3%

of freedom:

nxs
β2 = t2k ∼ χ2nxs (246)
k=1

The χ2 (chi-square)-distribution has the following probability density function

and is tabulated in statistical handbooks and subroutines.
nxs 2
−1
(β 2 ) 2 · exp(− β2 )
pdfχ2n (β 2 ) = nxs (247)
xs
2 2 · Γ( n2xs )
The yield value YE corresponding to Ts,E then is defined by:
 β2
W
Ts,E ↔ YE = pdfχ2n (β 2 ) · dβ 2 (248)
xs
0

From (247) and (248) we can see that the yield YE that corresponds to an
ellipsoid tolerance region is independent of the correlation values of the nor-
mal distribution of the parameters, contrary to a box tolerance region. But
(247) and (248) also show that YE depends on the dimension of the parameter
space, nxs .
Table 11 shows the yield value YE corresponding to a tolerance box Ts,E with
βW = 3 for different number of parameters. Obviously the yield YE within
a tolerance ellipsoid determined by βW strongly decreases with an increasing
number of parameters. This property is advantageous for the sampling proper-
ties of the normal distribution, as it adapts the spreading of sample elements to
Yield Analysis 121

the dimension of the scanned parameter space. But this property is disadvan-
tageous for a tolerance class, where it is unwanted if a tolerance region refers
to different yield values depending on the number of parameters.

6.2.4 Single-Plane-Bounded Tolerance Region

The single-plane-bounded tolerance class cuts the parameter space in two
halves by a given single plane.
The motivation for this type of tolerance class is the consideration of a sin-
gle performance-specification feature. The previous tolerance class types are
restricted to reproducing the parameter distribution. Contrary to that, the single-
plane-bounded tolerance class considers both the parameter distribution and the
performance specification.
Note that the following considerations are similar to those done previously for
a worst-case analysis. Yet, there are differences and advances in the following
description.
In Sections 5.2 and 5.4, we started from a tolerance region of statistical para-
meters and computed worst-case performance-feature values. We illustrated
that the worst-case distance as an input determining the size of the tolerance
region relates to the yield partition of an individual performance feature (Sec-
tions 5.3 and 5.5). In the following, we will reverse the direction and start
from a performance-specification-feature value and relate it to a yield partition
value and corresponding size of tolerance region. While in Sections 5.2 and
5.4, usually the same tolerance region and corresponding worst-case distance is
given for all performance features, now different worst-case distances and yield
partitions will result for the individual performance-specification features. In
addition, as we start from a performance-specification feature, we will face four
cases that emanate from having a lower or upper performance-feature bound,
which may be violated or satisfied at the nominal parameter vector. Last not
least, the following description will be based on geometric considerations, while
Sections 5.2 and 5.4 were based on optimality conditions.
In the following, the case of a tolerance region that results from an upper
bound fU on a single performance feature f is described. An index i denoting
the performance feature is left out for simplicity. The single-plane-bounded
tolerance region Ts,SP,U and corresponding yield partition YSP,U are defined
by:
Ts,SP,U = {xs | gT · (xs − xs,W U ) ≤ 0} (249)
 
↔ YSP,U = ... pdfN (xs ) · dxs
Ts,SP,U
 
= ... pdfN (xs ) · dxs (250)
gT ·(xs −xs,W U )≤0
122 ANALOG DESIGN CENTERING AND SIZING

Figure 49. Single-plane-bounded tolerance region Ts,SP,U and equidensity contours of normal
probability density function pdfN .

Figure 49 illustrates a single-plane-bounded tolerance region for two para-

meters. Assuming a linear performance model,

f = fU + gT · (xs − xs,W U ) (251)

the normally distributed parameters xs ∼ N (xs,0 , C) are transformed into a

normally distributed performance feature:

f ∼ N (fU + gT · (xs,0 − xs,W U ), σf2 ) (252)

σf2 = gT · C · g (253)

From (251) follows that

f ≤ fU ⇔ gT · (xs − xs,W U ) ≤ 0 (254)

In Figure 49, xs,W U being an upper bound means that the gradient g points
from the border of Ts,SP,U away from the tolerance region Ts,SP,U .
We can formulate an equivalent form of the single-plane-bounded tolerance
region Ts,SP,U and corresponding yield partition YSP,U to (249) and (250) in
the space of the single performance feature f :

Ts,SP,U = {xs | f (xs ) ≤ fU } (255)

Yield Analysis 123
 fU
↔ YSP,U = pdff (f ) · df (256)
−∞

pdff is a normal probability density function with mean value and covariance
as given in (252) and (253).
Let us express the difference between the given bound fU and the perfor-
mance value at xs,0 , f0 = f (xs,0 ), as a multiple of the performance variance
σf2 (253):

+ βW U · σf , f0 ≤ fU
fU − f0 = g · (xs,W U − xs,0 ) ≡
T
(257)
− βW U · σf , f0 ≥ fU

(257) considers that the nominal parameter vector xs,0 can be inside or outside
of the tolerance region Ts,SP,U according to (249). If xs,0 is outside of Ts,SP,U ,
then f0 > fU and the difference is negative.
(256) can be reformulated for a standardized normal distribution with zero
mean and unity variance using (257):
⎧  βW U
⎪
⎪ 1 − 12 t2
⎨ −∞ √2π e
⎪ · dt , f0 ≤ fU
YSP,U = (258)
⎪  − βW U 1
⎪
⎪
⎩ − 1 2
√ e 2 · dt , f0 ≥ fU
t
−∞ 2π
The equivalence of (249) and (255) means that a single-plane-bounded tolerance
region Ts,SP,U corresponds to the acceptance region of a single performance-
specification feature, i.e. a single bound on a performance feature.
The equivalence of (250) and (258) means that the corresponding yield parti-
tion YSP,U can be evaluated based on the standardized normal distribution with
zero mean and unity variance.
The same considerations can be done for a lower bound on a performance
feature, starting from the corresponding formulation of the tolerance region,
Ts,SP,L = {xs | gT · (xs − xs,W L ) ≥ 0}. The resulting yield partition is:
⎧  βW L
⎪
⎪ 1 1 2
√ e− 2 t · dt , f0 ≥ fL
⎪
⎨ −∞ 2π
YSP,L =  − βW L (259)
⎪
⎪ 1
⎪
⎩ − 1 2
√ e 2 · dt , f0 ≤ fL
t
−∞ 2π
From (258) and (259) follows that the worst-case distance gets a positive sign,
if the performance-specification feature is satisfied at the nominal parameter
vector, and gets a negative sign, if the performance-specification feature is
violated at the nominal parameter vector. Figure 50 illustrates the four cases
124 ANALOG DESIGN CENTERING AND SIZING

Figure 50. Single-plane-bounded tolerance regions Ts,SP,L , Ts,SP,U for a single performance
feature with either an upper bound (first row) or a lower bound (second row), which is either
satisfied (first column) or violated at the nominal parameter vector (second column).

that result from a lower or upper performance-feature bound violated or satisfied

by the nominal design.
Table 12 shows selected yield partition values YSP for some selected single-
plane-bounded tolerance regions TSP regardless if it originates from a lower
or upper performance-feature bound.
Obviously the single-plane-bounded tolerance class provides yield partition
values that can easily be evaluated using statistical tables and functions and that
additionally are independent of the number of parameters and of the variances
and correlations.
Yield Analysis 125

Table 12. Single-plane-bounded tolerance region Ts,SP and corresponding yield partition val-
ues YSP according to (258) or (259).

Ts,SP = [ −∞ , βW σf ] YSP
[ −∞ , −1σf ] 15.9%
[ −∞ , 0 ] 50.0%
[ −∞ , 1σf ] 84.1%
[ −∞ , 2σf ] 97.7%
[ −∞ , 3σf ] 99.9%

The single-plane-bounded tolerance class is suitable if the performance spe-

cification is partitioned into the individual performance-specification features,
which leads to a partitioning of the parameter acceptance region as illustrated
in Figure 32.
It can also be shown that the single-plane-bounded tolerance region defined
by βW determines a corresponding ellipsoid tolerance region. This is done in
the following.
Figure 49 indicates that there is an equidensity contour of the parameters’
probability density function that touches the plane (249) at a certain point, for
instance at xW . This means that the orthogonal on the quadratic form (24) at
xW is parallel to g:

−1 + λ · g , f0 ≤ fU
C · (xs,W U − xs,0 ) = (260)
− λ · g , f0 ≥ fU
Inserting xs,W U − xs,0 from (260) in the equivalence part of (257) results in:
βW U
λ= (261)
σf
Solving (260) for xs,W U − xs,0 and inserting λ from (261), and inserting the
resulting expression for xs,W U − xs,0 into (24) results in

(xs,W U − xs,0 )T · C−1 · (xs,W U − xs,0 ) = βW

2
U (262)
(262) shows that the βW -multiple of the difference between nominal perfor-
mance and performance bound corresponds to the tolerance ellipsoid touching
the given plane.
From Sections 5.2 and 5.5, it is known that the realistic and general worst-
case analysis, if they are based on the single-plane-bounded tolerance class,
126 ANALOG DESIGN CENTERING AND SIZING

Figure 51. Single-plane-bounded tolerance region Ts,SP with corresponding worst-case para-
meter vector xW,SP and worst-case distance βW for two parameters. βW denotes a ±βW times
the covariances tolerance ellipsoid. Tolerance box Ts,B determined by ±βW times the covari-
ances with corresponding worst-case parameter vector xW,B and worst-case distance βW,B .

lead to a high accuracy of the approximated yield partition value of a single

performance-specification feature. Among the mentioned tolerance classes, the
yield value of a single performance-specification feature fi ≥ fL,i or fi ≤ fU,i
is therefore geometrically optimally approximated based on a single-plane-
bounded tolerance region, i.e. according to Table 12.

6.2.5 Corner Worst Case vs. Realistic Worst Case

Corner parameter vectors obtained from a classical worst-case analysis on the
other hand result in exaggerated robustness estimations. Figure 51 illustrates the
resulting worst-case parameter vectors and yield approximations of a classical
worst-case analysis and a realistic worst-case analysis.
A classical worst-case analysis is based on a tolerance box of parameters, the
corresponding quantities in Figure 51 therefore have the index B. A realistic
worst-case analysis is based on a single-plane-bounded tolerance region, the
corresponding quantities in Figure 51 therefore have the index SP .
Both types of worst-case analysis are based on a tolerance region refer-
ing to the same multiple of the respective parameter variances. The realistic
Yield Analysis 127

Table 13. Exaggerated robustness βW,B represented by a corner worst-case parameter vector
of a classical worst-case analysis, for different correlation ∀k=l k,l =  among the parameters,
and for different numbers of parameters nxs .

→ 0 0.3 0.6 0.9

↓ nxs
2 4.2σf 5.1σf 6.7σf 13.4σf
3 5.2σf 6.1σf 7.8σf 15.5σf
4 6.0σf 7.2σf 9.5σf 19.0σf
βW = 3σf

worst-case analysis starts from an ellipsoid determined by βW . This ellipsoid

is within a range of ±βW · σk around the nominal parameter values x0,k . The
combination of these intervals of the individual parameters determines the tol-
erance box Ts,B of a classical worst-case analysis. Note that Ts,B holds for any
correlation according to (37) and as illustrated in Figure 22(d).
Due to these initial tolerance regions, the resulting worst-case parameter vec-
tor xW,B of a classical worst-case analysis will always refer to larger ellipsoid
and worst-case distance than the resulting worst-case parameter vector xW,SP
of a realistic worst-case analysis:

βW,B ≡ β(xs,W,B ) ≥ βW ≡ β(xs,W,SP ) (263)

The robustness that is represented by a corner worst case therefore is always

more exaggerated than that represented by a realistic worst case. The degree of
exaggeration depends on the number of parameters, on the correlation and on
the correlation in relation to the performance sensitivity.
Table 13 illustrates how exaggerated the robustness represented by the corner
worst case can be. The underlying classical worst-case analysis has a tolerance
box ±3σk around the nominal parameter vector, which refers to a worst-case
distance of βW = 3 and a yield of YW = 99.9% in a realistic worst-case
analysis.
The entry in the last row and last column for instance says that for 4 para-
meters and a correlation of 0.9 among the parameters, the corner worst case
obtained from a ±3σk may correspond to a safety margin of βW,B = 19.0σ,
which is excessive.
128 ANALOG DESIGN CENTERING AND SIZING

6.3 Geometric Yield Analysis

Unlike a statistical yield analysis, which is based on statistical sampling of
the parameter space, a geometric yield analysis is based on an approximation
of the parameter acceptance region.
Ageometric approximation of the parameter acceptance region is more easy if
a partitioning of the performance specification into the individual performance-
specification features is done. This partitioning of the performance specification
according to lower and upper bounds of individual performance features (123)–
(125) has been illustrated in Figure 32.
As before, we assume that the statistical parameters are normally distributed
or have been transformed into normally distributed parameters.

6.3.1 Problem Formulation

We can distinguish between four cases of a geometric yield analysis, which
arise because either a lower or an upper bound on the performance feature is
given, which is either satisfied or violated at the nominal parameter vector.
Figure 52 picks the lower bound fL,1 of performance f1 in Figure 32 and
illustrates the geometric analysis for the case that a lower bound fL on a per-
formance feature f is given, and that the nominal design xs,0 is inside the
parameter acceptance region partition. To simplify the description, the index i
denoting the ith performance feature will be left out.
Figure 52 shows the border of the parameter acceptance region partition
As,L , which is generally nonlinear. This nonlinear border cannot be formu-
lated analytically due to the nature of performance functions in analog design.
An approximation of the border shall be formulated that determines its most
important point and the corresponding performance gradient.
As a larger ellipsoid corresponds to a smaller probability density value in
Figure 52, it becomes visible that the statistical parameter vectors outside of
the acceptance region partition As,L or just on the border of As,L differ from
each other concerning their probability density value. Among all points outside
or on the border of the acceptance region, there exists a point that has a maxi-
mum probability density value. The infinitesimal region around this parameter
vector has therefore the highest probability of occurrence among all parameter
vectors on the border. We call this distinguished parameter vector “worst-case
parameter vector” and denote it as xs,W L .
If the nominal design would violate the lower bound on the performance
feature, then the center xs,0 of the ellipsoidal equidensity contours in Figure
52 would be outside of the parameter acceptance region partition, that means
on the non-gray side. In this case as well, there is a distinguished parameter
vector with highest probability of occurrence, but this time among all parameter
vectors inside or on the border of the parameter acceptance region partition.
Yield Analysis 129

Figure 52. Geometric yield analysis for a lower bound on a performance feature, f > fL ,
which is satisfied at the nominal parameter vector. Worst-case parameter vector xs,W L has the
smallest distance from the nominal statistical parameter vector xs,0 measured according to the
equidensity contours among all parameter vectors outside of or on the border of the acceptance
region partition As,L . The tangential plane to the tolerance ellipsoid through xs,W L as well as
to the border of As,L at xs,W L determines a single-plane-bounded tolerance region As,L .

