fm
340 61 28MB
Spanish Pages [740] Year 2009
DK3213_half 4/29/05 2:59 PM Page 1
Probability Models in Engineering and Science
© 2005 by Taylor & Francis Group, LLC
DK3151_series.qxd
4/20/05
2:26 PM
Page 1
MECHANICAL ENGINEERING A Series of Textbooks and Reference Books Founding Editor L. L. Faulkner Columbus Division, Battelle Memorial Institute and Department of Mechanical Engineering The Ohio State University Columbus, Ohio
1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 20. 21. 22. 23. 24.
Spring Designer’s Handbook, Harold Carlson ComputerAided Graphics and Design, Daniel L. Ryan Lubrication Fundamentals, J. George Wills Solar Engineering for Domestic Buildings, William A. Himmelman Applied Engineering Mechanics: Statics and Dynamics, G. Boothroyd and C. Poli Centrifugal Pump Clinic, Igor J. Karassik ComputerAided Kinetics for Machine Design, Daniel L. Ryan Plastics Products Design Handbook, Part A: Materials and Components; Part B: Processes and Design for Processes, edited by Edward Miller Turbomachinery: Basic Theory and Applications, Earl Logan, Jr. Vibrations of Shells and Plates, Werner Soedel Flat and Corrugated Diaphragm Design Handbook, Mario Di Giovanni Practical Stress Analysis in Engineering Design, Alexander Blake An Introduction to the Design and Behavior of Bolted Joints, John H. Bickford Optimal Engineering Design: Principles and Applications, James N. Siddall Spring Manufacturing Handbook, Harold Carlson Industrial Noise Control: Fundamentals and Applications, edited by Lewis H. Bell Gears and Their Vibration: A Basic Approach to Understanding Gear Noise, J. Derek Smith Chains for Power Transmission and Material Handling: Design a nd Applications Handbook, American Chain Association Corrosion and Corrosion Protection Handbook, edited by Philip A. Schweitzer Gear Drive Systems: Design and Application, Peter Lynwander Controlling InPlant Airborne Contaminants: Systems Design and Calculations, John D. Constance CAD/CAM Systems Planning and Implementation, Charles S. Knox Probabilistic Engineering Design: Principles and Applications, James N. Siddall Traction Drives: Selection and Application, Frederick W. Heilich III and Eugene E. Shube
© 2005 by Taylor & Francis Group, LLC
DK3151_series.qxd
25. 26. 27. 28. 29. 30. 31. 32. 33. 34. 35. 36. 37. 38. 39. 40. 41. 42. 43. 44. 45. 46. 47. 48. 49. 50. 51. 52. 53. 54. 55. 56.
4/20/05
2:26 PM
Page 2
Finite Element Methods: An Introduction, Ronald L. Huston and Chris E. Passerello Mechanical Fastening of Plastics: An Engineering Handbook, Brayton Lincoln, Kenneth J. Gomes, and James F. Braden Lubrication in Practice: Second Edition, edited by W. S. Robertson Principles of Automated Drafting, Daniel L. Ryan Practical Seal Design, edited by Leonard J. Martini Engineering Documentation for CAD/CAM Applications, Charles S. Knox Design Dimensioning with Computer Graphics Applications, Jerome C. Lange Mechanism Analysis: Simplified Graphical and Analytical Techniques, Lyndon O. Barton CAD/CAM Systems: Justification, Implementation, Productivity Measurement, Edward J. Preston, George W. Crawford, and Mark E. Coticchia Steam Plant Calculations Manual, V. Ganapathy Design Assurance for Engineers and Managers, John A. Burgess Heat Transfer Fluids and Systems for Process and Energy Applications, Jasbir Singh Potential Flows: Computer Graphic Solutions, Robert H. Kirchhoff ComputerAided Graphics and Design: Second Edition, Daniel L. Ryan Electronically Controlled Proportional Valves: Selection and Application, Michael J. Tonyan, edited by Tobi Goldoftas Pressure Gauge Handbook, AMETEK, U.S. Gauge Division, edited by Philip W. Harland Fabric Filtration for Combustion Sources: Fundamentals and Basic Technology, R. P. Donovan Design of Mechanical Joints, Alexander Blake CAD/CAM Dictionary, Edward J. Preston, George W. Crawford, and Mark E. Coticchia Machinery Adhesives for Locking, Retaining, and Sealing, Girard S. Haviland Couplings and Joints: Design, Selection, and Application, Jon R. Mancuso Shaft Alignment Handbook, John Piotrowski BASIC Programs for Steam Plant Engineers: Boilers, Combustion, Fluid Flow, and Heat Transfer, V. Ganapathy Solving Mechanical Design Problems with Computer Graphics, Jerome C. Lange Plastics Gearing: Selection and Application, Clifford E. Adams Clutches and Brakes: Design and Selection, William C. Orthwein Transducers in Mechanical and Electronic Design, Harry L. Trietley Metallurgical Applications of ShockWave and HighStrainRate Phenomena, edited by Lawrence E. Murr, Karl P. Staudhammer, and Marc A. Meyers Magnesium Products Design, Robert S. Busk How to Integrate CAD/CAM Systems: Management and Technology, William D. Engelke Cam Design and Manufacture: Second Edition; with cam design software for the IBM PC and compatibles, disk included, Preben W. Jensen SolidState AC Motor Controls: Selection and Application, Sylvester Campbell
© 2005 by Taylor & Francis Group, LLC
DK3151_series.qxd
57. 58. 59. 60. 61. 62. 63. 64. 65. 66. 67. 68. 69. 70. 71. 72. 73. 74. 75. 76. 77. 78. 79. 80. 81. 82. 83. 84. 85. 86. 87.
