Design of Experiments in Nonlinear Models Asymptotic

Normality, Optimality Criteria and Small Sample Properties

Visit the link below to download the full version of this book:

https://medipdf.com/product/design-of-experiments-in-nonlinear-models-asymptotic
-normality-optimality-criteria-and-small-sample-properties/

Click Download Now

Luc Pronzato • Andrej Pázman

Design of Experiments
in Nonlinear Models
Asymptotic Normality, Optimality Criteria
and Small-Sample Properties

123
Luc Pronzato Andrej Pázman
French National Center Department of Applied Mathematics
for Scientific Research (CNRS) and Statistics
University of Nice Comenius University
Sophia Antipolis, France Bratislava, Slovakia

ISSN 0930-0325
ISBN 978-1-4614-6362-7 ISBN 978-1-4614-6363-4 (eBook)
DOI 10.1007/978-1-4614-6363-4
Springer New York Heidelberg Dordrecht London
Library of Congress Control Number: 2013932292

© Springer Science+Business Media New York 2013

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed. Exempted from this legal reservation are brief excerpts in connection
with reviews or scholarly analysis or material supplied specifically for the purpose of being entered and
executed on a computer system, for exclusive use by the purchaser of the work. Duplication of this pub-
lication or parts thereof is permitted only under the provisions of the Copyright Law of the Publisher’s
location, in its current version, and permission for use must always be obtained from Springer. Permis-
sions for use may be obtained through RightsLink at the Copyright Clearance Center. Violations are liable
to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of publica-
tion, neither the authors nor the editors nor the publisher can accept any legal responsibility for any errors
or omissions that may be made. The publisher makes no warranty, express or implied, with respect to the
material contained herein.

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

To Barbara, Lilas, Zélie and Nora,
to Tatiana, Miriam, René, Klára, Kristı́na, Emma, Filip, David,
and all our friends.
Preface

The final form of this volume differs a lot from our initial project that was born
in 2003. Our collaboration that was initiated about 10 years before was then
at a pace and we formulated the objective of reviving it through a project
that could be conducted over several years (we thought three or four) and
could serve as a motivation for maintaining regular exchanges between us.
Our initial idea was to write a short monograph that would expose the con-
nections between the asymptotic properties of estimators and experimental
design. This corresponds basically to parts which are covered by Chaps. 2–4.
The deviation from this initial project was progressive. First, we realized that
we had more to say about regression models with heteroscedastic errors than
what we expected initially and we found that the investigation of asymptotic
normality in the case of singular designs required a rather fundamental re-
vision (Chap. 3). We then quickly agreed that we could not avoid writing
a chapter on optimality criteria and optimum experimental design based on
asymptotic normality (Chap. 5). Up to that point, the presentation was rather
standard, although we gave more emphasis than usual to some particular as-
pects, like the estimation of a nonlinear function of the model parameters
and models with heteroscedastic errors. The results obtained during our col-
laboration in the 1990s encouraged us to write a chapter on non-asymptotic
design approaches (Chap. 6). The motivation for exposing our views on the
specific difficulties caused by nonlinear models in LS estimation had always
been present in our mind; this project gave us the opportunity to develop and
present some of these ideas (Chap. 7). Since this book focused on nonlinear
models, having a chapter devoted to the problem raised by the dependency
of an optimal experiment on the value of the parameters to be estimated ap-
peared to be essential (Chap. 8). Here some kind of prior for design purposes
is unavoidable, with the property that an incorrect prior causes less damage
when used for design than for estimation. Finally, we hesitated about indi-
cating or not algorithms for the optimization of the different design criteria
that are presented throughout the chapters. This could have led us quite far
from the initial project but at the same time was essential for the practical

VII
VIII Preface

use of the methods suggested. We reached a compromise solution where all

algorithms are gathered in a specific chapter (Chap. 9), where the principles
are indicated, in connection with more classical optimization methods. The
result is thus much different from what we planned in 2003 and this book
covers the following aspects.

Asymptotic Normality. The first three chapters expose the necessary back-
ground on asymptotic properties of estimators in nonlinear models. The pre-
sentation is mathematically rigorous, with detailed proofs indicated in an
appendix to improve readability. The stress here is on deriving asymptotic
properties of estimators from properties of the experimental design, in par-
ticular the “design measure” which is a basic notion in classical experimental
design since the pioneering work of Jack Kiefer in the early 1960s. For non-
linear models, this is not covered in other books on design and considered
in a few research papers only; in general, the published proofs of asymptotic
properties of estimators require many assumptions of different types, which
are usually rather technical and not directly related to the design. Besides
that, some results in Chap. 3 are new, e.g., on singular designs and on models
with misspecification or with parameterized variance.