Therefore, a basic formulation of a geometric yield analysis is to compute

the worst-case statistical parameter vector among all parameter vectors which
are outside or on the border of the acceptance region partition, if the nominal
parameter vector satisfies the considered performance-specification feature (i.e.
is inside the acceptance region partition), or which are inside or on the border
of the acceptance region partition, if the nominal parameter vector violates the
considered performance-specification feature (i.e. is outside the acceptance
region partition):

xs,0 ∈ As,L/U : max pdfN (xs ) s.t. xs ∈ Ās,L/U (264)

xs

xs,0 ∈ Ās,L/U : max pdfN (xs ) s.t. xs ∈ As,L/U (265)

xs

Generally, (264) and (265) may have several solutions. In practice, a unique
solution as illustrated in Figure 52 can be observed for most performance fea-
tures. Exceptions are for instance mismatch-sensitive performance features for
which (264) and (265) often lead to two worst-case parameter vectors for each
mismatch-producing transistor pair, which are nearly symmetrical on both sides
of the nominal parameter vector. This special case requires special solution
algorithms.
Due to the presence of range-parameter tolerances, the constraints in (264)
and (265) have to be worked out in more detail. According to (123) and (124),
130 ANALOG DESIGN CENTERING AND SIZING

the parameter acceptance region partition of a performance-specification feature

is defined as the set of those statistical parameter vectors, for which all range-
parameter vectors satisfy the corresponding bound. For a lower performance-
feature bound, this means that even the smallest performance-feature value that
is obtained over all range-parameter vectors in their tolerance region must be
greater than the lower performance-feature bound. For an upper performance-
feature bound, it means that even the largest performance-feature value that is
obtained over all range-parameter vectors in their tolerance region must be less
than the upper performance-feature bound:
f ≥ fL : xs ∈ As,L ⇔ min f (xs , xr ) ≥ fL (266)
xr ∈Tr

f ≤ fU : xs ∈ As,U ⇔ max f (xs , xr ) ≤ fU (267)

xr ∈Tr

The parameter non-acceptance region partition of a performance-specification

feature is complementary to the acceptance region and was defined as the
set of those statistical parameter vectors, for which a range-parameter vec-
tor exists that violates the corresponding bound (131) and (132). For a lower
performance-feature bound, this means that already the smallest performance-
feature value that is obtained over all range-parameter vectors in their tolerance
region is less than the lower performance-feature bound. For an performance-
feature upper bound, it means that already the largest performance-feature value
that is obtained over all range-parameter vectors in their tolerance region is
greater than the upper performance-feature bound:
f ≥ fL : xs ∈ Ās,L ⇔ min f (xs , xr ) < fL (268)
xr ∈Tr

f ≤ fU : xs ∈ Ās,U ⇔ max f (xs , xr ) > fU (269)

xr ∈Tr

(266)–(269) are inserted into (264) and (265) to produce the problem formu-
lation of geometric yield analysis for the four cases of a lower/upper bound
that is satisfied/violated at the nominal statistical parameter vector. At the same
time, we replace the maximization of the probability density function by the
equivalent problem of minimizing β, which determines the equidensity contour
according to (24).
f ≥ fL and xs,0 ∈ As,L : min β 2 (xs ) s.t. min f (xs , xr ) ≤ fL (270)
xs ,xr xr ∈Tr

f ≥ fL and xs,0 ∈ Ās,L : min β 2 (xs ) s.t. min f (xs , xr ) ≥ fL (271)

xs ,xr xr ∈Tr

f ≤ fU and xs,0 ∈ As,U : min β 2 (xs ) s.t. max f (xs , xr ) ≥ fU (272)

xs ,xr xr ∈Tr

f ≤ fU and xs,0 ∈ Ās,U : min β 2 (xs ) s.t. max f (xs , xr ) ≤ fU (273)

xs ,xr xr ∈Tr
Yield Analysis 131

β 2 (xs ) = (xs − xs,0 )T · C−1 · (xs − xs,0 ) (274)

β represents the distance of a statistical parameter vector from the nominal

parameter vector, measured as a weighted l2 -norm according to the equidensity
contours of pdfN . The solution of (270)–(273) leads to worst-case distances
βW L/U , which are the basis of a geometric approximation of the yield partitions
and yields analogous to Sections 5.5 and 6.2.4.
(270)–(273) present the problem formulation of geometric yield analysis in
the form of two nested nonlinear optimization problems. The outer optimization
problem formulates the computation of the statistical parameter vector that can
be seen from the nominal parameter vector on the other side of the border of
the acceptance region partition and that has minimum weighted distance. The
objective function is quadratic in the statistical parameters.
The inner optimization problem considers the range parameters that deter-
mine the border of the acceptance region in the space of statistical parameters.
It is a single constraint with a nonlinear objective function. In what follows
we assume the usual box constraints for the range parameters (8). Then, the
inner optimization problem corresponds to a classical worst-case analysis as
described in Section 5.1. The difference to a classical worst-case analysis is
the nonlinear performance function in the constraints of (270)–(273).
An analytical solution of problems (270)–(273) cannot be formulated. The
solution is computed based on numerical optimization, for instance with a
deterministic approach based on Sequential Quadratic Programming. In the
following, we will prepare a solution approach and properties of the solution.

6.3.2 Lagrangian Function

The Lagrangian functions of (270)–(273) will be formulated according to
Appendix C. As there are two encapsulated optimization problems, we will
formulate two encapsulated Lagrangian functions.
In (270)–(273), the inner worst-case analysis problem minimizes the perfor-
mance feature if a lower bound is specified, and maximizes the performance
feature if an upper bound is specified. This leads to two Lagrangian functions
for the inner optimization problems of (270)–(273):

LI,L (xs , xr , λ1 , λ2 ) = f (xs , xr )

−λT1 · (xr − xr,L ) − λT2 · (xr,U − xr ) (275)
LI,U (xs , xr , λ1 , λ2 ) = −f (xs , xr )
−λT1 · (xr − xr,L ) − λT2 · (xr,U − xr ) (276)

As mentioned, we have assumed box constraints for the range parameters (8).
(275) applies in (270) and (271), (276) applies in (272) and (273). The negative
132 ANALOG DESIGN CENTERING AND SIZING

sign of the performance function in (276) results from replacing “max f ” in

(272) and (273) by “− min −f .” The Lagrangian functions of (270)–(273) can
now be formulated as follows:

f ≥ fL and xs,0 ∈ As,L : L(xs , xr , λ, λ1 , λ2 ) (277)

= β 2 (xs ) − λ · (fL − LI,L (xs , xr , λ1 , λ2 ))
f ≥ fL and xs,0 ∈ Ās,L : L(xs , xr , λ, λ1 , λ2 ) (278)
= β 2 (xs ) + λ · (fL − LI,L (xs , xr , λ1 , λ2 ))
f ≤ fU and xs,0 ∈ As,U : L(xs , xr , λ, λ1 , λ2 ) (279)
= β (xs ) + λ · (fU + LI,U (xs , xr , λ1 , λ2 ))
2

f ≤ fU and xs,0 ∈ Ās,U : L(xs , xr , λ, λ1 , λ2 ) (280)

= β (xs ) − λ · (fU + LI,U (xs , xr , λ1 , λ2 ))
2

6.3.3 First-Order Optimality Condition

We will formulate the optimality conditions of (270) and (277) respectively.
The other three cases can be formulated analogously.
The first-order optimality condition describing a stationary point of (277) and
a solution of (270), xs,W L , xr,W L , λW L , λ1,W L , λ2,W L and βW L = β(xs,W L )
is:

f ≥ fL and xs,0 ∈ As,L :

2C−1 · (xs,W L − xs,0 ) + λW L · ∇LI,L (xs,W L ) = 0 (281)

  
∇f (xs,W L )

λW L · (fL − LI,L (xs,W L , xr,W L , λ1,W L , λ2,W L )) = 0 (282)

  
f (xs,W L , xr,W L )

min f (xs,W L , xr ) ≤ fL (283)

xr ∈Tr

∇LI,L (xr,W L )
  
λW L · (∇f (xr,W L ) − λ1,W L + λ2,W L ) = 0 (284)
k = 1, . . . , nxr : λ1,W L,k · (xr,W L,k − xr,L,k ) = 0 (285)
k = 1, . . . , nxr : λ2,W L,k · (xr,U,k − xr,W L,k ) = 0 (286)
xr,L ≤ xr,W L ≤ xr,U (287)
Yield Analysis 133

(281) results from the condition that ∇L(xs ) = 0 must hold for a stationary
point of the outer optimization problem. (284) results from the condition that
∇L(xr ) = 0 must hold for a stationary point of the inner optimization problem.
(282) represents the complementarity condition of the outer optimization
problem, (285) and (286) represent the complementarity condition of the inner
optimization problem.
(283) is the constraint of the outer optimization problem, (287) is the con-
straint of the inner optimization

problem.

∂f
∂f

∇f (xs,W L ) = ∂xs
and ∇f (xr,W L ) = ∂x
hold.
xs,W L ,xr,W L r xs,W L ,xr,W L
Note that (281) and (282) are very similar to the first-oder optimality condi-
tion of the general worst-case analysis (190) and (191). And that (284)–(286)
have nearly the same form as the first-oder optimality condition of the classical
worst-case analysis (153) and (155).
We can verify by contradiction that the constraint (283) is active at the
solution, i.e. f (xs,W L , xr,W L ) = fL . If this constraint would be inactive
at the solution, then λW L = 0. Using (281), then xs,W L = xs,0 would be
the solution. It follows that f (xs,W L , xr,W L ) = f (xs,0 , xr,W L ) < fL would
hold, which is a contradiction to the initial situation that the nominal statistical
parameter vector is satisfying the performance-feature bound. Therefore, the
constraint (283) is active at the solution in all cases (277)–(280), i.e.:

λW L/U > 0 (288)

f (xs,W L/U , xr,W L/U ) = fL/U (289)

6.3.4 Second-Order Optimality Condition

The second-order optimality condition requires the Hessian matrix of the
Lagrange function ∇2 L(xs,W L , xr,W L ). ∇2 L(xs,W L , xr,W L ) of (270) and
(277) respectively in the case f ≥ fL and xs,0 ∈ As,L is:
0 ∂2L ∂2L
1
∂x2s ∂xs ∂xr
∇2 L(xs,W L , xr,W L ) = ∂2L ∂2L
∂xs ∂xr ∂x2r xs,W L ,xr,W L
⎡
⎤
2f

2C−1 + λW L · ∇2 f (xs,W L ) λW L · ∂x∂s ∂x

⎢
r x
s,W L ,xr,W L ⎥
=⎣ ∂2f
⎦ (290)
λW L · ∂xs ∂xr
λW L · ∇ f (xr,W L )
2
xs,W L ,xr,W L

2
Nonnegative curvature of ∂∂xL2 at the worst-case parameter vector for uncon-
r
strained directions describes that the performance function is bounded below
with respect to the range parameters. This is required for the existence of a
border of the acceptance region As as defined in (123).
134 ANALOG DESIGN CENTERING AND SIZING

Figure 53. Two stationary points xs,A and xs,W of (277), which satisfy the first-order opti-
mality condition. Only xs,W satisfies the second-order optimality condition and is therefore a
solution of (270).

2
Nonnegative curvature of ∂∂xL2 at the worst-case parameter vector for
s
unconstrained directions corresponds to the relation between the curvature of
the tolerance ellipsoid of the statistical distribution and the curvature of the
border of the acceptance region. For xs,W L to be a minimum of (270), the cur-
vature of the tolerance ellipsoid has to be stronger than the curvature of the
border of the acceptance region as in the example in Figure 52.
Figure 53 shows another example. Both parameter vectors xs,A and xs,W L
satisfy the first-order optimality condition, but only xs,W L satisfies the second-
order optimality condition as well. In xs,A , the curvature of the border of the
acceptance region is stronger than the curvature of the corresponding tolerance
region. Therefore, this tolerance region is not a subset of the acceptance region,
and there are points inside this tolerance region but outside of the acceptance
region for which the probability density value or the corresponding β value is
larger than at xs,A . Therefore, xs,A is not a solution of (270).
Note that there is a third stationary point of (277) in Figure 53, where the
thin circle touches the border of the acceptance region. This point satisfies the
second-order optimality condition and is a local minimum of (270). A similar
situation will occur for locally varying mismatch-producing parameters. From
a deterministic optimizer starting from xs,0 can be expected that it will find
the global minimum xs,W L , as xs,W L is closer to xs,0 due to the problem
formulation. But as there is no guaranty for that and as we might be interested
to know all local minima, suitable measures have to be taken.
Yield Analysis 135

6.3.5 Worst-Case Range-Parameter Vector

We can formulate the worst-case range-parameter vector xr,W L based on
(284)–(287). Due to (288), λW L can be eliminated from (284). (284)–(286)
then have the same form as (153)–(155) of a classical worst-case analysis.
If the worst-case range-parameter vector is in a corner of the tolerance region
Tr , it can therefore be formulated by replacing the performance gradients at
the nominal parameter vector with the performance gradients at the worst-
case parameter vector in the corresponding equations (158) and (159) from the
worst-case analysis in Section 5.1. The components of the worst-case parameter
vector xr,W L/U = [ . . . xr,W L/U,k . . . ]T then are determined by:

xr,L,k , ∇f (xr,W L,k ) > 0
xr,W L,k = (291)
xr,U,k , ∇f (xr,W L,k ) < 0

xr,U,k , ∇f (xr,W U,k ) > 0
xr,W U,k = (292)
xr,L,k , ∇f (xr,W U,k ) < 0
(291) and (292) can be applied to save computational cost of the iterative
solution of (270)–(273). Towards that, the worst-case range-parameter vector
is initialized based on (291), (292) and on the performance-feature gradient at
the nominal parameter vector. This initial worst-case range-parameter vector
is not included in the Sequential-Quadratic-Programming solution of (270)–
(273). A monitoring of the gradients with regard to the range parameters is
applied instead to iteratively update the worst-case range-parameter vector. It
has turned out in practice that this partitioning of the solution of (270)–(273)
leads to a saving in computational cost due to the often plain behavior of the
performance function with respect to range parameters.

6.3.6 Worst-Case Statistical Parameter Vector

From (281) follows in the case f ≥ fL and xs,0 ∈ As,L :
λW L
xs,W L − xs,0 = − · C · ∇f (xs,W L ) (293)
2
Inserting (293) in (274) leads to:

λ2W L
· ∇f (xs,W L )T · C · ∇f (xs,W L ) = βW
2
L
4
λW L βW L
=- (294)
2 ∇f (xs,W L )T · C · ∇f (xs,W L )
Inserting (294) in (293) yields the formulation of the statistical worst-case para-
meter vector for the case of a lower performance-feature bound that is satisfied
136 ANALOG DESIGN CENTERING AND SIZING

at the nominal parameter vector. The other cases are obtained analogously,
starting from the Lagrangian functions (277)–(280).

f ≥ fL and xs,0 ∈ As,L , f ≤ fU and xs,0 ∈ Ās,U : (295)

−βW L/U
xs,W L/U − xs,0 = . · C · ∇f (xs,W L/U )
∇f (xs,W L/U )T ·C·∇f (xs,W L/U )

f ≥ fL and xs,0 ∈ Ās,L , f ≤ fU and xs,0 ∈ As,U : (296)

+βW L/U
xs,W L/U − xs,0 = . · C · ∇f (xs,W L/U )
∇f (xs,W L/U )T ·C·∇f (xs,W L/U )

Note that (295), (296), which describe the worst-case statistical parameter vec-
tors of a geometric yield analysis, are identical to (194), (196), which describe
the worst-case statistical parameter vectors of a general worst-case analysis.

6.3.7 Worst-Case Distance

The worst-case statistical parameter vector xs,W L/U describes a tolerance
ellipsoid according to (274):
T −1
L/U = β (xs,W L/U ) = (xs,W L/U − xs,0 ) C (xs,W L/U − xs,0 ) (297)
2 2
βW

The worst-case distance βW 2

L/U determines the maximal tolerance ellipsoid of
statistical parameters that touches the boundary of the parameter acceptance
region for the considered performance-specification feature, i.e. performance-
feature bound. In Figure 52, this tolerance ellipsoid is marked with an increased
line thickness.
The linearization of the performance-feature function at the worst-case para-
meter vector is:

f¯(W L/U ) (xs ) = fL/U + ∇f (xs,W L/U )T · (xs − xs,W L/U ) (298)

From (298) we obtain:

−∇f (xs,W L/U )T · (xs,W L/U − xs,0 ) = f¯(W L/U ) (xs,0 ) − fL/U (299)

Inserting (295) or (296) in (299) leads to

f ≥ fL and xs,0 ∈ As,L , f ≤ fU and xs,0 ∈ Ās,U :

f¯(W L/U ) (xs,0 ) − fL/U
βW L/U = . (300)
∇f (xs,W L/U )T · C · ∇f (xs,W L/U )
Yield Analysis 137

∇f (xs,W L/U )T · (xs,0 − xs,W L/U )

=
σf¯(W L/U )

f ≥ fL and xs,0 ∈ Ās,L , f ≤ fU and xs,0 ∈ As,U :

fL/U − f¯(W L/U ) (xs,0 )
βW L/U = . (301)
∇f (xs,W L/U )T · C · ∇f (xs,W L/U )

∇f (xs,W L/U )T · (xs,W L/U − xs,0 )

=
σf¯(W L/U )

Note that (300) and (301), which describe the worst-case distance from a
geometric yield analysis, are identical to the worst-case distance from a single-
plane-bounded tolerance region (257), and are identical to the worst-case
distance from a general worst-case analysis (200)–(203).
In all cases, a worst-case distance is defined as a multiple of a performance
standard deviation σf¯(W L) . Specifically, it is the standard deviation of the
linearized performance at the worst-case parameter vector.
Moreover, (300) and (301) show that a change in the worst-case distance
consists of two parts. On the one hand, the distance between the performance-
feature bound and the nominal performance value has to be changed according
to the nominator in (300) and (301). This corresponds to performance centering
as described in Sections 2.4 and 2.5. On the other hand, the performance
sensitivity with regard to the statistical parameters has to be changed according
to the denominator in (300) and (301). The appropriate combination of both
parts constitutes yield optimization/design centering. Note that we are aiming
at increasing the worst-case distance if the nominal parameter vector is inside
the parameter acceptance region partition, and that we are aiming at decreasing
the worst-case distance if the nominal parameter vector is outside the parameter
acceptance region partition.

6.3.8 Geometric Yield Partition

Based on the worst-case parameter vector according to (295), (296) and
the performance-feature function linearized at the worst-case parameter vector
(298), an approximation As,L/U,i of the parameter acceptance region partition
(123), (124) is given (Figure 52).
This approximation corresponds to the single-plane-bounded tolerance class
according to (258) and (259). Therefore, the worst-case distance either accord-
ing to (297), using the worst-case parameter vector according to (295), (296),
or according to (300), (301), using the performance-feature function linearized
at the worst-case parameter vector (298), is directly related to the yield partition
138 ANALOG DESIGN CENTERING AND SIZING

for one performance-specification feature:

⎧  β
⎪
⎪
W L/U,i 1 1 2
⎪
⎪ √ e− 2 t · dt , fi (xs,0 ) ∈ As,L/U,i
⎨ −∞ 2π
ȲL/U,i = (302)
⎪  − βW L/U,i 1
⎪
⎪
⎪ 1 2
⎩ √ e− 2 t · dt , fi (xs,0 ) ∈ As,L/U,i
−∞ 2π
As,L/U,i are defined according to (123) and (124).
The corresponding yield values can be obtained from statistical tables or
functions. Some values are given in Table 12. The accuracy of this geometric
yield partition has been described for the general worst-case analysis in Section
5.5. In summary, the following can be stated about the accuracy of the geometric
yield approximation with respect to one performance-specification feature:
According to practical experience, the absolute yield error is around 1%–3%.
The larger the actual yield value is, the smaller is the approximation error.
The weaker the curvature of the true border of the specification-feature
parameter acceptance region is compared to the curvature of the equiden-
sity contours of the parameter distribution, the higher is the approximation
quality.
The worst-case distance inherently is suitable for yield values beyond 99.9%.
Among all tangential planes of the parameter acceptance region partition,
the one in the worst-case parameter vector provides a greatest lower bound
or least upper bound on the yield partition approximation.