4/20/05
2:26 PM
Page 3
Fundamentals of Robotics, David D. Ardayfio Belt Selection and Application for Engineers, edited by Wallace D. Erickson Developing ThreeDimensional CAD Software with the IBM PC, C. Stan Wei Organizing Data for CIM Applications, Charles S. Knox, with contributions by Thomas C. Boos, Ross S. Culverhouse, and Paul F. Muchnicki ComputerAided Simulation in Railway Dynamics, by Rao V. Dukkipati and Joseph R. Amyot FiberReinforced Composites: Materials, Manufacturing, and Design, P. K. Mallick Photoelectric Sensors and Controls: Selection and Application, Scott M. Juds Finite Element Analysis with Personal Computers, Edward R. Champion, Jr. and J. Michael Ensminger Ultrasonics: Fundamentals, Technology, Applications: Second Edition, Revised and Expanded, Dale Ensminger Applied Finite Element Modeling: Practical Problem Solving for Engineers, Jeffrey M. Steele Measurement and Instrumentation in Engineering: Principles and Basic Laboratory Experiments, Francis S. Tse and Ivan E. Morse Centrifugal Pump Clinic: Second Edition, Revised and Expanded, Igor J. Karassik Practical Stress Analysis in Engineering Design: Second Edition, Revised and Expanded, Alexander Blake An Introduction to the Design and Behavior of Bolted Joints: Second Edition, Revised and Expanded, John H. Bickford High Vacuum Technology: A Practical Guide, Marsbed H. Hablanian Pressure Sensors: Selection and Application, Duane Tandeske Zinc Handbook: Properties, Processing, and Use in Design, Frank Porter Thermal Fatigue of Metals, Andrzej Weronski and Tadeusz Hejwowski Classical and Modern Mechanisms for Engineers and Inventors, Preben W. Jensen Handbook of Electronic Package Design, edited by Michael Pecht ShockWave and HighStrainRate Phenomena in Materials, edited by Marc A. Meyers, Lawrence E. Murr, and Karl P. Staudhammer Industrial Refrigeration: Principles, Design and Applications, P. C. Koelet Applied Combustion, Eugene L. Keating Engine Oils and Automotive Lubrication, edited by Wilfried J. Bartz Mechanism Analysis: Simplified and Graphical Techniques, Second Edition, Revised and Expanded, Lyndon O. Barton Fundamental Fluid Mechanics for the Practicing Engineer, James W. Murdock FiberReinforced Composites: Materials, Manufacturing, and Design, Second Edition, Revised and Expanded, P. K. Mallick Numerical Methods for Engineering Applications, Edward R. Champion, Jr. Turbomachinery: Basic Theory and Applications, Second Edition, Revised and Expanded, Earl Logan, Jr. Vibrations of Shells and Plates: Second Edition, Revised and Expanded, Werner Soedel Steam Plant Calculations Manual: Second Edition, Revised and Expanded, V. Ganapathy
© 2005 by Taylor & Francis Group, LLC
DK3151_series.qxd
88. 89. 90. 91. 92. 93. 94. 95. 96. 97. 98. 99. 100. 101. 102. 103. 104. 105. 106. 107. 108. 109. 110. 111. 112. 113. 114. 115. 116. 117. 118. 119.