Optimality Criteria. The next chapters concern optimum design more di-
rectly. Readers only interested in the application of optimal design method-
ology can possibly start by reading Chap. 5, where the classical theory is
presented together with several new aspects and results. The optimality cri-
teria considered in Chap. 5 are related to the asymptotic behavior of es-
timators. Optimality criteria obtained under non-asymptotic considerations
(small-sample situation) are considered in Chap. 6, while Chap. 7 concerns
the connection between design and identifiability/estimability issues, includ-
ing new extensions of some classical optimality criteria. Nonlinear models have
the particularity that an optimal design depends on the value of the parame-
ters to be estimated; this is considered in detail in Chap. 8. Once an optimality
criterion is chosen, we still need to optimize it; algorithms are presented in
Chap. 9 that cover all situations presented in Chaps. 5–8.

Small-sample Properties. The small-sample differential-geometric approach

to the subject is considered in Chap. 6; it mainly corresponds to results ob-
tained by the second author in a series of research papers. Chapter 7 contains
results on situations when the nonlinearity of the model is such that the es-
timator can be totally erroneous because of too small a sample at hand. This
subject is not much considered in the literature, even in research papers, al-
though it may be of crucial importance in applications.

Several academic friends gave a substantial support through exchanges of

different forms; a special mention goes to Radoslav Harman, Werner Müller,
Éric Thierry, Henry Wynn, and Anatoly Zhigljavsky. We thank Jean-Pierre
Gauchi, Rainer Schwabe, and particularly Henry Wynn, for their help and
 Preface IX

encouragement in reviewing our manuscript. Of course all possible mistakes

remain ours. Our friends and families gave a remarkable level of support. The
first author expresses his deepest thanks to the Bittner clan in St. Peter in
der Au (Austria) whose kind hospitality provided a perfect environment for
precious periods of intensive work during winter and summer holidays. The
second author thanks the I3S Laboratory for the friendly working atmosphere
during his many visits at Sophia Antipolis.
A long-term project like this one necessarily involves supports and grants of
different sources. In particular, the research of Andrej Pázman has been partly
supported by the VEGA grants Nb. 1/3016/06, 1/0077/09, and 2/0038/12.
The work of Luc Pronzato was partially supported by the IST Programme of
the European Community, under the PASCAL Network of Excellence, IST-
2002-506778; he also benefited from invitations by the Isaac Newton Institute
for Mathematical Sciences, Cambridge, UK, in summer 2008 and 2011. Andrej
Pázman benefited from several invitations from the University of Nice Sophia
Antipolis. The institutions of both authors, CNRS for Luc Pronzato and the
Comenius University for Andrej Pázman, are gratefully acknowledged for the
freedom they allowed through the years.

Sophia Antipolis, France Luc Pronzato

Bratislava, Slovakia Andrej Pázman
Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Experiments and Their Designs . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Information and Design Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2 Asymptotic Designs and Uniform Convergence . . . . . . . . . . . . 11

2.1 Asymptotic Designs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Uniform Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3 Bibliographic Notes and Further Remarks . . . . . . . . . . . . . . . . . . 19

3 Asymptotic Properties of the LS Estimator . . . . . . . . . . . . . . . . 21

3.1 Asymptotic Properties of the LS Estimator in Regression
Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.1.1 Consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.1.2 Consistency Under a Weaker LS Estimability
Condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.1.3 Asymptotic Normality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.1.4 Asymptotic Normality of a Scalar Function of the LS
Estimator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.2 Asymptotic Properties of Functions of the LS Estimator
Under Singular Designs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2.1 Singular Designs in Linear Models . . . . . . . . . . . . . . . . . . . 38
3.2.2 Singular Designs in Nonlinear Models . . . . . . . . . . . . . . . . 38
3.3 LS Estimation with Parameterized Variance . . . . . . . . . . . . . . . . 48
3.3.1 Inconsistency of WLS with Parameter-Dependent
Weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.3.2 Consistency and Asymptotic Normality of Penalized
WLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.3.3 Consistency and Asymptotic Normality
of Two-stage LS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