6.3.9 Geometric Yield

A geometric yield analysis is done for each individual performance-specifica-
tion feature. As a result, for each individual performance-specification feature,
a worst-case parameter vector, a worst-case distance, a performance-feature
function linearization and an approximate yield partition value are obtained.
For an operational amplifier like that in Figure 3, Table 14 shows five perfor-
mance features with an either lower or upper bound, the nominal performance
feature values and the worst-case distances for each of the five performance-
specification features, which result from a geometric yield analysis.
While it is not possible to judge whether a performance safety margin of 11dB
for the gain is better or worse than a performance safety margin of 37M Hz for
the transit frequency, the worst-case distances computed through a geometric
yield analysis immediately show that a worst-case distance of 2.5 for the gain
means less robustness than a worst-case distance of 7.7 for the transit frequency.
Figure 54 shows the worst-case distances as a chart. The left axis in Figure 54
Yield Analysis 139

Table 14. Geometric yield analysis of an operational amplifier.

Performance Specification Nominal Worst-case

feature feature performance distance
Gain ≥ 65dB 76dB 2.5
Transit frequency ≥ 30M Hz 67M Hz 7.7
Phase margin ≥ 60◦ 68◦ 1.8
Slew rate ≥ 32V /µs 67V /µs 6.3
DC power ≤ 3.5µW 2.6µW 1.1
Ȳ = 82.9%

Figure 54. Worst-case distances and approximate yield values from a geometric yield analysis
of the operational amplifier from Table 14.
140 ANALOG DESIGN CENTERING AND SIZING

Figure 55. Parameter acceptance region As (gray area) originating from four performance-
specification features, f1 ≥ fL,1 , f1 ≤ fU,1 , f2 ≥ fL,2 , f2 ≤ fU,2 (Figure 32). A geometric
yield analysis leads to four worst-case parameter vectors xW L,1 , xW U,1 , xW L,2 , xW U,2 and
four single-plane-bounded tolerance regions. The intersection of these single-plane-bounded
tolerance regions forms the approximate parameter acceptance region As (linen-pattern-filled
area).

is scaled according to the worst-case distances. The unit is σf¯, as a worst-case

distance represents the performance safety margin as a multiple of the standard
deviation of the linearized performance feature (298). The right axis in Figure
54 is scaled according to the yield partition value (302).
Figure 54 illustrates that the performance features gain, phase margin and
DC power have relatively small worst-case distances leading to yield losses.
The smallest worst-case distance with 1.1 is that of the DC power. Design
centering obviously has to increase all these worst-case distances as much as
possible, especially the smallest one.
The overall yield cannot be larger than the smallest yield value of the perform-
ance-specification features. The overall yield can be approximated by using the
intersection of the parameter acceptance regions of the individual performance-
specification features. This is illustrated in Figure 55. A Monte-Carlo analysis
using the approximate parameter acceptance region As can be performed at no
additional simulation cost and is therefore very fast. The yield is approximated
as given in the last row of Table 14. This value has been confirmed with a 2%
accuracy by a Monte-Carlo analysis based on numerical simulation.
Yield Analysis 141

Practical experience shows that a geometric yield analysis requires k  =

1 . . . 7 iterative steps of an SQP-based solution algorithm. In each iteration step,
a sensitivity analysis has to be performed, whose complexity is proportional to
the number of parameters nxs . As sensitivity analyses have to be done for each
performance-specification feature, the total number of analyses is proportional
to the number of performance features nf .
Overall we have for a geometric yield analysis:
<
Accuracy ∼ 3%
Complexity ∼ k · nf · nxs

Compared to a Monte-Carlo analysis, the accuracy depends on the curvature of

the performance function in the worst-case parameter vector.
The complexity depends on the number of parameters and the number of
performance-specification features.
For problems with around 50 parameters, a whole yield optimization/design
centering process based on geometric yield analyses can be done at the cost of
one Monte-Carlo analysis.

6.3.10 General Worst-Case Analysis/Geometric Yield Analysis

In Sections 4.7 and 4.8, the input and output quantities of a worst-case ana-
lysis (see Figure 28 and (112)) and a yield analysis (see Figure 33 and (140))
have been discussed. In Sections 5.4 and 6.3, the general worst-case analysis
and the geometric yield analysis have been formulated and analyzed.
Based on the probability density function of statistical parameters, on a toler-
ance region of range parameters and on a nominal parameter vector, the general
worst-case analysis basically maps a worst-case distance onto worst-case per-
formance values:

General worst-case analysis: βW → fW L/U,i , i = 1, . . . , nf (303)

The worst-case distance corresponds to a minimum yield value based on the

single-plane-bounded tolerance class (Section 6.2.4). The geometric yield ana-
lysis has the same basis as the general worst-case analysis and basically maps
performance-specification features onto worst-case distances:

Geometric yield analysis: fW L/U,i → βW L/U,i , i = 1, . . . , nf (304)

The obtained worst-case distances are transformed into performance-

specification-feature yield values based on single-plane-bounded tolerance class
(Section 6.2.4). Both the general worst-case analysis and the geometric yield
analysis additionally compute worst-case statistical parameter vectors and worst-
case range-parameter vectors. Obviously, the general worst-case analysis is the
142 ANALOG DESIGN CENTERING AND SIZING

Figure 56. General worst-case analysis and geometric yield analysis as inverse mappings
exchanging input and output.

inverse mapping of the geometric yield analysis. This mapping is bijective if

the worst-case parameter vectors are unique. Figure 56 illustrates the inverse
character of the two tasks, where the output worst-case performance values of a
general worst-case analysis turn into input performance-specification features
of a geometric yield analysis.

6.3.11 Approximate Geometric Yield Analysis

An iterative deterministic solution algorithm for a geometric yield is start-
ing from a sensitivity analysis of the performance features with regard to the
parameters at the nominal parameter vector.
Yield Analysis 143

An approximate geometric yield analysis could be performed by restriction

to this first step. This compares to a general worst-case analysis, where the first
step corresponds to a realistic worst-case analysis.

6.4 Exercise
Given is the example of Section 5.6 with a single performance function of
two parameters:
f = xs,1 · xs,2 (305)
f could be for instance the time constant of the RC circuit in Chapter 2. The
nominal parameter vector is:
xs,0 = [ 1 1 ]T (306)
The parameters are normally distributed with the covariance matrix:
 
0.22 0
C= (307)
0 0.82
Two performance-specification features are given:
f ≥ fL ≡ 0.5 (308)
f ≤ fU ≡ 2.0 (309)
Perform a geometric yield analysis for the two performance-specification
features. Apply the optimality conditions (Appendix C) to calculate a
solution.
Given is the following single performance function of two parameters:
1
f = x2s,1 · x2s,2 (310)
4
The nominal parameter vector is:
xs,0 = [ 0 0 ]T (311)
The parameters are normally distributed with the covariance matrix:
 
1 0
C= (312)
0 1
One performance-specification feature is given:
f ≤ fU ≡ 1.0 (313)
Perform a geometric yield analysis. Apply the optimality conditions
(Appendix C) to calculate a solution. Check the positive definiteness of
∇2 L(xr,W L ) and the second-order optimality condition (290) to verify the
solution.
Chapter 7

YIELD OPTIMIZATION/DESIGN CENTERING

In this chapter, the problem formulation of yield optimization/design center-

ing of Section 4.8.5 will be further developed. Two basic directions to approach
yield optimization/design centering will be explained.
The first approach is based on a statistical yield analysis with a Monte-Carlo
analysis described in Section 6.1. The gradient and the Hessian matrix of the
statistically determined yield with regard to the nominal values of statistical
parameters will be derived [8, 11, 10]. The statistical yield gradient and Hes-
sian is calculated based on the results of a Monte-Carlo analysis and enable a
Newton-type deterministic solution approach to statistical-yield optimization.
The second approach is based on a geometric yield analysis described in
Section 6.3. The gradients of the worst-case distances, which result from a
geometric yield analysis, with regard to any parameter will be derived [6].
Worst-case distances and their gradients lead to a geometric-yield optimization
approach, which can be solved with multiple-objective optimization methods
developed for nominal design.

7.1 Statistical-Yield Optimization

7.1.1 Acceptance-Truncated Distribution
Figures 18 and 46 illustrated that the probability density function of statistical
parameters is truncated by the acceptance function (134). All those parts of the
probability density function are cut away which correspond to parameter vectors
that violate at least one of the performance-specification features according
to (127), (128), or (131), (132). The resulting truncated probability density
function pdfδ is not a normal distribution:

1
pdfδ (xs ) = · δ(xs ) · pdf(xs ) (314)
Y
146 ANALOG DESIGN CENTERING AND SIZING

The factor Y1 makes pdfδ satisfy (15):

∞ ∞ ∞ ∞
· · · −∞ pdfδ (xs ) · dxs = Y1 · −∞
−∞ · · · −∞ δ(xs ) · pdf(xs ) · dxs = 1
Y ·Y =1
The mean value xs,0,δ and the covariance matrix Cδ of the truncated probability
density function pdfδ can be formulated according to Appendix A:

 ∞  ∞
1
xs,0,δ = E {xs } = · ··· xs · δ(xs ) · pdf(xs ) · dxs (315)
pdf Y −∞ −∞
δ

Cδ = E {(xs − xs,0,δ ) · (xs − xs,0,δ )T }

pdfδ
  ∞
1 ∞
= · · · (xs −xs,0,δ )(xs −xs,0,δ )T δ(xs )pdf(xs )dxs (316)
Y −∞ −∞

An estimator of the mean value x /s,0,δ and an estimator of the covariance

/
matrix Cδ of the truncated probability density function pdfδ can be formulated
according to Appendix B:

1 
nM C
/s,0,δ =
x s ) · xs
δ(x(µ) (µ)
(317)
nok
µ=1

n
MC
/δ = 1
C s )(xs −/
δ(x(µ) (µ)
xs,0,δ )(x(µ)
s −/xs,0,δ )T (318)
nok − 1
µ=1

n
MC
/
nok = s ) = Y · nM C ≈ Y · nM C
δ(x(µ) (319)
µ=1

These estimators can be computed within a Monte-Carlo analysis according to

Section 6.1.

7.1.2 Statistical Yield Gradient

The gradient of the statistically estimated yield with regard to the nominal
statistical parameter vector ∇Y (xs,0 ) can be calculated in the following way:
 +∞  +∞
(137)
∇Y (xs,0 ) = ... δ(xs ) · ∇pdfN (xs,0 ) · dxs (320)
−∞ −∞
 
(A.1) δ(xs )
= E · ∇pdfN (xs,0 ) (321)
pdfN (xs )
Yield Optimization/Design Centering 147
 +∞  +∞
(23),(24)
= ... δ(xs )C−1 (xs − xs,0 )pdfN (xs )dxs (322)
−∞ −∞
 +∞  +∞
= C−1 · ... xs · δ(xs ) · pdfN (xs ) · dxs
−∞ −∞
 +∞  +∞ 
−xs,0 · ... δ(xs ) · pdfN (xs ) · dxs (323)
−∞ −∞

= C−1 · [Y · xs,0,δ − xs,0 · Y ]

∇Y (xs,0 ) = Y · C−1 · (xs,0,δ − xs,0 ) (324)

xs,0,δ denotes the mean value of the truncated probability density function pdfδ
according to (315).
From (324), the first-order optimality condition for a yield maximum Y ∗ =
Y (x∗s,0 ) follows immediately:

x∗s,0 = x∗s,0,δ (325)

(325) says that the optimal yield is achieved when the mean value of the trun-
cated probability density function equals that of the original probability density
function.
The mean value of a probability density function can be interpreted as the
center of gravity of the mass represented by the volume under the probability
density function. The first-order optimality condition (325) can therefore be
interpreted in the sense, that the truncation of the probability density function
due to the performance specification does not change the center of gravity
in the optimum. This is the motivation for the term design centering. In the
optimum nominal statistical parameter vector x∗s,0,δ , the design is centered with
regard to the performance specification in the sense that the center of gravity
of the probability density function is not affected by the truncations due to
the performance specification. Design centering means to find a sizing which
represents an equilibrium concerning the center of gravity of the manufactured
and tested “mass.”
Note that a centering of the performance-feature values between their bounds
is not a centered design according to this interpretation. Note also that geomet-
rically inscribing a maximum tolerance ellipsoid in the parameter acceptance
region neither is a centered design according to this interpretation.
Figure 57 illustrates the situation before and after having reached the equi-
librium concerning the centers of gravity of original and truncated probability
density function for two statistical parameters. Those parts of the equidensity
contours that belong to the truncated probability density function are drawn
in bold.
148 ANALOG DESIGN CENTERING AND SIZING

Figure 57. (a) Statistical-yield optimization before having reached the optimum. Center of
gravity xs,0,δ of the probability density function truncated due to the performance specification
(remaining parts drawn as bold line) differs from the center of gravity of the original probability
density function. A nonzero yield gradient ∇Y (xs,0 ) results. (b) After statistical-yield opti-
mization having reached the optimum. Centers of gravity of original and truncated probability
density function are identical.

The center of gravity xs,0,δ of the truncated probability density function is

estimated by (317). It can be imagined as the point where the mass of the
volume under the truncated probability density function can be balanced.
The gradient ∇Y (xs,0 ) according to (324) has been drawn using the property
that it is tangential to the equidensity contour through xs,0,δ :


1 ∂β 2

(24)

= C−1 · (xs,0,δ − xs,0 ) (326)
2 ∂xs xs =xs,0,δ

7.1.3 Statistical Yield Hessian

In the same way as the gradient, the Hessian matrix of the yield with regard
to the nominal statistical parameter vector ∇2 Y (xs,0 ) can be calculated:
 +∞  +∞
(137)
∇2 Y (xs,0 ) = ... δ(xs ) · ∇2 pdfN (xs,0 ) · dxs (327)
−∞ −∞
The second-order derivative of the probability density function with regard to
xs,0 can be calculated based on (23), (24) and using the first-order derivative
in (322):
∇2 pdfN (xs,0 ) = C−1 (xs − xs,0 ) · ∇pdfN (xTs,0 ) − C−1 · pdfN (xs )
 
= C−1 (xs − xs,0 )(xs − xs,0 )T C−1 − C−1 pdfN (xs )
 
= C−1 (xs − xs,0 )(xs − xs,0 )T − C C−1 pdfN (xs ) (328)
Yield Optimization/Design Centering 149

We insert (328) into (327) and extend the terms xs − xs,0 to (xs − xs,0,δ ) +
(xs,0,δ − xs,0 ):
∇2 Y (xs,0 ) =

  +∞ 
C−1 · · · δ(xs )[(xs −xs,0,δ )+(xs,0,δ −xs,0 )][”]T pdfN (xs )dxs − Y C C−1 (329)
−∞

Using (315), (316) and (A.12)–(A.14), we obtain from (329) the Hessian matrix
of the yield with regard to the nominal values of statistical parameters:
 
∇2 Y (xs,0 ) = Y C−1 Cδ + (xs,0,δ − xs,0 )(xs,0,δ − xs,0 )T − C C−1 (330)
xs,0,δ denotes the mean value of the truncated probability density function pdfδ
according to (315), and Cδ denotes the covariance matrix of the truncated
probability density function pdfδ according to (316).
xs,0,δ and Cδ can be estimated as part of a Monte-Carlo analysis using (317)
and (318).
From (330) and the first-order optimality condition (325), the necessary
second-order optimality condition for a yield maximum Y ∗ = Y (x∗s,0 ) follows
immediately:
Cδ − C is negative semidefinite (331)
(331) says that the yield is maximum if the variability of the truncated prob-
ability density function is smaller than that of the original probability density
function. This expresses the property that the performance specification cuts
away a part of the volume under the probability density function, and does this
still in the optimum.