4/20/05
2:26 PM
Page 4
Industrial Noise Control: Fundamentals and Applications, Second Edition, Revised and Expanded, Lewis H. Bell and Douglas H. Bell Finite Elements: Their Design and Performance, Richard H. MacNeal Mechanical Properties of Polymers and Composites: Second Edition, Revised and Expanded, Lawrence E. Nielsen and Robert F. Landel Mechanical Wear Prediction and Prevention, Raymond G. Bayer Mechanical Power Transmission Components, edited by David W. South and Jon R. Mancuso Handbook of Turbomachinery, edited by Earl Logan, Jr. Engineering Documentation Control Practices and Procedures, Ray E. Monahan Refractory Linings Thermomechanical Design and Applications, Charles A. Schacht Geometric Dimensioning and Tolerancing: Applications and Techniques for Use in Design, Manufacturing, and Inspection, James D. Meadows An Introduction to the Design and Behavior of Bolted Joints: Third Edition, Revised and Expanded, John H. Bickford Shaft Alignment Handbook: Second Edition, Revised and Expanded, John Piotrowski ComputerAided Design of PolymerMatrix Composite Structures, edited by Suong Van Hoa Friction Science and Technology, Peter J. Blau Introduction to Plastics and Composites: Mechanical Properties and Engineering Applications, Edward Miller Practical Fracture Mechanics in Design, Alexander Blake Pump Characteristics and Applications, Michael W. Volk Optical Principles and Technology for Engineers, James E. Stewart Optimizing the Shape of Mechanical Elements and Structures, A. A. Seireg and Jorge Rodriguez Kinematics and Dynamics of Machinery, Vladimír Stejskal and Michael Valásek Shaft Seals for Dynamic Applications, Les Horve ReliabilityBased Mechanical Design, edited by Thomas A. Cruse Mechanical Fastening, Joining, and Assembly, James A. Speck Turbomachinery Fluid Dynamics and Heat Transfer, edited by Chunill Hah HighVacuum Technology: A Practical Guide, Second Edition, Revised and Expanded, Marsbed H. Hablanian Geometric Dimensioning and Tolerancing: Workbook and Answerbook, James D. Meadows Handbook of Materials Selection for Engineering Applications, edited by G. T. Murray Handbook of Thermoplastic Piping System Design, Thomas Sixsmith and Reinhard Hanselka Practical Guide to Finite Elements: A Solid Mechanics Approach, Steven M. Lepi Applied Computational Fluid Dynamics, edited by Vijay K. Garg Fluid Sealing Technology, Heinz K. Muller and Bernard S. Nau Friction and Lubrication in Mechanical Design, A. A. Seireg Influence Functions and Matrices, Yuri A. Melnikov
© 2005 by Taylor & Francis Group, LLC
DK3151_series.qxd
120. 121. 122. 123. 124. 125. 126. 127. 128. 129. 130. 131. 132. 133. 134. 135. 136. 137. 138. 139. 140. 141. 142. 143. 144. 145. 146. 147. 148. 149. 150.
4/20/05
2:26 PM
Page 5
Mechanical Analysis of Electronic Packaging Systems, Stephen A. McKeown Couplings and Joints: Design, Selection, and Application, Second Edition, Revised and Expanded, Jon R. Mancuso Thermodynamics: Processes and Applications, Earl Logan, Jr. Gear Noise and Vibration, J. Derek Smith Practical Fluid Mechanics for Engineering Applications, John J. Bloomer Handbook of Hydraulic Fluid Technology, edited by George E. Totten Heat Exchanger Design Handbook, T. Kuppan Designing for Product Sound Quality, Richard H. Lyon Probability Applications in Mechanical Design, Franklin E. Fisher and Joy R. Fisher Nickel Alloys, edited by Ulrich Heubner Rotating Machinery Vibration: Problem Analysis and Troubleshooting, Maurice L. Adams, Jr. Formulas for Dynamic Analysis, Ronald L. Huston and C. Q. Liu Handbook of Machinery Dynamics, Lynn L. Faulkner and Earl Logan, Jr. Rapid Prototyping Technology: Selection and Application, Kenneth G. Cooper Reciprocating Machinery Dynamics: Design and Analysis, Abdulla S. Rangwala Maintenance Excellence: Optimizing Equipment LifeCycle Decisions, edited by John D. Campbell and Andrew K. S. Jardine Practical Guide to Industrial Boiler Systems, Ralph L. Vandagriff Lubrication Fundamentals: Second Edition, Revised and Expanded, D. M. Pirro and A. A. Wessol Mechanical Life Cycle Handbook: Good Environmental Design a nd Manufacturing, edited by Mahendra S. Hundal Micromachining of Engineering Materials, edited by Joseph McGeough Control Strategies for Dynamic Systems: Design and Implementation, John H. Lumkes, Jr. Practical Guide to Pressure Vessel Manufacturing, Sunil Pullarcot Nondestructive Evaluation: Theory, Techniques, and Applications, edited by Peter J. Shull Diesel Engine Engineering: Thermodynamics, Dynamics, Design, and Control, Andrei Makartchouk Handbook of Machine Tool Analysis, Ioan D. Marinescu, Constantin Ispas, and Dan Boboc Implementing Concurrent Engineering in Small Companies, Susan Carlson Skalak Practical Guide to the Packaging of Electronics: Thermal and Mechanical Design and Analysis, Ali Jamnia Bearing Design in Machinery: Engineering Tribology and Lubrication, Avraham Harnoy Mechanical Reliability Improvement: Probability and Statistics for Experimental Testing, R. E. Little Industrial Boilers and Heat Recovery Steam Generators: Design, Applications, and Calculations, V. Ganapathy The CAD Guidebook: A Basic Manual for Understanding and Improving ComputerAided Design, Stephen J. Schoonmaker
© 2005 by Taylor & Francis Group, LLC
DK3151_series.qxd
151. 152. 153. 154. 155. 156. 157. 158. 159. 160. 161. 162. 163. 164. 165. 166. 167. 168. 169. 170. 171. 172. 173. 174. 175. 176. 177. 178. 179. 180. 181.