XI
XII Contents

3.3.4 Consistency and Asymptotic Normality of Iteratively

Reweighted LS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.3.5 Misspecification of the Variance Function . . . . . . . . . . . . . 57
3.3.6 Different Parameterizations for the Mean
and Variance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.3.7 Penalized WLS or Two-Stage LS? . . . . . . . . . . . . . . . . . . . 65
3.3.8 Variance Stabilization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.4 LS Estimation with Model Error . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.5 LS Estimation with Equality Constraints . . . . . . . . . . . . . . . . . . . 74
3.6 Bibliographic Notes and Further Remarks . . . . . . . . . . . . . . . . . . 76

4 Asymptotic Properties of M, ML, and Maximum

A Posteriori Estimators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.1 M Estimators in Regression Models . . . . . . . . . . . . . . . . . . . . . . . . 79
4.2 The Maximum Likelihood Estimator . . . . . . . . . . . . . . . . . . . . . . . 83
4.2.1 Regression Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.2.2 General Situation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.3 Generalized Linear Models and Exponential Families . . . . . . . . . 89
4.3.1 Models with a One-Dimensional Sufficient Statistic . . . . 90
4.3.2 Models with a Multidimensional Sufficient Statistic . . . . 92
4.4 The Cramér–Rao Inequality: Efficiency of Estimators . . . . . . . . 94
4.4.1 Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.4.2 Asymptotic Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.5 The Maximum A Posteriori Estimator . . . . . . . . . . . . . . . . . . . . . 98
4.6 Bibliographic Notes and Further Remarks . . . . . . . . . . . . . . . . . . 100

5 Local Optimality Criteria Based on Asymptotic

Normality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.1 Design Criteria and Their Properties . . . . . . . . . . . . . . . . . . . . . . . 108
5.1.1 Ellipsoid of Concentration . . . . . . . . . . . . . . . . . . . . . . . . . . 108
5.1.2 Classical Design Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.1.3 Positive Homogeneity, Concavity, and Isotonicity . . . . . . 114
5.1.4 Equivalence Between Criteria . . . . . . . . . . . . . . . . . . . . . . . 115
5.1.5 Concavity and Isotonicity of Classical Criteria . . . . . . . . 116
5.1.6 Classification into Global and Partial Optimality
Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
5.1.7 The Upper Semicontinuity of the c-Optimality
Criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
5.1.8 Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
5.1.9 Combining Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
5.1.10 Design with a Cost Constraint . . . . . . . . . . . . . . . . . . . . . . 124
5.2 Derivatives and Conditions for Optimality of Designs . . . . . . . . 125
5.2.1 Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
5.2.2 The Equivalence Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 131
5.2.3 Number of Support Points . . . . . . . . . . . . . . . . . . . . . . . . . . 139
5.2.4 Elfving’s Set and Some Duality Properties . . . . . . . . . . . . 140
 Contents XIII

5.3 c-Optimum Design in Linearized Nonlinear Models . . . . . . . . . . 142

5.3.1 Elfving’s Theorem and Related Properties . . . . . . . . . . . . 142
5.3.2 c-Maximin Efficiency and D-Optimality . . . . . . . . . . . . . . 147
5.3.3 A Duality Property for c-Optimality . . . . . . . . . . . . . . . . . 148
5.3.4 Equivalence Theorem for c-Optimality . . . . . . . . . . . . . . . 149
5.4 Specific Difficulties with c-Optimum Design in Presence
of Nonlinearity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
5.5 Optimality Criteria for Asymptotic Variance–Covariance
Matrices in Product Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
5.5.1 The WLS Estimator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
5.5.2 The Penalized WLS Estimator . . . . . . . . . . . . . . . . . . . . . . 156
5.5.3 The LS Estimator with Model Error . . . . . . . . . . . . . . . . . 157
5.5.4 The M Estimator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
5.6 Bibliographic Notes and Further Remarks . . . . . . . . . . . . . . . . . . 160