7.1.4 Solution Approach to Statistical-Yield Optimization

Based on the statistical gradient and Hessian matrix of the yield, (324) and
(330), a quadratic model of the yield with respect to the nominal values of
statistical parameters can be formulated:
(next) (next)
Y (2) (xs,0 ) = Y (xs,0 ) + ∇Y (xs,0 ) · (xs,0 − xs,0 )
1 (next) (next)
+ (xs,0 − xs,0 )T · ∇2 Y (xs,0 ) · (xs,0 − xs,0 ) (332)
2
(next)
(332) has a stationary point according to Appendix C, where ∇Y (2) (xs,0 )≡
0 holds:
(next) (next)
∇Y (2) (xs,0 ) ≡ 0 : ∇Y (xs,0 ) + ∇2 Y (xs,0 ) · (xs,0 − xs,0 ) = 0 (333)
The solution of this equation system produces a search direction
(next)
r = xs,0 − xs,0 (334)
150 ANALOG DESIGN CENTERING AND SIZING

for a Newton-type optimization approach. In the case that ∇2 Y (xs,0 ) is not

positive semidefinite, special measures have to be taken. Such measures are:
Ignore the Hessian matrix and use the gradient as search direction.
Switch the signs of the negative eigenvalues of the Hessian.
Bias the diagonal of the Hessian matrix in positive direction.
The quadratic model (332) reflects the behavior of the yield in a limited range
of parameter values. To cope with the limited accuracy of the quadratic model,
a line search along r is performed in a Newton-type optimization approach. It
results in a new parameter vector:
(new)
xs,0 = xs,0 + α · r (335)
At the new parameter vector, another quadratic model according to (332) is
computed, and the process described above restarts. This process is iteratively
repeated until convergence is achieved.
For the line search, a quadratic model of the variance of the yield estimator
with respect to the nominal values of statistical parameters can be formulated
[8, 10]. An increase in the variance of the yield estimator goes together with
the predicted yield improvement. At a certain point, the yield improvement that
can be predicted with a required confidence reaches a maximum. This point
can be used to determine the step length in (335).

7.1.5 Tolerance Assignment

For an optimization of yield by tuning of the covariance matrix, i.e. tolerance
assignment (Section 4.8.6), the derivative of yield with respect to the covariance
matrix, ∇Y (C) is required:
1
∇Y (C) = · ∇2 Y (xs,0 ) (336)
2
A solution approach to tolerance assignment based on (336) is presented in [11].

7.1.6 Deterministic Design Parameters

The yield optimization/design centering based on (324) and (330) works if
the design parameters are at the same time statistical parameters. If design
parameters have a deterministic character, a statistical estimation of yield by a
Monte-Carlo analysis does not lead to yield derivatives.
A crude solution in that case would be a finite-difference approximation by
repeated Monte-Carlo analyses:
Y (xd ) − Y (xd )
∇Y (xd ) ≈ (337)
xd − xd
Yield Optimization/Design Centering 151

Computing the statistical yield gradient with regard to deterministic design para-
meters according to (337) requires nxd Monte-Carlo analyses. The resulting
computational costs are prohibitive if numerical simulation is applied.
An alternative can be developed based on selecting a single statistical
parameter xs,k ,
⎡ ⎤
xs,1 ⎡ ⎤
xs,1
⎢ .. ⎥
⎢ . ⎥ ⎢ .. ⎥
⎢ ⎥ ⎢ . ⎥
⎢ xs,k−1 ⎥ ⎢ ⎥
⎢ ⎥ ⎢ x ⎥
xs = ⎢ xs,k ⎥ ∈ Rnxs −→ xs = ⎢ s,k−1 ⎥ ∈ Rnxs −1 , xs,k (338)
⎢ ⎥ ⎢ xs,k+1 ⎥
⎢ xs,k+1 ⎥ ⎢ .. ⎥
⎢ .. ⎥ ⎣ ⎦
⎣ . ⎦ .
xs,nxs
xs,nxs
and formulating the yield (116) via the marginal distribution of the remaining
statistical parameter xs :
 +∞  +∞ 0 xs,k,U (xs ) 1
Y = ... pdfN (xs,k ) · dxs,k pdfN (xs ) · dxs (339)
−∞ −∞ xs,k,L (xs )

 
= E cdfN (xs,k,U (xs )) − cdfN (xs,k,L (xs )) (340)
pdfN (xs )
Here, we have assumed that the parameters have been transformed into stan-
dardized normally distributed random variables xs with zero mean and unity
covariance matrix, i.e. xs ∼ N (0, I). pdfN and cdfN are the probability
density function and cumulative distribution function according to (16), (17)
and (23).
xs,k,U and xs,k,L represent the borders of the parameter acceptance region
As (122) projected onto the axis of parameter xs,k
A statistical yield estimator based on (340) is determined using (B.1):
n
MC
1
Y/ = (cdfN (xs,k,U (xs(µ) )) − cdfN (xs,k,L (xs(µ) ))) (341)
nM C
µ=1

xs(µ) ∼ N (0, I), µ = 1, . . . , nM C

A higher accuracy is achieved if (341) is averaged over all available statistical
parameters, xs,k , k = 1, . . . , nxs :

1  1 
nxs nM C
/
Y/ = (cdfN (xs,k,U (xs(µ) )) − cdfN (xs,k,L (xs(µ) ))) (342)
nxs nM C
k=1 µ=1

xs(µ) ∼ N (0, I), µ = 1, . . . , nM C

152 ANALOG DESIGN CENTERING AND SIZING

(341) and (342) are evaluated based on a Monte-Carlo analysis. This requires
(µ) (µ) (µ)
the computation of xs,k,U (xs ) and xs,k,L (xs ) in each sample element xs
by solving the following optimization problems, if the nominal statistical para-
meter vector is inside the acceptance region, xs,0 ∈ AsL/U :
⎧
⎪ (µ)
⎪
⎪ min fi (xs , xs,k , xr ) ≤ fL,i
⎪
⎪ x ∈ T
⎪
⎨
r r
(µ)
min (xs,k − xs,k,0 ) s.t.
2 max fi (xs , xs,k , xr ) ≥ fU,i (343)
⎪ x ∈ T
⎪
xs,k ,xr r r
⎪
⎪ i = 1, . . . , nf
⎪
⎪
⎩ x ≤x
s,k s,k,0
⎧
⎪ (µ)
⎪
⎪ min fi (xs , xs,k , xr ) ≤ fL,i
⎪
⎪ x ∈ T
⎪
⎨
r r
(µ)
min (xs,k − xs,k,0 ) s.t.
2 max fi (xs , xs,k , xr ) ≥ fU,i (344)
⎪ x ∈ T
⎪
xs,k ,xr r r
⎪
⎪ i = 1, . . . , nf
⎪
⎪
⎩ x ≥x
s,k s,k,0

(343) and (344) compare to the geometric yield analysis (264). The difference
is that all performance-specification features are considered simultaneously and
that only one parameter is considered in the objective function.
The solution of (343) and (344) becomes a line search along the parameter
xs,k if the worst-case range-parameter vector can be predetermined as described
in Section 6.3.5.
The statistical yield gradient with regard to deterministic design parameters
is formulated starting from (340):

∇Y (xd ) = E {cdfN (xs,k,U ) · ∇xs,k,U (xd )

pdfN (xs )
− cdfN (xs,k,L ) · ∇xs,k,L (xd )} (345)

∇fi (xd )
= E cdfN (xs,k,U ) ·
pdfN (xs ) ∇fi (xs,k,U )

∇fj (xd )
− cdfN (xs,k,L ) · (346)
∇fj (xs,k,L )

Here fi and fj are the performance features whose bounds are active in the
solution of (343) and (344).
In addition to the nM C simulations of a Monte-Carlo analysis, the com-
putation of the statistical yield gradient with respect to deterministic design
Yield Optimization/Design Centering 153

parameters requires at least 2nM C line searches to solve (343) and (344) plus
2nM C sensitivity analyses to solve (346).
The resulting simulation costs may still be prohibitive in practice. Methods
based on statistical estimation of the yield gradient for deterministic design
parameters like [43, 63] therefore fall back on response surface models.

7.2 Geometric-Yield Optimization

According to (302), a yield partition value is improved either by increasing
the corresponding worst-case distance if the nominal parameter vector is inside
its parameter acceptance region partition, or by decreasing the corresponding
worst-case distance if the nominal parameter vector is outside its parameter
acceptance region partition.
According to Figure 36, yield optimization/design centering by tuning of
deterministic design parameters changes the shape of the parameter acceptance
region in such a way that the truncation of the probability density function is
minimized. According to Figure 37, yield optimization/design centering by
tuning of the nominal values of the statistical parameters shifts the probability
density function within the parameter acceptance region such that its truncation
is minimized.
By partitioning of the parameter acceptance region according to individual
performance-specification features as illustrated in Figure 32, a set of worst-case
distances is obtained, each of which represents a partial amount of truncation
due to the respective performance-specification feature.
Figure 52 indicates that the worst-case distances depend on both the shape of
the parameter acceptance region and the nominal value of statistical parameters.
In the following, we will derive the worst-case-distance gradient, which
has the same form and same computational cost for statistical and determin-
istic design parameters. This is in contrast to the statistical yield gradient,
which is complicated for deterministic design parameters as has been shown in
Section 7.1.6.
Based on worst-case distances, geometric-yield optimization will be formu-
lated as a multiple-objective optimization problem. Two solution approaches
for the geometric-yield optimization problem will be described.

7.2.1 Worst-Case-Distance Gradient

The starting point to calculate the worst-case-distance gradient is an exten-
sion of (300) and (301), which simultaneously shows the first-order dependence
of the worst-case distance βW L/U on the mean values of statistical parameters
xs,0 , and on the (deterministic) design parameters xd .
Deterministic design parameters are included in the geometric yield analysis
problem (270)-(273) by expanding the performance function f (xs , xr ) with a
linear term with respect to the design parameters, which has been developed at
154 ANALOG DESIGN CENTERING AND SIZING

xd,µ :

f (xs , xr ) → f (xs , xr ) + ∇f (xd,µ )T · (xd − xd,µ ) (347)

As this extension neglects second-order effects, it represents a constant for the

inner optimization problem in (270)–(273). It can be passed on to the outer
optimization problem and becomes a part of its constraint in that the bounds
fL/U in (270)–(273) are extended by:

fL/U → fL/U − ∇f (xd,µ )T · (xd − xd,µ ) (348)

The replacement (348) proceeds to (300) and (301), which become:

f ≥ fL and xs,0 ∈ As,L , f ≤ fU and xs,0 ∈ Ās,U : (349)

∇f (xs,W L )T · (xs,0 − xs,W L ) + ∇f (xd,µ )T · (xd − xd,µ )

βW L/U = -
∇f (xs,W L )T · C · ∇f (xs,W L )

f ≥ fL and xs,0 ∈ Ās,L , f ≤ fU and xs,0 ∈ As,U : (350)

∇f (xs,W L )T · (xs,W L − xs,0 ) − ∇f (xd,µ )T · (xd − xd,µ )

βW L/U = -
∇f (xs,W L )T · C · ∇f (xs,W L )

From (349) and (350), the worst-case-distance gradients with respect to the
mean values of statistical parameters xs,0 and with respect to (deterministic)
design parameters xd follow:

f ≥ fL and xs,0 ∈ As,L , f ≤ fU and xs,0 ∈ Ās,U :

+1
∇βW L/U (xs,0 )= . ·∇f (xs,W L/U ) (351)
∇f (xs,W L/U )T ·C· ∇f (xs,W L/U )

+1
∇βW L/U (xd )= . ·∇f (xd,µ ) (352)
∇f (xs,W L/U )T ·C· ∇f (xs,W L/U )

f ≥ fL and xs,0 ∈ Ās,L , f ≤ fU and xs,0 ∈ As,U :

−1
∇βW L/U (xs,0 )= . ·∇f (xs,W L/U ) (353)
∇f (xs,W L/U )T ·C· ∇f (xs,W L/U )

−1
∇βW L/U (xd )= . ·∇f (xd,µ ) (354)
∇f (xs,W L/U )T ·C· ∇f (xs,W L/U )
Yield Optimization/Design Centering 155

(351)–(354) show that the worst-case-distance gradient has the same form con-
cerning statistical parameters and (deterministic) design parameters.
The worst-case-distance gradient corresponds to the performance gradient at
the worst-case parameter vector. Its length is scaled according to the variance
of the linearized performance (198), its direction depends on whether a lower
or upper performance-feature bound is specified and whether it is satisfied or
violated.
The equivalence of statistical parameters and deterministic design parameters
in the worst-case distance (349) and (350) and in the worst-case-distance gra-
dient (351)–(354) can also be interpreted using Figure 52.
The worst-case distance is increased by increasing the difference between the
performance-feature value at the nominal statistical parameter vector and the
performance-feature bound. In a linear performance model, this increase can
be achieved by shifting any parameter of any kind. In Figure 52, which shows
the subspace of parameters that are statistical, the performance gradient at the
worst-case parameter vector shows the direction in which the nominal statistical
parameter vector xs,0 has to be shifted for a steepest ascent in the worst-case
distance according to (351). The same effect is achieved by a change in the
(deterministic) design parameter vector xd according to (352), which shifts the
boundary of As,L away from xs,0 according to (348).

Worst-Case-Distance Gradient by Lagrange Factor. From Appendix C

follows that the sensitivity of the solution of (270)–(273) with respect to a
perturbation  in the constraint is determined by the Lagrange factor λW/L :

∇βW
2
L/U () = λW L/U (355)

Inserting (348) into (277)–(280), we can find that  is defined as:

f ≥ fL and xs,0 ∈ As,L , f ≤ fU and xs,0 ∈ Ās,U :

 = ∇f (xd,µ )T · (xd − xd,µ ) (356)

f ≥ fL and xs,0 ∈ Ās,L , f ≤ fU and xs,0 ∈ As,U :

 = −∇f (xd,µ )T · (xd − xd,µ ) (357)

By the chain rule of differentiation and using (355), we obtain:

1
∇βW L/U (xd ) = · ∇βW
2
L/U () · ∇(xd )
∇βW
2
L/U (βW L/U )
1
= · λW L/U · ∇(xd ) (358)
2 · βW L/U
156 ANALOG DESIGN CENTERING AND SIZING

With (294) and based on differentiating (356) and (357) with respect to xd , we
obtain (352) and (354).

7.2.2 Solution Approaches to Geometric-Yield Optimization

According to (302) and Section 6.3.8, each worst-case distance corresponds
to an approximate yield partition value for the respective performance-
specification feature if single-plane bounded tolerance classes (Section 6.2.4)
are applied. Each worst-case distance thus represents a partition of the overall
yield as illustrated in Figure 54. Yield optimization/design centering is achieved
by maximizing those worst-case distances as much as possible, for which the
nominal statistical parameter vector is inside the parameter acceptance region
partition. If the nominal statistical parameter vector is outside the parameter
acceptance region partition, then the corresponding worst-case distance is a
measure for the degree of violation of the respective performance-specification
feature. In this case, the worst-case distance has to be decreased. Once it
reaches zero, the situation switches in that the nominal parameter vector gets
inside the acceptance region partition and the worst-case distance becomes a
robustness measure that shall be maximized.
Yield optimization/design centering in this way becomes a multiple-objective
optimization problem according to (142) with worst-case distances as
objectives:
 
αi (xd ) · βW L/U,i (xd )
max s.t. c(xd ) ≥ 0 (359)
xd i = 1, . . . , nf

+1, xs,0 ∈ As,L/U,i (xd )
αi = (360)
−1, xs,0 ∈ Ās,L/U,i (xd )
The sign αi considers the two cases of the nominal statistical parameter vector
lying inside or outside of the corresponding acceptance region partition.
Worst-case distance gradients (352), (354) are applied to solve (359).
Note that second-order derivatives are available for the statistical-yield opti-
mization approach (332), while only first-order derivatives are available for the
geometric-yield optimization approach (359). On the other hand, the statistical
yield gradient with regard to deterministic parameters (346) involves imprac-
tical computational cost, while the worst-case distance gradient with regard to
deterministic parameters (352), (354) is cheap.
For the example in Section 6.3.9, with the initial situation previously illus-
trated in Table 14 and Figure 54, yield optimization/design centering according
to (359) leads to an optimal situation illustrated in the following Table 15 and
Figure 58.
Compared to the initial situation after nominal design, the worst-case dis-
tances with smaller values have been maximized at the cost of those with larger
Yield Optimization/Design Centering 157

Table 15. Geometric-yield optimization of an operational amplifier from Section 6.3.9.

After After geometric-

nominal design yield optimization
Performance-specification Nominal Worst-case Nominal Worst-case
feature performance distance performance distance
Gain ≥ 65dB 76dB 2.5 76dB 4.2
Transit frequency ≥ 30M Hz 67M Hz 7.7 58M Hz 4.5
Phase margin ≥ 60◦ 68◦ 1.8 71◦ 3.9
Slew rate ≥ 32V /µs 67V /µs 6.3 58V /µs 3.9
DC power ≤ 3.5µW 2.6µW 1.1 2.3µW 4.2
Ȳ = 82.9% Ȳ = 99.9%

values by geometric-yield optimization. The smallest worst-case distances after

geometric-yield optimization is 3.9 and means that this is nearly a four-sigma
design. The smallest worst-case distance is obtained for the phase margin
and the slew rate. The achieved value represents a Pareto point concerning
the worst-case distances of these two performance features. None of them
can be improved without worsening the other one. 3.9-sigma is the maximum
robustness of speed and stability that can be achieved for the underlying process
technology.
Note that the nominal performance value of the gain is the same before and
after geometric-yield optimization, but the worst-case distance has increased
from 2.5 to 4.2. According to (300) and (301), this must have been achieved
by decreasing the sensitivity of gain with regard to the statistical parameters.