4/20/05
2:26 PM
Page 6
Industrial Noise Control and Acoustics, Randall F. Barron Mechanical Properties of Engineered Materials, Wolé Soboyejo Reliability Verification, Testing, and Analysis in Engineering Design, Gary S. Wasserman Fundamental Mechanics of Fluids: Third Edition, I. G. Currie Intermediate Heat Transfer, KauFui Vincent Wong HVAC Water Chillers and Cooling Towers: Fundamentals, Application, and Operation, Herbert W. Stanford III Gear Noise and Vibration: Second Edition, Revised and Expanded, J. Derek Smith Handbook of Turbomachinery: Second Edition, Revised and Expanded, edited by Earl Logan, Jr. and Ramendra Roy Piping and Pipeline Engineering: Design, Construction, Maintenance, Integrity, and Repair, George A. Antaki Turbomachinery: Design and Theory, Rama S. R. Gorla and Aijaz Ahmed Khan Target Costing: MarketDriven Product Design, M. Bradford Clifton, Henry M. B. Bird, Robert E. Albano, and Wesley P. Townsend Fluidized Bed Combustion, Simeon N. Oka Theory of Dimensioning: An Introduction to Parameterizing Geometric Models, Vijay Srinivasan Handbook of Mechanical Alloy Design, edited by George E. Totten, Lin Xie, and Kiyoshi Funatani Structural Analysis of Polymeric Composite Materials, Mark E. Tuttle Modeling and Simulation for Material Selection and Mechanical Design, edited by George E. Totten, Lin Xie, and Kiyoshi Funatani Handbook of Pneumatic Conveying Engineering, David Mills, Mark G. Jones, and Vijay K. Agarwal Clutches and Brakes: Design and Selection, Second Edition, William C. Orthwein Fundamentals of Fluid Film Lubrication: Second Edition, Bernard J. Hamrock, Steven R. Schmid, and Bo O. Jacobson Handbook of LeadFree Solder Technology for Microelectronic Assemblies, edited by Karl J. Puttlitz and Kathleen A. Stalter Vehicle Stability, Dean Karnopp Mechanical Wear Fundamentals and Testing: Second Edition, Revised and Expanded, Raymond G. Bayer Liquid Pipeline Hydraulics, E. Shashi Menon Solid Fuels Combustion and Gasification, Marcio L. de SouzaSantos Mechanical Tolerance Stackup and Analysis, Bryan R. Fischer Engineering Design for Wear, Raymond G. Bayer Vibrations of Shells and Plates: Third Edition, Revised and Expanded, Werner Soedel Refractories Handbook, edited by Charles A. Schacht Practical Engineering Failure Analysis, Hani M. Tawancy, Anwar UlHamid, and Nureddin M. Abbas Mechanical Alloying and Milling, C. Suryanarayana Mechanical Vibration: Analysis, Uncertainties, and Control, Second Edition, Revised and Expanded, Haym Benaroya
© 2005 by Taylor & Francis Group, LLC
DK3151_series.qxd
182. 183. 184. 185. 186. 187. 188. 189. 190. 191. 192. 193. 194.
4/20/05
2:26 PM
Page 7
Design of Automatic Machinery, Stephen J. Derby Practical Fracture Mechanics in Design: Second Edition, Revised and Expanded, Arun Shukla Practical Guide to Designed Experiments, Paul D. Funkenbusch Gigacycle Fatigue in Mechanical Practive, Claude Bathias and Paul C. Paris Selection of Engineering Materials and Adhesives, Lawrence W. Fisher Boundary Methods: Elements, Contours, and Nodes, Subrata Mukherjee and Yu Xie Mukherjee Rotordynamics, Agnieszka (Agnes) Muszn´yska Pump Characteristics and Applications: Second Edition, Michael W. Volk Reliability Engineering: Probability Models and Maintenance Methods, Joel A. Nachlas Industrial Heating: Principles, Techniques, Materials, Applications, and Design, Yeshvant V. Deshmukh Micro Electro Mechanical System Design, James J. Allen Probability Models in Engineering and Science, Haym Benaroya and Seon Mi Han Damage Mechanics, George Z. Voyiadjis and Peter I. Kattan
© 2005 by Taylor & Francis Group, LLC
DK3213_title 5/16/05 8:52 AM Page 1
Probability Models in Engineering and Science Haym Benaroya and Seon Mi Han
Boca Raton London New York Singapore
A CRC title, part of the Taylor & Francis imprint, a member of the Taylor & Francis Group, the academic division of T&F Informa plc.