6 Criteria Based on the Small-Sample Precision of the LS

Estimator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
6.1 The Geometry of the Regression Model . . . . . . . . . . . . . . . . . . . . 168
6.1.1 Basic Notions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
6.1.2 A Classification of Nonlinear Regression Models . . . . . . . 169
6.1.3 Avoiding Failures of LS Estimation . . . . . . . . . . . . . . . . . . 172
6.2 The Probability Density of the LS Estimator in Nonlinear
Models with Normal Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
6.2.1 Intrinsically Linear Models . . . . . . . . . . . . . . . . . . . . . . . . . 174
6.2.2 Models with dim(θ) = 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
6.2.3 Flat Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
6.2.4 Models with Riemannian Curvature Tensor R(θ) ≡ 0 . . 176
6.2.5 Density of the Penalized LS Estimator . . . . . . . . . . . . . . . 176
6.2.6 Marginal Densities of the LS Estimator . . . . . . . . . . . . . . 177
6.3 Optimality Criteria Based on the p.d.f. of the LS Estimator . . . 178
6.4 Higher-Order Approximations of Optimality Criteria . . . . . . . . . 180
6.4.1 Approximate Bias and Mean-squared Error . . . . . . . . . . . 182
6.4.2 Approximate Entropy of the p.d.f. of the LS Estimator . 183
6.5 Bibliographic Notes and Further Remarks . . . . . . . . . . . . . . . . . . 184

7 Identifiability, Estimability, and Extended Optimality

Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
7.1 Identifiability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
7.2 LS Estimability of Regression Models . . . . . . . . . . . . . . . . . . . . . . 189
7.3 Numerical Issues Related to Estimability in Regression
Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
7.4 Estimability Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
7.4.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
7.4.2 Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
XIV Contents

7.4.3 Replications and Design Measures . . . . . . . . . . . . . . . . . . . 201

7.4.4 Estimability for Parametric Functions . . . . . . . . . . . . . . . . 203
7.5 An Extended Measure of Intrinsic Nonlinearity . . . . . . . . . . . . . . 207
7.6 Advantages and Drawbacks of Using p-point Designs . . . . . . . . . 211
7.7 Design of Experiments for Improving Estimability . . . . . . . . . . . 215
7.7.1 Extended (Globalized) E-Optimality . . . . . . . . . . . . . . . . . 215
7.7.2 Extended (Globalized) c-Optimality . . . . . . . . . . . . . . . . . 219
7.7.3 Maximum-Entropy Regularization of Estimability
Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
7.7.4 Numerical Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
7.8 Remarks on Estimability for Estimators Other than LS . . . . . . 231
7.9 Bibliographic Notes and Further Remarks . . . . . . . . . . . . . . . . . . 233

8 Nonlocal Optimum Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235

8.1 Average-Optimum Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
8.1.1 Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
8.1.2 A Bayesian Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . 238
8.2 Maximin-Optimum Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
8.3 Regularization of Maximin Criteria via Average Criteria . . . . . . 248
8.3.1 Regularization via Lq Norms . . . . . . . . . . . . . . . . . . . . . . . 248
8.3.2 Maximum-Entropy Regularization . . . . . . . . . . . . . . . . . . . 254
8.4 Probability Level and Quantile Criteria . . . . . . . . . . . . . . . . . . . . 259
8.5 Sequential Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
8.5.1 Two-Stage Allocation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
8.5.2 Full-Sequential D-Optimum Design for LS Estimation
in Nonlinear Regression Models . . . . . . . . . . . . . . . . . . . . . 271

9 Algorithms: A Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277

9.1 Maximizing a Concave Differentiable Functional
of a Probability Measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
9.1.1 Vertex-Direction Algorithms . . . . . . . . . . . . . . . . . . . . . . . . 279
9.1.2 Constrained Gradient and Gradient Projection . . . . . . . . 286
9.1.3 Multiplicative Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . 291
9.1.4 D-optimum Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
9.2 Exact Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
9.2.1 Exchange Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
9.2.2 Branch and Bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
9.3 Maximin-Optimum Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
9.3.1 Non-Differentiable Optimization of a Design Measure . . 302
9.3.2 Maximin-Optimum Exact Design . . . . . . . . . . . . . . . . . . . . 311
9.4 Average-Optimum Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
9.4.1 Average-Optimal Design Measures and Stochastic
Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
9.4.2 Average-Optimum Exact Design . . . . . . . . . . . . . . . . . . . . 313
9.5 Two Methods for Convex Programming . . . . . . . . . . . . . . . . . . . . 314
 Contents XV

9.5.1 Principles for Cutting Strategies and Interior-Point

Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
9.5.2 The Ellipsoid Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
9.5.3 The Cutting-Plane Method . . . . . . . . . . . . . . . . . . . . . . . . . 326

Subdifferentials and Subgradients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335

Computation of Derivatives Through Sensitivity Functions . . . . 339

Proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343

Symbols and Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361

List of Labeled Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389

Subject Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395