7.2.3 Least-Squares/Trust-Region Solution Approach

We will not consider the constraints c(xd ) ≥ 0 for reasons of simplicity in
the following.
(359) can be solved by formulating target values β W,target for all worst-
case distances, which have been collected in a vector β W L/U , and scalarizing
the multiple-objective problem (106). The vector norm . is meant to be the
l2 -norm .2 in the following.

min β W L/U (xd ) − β W,target 2 (361)

xd

β W L/U = [ . . . αi · βW L/U,i . . . ]T
158 ANALOG DESIGN CENTERING AND SIZING

Figure 58. Worst-case distances and approximate yield partition values before and after
geometric-yield optimization of the operational amplifier from Table 15.

(361) is a least-squares problem due to the objective function.

It is easy to formulate target values for the worst-case distances. For a three-
sigma design for instance, all worst-case distances should be 3, for a six-sigma
design, they should be 6.
Using the worst-case distance gradients (352), (354), a linear model of the
objective function in (361) is established. First, the linear model of the vector
of worst-case distances β W L/U with regard to a search direction r = xd − xd,µ
starting from the current point xd,µ is formulated:

β̄ W L/U (r) = β W L/U (xd,µ ) + ∇β W L/U (xd,µ )T · r (362)

  
J
with r = xd − xd,µ

The matrix J in (362) contains all individual worst-case-distance gradients

according to (352), (354) as rows.
Yield Optimization/Design Centering 159

Based on (362), the gradient of the linearized objective function in (361) is

calculated:

¯F (r)2 = β̄ W L/U (r) − β W,target 2

= TF,0 · F,0 + 2 · rT · JT · F,0 + rT · JT · J · r (363)

with F,0 = β W L/U (xd,µ ) − β W,target

Inserting (363) in (361) leads to a linear least-squares optimization problem,

min ¯F (r)2 (364)

r

which has a quadratic objective function due to the last term in (363). The
stationary point of (364) is calculated as:

∇¯F 2 (r) ≡ 0 : JT · J · r = −J · F,0 (365)

The solution of (365) is known as Gauss-Newton direction.

An improved computation of a search direction r considers the limited
accuracy of the linearized model (362) by introducing a trust-region ∆ that
the search direction must not leave:

min ¯F (r)2 s.t. r2 ≤ ∆2 (366)

r

(366) represents a least-squares/trust-region approach to geometric-yield opti-

mization. The Lagrangian function of (366) is

L(r, λ) = ¯F (r)2 − λ · (∆2 − rT · r) (367)

The stationary point of (367) is calculated as:

∇L(r, λ) ≡ 0 : (JT · J + λ · I) · r = −J · F,0 (368)

The solution of (368) is known as Levenberg-Marquardt direction.

An actual least-squares/trust-region step r∗ is determined by the following
procedure:
1. Solve (368) starting from λ = 0 with increasing values of λ including the
limit λ → ∞.
2. Visualize the resulting Pareto front P Fr∗ of all optimal objective value for
all trust regions, ¯F (r∗ )2 vs. r∗ 2 .
3. Select a step from the Pareto front P Fr∗ utilizing its bend and additional
simulations.
160 ANALOG DESIGN CENTERING AND SIZING

Figure 59. Least-squares/trust-region approach to geometric-yield optimization according to

(366). Quarter-circles represent trust-regions of a step. Ellipses represent level contours of the
least-squares optimization objective (363). r∗ is the optimal step for the respective trust region
determined by ∆.

The idea behind this procedure is to include the computation of a suitable

trust region into the computation of a search direction. Figure 59 illustrates
the procedure in the space of two design parameters xd,1 and xd,2 . The two
axes are shifted so that the origin lies in the actual design parameter vector
xd,µ . The quarter-circles represent trust regions for a length of step r not larger
than ∆. The ellipses represent level contours of the least-squares optimization
objective (363).
If the allowed step length ∆ is zero, no step is allowed and the resulting step
r∗ has length zero. This corresponds to a Lagrange factor λ going to infinity in
the Lagrangian function (368).
If the allowed step length is arbitrarily large, i.e. there is no restriction on the
step length, this means that ∆ is infinite and that λ is zero and (368) becomes
(365). In this case, the minimum of ¯F (r)2 will determine the step r∗ as
illustrated in Figure 59.
For an allowed step length in between, i.e. 0 < ∆ < ∞, the resulting
problem resembles the problem of a geometric yield analysis. The optimum step
r∗ results from the point where the bounding circle of the trust region touches
Yield Optimization/Design Centering 161

Figure 60. Pareto front of optimal solutions of least-squares/trust-region approach (366) in

dependence of maximum steplength ∆, i.e. Lagrange factor λ (367). A point in the bend
corresponds to a step with a grand progress towards the target worst-case distances at a small
step length.

a level contour of the objective ¯F (r)2 . Figure 59 illustrates optimum steps
for some trust regions.
Due to the quadratic nature of the objective ¯F (r)2 , the amount of
additional decrease in the objective that can be obtained by an increase of
the allowed step length r ≤ ∆ is decreasing. This is even more so with a
worsening problem condition of the Jacobian matrix J.
Therefore, the Pareto front of objectives ¯F (r∗ )2 versus the step r∗
according to (366) acquires a typical shape as illustrated in Figure 60.
A point in the bend of this curve is preferable, because it leads to a grand
progress towards the target worst-case distances at a small step length. The
additional progress towards the target worst-case distances beyond the bend
is rather small. Additionally, the required step length for additional progress
beyond the bend becomes large, and the step will be more likely to leave the
region of validity of the linearized model (362). Therefore, a step in the bend
will be selected. Additional simulations have to be spent to verify that the
linearized model holds in the corresponding trust region.
Note that the described approach is a part of a complete optimization algo-
rithm. Other algorithmic components are required that deal with updating the
target values of the optimization objectives, or with determining bends in the
Pareto curve.
162 ANALOG DESIGN CENTERING AND SIZING

Efficient methods to calculate the Pareto curve of progress in the optimization

objective versus step length according to problem (366) have been developed
[7, 9]. They include the additional constraints in the problem formulation (359).

Special Case: Nominal Design. The least-squares/trust-region solution

approach described in this section can be applied to nominal design as well.
The worst-case distances then are simply replaced by the performance fea-
tures and corresponding target values.

7.2.4 Min-Max Solution Approach

According to the definition of yield and yield partitions, (137)–(139), the
smallest yield partition value is an upper bound for the overall yield.
The multiple-objective geometric-yield optimization problem (359) can there-
fore be transformed into a single-objective optimization by maximization of the
minimum yield partition:
max min αi (xd )·βW L/U,i (xd ) s.t. c(xd ) ≥ 0 (369)
xd i

The optimization objective of problem formulation (369) is also obtained by

using the l∞ -norm in (366).
(369) describes a maximum tolerance ellipsoid inside the parameter accep-
tance region as illustrated in Figure 61.
Sometimes, the min-max approach to geometric-yield optimization is denoted
as design centering. Design centering in the sense of finding an equilibrium of
the center of gravity of the truncated probability density function however is
similar but not equal to geometric-yield optimization.

7.2.5 Linear-Programming Solution Approach

Using the linearization of the worst-case distances (362), the min-max for-
mulation (369) can be transformed into a linear programming problem.
*
β W L/U (xd,µ ) + J(xd,µ ) · (xd − xd,µ ) ≥ β · 1
max β s.t. (370)
xd ,β c(xd,µ ) + ∇c(xTd,µ ) · (xd − xd,µ ) ≥ 0

1 denotes a vector with “1” at each position. (370) formulates a worst-case

distance value β that is to be maximized such that the linear models of all
worst-case distances obtained from a geometric analysis have at least this value
β. The resulting value of β will describe the largest tolerance ellipsoid that can
be inscribed in the overall parameter acceptance region.
(370) describes a linear programming problem.
The geometric-yield optimization problem then consists in a sequence of
linear programming problems (370), where the linearization of the worst-case
distances is iteratively updated.
Yield Optimization/Design Centering 163

Figure 61. Min-max solution to geometric-yield optimization.

Note that any piecewise linear approximation of the parameter acceptance

region can be applied within (370).
In [38], (370) has been presented for the mean values of statistical parameters
as optimization parameters and coordinate searches to compute border points
of the parameter acceptance region.
Ellipsoidal methods can be applied to solve (370) [1, 123].

7.2.6 Tolerance Assignment, Other Optimization Parameters

From (300) and (301), the first-order derivatives of the worst-case distance
with regard to components of the covariance matrix C, i.e. standard devia-
tions and correlations, which are the basis for tolerance assignment, can be
derived [122].
In the same manner as described in Section 7.2.1, first-order derivatives of
the worst-case distance with regard to any other quantity in problem (264)
and (265), i.e. performance-feature bound, range-parameter bounds, can be
derived [122].
Appendix A
Expectation Values

A.1 Expectation Value

The expectation value of a vector function1 h(z) of a random vector2 z,
h ∈ Rnh , z ∈ Rnz , that originates from a statistical distribution with the
probability density function pdf(z) is denoted by and defined as:
 +∞  +∞
E {h(z)} = E {h(z)} = ... h(z) · pdf(z) · dz (A.1)
pdf(z) −∞ −∞

dz = dz1 · dz2 · . . . · dznz

Given a probability density function pdf(z), the expectation value is assumed
to refer to this probability density function without explicitly mentioning it.

A.2 Moments
The moment of order κ, m(κ) , of a single random variable z results from
setting h(z) = z κ in (A.1):
m(κ) = E {z κ } (A.2)

A.3 Mean Value

The moment of order 1, i.e. the expectation value of z, or z respectively, is
denoted as mean value m, or m respectively:
m(1) = m = E {z} (A.3)

1 See Note 2 in Section 3.7.

2 See Note 1 in Section 3.5.
166 ANALOG DESIGN CENTERING AND SIZING

m = E {z} = [ E {z1 } . . . E {znz } ]T (A.4)

A.4 Central Moments

The central moment of order κ, c(κ) , of a single random variable z results
from setting h(z) = (z − m)κ in (A.1):
c(κ) = E {(z − m)κ } (A.5)

A.5 Variance
The variance of a single random variable z is defined as the central moment
of order 2 of z and denoted by σ 2 or V {z}:
 
c(2) = σ 2 = V {z} = E (z − µ)2 (A.6)
σ denotes the standard deviation of a single random variable z:
-
σ = V {z} (A.7)

A.6 Covariance
The covariance, cov {zk , zl }, of two random variables zk and zl is defined
as a mixed central moment of order 2:
cov {zk , zl } = E {(zk − mk ) · (zl − ml )} (A.8)

A.7 Correlation
The correlation, corr {zk , zl }, of two random variables zk and zl is defined as
their covariance normalized with respect to their individual standard deviations:
cov {zk , zl }
corr {zk , zl } = (A.9)
σk · σl

A.8 Variance/Covariance Matrix

The central moments of order 2 of a vector z are defined component-wise
according to (A.6) and (A.8) and combined in the variance/covariance matrix
or simply covariance matrix C = V {z}:
 
C = V {z} = E (z − m) · (z − m)T
⎡ ⎤
V {z1 } cov {z1 , z2 } . . . cov {z1 , znz }
⎢ cov {z2 , z1 } V {z2 } . . . cov {z2 , znz } ⎥
⎢ ⎥
= ⎢ .. .. ⎥ (A.10)
⎣ . . ⎦
cov {znz , z1 } ... V {znz }
Appendix A: Expectation Values 167

Correspondingly, the covariance matrix of a vector function h(z) of a random

vector z is defined as:
2 
V {h(z)} = E (h(z) − E {h(z)}) · (h(z) − E {h(z)})T (A.11)

A.9 Calculation Formulas

In the following, 3 calculation formulas are given together with respective
special cases.

E {A · h(z) + b} = A · E {h(z)} + b (A.12)

E {const} = const
E {c · h(z)} = c · E {h(z)}
E {h1 (z) + h2 (z)} = E {h1 (z)} + E {h2 (z)}

V {A · h(z) + b} = A · V {h(z)} · AT (A.13)

 
V aT · h(z) + b = aT · V {h(z)} · a
V {a · h(z) + b} = a2 · V {h(z)}

2 
V {h(z)} = E (h(z) − a) · (h(z) − a)T
− (E {h(z)} − a) · (E {h(z)} − a)T (A.14)
2 
V {h(z)} = E (h(z) − a)2 − (E {h(z)} − a)2
   
V {h(z)} = E h(z) · hT (z) − E {h(z)} · E hT (z)
 
V {h(z)} = E h2 (z) − (E {h(z)})2

(A.12) is the linear transformation formula for expectation values. It says that
the expectation value of the linear transformation of a random vector equals the
corresponding linear transformation of the vector’s expectation value.
(A.13) is the linear transformation formula for variances.
(A.14) is the translation formula for variances. It relates the variance as a
second-order central moment to the second-order moment and the quadratic
expectation value.
From (A.13) follows the Gaussian error propagation:

V{aT · z + b} = aT · V {z} · a = aT · C · a
 k,l =0  2 2
= ak al σk k,l σl = ak σk (A.15)
k,l k
168 ANALOG DESIGN CENTERING AND SIZING

A.10 Standardization of Random Variables

A random variable z originating from a statistical distribution corresponds
to a standardized random variable z  with a mean value of zero and a variance
of one.
The standardization is done with the following formula:
z − E {z} z − mz
z = - = (A.16)
V {z} σz
 
E z =0
 
V z = 1

A.11 Exercises
1. Prove (A.12). (Apply (A.1), (15).)
2. Prove (A.13). (Apply (A.11), (A.12), (A.1).)
3. Prove (A.14). (Apply (A.11), (A.12).)
4. Prove that z  according to (A.16) has a mean value of zero and a variance
of one. (Apply calculation formulas from Section A.9.)
5. Show that (52) holds for a multinormal distribution according to (23) and
(24). (Insert (23) and (24) in (A.1). Use (57) and (58) for a variable sub-
stitution. Consider that the corresponding integral over an even function is
zero.)
6. Show that (53) holds for a multinormal distribution according to (23) and
(24). (Insert (23) and (24) in (A.10). Use (57) and (58) for a variable
substitution.)
Appendix B
Statistical Estimation of Expectation Values

B.1 Expectation-Value Estimator

A consistent, unbiased of the expectation value (A.1) is
n
MC
/ {h(x)} = m 1
E /h = h(x(µ) ) (B.1)
nM C
µ=1

x(µ) ∼ D (pdf(x)) , µ = 1, . . . , nM C
A statistical estimation is based on a sample of sample elements.
A widehat is used to denote an estimator function Φ(x/ (1) , . . . , x(nM C ) ) for
a function Φ(x).
Sample elements x(µ) are independently and identically distributed accord-
ing to the given statistical distribution D with the probability density function
pdf(x). Therefore:
2 
E h(x(µ) ) = E {h(x} = mh (B.2)
2  
V [h(x(1) ) . . . h(x(nM C ) )]T = diag V{h(x(1) )} . . .

V{h(x(nM C ) )} (B.3)
2 
V h(x(µ) ) = V {h(x)} (B.4)

B.2 Variance Estimator

A consistent, unbiased estimator of the variance/covariance matrix (A.11) is
nMC    T
/ {h(x)} = 1
V h(x(µ) ) − m/ h · h(x(µ) ) − m/h (B.5)
nM C − 1
µ=1

x(µ) ∼ D (pdf(x)) , µ = 1, . . . , nM C
170 ANALOG DESIGN CENTERING AND SIZING

/ h is used. If the expectation value

In (B.5), the estimated expectation value m
mh would be used, the consistent, unbiased estimator would be:

1 
nM C  T

{h(x)} =
V h(x(µ) ) − mh · h(x(µ) ) − mh (B.6)
nM C µ=1

B.3 Estimator Bias

The bias b of an estimator is the difference between the expectation value
of the estimator function and the original function:
2 
/
bΦ
= E Φ(x) − Φ(x) (B.7)

The expectation value of an unbiased estimator function is the original function:

2 
/
Φ(x) is unbiased ⇔ bΦ
= 0 ⇔ E Φ(x) / = Φ(x) (B.8)

A consistent estimator function certainly approaches the original function if the

number of sample elements nM C approaches ∞:
 
/
Φ(x) is consistent ⇔ prob /
lim Φ = Φ = 1 (B.9)
nM C →∞

B.4 Estimator Variance

A measure for the quality of an estimator function is the 2nd-order moment
of the estimator function around the original function:
2  2 
QΦ
= E (Φ / − Φ) · (Φ/ − Φ)T = V Φ / + b
· bT (B.10)
Φ

From (B.10) follows that the quality measure Q is the estimator variance, i.e.
the variance of the estimation function, for an unbiased estimator:
2 
bΦ
= 0 : QΦ
= V Φ / (B.11)

B.5 Expectation-Value-Estimator Variance

To determine the quality of the expectation-value estimator of (B.1), (B.11)
can be applied, because (B.1) provides an unbiased estimator (see Exercise
B.7.1). 2 
Qµ
h = V E / {h(x)} (B.12)
Applying (B.1) and (A.13) leads to
1 2 
Qµ
h = · V [ 1 1 . . . 1 ] · [ h(1) h(2) . . . h(nM C ) ]T , (B.13)
nM C
Appendix B: Statistical Estimation of Expectation Values 171

with h(µ) = h(x(µ) ). Applying (A.13), (B.3) and (B.4) finally leads to:
2  1
/
Qµ
h = V E {h(x)} = · V {h(x)} (B.14)
nM C
Using (B.16) and (B.17) in the above proof leads to the corresponding formula
for an estimator of the quality:
2 
/ µ
= V
Q / E / {h(x)} = 1 · V / {h(x)} (B.15)
h
nM C

B.6 Estimator Calculation Formulas

In the following, estimator formulas corresponding to the calculation formu-
las (A.12), (A.13) and (A.14) are given.