© 2005 by Taylor & Francis Group, LLC
DK3213_Discl.fm Page 1 Friday, May 20, 2005 9:29 AM
Cover: The autocorrelation function when 1 S 0 ( w ) = for w 2 < w < w 2 w2 – w1
and zero elsewhere. When w2 w1, the process X(t) is a narrow band process, and when w2 is far from w1, the process X(t) is a broad band process. See also Figures 5.28 and 5.29.
Published in 2005 by CRC Press Taylor & Francis Group 6000 Broken Sound Parkway NW, Suite 300 Boca Raton, FL 334872742 © 2005 by Taylor & Francis Group, LLC CRC Press is an imprint of Taylor & Francis Group No claim to original U.S. Government works Printed in the United States of America on acidfree paper 10 9 8 7 6 5 4 3 2 1 International Standard Book Number10: 0824723155 (Hardcover) International Standard Book Number13: 9780824723156 (Hardcover) Library of Congress Card Number 2005043711 This book contains information obtained from authentic and highly regarded sources. Reprinted material is quoted with permission, and sources are indicated. A wide variety of references are listed. Reasonable efforts have been made to publish reliable data and information, but the author and the publisher cannot assume responsibility for the validity of all materials or for the consequences of their use. No part of this book may be reprinted, reproduced, transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying, microfilming, and recording, or in any information storage or retrieval system, without written permission from the publishers. For permission to photocopy or use material electronically from this work, please access www.copyright.com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC) 222 Rosewood Drive, Danvers, MA 01923, 9787508400. CCC is a notforprofit organization that provides licenses and registration for a variety of users. For organizations that have been granted a photocopy license by the CCC, a separate system of payment has been arranged. Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for identification and explanation without intent to infringe.
Library of Congress CataloginginPublication Data Benaroya, Haym, 1954Probability models in engineering and science / Haym Benaroya, Seon Han. p. cm.  (Mechanical engineering ; v. 192) Includes bibliographical references and index. ISBN 0824723155 1. Reliability (Engineering)Mathematical models. I. Han, Seon Mi. II. Title. III. Mechanical engineering series (Boca Raton, Fla.) ; v. 192. TA169.B464 2005 620'.000452'015118dc22
2005043711
Visit the Taylor & Francis Web site at http://www.taylorandfrancis.com Taylor & Francis Group is the Academic Division of T&F Informa plc.
© 2005 by Taylor & Francis Group, LLC
and the CRC Press Web site at http://www.crcpress.com
The first author dedicates this book to his parents Esther and Alfred Benaroya, to whom he is grateful for much. The second author dedicates this book to her son James, who brought so much joy to her life.
© 2005 by Taylor & Francis Group, LLC
Preface There are not many texts that develop applied probability for engineering and scientific applications. Our intention in writing this text is to provide engineers and scientists a selfcontained introduction to probabilistic modeling. This book can be studied in two semesters, one undergraduate and one graduate. Distinguishing features worth noting include: numerous example problems from simple to challenging, short biographies and portraits of some of the key “names” mentioned, and the completely selfcontained nature of the presentation, thus making this book very suitable for selfstudy. Also included are endofchapter problems, plus footnotes to the literature. These footnotes have two purposes. The first is to provide proper citation and expanded discussion. There is no separate list of references in this text, and the footnotes serve as such attribution. The second is to introduce the reader to the relevant journal literature and to some of the very useful texts. The biographical summaries provided cannot do justice to the richness of the history of the field or its applications. Rather, the intent is to add, for the reader, the essential human connection to this subject. The history of the subject provides valuable insights to the subject, even though we rarely spend much time in its consideration when we are introduced to the field. These biographies, and portraits, downloaded from the World Wide Web, are included here by courtesy and permission of Professors E.F. Robertson and J.J. O’Connor, School of Mathematical and Computer Sciences, University of St. Andrews, St. Andrews, Scotland. The Web site is at :// . .  ../ /.. Finally, that this text is selfcontained means that the student may start at the beginning and continue to the end with rare need to refer to other works, except to find additional perspectives on the subject. But then, no one text can cover all aspects of a subject as broad as probability. Where more details become necessary, other works are cited where the reader will find additional information. The instructor may choose a variety of options regarding the use of this © 2005 by Taylor & Francis Group, LLC
v
vi
PREFACE
text. Generally, an undergraduate introduction will include Chapters 14 on probability, plus parts of Chapter 5 on random processes, Chapter 6 on singledegreeoffreedom random vibration, and Chapter 9 on reliability. A graduate, or second, course will include much of the second part of the text, Chapters 513. We have decided to create a Web page for this text where corrections and additional notes will be placed. On occasion interesting items will also be added. That site is found listed in :// . ./ .. Readers’ comments are most welcome, as are any suggestions and corrections. All of these would be most appreciated. The authors may be reached at @ . ., and @ ... All messages will be acknowledged.