/ {A · h(z) + b} = A · E
E / {h(z)} + b (B.16)

/ {A · h(z) + b} = A · V
V / {h(z)} · AT (B.17)
    
/ {h(z)} = nM C
V / h(z) · hT (z) − E
E / {h(z)} · E
/ hT (z) (B.18)
nM C − 1

B.7 Exercises
1. Prove that the expectation-value estimator according to (B.1) is unbiased.
(Check if (B.8) is true. Apply (A.12) and (B.2).)
2. Prove the second part of (B.10). (Express Φ by using (B.7) and insert in
(A.14). Combine terms with Φ / as h.)

3. Prove that the variance estimator according to (B.5) is unbiased.

Check if (B.8) is true. Apply (B.5). Within the sum extend each product
term by −mh + mh . Apply (B.1) to get to a form with a sum of the dyadic
product of h(µ) − mh and nM C times the dyadic product of m / h − mh . Then
apply (B.4) and (B.14).
4. Prove (B.16).
5. Prove (B.17).
6. Prove (B.18).
Appendix C
Optimality Conditions of Nonlinear
Optimization Problems

Optimality conditions of optimization problems can be found in the literature

[56, 45, 91, 19]. The following summary adopts the descriptions in [45].
The statements in the following hold for a local optimum. Global optimiza-
tion requires a preceding stage to identify regions of local optima.

C.1 Unconstrained Optimization

Without loss of generality, we will assume that a minimum of a function
f (x) shall be targeted in the following:

min f (x) (C.1)

x

Starting point is the Taylor expansion of the objective function f (x) at the
optimum solution x∗ and f ∗ = f (x∗ ) :

f (x) = f (x∗ ) + ∇f (x∗ )T ·(x − x∗ )

     
f∗ gT
1
+ · (x − x∗ )T · ∇2 f (x∗ ) ·(x − x∗ ) + . . . (C.2)
2   
H

With x = x∗ + r, a point x in the direction r from the optimum solution x∗ has

the objective value:

1 T
f (x∗ + r) = f ∗ + gT · r + · r · H · r + ... (C.3)
2
174 ANALOG DESIGN CENTERING AND SIZING

Figure C1. Descent directions and steepest descent direction of a function of two parameters.

C.2 First-Order Unconstrained Optimality Condition

Descent Direction. A descent direction r at any point x is defined by:

∆f (r) = ∇f (x)T · r < 0 (C.4)

Figure C1 illustrates the range of possible descent directions for an example

with two parameters. All vectors starting from x that are in the gray region are
descent directions. The steepest descent direction is the direction in which the
amount of decrease in the objective function is maximum. From (C.4) follows
that the steepest descent is along the negative gradient (Figure C1):

min ∆f (r) = ∇f (x)T · (−∇f (x)) (C.5)

r

Minimum: No More Descent. x∗ being the solution with minimum objec-

tive value f ∗ requires that there exists no descent direction r starting from x∗ .
Only then, no further decrease in f is achievable and f ∗ is the minimum:

∀ gT · r ≥ 0 (C.6)
r = 0
Appendix C: Optimality Conditions of Nonlinear Optimization Problems 175

Necessary Condition. The only way to satisfy (C.6) is that the gradient g is
zero. That represents the necessary first-order optimality condition for a local
minimum solution x∗ with f ∗ = f (x∗ ):

∇f (x∗ ) = 0 (C.7)

A point that satisfies (C.7) is also called stationary point.

C.3 Second-Order Unconstrained Optimality Condition

Inserting (C.7) into (C.3), we can:
see that another condition for the quadratic term has to be formulated, in
order to guarantee that no decrease in f is achievable at x∗ . This condition says
that no direction r may exist, in which the quadratic term in (C.3) leads to a
reduction in the objective value f ∗ = f (x∗ ).

Necessary Condition. The requirement of no further reduction is a necessary

second-order optimality condition for a local minimum solution x∗ with f ∗ =
f (x∗ ).

∀ rT · ∇2 f (x∗ ) · r ≥ 0 ⇔ ∇2 f (x∗ ) is positive semidefinite (C.8)

r = 0

Figure C2 illustrates examples of functions of two parameters featuring dif-

ferent types of definiteness of their second-order derivatives.
We can see that positive definiteness refers to a minimum, negative definite-
ness to a maximum. If the Hessian is indefinite, no finite optimum exists. If the
Hessian is positive semidefinite, no more descent can happen, but there may be
no unique solution.
The type of definiteness of the Hessian corresponds to the signs of its eigen-
values. Positive definiteness corresponds to positive eigenvalues, negative def-
initeness to negative eigenvalues. Positive semidefiniteness corresponds to
eigenvalues greater or equal zero. Indefiniteness corresponds to eigenvalues
both negative and positive.

Sufficient Condition. A sufficient second-order optimality condition for a

local minimum solution x∗ with f ∗ = f (x∗ ) is that only an increase in the
objective around the minimum is possible:

∀ rT · ∇2 f (x∗ ) · r > 0 ⇔ ∇2 f (x∗ ) is positive definite (C.9)

r = 0
176 ANALOG DESIGN CENTERING AND SIZING

x**2+y**2 -x**2-y**2
200 -20
180 -40
160 -60
140 -80
120 -100
100 -120
200 80 0 -140
180 60 -20 -160
160 40 -40 -180
140 20 -60
120 -80
100 -100
80 -120
60 -140
40 -160
20 -180
0 -200

10 10
5 5
-10 0 -10 0
-5 -5
0 -5 0 -5
5 5
10 -10 10 -10

x**2 x**2-y**2
100 100
90 80
80 60
70 40
60 20
50 0
100 40 100 -20
30 -40
80 20 -60
10 50 -80
60
0
40
20 -50

0 -100

10 10
5 5
-10 0 -10 0
-5 -5
0 -5 0 -5
5 5
10 -10 10 -10

Figure C2. Different types of definiteness of the second-derivative of a function of two para-
meters: positive definite (upper left), positive semidefinite (upper right), indefinite (lower left),
negative definite (lower right).

C.4 Constrained Optimization

Without loss of generality, we will assume that a minimum of a function f (x)
shall be targeted under equality and inequality constraints in the following form

cµ (x) = 0, µ ∈ EC
min f (x) subject to (C.10)
x cµ (x) ≥ 0, µ ∈ IC

EC and IC denote the index sets of equality and inequality constraints.

The optimality conditions will be based on the Lagrangian function of (C.10)
that combines objective and constraints into one new objective:

L(x, λ) = f (x) − λµ · cµ (x) (C.11)
µ∈EC∪IC

λµ is called Lagrange factor. It is part of the optimization parameters and has

a certain optimum value.
Appendix C: Optimality Conditions of Nonlinear Optimization Problems 177

Note that in (C.10) we have chosen to formulate “greater-than” constraints

and to subtract the constraint components in the Lagrangian function (C.11).
This determines the sign of the optimum Lagrange factor.
As in the unconstrained case, conditions will be formulated for the first- and
second-order derivative that ensure that no further descent can happen in the
surrounding of a minimum. But now, they are extended by the consideration
of the constraints in the Lagrangian function.

C.5 First-Order Constrained Optimality Condition

Unconstrained Descent Direction. In constrained optimization, the con-
cept of descent direction is extended to unconstrained descent directions. An
unconstrained descent direction r with respect to one constraint cµ at any point
x is defined by:

∆f (r) = ∇f (x)T · r < 0 (C.12)

cµ (x + r) ≈ cµ (x) + ∇cµ (x)T · r ≥ 0 (C.13)

(C.12) is identical to the unconstrained case (C.4). It describes the directions r

in which the objective will decrease.
(C.13) describes the restrictions due to the constraint cµ . If the constraint is
not active at point x, then there is no restriction on the direction r, only on the
length of it. The constraint being inactive means that it is still away from the
bound. Then, a descent direction may eventually result in the constraint value
reaching the bound. In order not to violate the constraint, the descent direction
may not exceed a certain length :

cµ (x) > 0 : r ≤  (C.14)

If the constraint is active at point x, then the value of cµ is at its limit, cµ = 0.

Equality constraints hence are permanently active, inequality constraints may
become active or inactive again. Then, there is a restriction on the direction r.
Only those directions are allowed according to (C.13) in which the constraint
value stays at its limit or in which the constraint value is increased such that the
constraint will become inactive again.
Figure C3 illustrates the range of possible descent directions for an example
with two parameters with a gray block sector of larger radius.
The performance contour f = const is the same as in Figure C1. The
performance gradient ∇f (x) defines a range of directions that are descent
directions. In addition, a constraint is active, c(x) = 0. The gradient of the
constraint function, ∇c(x) defines a range of directions that are allowed. This
range is illustrated with a gray block sector of smaller radius in Figure C3.
The intersection of the region of descent directions and the region of uncon-
strained directions is the region of unconstrained descent directions. In Figure
178 ANALOG DESIGN CENTERING AND SIZING

Figure C3. Descent directions and unconstrained directions of a function of two parameters
with one active constraint.

C3, this corresponds to the region where the two gray block sectors overlap and
is illustrated with a dark gray block sector.
Obviously, each active constraint reduces the range of unconstrained descent
directions.

Minimum: No More Unconstrained Descent. If x∗ is the solution that

achieves the minimum objective value f ∗ without violating any constraint,
then there must not be any unconstrained descent direction r starting from x∗ .
Only then, no further decrease in f is achievable and f ∗ is the minimum.
From Figure C3, we can motivate that no unconstrained descent region exists,
if the dark gray region is empty. This is the case, if the gradients of objective
and constraint are parallel and have the same direction:
∇f (x∗ ) = λ∗ · ∇c(x∗ ) ∧ λ∗ > 0 (C.15)
If λ∗ would have a negative sign, the gradients of objective and constraint would
point in opposite directions. According to Figure C3, then the constraint would
impose no restriction at all on the descent direction. Note that the positive
sign is due to fact that we have formulated a minimization problem with lower
bounds as constraints.
The first part of (C.15) could be derived by starting from the Lagrangian func-
tion (C.11) with only one constraint and treating it as done in the unconstrained
Appendix C: Optimality Conditions of Nonlinear Optimization Problems 179

case. Then, we would formulate the requirement that the first-order derivative
of the gradient of the Lagrangian function is zero:
∇L(x) ≡ 0 : ∇f (x∗ ) − λ∗ · ∇c(x∗ ) = 0 (C.16)

Necessary Condition. We can now formulate the necessary first-order opti-

mality condition of: a constrained optimization problem according to (C.10) by
formulating the no-descent condition through (C.15) and (C.16), and by adding
the requirements on the constraint satisfaction:

∇L(x∗ ) = ∇f (x∗ ) − λ∗µ · ∇cµ (x∗ ) = 0 (C.17)
µ∈A(x∗ )
A(x ) = EC ∪ {µ ∈ IC | cµ (x∗ ) = 0 }
∗

cµ (x∗ ) = 0, µ ∈ EC (C.18)
cµ (x∗ ) ≥ 0, µ ∈ IC (C.19)
λ∗µ ≥ 0, µ ∈ IC (C.20)
λ∗µ · cµ (x∗ ) = 0, µ ∈ EC ∪ IC (C.21)
(C.17) and (C.20) are explained through the condition that no unconstrained
descent direction exists in the minimum. The restrictions in (C.17) are defined
for all active constraints A(x∗ ).
No statement about the sign of the Lagrange factor can be made for an
equality constraint.
(C.18) and (C.19) formulate that the constraint must not be violated in the
minimum.
(C.21) is the so-called complementarity condition. It expresses that in the
minimum either a constraint is 0 (that means “active”) or the corresponding
Lagrange factor is 0. If both are zero at the same time, this corresponds to a
minimum of the objective function where the constraint just got active. Deleting
this constraint would not change the solution of the optimization problem.
A Lagrange factor of zero for an inactive constraint corresponds to deleting
it from the Lagrange function. Therefore we have that:
L(x∗ , λ∗ ) = f (x∗ ) (C.22)
(C.17)–(C.21) are known as Karush-Kuhn-Tucker(KKT) conditions.

C.6 Second-Order Constrained Optimality Condition

The second-order optimality condition for the minimum f ∗ of a constrained
optimization problem can be explained based on the quadratic form of the
objective around the optimum x∗ in a direction r:
(C.22)
f (x∗ + r) = L(x∗ + r, λ∗ )
180 ANALOG DESIGN CENTERING AND SIZING

1
= L(x∗ , λ∗ ) + ∇L(x∗ )T · r + rT ·∇2 L(x∗ )·r + . . . (C.23)
   2
0 (C.17)


∇2 L(x∗ ) = ∇2 f (x∗ ) − λ∗µ · ∇2 cµ (x∗ ) (C.24)
µ∈A(x∗ )

As in the unconstrained case, we require that no direction r may exist in which

the quadratic term in (C.23) leads to a reduction in the objective value f ∗ =
f (x∗ ). The difference to the unconstrained case is that we only have to consider
unconstrained directions, but that the curvatures of both the objective function
and the constraint functions are considered.

Necessary Condition. For any unconstrained stationary direction r from the

stationary point x∗ , no descent is achievable through the second-order derivative
of the Lagrangian function:

∗ T
∀ ∗
rT · ∇2 L(x∗ ) · r ≥ 0 (C.25)
∇cµ (x ) · r = 0 , µ ∈ A+ (x )
∇cµ (x∗ )T · r ≥ 0, µ ∈ A(x∗ ∗
 ) \ A+ (x ) 
A+ (x∗ ) = EC ∪ µ ∈ IC  cµ (x∗ ) = 0 ∧ λ∗µ > 0

Note that this is a weaker requirement than positive semidefiniteness, because

not all directions r are included in (C.25).

Sufficient Condition. The corresponding sufficient condition is that only an

increase in the objective function is obtained into unconstrained stationary
directions around the minimum:
∀ rT · ∇2 L(x∗ ) · r > 0 (C.26)
∇cµ (x∗ )T · r = 0 , µ ∈ A+ (x∗ )
∇cµ (x∗ )T · r ≥ 0, µ ∈ A(x ∗ ∗
 ) \∗A+ (x ) ∗ 
∗ 
A+ (x ) = EC ∪ µ ∈ IC cµ (x ) = 0 ∧ λµ > 0

Note that this is a weaker requirement than positive definiteness, because not
all directions r are included in (C.25).

C.6.1 Lagrange-Factor and Sensitivity to Constraint

Let the bound of the constraint cµ ≥ 0 in (C.10) be increased by µ : cµ ≥ µ .
Then the Lagrange function (C.11) becomes:

L(x, λ, ) = f (x) − λµ · (cµ (x) − µ ) (C.27)
µ

From the first-order optimality condition we have that ∇L(x∗ ) = 0 and

∇L(λ∗ ) = 0. Therefore the sensitivity of the objective function at the min-
imum is identical to the sensitivity of the Lagrange function at the minimum
Appendix C: Optimality Conditions of Nonlinear Optimization Problems 181

(C.22):
(C.27)
∇f ∗ (µ ) = ∇L∗ (µ ) = λ∗µ (C.28)
(C.28) says that the sensitivity of the minimum objective with respect to an
increase in the constraint boundary is equal to the corresponding Lagrange
factor.

C.7 Bounding-Box-of-Ellipsoids Property (37)

The bounding-box-of-ellipsoids property in (37) can be motivated for a para-
meter xk based on the following optimization problem:
max |xk | s.t. xT C−1 x = β 2 (C.29)
xk

Without loss of generality, we have assumed that x0 = 0.

The first-order optimality condition (Appendix C) for a solution x∗k of (C.29)
based on a corresponding Lagrangian function,
1
L(xk , λ) = xk − λ · (xT C−1 x − β 2 ) (C.30)
2
is:
∇L(x∗ ) ≡ ek − λ∗ C−1 x∗ = 0 (C.31)
∗T
x C−1 x∗ = β 2 (C.32)
ek is a vector with a 1 at the kth position and 0s at the other positions. Note that
(C.30) can be written with any sign of λ due to the equality constraint in (C.29).
Therefore, (C.30) covers both cases of (C.30), i.e., max xk ≡ − min −xk , and,
min xk .
With the following decomposition of the covariance matrix C,

C−1 = C / ⇔C=C
/T · C / −1 · C
/ −T , (C.33)
which can be obtained by a Cholesky decomposition or an eigenvalue decom-
position, (C.31) can be transformed into:

/ · x∗ = 1 · C
C / −T · ej (C.34)
λ∗
Applying (C.33) and two times (C.34) in (C.32) leads to
σk = |λ∗ | · β (C.35)
Applying (C.33) and one time (C.34) in (C.32) leads to
x∗k = λ∗ · β 2 (C.36)
182 ANALOG DESIGN CENTERING AND SIZING

From (C.35) and (C.36) and the Hessian matrix of the Lagrangian function
(C.30), ∇2 L(x) = −λ · C−1 , follows that the Lagrangian function (C.30) has
a maximum and a minimum with the absolute value |x∗k | = β · σj .
(C.29) therefore says that any ellipsoid with any correlation value leads to a
minimum and maximum value of x∗k = ±β · σk . This shows the lub property
of (37). For a complete proof we additionally have to show that in direction
of other parameters all values on the bounding box are reached by varying the
correlation. Instead we refer to the visual inspection of Figure 22(d).
References

[1] H. Abdel-Malek and A. Hassan. The ellipsoidal technique for design centering and
region approximation. IEEE Transactions on Computer-Aided Design of Circuits and
Systems, 10:1006–1013, 1991.

[2] D. Agnew. Improved minimax optimization for circuit design. IEEE Transactions on
Circuits and Systems CAS, 28:791–803, 1981.

[3] G. Alpaydin, S. Balkir, and G. Dundar. An evolutionary approach to automatic synthesis

of high-performance analog integrated circuits. IEEE Transactions on Evolutionary
Computation, 7(3):240–252, June 2003.

[4] Antonio R. Alvarez, Behrooz L. Abdi, Dennis L. Young, Harrison D. Weed, Jim Teplik,
and Eric R. Herald. Application of statistical design and response surface methods to
computer-aided VLSI device design. IEEE Transactions on Computer-Aided Design of
Circuits and Systems, 7(2):272–288, February 1988.

[5] T. Anderson. An Introduction to Multivariate Statistical Analysis. Wiley, New York,

1958.

[6] K. Antreich, H. Graeb, and C. Wieser. Circuit analysis and optimization driven by worst-
case distances. IEEE Transactions on Computer-Aided Design of Circuits and Systems,
13(1):57–71, January 1994.

[7] K. Antreich and S. Huss. An interactive optimization technique for the nominal design
of integrated circuits. IEEE Transactions on Circuits and Systems CAS, 31:203–212,
1984.

[8] K. Antreich and R. Koblitz. Design centering by yield prediction. IEEE Transactions
on Circuits and Systems CAS, 29:88–95, 1982.

[9] K. Antreich, P. Leibner, and F. Poernbacher. Nominal design of integrated circuits on

circuit level by an interactive improvement method. IEEE Transactions on Circuits and
Systems CAS, 35:1501–1511, 1988.

[10] Kurt J. Antreich, Helmut E. Graeb, and Rudolf K. Koblitz. Advanced Yield Optimization
Techniques, Volume 8 (Statistical Approach to VLSI) of Advances in CAD for VLSI.
Elsevier Science Publishers, Amsterdam, 1994.
184 REFERENCES

[11] J. Armaos. Zur Optimierung der Fertigungsausbeute elektrischer Schaltungen unter

Beruecksichtigung der Parametertoleranzen. PhD thesis, Technische Universitaet
Muenchen, 1982.

[12] J. Bandler and S. Chen. Circuit optimization: The state of the art. IEEE Transactions
on Microwaves Theory Techniques (MTT), 36:424–442, 1988.

[13] J. Bandler, S. Chen, S. Daijavad, and K. Madsen. Efficient optimization with integrated
gradient approximation. IEEE Transactions on Microwaves Theory Techniques (MTT),
36:444–455, 1988.

[14] T. Barker. Quality engineering by design: Taguchi’s philosophy. Quality Assurance,

13:72–80, 1987.

[15] Kamel Benboudjema, Mounir Boukadoum, Gabriel Vasilescu, and Georges Alquie.
Symbolic analysis of linear microwave circuits by extension of the polynomial interpo-
lation method. IEEE Transactions on Circuits and Systems I: Fundamental Theory and
Applications, 45(9):936, 1998.

[16] M. Bernardo, R. Buck, L. Liu, W. Nazaret, J. Sacks, and W. Welch. Integrated circuit
design optimization using a sequential strategy. IEEE Transactions on Computer-Aided
Design of Circuits and Systems, 11:361–372, 1992.

[17] R. Biernacki, J. Bandler, J. Song, and Q. Zhang. Efficient quadratic approximation

for statistical design. IEEE Transactions on Circuits and Systems CAS, 36:1449–1454,
1989.

[18] C. Borchers. Symbolic behavioral model generation of nonlinear analog circuits.

IEEE Transactions on Circuits and Systems II: Analog and Digital Signal Processing,
45(10):1362, 1998.

[19] Stephen Boyd and Lieven Vandenberghe. Convex Optimization. Cambridge University
Press, 2004.

[20] Graeme R. Boyle, Barry M Cohn, Danald O. Pederson, and James E. Solomon. Macro-
modeling of integrated operational amplifiers. IEEE Journal of Solid-State Circuits SC,
9(6):353–364, December 1974.

[21] R. Brayton, G. Hachtel, and A. Sangiovanni-Vincentelli. A survey of optimization tech-

niques for integrated-circuit design. Proceedings of the IEEE, 69:1334–1363, 1981.

[22] G. Casinovi and A. Sangiovanni-Vincentelli. A macromodeling algorithm for analog

circuits. IEEE Transactions on Computer-Aided Design of Circuits and Systems, 10:150–
160, 1991.

[23] R. Chadha, K. Singhal, J. Vlach, and E. Christen. WATOPT - an optimizer for circuit
applications. IEEE Transactions on Computer-Aided Design of Circuits and Systems,
6:472–479, 1987.

[24] H. Chang, E. Charbon, U. Choudhury, A. Demir, E. Felt, E. Liu, E. Malavasi,

A. Sangiovanni-Vincentelli, and I. Vassiliou. A Top-Down, Constraint-Driven Design
Methodology for Analog Integrated Circuits. Kluwer Academic Publishers, 1997.
References 185

[25] E. Christensen and J. Vlach. NETOPT – a program for multiobjective design of linear
networks. IEEE Transactions on Computer-Aided Design of Circuits and Systems,
7:567–577, 1988.

[26] M. Chu and D. J. Allstot. Elitist nondominated sorting genetic algorithm based rf ic
optimizer. IEEE Transactions on Circuits and Systems CAS, 52(3):535–545, March
2005.

[27] L. Chua. Global optimization: a naive approach. IEEE Transactions on Circuits and
Systems CAS, 37:966–969, 1990.

[28] Andrew R. Conn, Paula K. Coulman, Ruud A. Haring, Gregory L. Morill, Chandu
Visweswariah, and Chai Wah Wu. JiffyTune: Circuit optimization using time-domain
sensitivities. IEEE Transactions on Computer-Aided Design of Circuits and Systems,
17(12):1292–1309, December 1998.

[29] P. Cox, P. Yang, S. Mahant-Shetti, and P. Chatterjee. Statistical modeling for efficient
parametric yield estimation of MOS VLSI circuits. IEEE Transactions on Electron
Devices ED, 32:471–478, 1985.

[30] Walter Daems, Georges Gielen, and Willy Sansen. Circuit simplification for the symbolic
analysis of analog integrated circuits. IEEE Transactions on Computer-Aided Design of
Circuits and Systems, 21(4):395–407, April 2002.

[31] Walter Daems, Georges Gielen, and Willy Sansen. Simulation-based generation of
posynomial performance models for the sizing of analog integrated circuits. IEEE
Transactions on Computer-Aided Design of Circuits and Systems, 22(5):517–534, May
2003.

[32] Walter Daems, Wim Verhaegen, Piet Wambacq, Georges Gielen, and Willy Sansen.
Evaluation of error-control strategies for the linear symbolic analysis of analog inte-
grated circuits. IEEE Transactions on Circuits and Systems I: Fundamental Theory and
Applications, 46(5):594–606, May 1999.

[33] Nader Damavandi and Safieddin Safavi-Naeini. A hybrid evolutionary programming

method for circuit optimization. IEEE Transactions on Circuits and Systems CAS, 2005.

[34] Indraneel Das and J. E. Dennis. Normal-boundary intersection: A new method for
generating the Pareto surface in nonlinear multicriteria optimization problems. SIAM
Journal on Optimization, 8(3):631–657, August 1998.

[35] Bart De Smedt and Georges G. E. Gielen. WATSON: Design space boundary exploration
and model generation for analog and RF IC design. IEEE Transactions on Computer-
Aided Design of Circuits and Systems, 22(2):213–223, February 2003.

[36] Kalyanmoy Deb. Multi-objective optimization using evolutionary algorithms. Wiley-

Interscience Series in Systems and Optimizat. Wiley, 2001.

[37] Maria del Mar Hershenson, Stephen P. Boyd, and Thomas H. Lee. Optimal design of a
CMOS Op-Amp via geometric programming. IEEE Transactions on Computer-Aided
Design of Circuits and Systems, 20(1):1–21, January 2001.

[38] S. Director and G. Hachtel. The simplicial approximation approach to design centering.
IEEE Transactions on Circuits and Systems CAS, 24:363–372, 1977.
186 REFERENCES

[39] Alex Doboli and Ranga Vemuri. Behavioral modeling for high-level synthesis of ana-
log and mixed-signal systems from vhdl-ams. IEEE Transactions on Computer-Aided
Design of Circuits and Systems, 2003.

[40] K. Doganis and D. Scharfetter. General optimization and extraction of IC device model
parameters. IEEE Transactions on Electron Devices ED, 30:1219–1228, 1983.

[41] Hans Eschenauer, Juhani Koski, and Andrzej Osyczka. Multicriteria design optimiza-
tion: procedures and applications. Springer-Verlag, 1990.

[42] Mounir Fares and Bozena Kaminska. FPAD: A fuzzy nonlinear programming approach
to analog circuit design. IEEE Transactions on Computer-Aided Design of Circuits and
Systems, 14(7):785–793, July 1995.

[43] P. Feldmann and S. Director. Integrated circuit quality optimization using surface inte-
grals. IEEE Transactions on Computer-Aided Design of Circuits and Systems, 12:1868–
1879, 1993.

[44] F. V. Fernandez, O. Guerra, J. D. Rodriguez-Garcia, and A. Rodriguez-Vazquez. Sym-

bolic analysis of large analog integrated circuits: The numerical reference generation
problem. IEEE Transactions on Circuits and Systems II: Analog and Digital Signal
Processing, 45(10):1351, 1998.

[45] Roger Fletcher. Practical Methods of Optimization. John Wiley & Sons, 1987.

[46] Kenneth Francken and Georges G. E. Gielen. A high-level simulation and synthesis
environment for sigma delta modulators. IEEE Transactions on Computer-Aided Design
of Circuits and Systems, 22(8):1049–1061, August 2003.

[47] D.D. Gajski and R.H. Kuhn. Guest editor’s introduction: New VLSI tools. ieeecomputer,
16:11–14, 1983.

[48] Floyd W. Gembicki and Yacov Y. Haimes. Approach to performance and sensitivity mul-
tiobjective optimization: The goal attainment method. IEEE Transactions on Automatic
Control, 20(6):769–771, December 1975.

[49] Ian E. Getreu, Andreas D. Hadiwidjaja, and Johan M. Brinch. An integrated-circuit

comperator macromodel. IEEE Journal of Solid-State Circuits SC, 11(6):826–833,
December 1976.

[50] G. Gielen and W. Sansen. Symbolic Analysis for Automated Design of Analog Integrated
Circuits. Kluwer Academic Publishers, Dordrecht, 1991.

[51] G. Gielen, P. Wacambacq, and W. Sansen. Symbolic analysis methods and applications
for analog circuits: A tutorial overview. Proceedings of the IEEE, 82, 1994.

[52] G. Gielen, H. C. Walscharts, and W. C. Sansen. ISAAC: A symbolic simulation for analog
integrated circuits. IEEE Journal of Solid-State Circuits SC, 24:1587–1597, December
1989.

[53] G. Gielen, H. C. Walscharts, and W. C. Sansen. Analog circuit design optimization based
on symbolic simulation and simulated annealing. IEEE Journal of Solid-State Circuits
SC, 25:707–713, June 1990.
References 187

[54] Georges G. E. Gielen, Kenneth Francken, Ewout Martens, and Martin Vogels. An analyt-
ical integration method for the simulation of continuous-time delta-sigma modulators.
IEEE Transactions on Computer-Aided Design of Circuits and Systems, 2004.

[55] Georges G. E. Gielen and Rob A. Rutenbar. Computer-aided design of analog and mixed-
signal integrated circuits. Proceedings of the IEEE, 88(12):1825–1852, December 2000.

[56] Philip E. Gill, Walter Murray, and Margaret H. Wright. Practical Optimization. Aca-
demic Press. Inc., London, 1981.

[57] G. J. Gomez, S. H. K. Embabi, E. Sanchez-Sinencio, and M. C. Lefebvre. A nonlin-

ear macromodel for CMOS OTAs. In IEEE International Symposium on Circuits and
Systems (ISCAS), Volume 2, pages 920–923, 1995.

[58] H. Graeb, S. Zizala, J. Eckmueller, and K. Antreich. The sizing rules method for analog
integrated circuit design. In IEEE/ACM International Conference on Computer-Aided
Design (ICCAD), pages 343–349, 2001.

[59] A. Groch, L. Vidigal, and S. Director. A new global optimization method for electronic
circuit design. IEEE Transactions on Circuits and Systems CAS, 32:160–169, 1985.

[60] G. D. Hachtel and P. Zug. APLSTAP – circuit design and optimization system – user’s
guide. Technical report, IBM Yorktown Research Facility, Yorktown, New York, 1981.

[61] R. Hanson and C. Lawson. Solving Least Squares Problems. Prentice-Hall, New Jersey,
1974.

[62] R. Harjani, R. Rutenbar, and L. Carley. OASYS: A framework for analog circuit syn-
thesis. IEEE Transactions on Computer-Aided Design of Circuits and Systems, 8:1247–
1266, 1989.

[63] D. Hocevar, P. Cox, and P. Yang. Parametric yield optimization for MOS circuit blocks.
IEEE Transactions on Computer-Aided Design of Circuits and Systems, 7:645–658,
1988.

[64] B. Hoppe, G. Neuendorf, D. Schmitt-Landsiedel, and W. Specks. Optimization of

high-speed CMOS logic circuits with analytical models for signal delay, chip area,
and dynamic power consumption. IEEE Transactions on Computer-Aided Design of
Circuits and Systems, 9:236–247, 1990.

[65] Xiaoling Huang, Chris S. Gathercole, and H. Alan Mantooth. Modeling nonlinear
dynamics in analog circuits via root localization. IEEE Transactions on Computer-
Aided Design of Circuits and Systems, 2003.

[66] Ching-Lai Hwang and Abu Syed Md. Masud. Multiple Objective Decision Making.
Springer, 1979.

[67] Jacob Katzenelson and Aharon Unikovski. Symbolic-numeric circuit analysis or sym-
bolic ciruit analysis with online approximations. IEEE Transactions on Circuits and
Systems I: Fundamental Theory and Applications, 46(1):197–207, January 1999.

[68] S. Kirkpatrick, C. D. Gelatt, Jr., and M. P. Vecchi. Optimization by simulated annealing.

Technical Report RC 9355 (#41093), IBM Thomas J. Watson Research Center, IBM
Research Division, San Jose, Yorktown, Zurich, February 1982.
188 REFERENCES

[69] G. Kjellstroem and L. Taxen. Stochastic optimization in system design. IEEE Transac-
tions on Circuits and Systems CAS, 28:702–715, 1981.

[70] Ken Kundert, Henry Chang, Dan Jefferies, Gilles Lamant, Enrico Malavasi, and Fred
Sendig. Design of mixed-signal systems-on-a-chip. IEEE Transactions on Computer-
Aided Design of Circuits and Systems, 19(12):1561–1571, December 2000.

[71] Francky Leyn, Georges Gielen, and Willy Sansen. Analog small-signal modeling – part
I: Behavioral signal path modeling for analog integrated circuits. IEEE Transactions
on Circuits and Systems II: Analog and Digital Signal Processing, 48(7):701–711, July
2001.

[72] Francky Leyn, Georges Gielen, and Willy Sansen. Analog small-signal modeling –
part II: Elementary transistor stages analyzed with behavioral signal path modeling.
IEEE Transactions on Circuits and Systems II: Analog and Digital Signal Processing,
48(7):712–721, July 2001.

[73] M. Lightner, T. Trick, and R. Zug. Circuit optimization and design. Circuit Analysis,
Simulation and Design, Part 2 (A. Ruehli). Advances in CAD for VLSI 3, pages 333–391,
1987.

[74] M. R. Lightner and S. W. Director. Multiple criterion optimization for the design of
electronic circuits. IEEE Transactions on Circuits and Systems CAS, 28(3):169–179,
March 1981.

[75] V. Litovski and M. Zwolinski. VLSI Circuit Simulation and Optimization. Chapman
Hall, 1997.

[76] Hongzhou Liu, Amit Singhee, Rob A. Rutenbar, and L. Richard Carley. Remebrance
of circuits past: Macromodeling by data mining in large analog design spaces. In
ACM/IEEE Design Automation Conference (DAC), pages 437–442, 2002.