© 2005 by Taylor & Francis Group, LLC
Authors Dr. Haym Benaroya received the B.E. from The Cooper Union for the Advancement of Science and Art, in 1976, and his M.S. and Ph.D. from the University of Pennsylvania, in 1977 and 1981, respectively. He worked for Weidlinger Associates, Consulting Engineers, New York, between 1981 and 1989, at which time he joined Rutgers University. He is currently a Professor of Mechanical and Aerospace Engineering at Rutgers, the State University of New Jersey. His research interests include structures and vibration, offshore structural dynamics, fluidstructure interaction, aircraft structures, and the development of concepts for lunar structures. Related interests include science, space and defense policy, and educational methods and policy. He is the director of the newlyformed Center for Structures in Extreme Environments, about which more information can be obtained through :// . ./.. Professor Benaroya is the author of numerous publications and other books, and is a Fellow of the American Society of Mechanical Engineers, a Fellow of the British Interplanetary Society, an Associate Fellow of the American Institute of Aeronautics and Astronautics, and a Corresponding Member of the International Academy of Astronautics. Dr. Seon Mi Han received the B.E. from The Cooper Union for the Advancement of Science and Art in 1996, and her M.S. and Ph.D. from Rutgers, the State University of New Jersey, in 1998 and 2001, respectively. She was a National Science Foundation summer intern at Korea Institute of Technology and Tokyo Institute of Technology. She received the Woods Hole Oceanographic Institution Postdoctoral Scholarship between 2001 and 2003, and she is currently an Assistant Professor of Mechanical Engineering at Texas Tech University. Her research interests include vibration and dynamics of offshore and marine structures.
© 2005 by Taylor & Francis Group, LLC
vii
Acknowledgments No project of this magnitude is completed without the explicit and implicit assistance of others. We acknowledge the input and efforts made by Professors M. Nagurka, M. Noori, and A. Zerva, who spent their valuable time reviewing an early manuscript. This is very much appreciated. We thank Professors E.F. Robertson and J.J. O’Connor who graciously allowed us to use the fruits of their labor, the biographical sketches included here. We are also grateful to John Corrigan, who originally signed this book for Marcel Dekker. We thank our students who have read sections of the text and provided useful feedback: Rene Gabbai, Mangala Gadagi, Subramanian Ramakrishnan, and Seyul Son. The first author wishes to acknowledge the Department of Mechanical and Aerospace Engineering, at the School of Engineering at Rutgers University, that provides an excellent scholarly environment, and especially Professor A. Zebib for many years of encouragement and support. The second author would like to acknowledge the Mechanical Engineering Department at Texas Tech University, and the support and encouragement of its former chairman, Dr. Thomas Burton. The first author is also pleased to thank Dr. T. Swean, of the Office of Naval Research, for the continued support of his research activity. Such research permits the refinement and development of concepts necessary for the advancement of engineering, in particular, and society, in general. It also is needed for the creation of texts such as this.
© 2005 by Taylor & Francis Group, LLC
ix
Contents 1 Introduction
1.1 Applications . . . . . . . . . . . . 1.1.1 Random Vibration . . . . 1.1.2 Fatigue Life . . . . . . . . 1.1.3 OceanWave Forces . . . . 1.1.4 Wind Forces . . . . . . . 1.1.5 Material Properties . . . . 1.1.6 Statistics and Probability 1.2 Units . . . . . . . . . . . . . . . . 1.3 Organization of the Text . . . . . 1.4 Problems . . . . . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
2.1 Sets . . . . . . . . . . . . . . . . . . 2.1.1 Basic Events . . . . . . . . . 2.1.2 Operational Rules . . . . . . 2.2 Probability . . . . . . . . . . . . . . 2.2.1 Axioms of Probability . . . . 2.2.2 Extensions from the Axioms . 2.2.3 Conditional Probability . . . 2.2.4 Statistical Independence . . . 2.2.5 Total Probability . . . . . . . 2.2.6 Bayes’ Theorem . . . . . . . 2.3 Concluding Summary . . . . . . . . 2.4 Problems . . . . . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
2 Events and Probability
3 Random Variable Models
3.1 Probability Distribution Function . 3.2 Probability Density Function . . . 3.3 Mathematical Expectation . . . . . 3.3.1 Variance . . . . . . . . . . .