[77] Arun N. Lokanathan and Jay B. Brockman. A methodology for concurrent process-
circuit optimization. IEEE Transactions on Computer-Aided Design of Circuits and
Systems, 18(7):889–902, July 1999.

[78] K. Low and S. Director. An efficient methodology for building macromodels of IC

fabrication processes. IEEE Transactions on Computer-Aided Design of Circuits and
Systems, 8:1299–1313, 1989.

[79] D. Luenberger. Optimization By Vector Space Methods. John Wiley, New York, 1969.

[80] David G. Luenberger. Linear and Nonlinear Programming. Addison-Wesley Publishing

Company, 2 edition, May 1989.

[81] Pradip Mandal and V. Visvanathan. CMOS Op-Amp sizing using a geometric pro-
gramming formulation. IEEE Transactions on Computer-Aided Design of Circuits and
Systems, 20(1):22–38, January 2001.

[82] H. Alan Mantooth and Mike F. Fiegenbaum. Modeling with an Analog Hardware
Description Language. Kluwer Academic Publishers, November 1994.

[83] P. Maulik, L. R. Carley, and R. Rutenbar. Integer programming based topology selection
of cell-level analog circuits. IEEE Transactions on Computer-Aided Design of Circuits
and Systems, 14(4):401ff, April 1995.
References 189

[84] Petra Michel, Ulrich Lauther, and Peter Duzy. The Synthesis Approach to Digital System
Design. Kluwer Academic Publishers, Boston, 1992.

[85] Gordon E. Moore. Cramming more components onto integrated circuits. Electronics,
38(8), April 1965.

[86] Daniel Mueller, Guido Stehr, Helmut Graeb, and Ulf Schlichtmann. Deterministic
approaches to analog performance space exploration (PSE). In ACM/IEEE Design
Automation Conference (DAC), June 2005.

[87] G. Mueller-Liebler. PASTA – The characterization of the inherent fluctuations in the

fabrication process for circuit simulation. International Journal of Circuit Theory and
Applications, 23:413–432, 1995.

[88] MunEDA. WiCkeD – Design for Manufacturability and Yield. www.muneda.com, 2001.

[89] L. Nagel. SPICE2: A computer program to simulate semiconductor circuits. Ph. D.

dissertation, Univ. of California, Berkeley, 1975.

[90] Dongkyung Nam and Cheol Hoon Park. Multiobjective simulated annealing: A compar-
ative study to evolutionary algorithms. International Journal of Fuzzy Systems, pages
87–97, June 2000.

[91] Jorge Nocedal and Stephen J. Wright. Numerical Optimization. Springer, 1999.

[92] W. Nye, D. Riley, A. Sangiovanni-Vincentelli, and A. Tits. DELIGHT.SPICE: An

optimization-based system for the design of integrated circuits. IEEE Transactions
on Computer-Aided Design of Circuits and Systems, 7:501–519, 1988.

[93] E. Ochotta, T. Mukherjee, R.A. Rutenbar, and L.R. Carley. Practical Synthesis of High-
Performance Analog Circuits. Kluwer Academic Publishers, 1998.

[94] Emil S. Ochotta, Rob A. Rutenbar, and L. Richard Carley. Synthesis of high-performance
analog circuits in ASTRX/OBLX. IEEE Transactions on Computer-Aided Design of
Circuits and Systems, 15(3):273–294, March 1996.

[95] M. Pelgrom, A. Duinmaijer, and A. Welbers. Matching properties of MOS transistors.

IEEE Journal of Solid-State Circuits SC, 24:1433–1440, 1989.

[96] Rodney Phelps, Michael Krasnicki, Rob A. Rutenbar, L. Richard Carley, and James R.
Hellums. ANACONDA: Simulation-based synthesis of analog circuits via stochastic
pattern search. IEEE Transactions on Computer-Aided Design of Circuits and Systems,
19(6):703–717, June 2000.

[97] Lawrence T. Pillage, Ronald A. Rohrer, and Chandramouli Visweswariah. Electronic

Circuit and System Simulation Methods. McGraw-Hill, Inc., 1995.

[98] Ming Qu and M. A. Styblinski. Parameter extraction for statistical ic modeling based
on recursive inverse approximation. IEEE Transactions on Computer-Aided Design of
Circuits and Systems, 16(11):1250–1259, 1997.

[99] Joao Ramos, Kenneth Francken, Georges G. E. Gielen, and Michiel S. J. Steyaert. An
efficient, fully parasitic-aware power amplifier design optimization tool. IEEE Trans-
actions on Circuits and Systems I: Fundamental Theory and Applications, 2005.
190 REFERENCES

[100] Carl R. C. De Ranter, Geert Van der Plas, Michiel S. J. Steyaert, Georges G. E. Gielen,
and Willy M. C. Sansen. CYCLONE: Automated design and layout of RF LC-oscillators.
IEEE Transactions on Computer-Aided Design of Circuits and Systems, 21(10):1161–
1170, October 2002.

[101] A. Ruehli(Editor). Circuit Analysis, Simulation and Design. Advances in CAD for VLSI.
North-Holland, 1986.

[102] Youssef G. Saab and Vasant B. Rao. Combinatorial optimization by stochastic evolution.
IEEE Transactions on Computer-Aided Design of Circuits and Systems, 10(4):525–535,
April 1991.

[103] T. Sakurai, B. Lin, and A. Newton. Fast simulated diffusion: an optimization algorithm
for multiminimum problems and its application to MOSFET model parameter extraction.
IEEE Transactions on Computer-Aided Design of Circuits and Systems, 11:228–233,
1992.

[104] Sachin S. Sapatnekar, Vasant B. Rao, Pravin M. Vaidya, and Sung-Mo Kang. An exact
solution to the transistor sizing problem for CMOS circuits using convex optimization.
IEEE Transactions on Computer-Aided Design of Circuits and Systems, 12(11):1621–
1634, November 1993.

[105] M. Sharma and N. Arora. OPTIMA: A nonlinear model parameter extraction program
with statistical confidence region algorithms. IEEE Transactions on Computer-Aided
Design of Circuits and Systems, 12:982–987, 1993.

[106] C.-J. Richard Shi and Xiang-Dong Tan. Canonical symbolic analysis of large analog
circuits with determinant decision diagrams. IEEE Transactions on Computer-Aided
Design of Circuits and Systems, 19(1):1–18, January 2000.

[107] C.-J. Richard Shi and Xiang-Dong Tan. Compact representation and efficient generation
of s-expanded symbolic network functions for computer-aided analog circuit design.
IEEE Transactions on Computer-Aided Design of Circuits and Systems, 20(7):813, July
2001.

[108] Guoyong Shi, Bo Hu, and C.-J. Richard Shi. On symbolic model order reduction. IEEE
Transactions on Computer-Aided Design of Circuits and Systems, 2006.

[109] C. Spanos and S. Director. Parameter extraction for statistical IC process characteriza-
tion. IEEE Transactions on Computer-Aided Design of Circuits and Systems, 5:66–78,
1986.

[110] Thanwa Sripramong and Christofer Toumazou. The invention of cmos amplifiers using
genetic programming and current-flow analysis. IEEE Transactions on Computer-Aided
Design of Circuits and Systems, 2002.

[111] H.H. Szu and R.L. Hartley. Nonconvex optimization by fast simulated annealing. Pro-
ceedings of the IEEE, 75:1538–1540, 1987.

[112] Sheldon X.-D. Tan. A general hierarchical circuit modeling and simulation algorithm.
IEEE Transactions on Computer-Aided Design of Circuits and Systems, 2005.
References 191

[113] Sheldon X.-D. Tan and C.-J. Richard Shi. Efficient approximation of symbolic expres-
sions for analog behavioral modeling and analysis. IEEE Transactions on Computer-
Aided Design of Circuits and Systems, 2004.

[114] Xiangdong Tan and C.-J. Richard Shi. Hierarchical symbolic analysis of large analog
circuits with determinant decision diagrams. In IEEE International Symposium on
Circuits and Systems (ISCAS), page VI/318, 1998.

[115] Hua Tang, Hui Zhang, and Alex Doboli. Refinement-based synthesis of continuous-time
analog filters through successive domain pruning, plateau search, and adaptive sampling.
IEEE Transactions on Computer-Aided Design of Circuits and Systems, 2006.

[116] Antonio Torralba, Jorge Chavez, and Leopoldo G. Franquelo. FASY: A fuzzy-logic based
tool for analog synthesis. IEEE Transactions on Computer-Aided Design of Circuits and
Systems, 15(7):705–715, July 1996.

[117] Wim M. G. van Bokhoven and Domine M. W. Leenaerts. Explicit formulas for the
solutions of piecewise linear networks. IEEE Transactions on Circuits and Systems I:
Fundamental Theory and Applications, 46(9):1110ff., September 1999.

[118] Geert Van der Plas, Geert Debyser, Francky Leyn, Koen Lampaert, Jan Vandenbussche,
Georges Gielen, Willy Sansen, Petar Veselinovic, and Domine Leenaerts. AMGIE–A
synthesis environment for CMOS analog integrated circuits. IEEE Transactions on
Computer-Aided Design of Circuits and Systems, 20(9):1037–1058, September 2001.

[119] P. Wambacq, P. Dobrovolny, G. G. E. Gielen, and W. Sansen. Symbolic analysis of

large analog circuits using a sensitivity-driven enumeration of common spanning trees.
IEEE Transactions on Circuits and Systems II: Analog and Digital Signal Processing,
45(10):1341, 1998.

[120] P. Wambacq, G. G. E. Gielen, and W. Sansen. Symbolic network analysis methods

for practical analog integrated circuits: A survey. IEEE Transactions on Circuits and
Systems II: Analog and Digital Signal Processing, 45(10):1331, 1998.

[121] Jacob K. White and Alberto Sangiovanni-Vincentelli. Relaxation Techniques for the
Simulation of VLSI Circuits. Kluwer Academic Publishers, 1987.

[122] Claudia Wieser. Schaltkreisanalyse mit Worst-Case Abstaenden. PhD thesis, Technische
Universitaet Muenchen, 1994.

[123] J. Wojciechowski and J. Vlach. Ellipsoidal method for design centering and yield estima-
tion. IEEE Transactions on Computer-Aided Design of Circuits and Systems, 12:1570–
1579, 1993.

[124] X. Xiangming and R. Spence. Trade-off prediction and circuit performance optimization
using a second-order model. International Journal of Circuit Theory and Applications,
20:299–307, 1992.

[125] D. Young, J. Teplik, H. Weed, N. Tracht, and A. Alvarez. Application of statistical design
and response surface methods to computer-aided VLSI device design II: desirability
functions and Taguchi methods. IEEE Transactions on Computer-Aided Design of
Circuits and Systems, 10:103–115, 1991.
192 REFERENCES

[126] T. Yu, S. Kang, I. Hajj, and T. Trick. Statistical performance modeling and parametric
yield estimation of MOS VLSI. IEEE Transactions on Computer-Aided Design of
Circuits and Systems, 6:1013–1022, 1987.

[127] T. Yu, S. Kang, J. Sacks, and W. Welch. Parametric yield optimization of CMOS analogue
circuits by quadratic statistical circuit performance models. International Journal of
Circuit Theory and Applications, 19:579–592, 1991.

[128] J. Zou, D. Mueller, H. Graeb, and U. Schlichtmann. A CPPLL hierarchical optimiza-

tion methodology considering jitter, power and locking time. In ACM/IEEE Design
Automation Conference (DAC), pages 19–24, 2006.
Index

Acceptance function, 75, 107, 145 variance, 170

Acceptance region Expectation value, 32, 75, 108, 165
parameter, 73 estimator, 169
performance, 48, 72 variance, 170
AC simulation, 46 linear transformation, 167
Analog circuit, 4 First-order optimality condition
Analog design, 6 classical worst-case analysis, 87
Analog sizing, 10, 12 constrained optimization, 179
Analog synthesis, 9 general worst-case analysis, 97
parametric, 10–11 geometric yield analysis, 132
path, 12 realistic worst-case analysis, 92
structural, 10–11 statistical-yield optimization, 147
Architecture level, 11 unconstrained optimization, 174
Center of gravity, 147 Gaussian error propagation, 167
Central moment, 166 Gauss-Newton direction, 159
χ2 (chi-square)-distribution, 42, 120 Geometric yield analysis, 14, 128
Circuit level, 11
accuracy, 141
Circuit netlist, 8
complexity, 141
Condition number, 55
Geometric-yield optimization, 14, 153, 156
Confidence level, 114
least-squares/trust-region, 158
Corner worst case, 90
linear programming, 162
excessive robustness, 127
min-max, 162
Correlation, 35, 166
Gradient, 16, 50
Covariance matrix, 35, 146, 166
linear transformation, 167 statistical-yield optimization, 146
Cumulative distribution function, 31 deterministic design parameter, 152
Cumulative frequency function, 31 worst-case distance, 154
DC simulation, 46 Hessian matrix
Descent direction, 174, 177 statistical-yield optimization, 148
Design centering, 13, 76 Importance sampling, 110
Design flow, 10 Jacobian matrix, 50
Design level, 6 Karush-Kuhn-Tucker(KKT) conditions, 179
Design partitioning, 7 Lagrangian function
Design technology, 2 classical worst-case analysis, 87
Design view, 8 general worst-case analysis, 97
Digital design, 6 geometric yield analysis, 131
Discrete parameter, 89 realistic worst-case analysis, 92
Error function, 33 Least-squares optimization, 159
Estimator, 108 Levenberg-Marquardt direction, 159
bias, 170 Linear programming, 87, 162
194 INDEX

Lognormal distribution, 41 covered range, 54

Macromodels, 47 reference point, 54
Manufacturing tolerances, 13 sensitivity, 55
Mean value, 32, 146, 165 Second-order optimality condition
Min-max optimization, 162 constrained optimization, 180
Mismatch, 45 general worst-case analysis, 97
Mixed-signal circuit, 5 geometric yield analysis, 133
Moment, 165 realistic worst-case analysis, 92
Monte-Carlo analysis, 14, 109 statistical-yield optimization, 149
accuracy, 115 unconstrained optimization, 175
complexity, 115 Sensitivity, 49
Multiple-criteria optimization, 56 computation, 53, 69
Multiple-objective optimization, 15, 56, 69, 77, 156 finite-difference approximation, 53, 150
Multivariate normal distribution, 34 Lagrange factor, 180
Newton-type optimization, 150 matrix, 50
Node voltages, 46 Sequential Quadratic Programming, 97, 131
Nominal design, 13, 20, 56, 77, 162 Single-objective optimization, 15, 61, 76, 162
Numerical integration, 46 Singular distribution, 36
Numerical simulation, 16, 46, 69, 78, 108 Six-sigma design, 94, 100
Operating tolerances, 13 Smoothness, 49
Operational amplifier, 4, 8, 16, 156 Standard deviation, 32, 35, 166
Parameter, 27 Standard normal distribution, 33
design parameter, 27, 78 Statistical yield analysis, 14
range parameter, 28 Statistical-yield optimization, 14, 150
significance, 51 Symbolic analysis, 47
similarity, 51 Three-sigma design, 94, 100
statistical parameter, 27 Tolerance assignment, 80, 150
tolerances, 28–29 Tolerance class, 115
global, 13, 44 box, 29, 118
local, 13, 44 ellipsoid, 29, 119
Pareto front, 59, 159 interval, 116
Pareto optimality, 57 polytope, 29
Pareto optimization, 61 single-plane-bounded, 122, 128
Pareto point, 59 Tolerance design, 13
Performance feature, 45, 51–52 Trade-off, 57
Performance-feature bound, 48, 63, 65, 86, 91, 95 Transformation of statistical distributions, 40
Performance-specification feature, 75 TR simulation, 46
Performance specification, 47, 70 Truncated distribution, 89
Performance-specification feature, 47, 73, 108, 121, Trust-region optimization, 159
129 Unconstrained direction, 177
Phase-locked loop, 5, 10 Unconstrained optimization, 173
Probability density function, 22, 31, 37, 71 Uniform distribution, 41
truncated, 145 Univariate normal distribution, 32
Random variable, 32 Variance, 33, 166
standardization, 168 Variance/covariance matrix, 35, 166
RC circuit, 19, 105, 143 estimator, 169
nominal design, 20 Vector norm, 63
performance specification, 20 l1 , 63
yield optimization/design centering, 22 l2 , 63
Relative frequency function, 31 l∞ , 59, 63
Response surface modeling, 47 Worst-case analysis, 64, 66, 69, 76
Sample, 108 classical, 85–86
element, 169 general, 95
generation, 43 realistic, 85, 90, 105
size, 114 Worst-case distance, 94, 100, 131, 136, 156
Scaling, 54 target, 158
affine transformation, 55 Worst-case optimization, 13, 64, 68, 77
Index 195
Worst-case parameter vector, 57, 65–66 Yield approximation error, 103
classical, 88 Yield
general, 98 “cake”, 25
geometric yield analysis, 135 catastrophic, 23
realistic, 93 Yield estimator, 108
Worst-case performance, 65–66 variance, 111
classical, 89 Yield optimization, 13, 76
general, 98 Yield
realistic, 93 parametric, 23
Y-chart, 8 Yield partition, 75, 108, 121, 137
Yield, 65, 70, 72, 108 Z-branch currents, 46
Yield analysis, 75, 78