© 2005 by Taylor & Francis Group, LLC
xi
. . . .
1
1 2 3 5 8 8 11 12 13 13
15 15 16 21 26 29 30 31 33 35 37 45 45
49
50 52 57 60
xii
CONTENTS
3.4 Useful Probability Densities . . . . . . . . . . . . . . . . . . . 66 3.4.1 The Uniform Density . . . . . . . . . . . . . . . . . . . 66 3.4.2 The Exponential Density . . . . . . . . . . . . . . . . 69 3.4.3 The Normal (Gaussian) Density . . . . . . . . . . . . 71 3.4.4 The Lognormal Density . . . . . . . . . . . . . . . . . 84 3.4.5 The Rayleigh Density . . . . . . . . . . . . . . . . . . 87 3.4.6 Probability Density Functions of a Discrete Random Variable . . . . . . . . . . . . . . . . . . . . . . . . . . 90 3.4.7 MomentGenerating Functions . . . . . . . . . . . . . 99 3.5 Two Random Variables . . . . . . . . . . . . . . . . . . . . . 101 3.5.1 Covariance and Correlation . . . . . . . . . . . . . . . 110 3.6 Concluding Summary . . . . . . . . . . . . . . . . . . . . . . 117 3.7 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
4 Functions of Random Variables
125
4.1 Exact Functions of One Variable . . . . . . . . . . . . . . . . 125 4.2 Functions of Two or More RVs . . . . . . . . . . . . . . . . . 132 4.2.1 General Case . . . . . . . . . . . . . . . . . . . . . . . 149 4.3 Approximate Analysis . . . . . . . . . . . . . . . . . . . . . . 161 4.3.1 Direct Methods . . . . . . . . . . . . . . . . . . . . . . 161 4.3.2 Mean and Variance of a General Function of X to Order σ2X . . . . . . . . . . . . . . . . . . . . . . . . . 165 4.3.3 Mean and Variance of a General Function of n RVs . . 168 4.4 Monte Carlo Methods . . . . . . . . . . . . . . . . . . . . . . 179 4.4.1 Independent Uniform Random Numbers . . . . . . . . 179 4.4.2 Independent Normal Random Numbers . . . . . . . . 183 4.4.3 A Discretization Procedure . . . . . . . . . . . . . . . 185 4.5 Concluding Summary . . . . . . . . . . . . . . . . . . . . . . 188 4.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188 4.7 The Standard Normal Table . . . . . . . . . . . . . . . . . . . 192
5 Random Processes 5.1 5.2 5.3 5.4 5.5 5.6 5.7
Basic Random Process Descriptors . . . . . Ensemble Averaging . . . . . . . . . . . . . Stationarity . . . . . . . . . . . . . . . . . . Derivatives of Stationary Processes . . . . . Fourier Series and Fourier Transforms . . . Harmonic Processes . . . . . . . . . . . . . Power Spectra . . . . . . . . . . . . . . . . . 5.7.1 Narrow and BroadBand Processes 5.7.2 White Noise Processes . . . . . . . .
© 2005 by Taylor & Francis Group, LLC
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
195
. 195 . 196 . 201 . 208 . 210 . 229 . 230 . 254 . 257
xiii
CONTENTS
5.7.3 Spectral Densities of Derivatives of Stationary Random Processes . . . . . . . . . . . . . . . . . . . . . . 259 5.8 Fourier Representation of a Random Process . . . . . . . . . 261 5.8.1 Borgman’s Method of Frequency Discretization . . . . 267 5.9 Concluding Summary . . . . . . . . . . . . . . . . . . . . . . 271 5.10 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
6 SingleDegreeofFreedom Dynamics
6.1 Motivating Examples . . . . . . . . . . . . . . . . . . . 6.1.1 Transport of a Satellite . . . . . . . . . . . . . 6.1.2 Rocket Ship . . . . . . . . . . . . . . . . . . . . 6.2 Deterministic SDoF Vibration . . . . . . . . . . . . . . 6.2.1 Free Vibration With No Damping . . . . . . . 6.2.2 Harmonic Forced Vibration With No Damping 6.2.3 Free Vibration With Damping . . . . . . . . . 6.2.4 Forced Vibration With Damping . . . . . . . . 6.2.5 Impulse Excitation . . . . . . . . . . . . . . . . 6.2.6 Arbitrary Loading: Convolution . . . . . . . . 6.2.7 Frequency Response Function . . . . . . . . . . 6.3 SDoF: The Response to Random Loads . . . . . . . . 6.3.1 Formulation . . . . . . . . . . . . . . . . . . . . 6.3.2 Derivation of Equations . . . . . . . . . . . . . 6.3.3 Response Correlations . . . . . . . . . . . . . . 6.3.4 Response Spectral Density . . . . . . . . . . . . 6.4 Response to Two Random Loads . . . . . . . . . . . . 6.5 Concluding Summary . . . . . . . . . . . . . . . . . . 6.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . .
7 MultidegreeofFreedom Vibration
7.1 Deterministic Vibration . . . . . . . . . . . . . . 7.1.1 Solution by Frequency Response Function 7.1.2 Modal Analysis . . . . . . . . . . . . . . . 7.1.3 Advantages of Modal Analysis . . . . . . 7.2 Response to Random Loads . . . . . . . . . . . . 7.2.1 Response due to a Single Random Force . 7.2.2 Response to Multiple Random Forces . . 7.3 Periodic Structures . . . . . . . . . . . . . . . . . 7.3.1 Perfect Lattice Models . . . . . . . . . . . 7.3.2 Effects of Imperfection . . . . . . . . . . . 7.4 Inverse Vibration . . . . . . . . . . . . . . . . . . 7.4.1 Deterministic Inverse Vibration Problem . 7.4.2 Effect of Uncertain Data . . . . . . . . . .
© 2005 by Taylor & Francis Group, LLC
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
275
. . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
. 277 . 277 . 277 . 278 . 288 . 289 . 291 . 292 . 296 . 299 . 303 . 307 . 307 . 308 . 309 . 312 . 325 . 333 . 333
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. 339 . 341 . 343 . 349 . 351 . 353 . 356 . 373 . 375 . 378 . 378 . 381 . 384
339
xiv
CONTENTS
7.5 Random Eigenvalues . . . . . . . . . . . . 7.5.1 A TwoDegreeofFreedom Model . 7.6 Concluding Summary . . . . . . . . . . . 7.7 Problems . . . . . . . . . . . . . . . . . .
. . . .
. 389 . 393 . 395 . 395
8.1 Deterministic Continuous Systems . . . . . . . . . . . . . . 8.1.1 Strings . . . . . . . . . . . . . . . . . . . . . . . . . . 8.1.2 Axial Vibration of Beams . . . . . . . . . . . . . . . 8.1.3 Transversely Vibrating Beams . . . . . . . . . . . . . 8.2 SturmLiouville Eigenvalue Problem . . . . . . . . . . . . . 8.2.1 Orthogonality . . . . . . . . . . . . . . . . . . . . . . 8.2.2 Natural Frequencies and Mode Shapes . . . . . . . . 8.3 Deterministic Vibration . . . . . . . . . . . . . . . . . . . . 8.3.1 Free Response . . . . . . . . . . . . . . . . . . . . . 8.3.2 Forced Response via Eigenfunction Expansion . . . . 8.4 Random Vibration of Continuous Systems . . . . . . . . . . 8.4.1 Derivation of Response Spectral Density . . . . . . . 8.5 Beams with Complex Loading . . . . . . . . . . . . . . . . . 8.5.1 Transverse Vibration of Beam with Axial Force . . . 8.5.2 Transverse Vibration of Beam on Elastic Foundation 8.5.3 Response of a Beam to a Traveling Force . . . . . . 8.6 Concluding Summary . . . . . . . . . . . . . . . . . . . . . 8.7 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. 404 . 404 . 406 . 408 . 411 . 413 . 414 . 420 . 420 . 422 . 428 . 429 . 439 . 439 . 442 . 446 . 451 . 452
8 Continuous System Vibration
9 Reliability
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
403
455
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455 9.2 First Excursion Failure . . . . . . . . . . . . . . . . . . . . . 457 9.2.1 Exponential Failure Law . . . . . . . . . . . . . . . . . 462 9.2.2 Modified Exponential Failure Law . . . . . . . . . . . 465 9.2.3 Calculation of UpCrossing Rate . . . . . . . . . . . . 466 9.2.4 NarrowBand Process — Envelope Function . . . . . . 472 9.2.5 Rice’s Envelope Function for Gaussian NarrowBand Process X (t) . . . . . . . . . . . . . . . . . . . . . . . 474 9.2.6 Other Failure Laws . . . . . . . . . . . . . . . . . . . . 488 9.3 Fatigue Life Prediction . . . . . . . . . . . . . . . . . . . . . . 493 9.3.1 Failure Curves . . . . . . . . . . . . . . . . . . . . . . 495 9.3.2 Peak Distribution for Stationary Random Process . . 497 9.3.3 Peak Distribution of a Gaussian Process . . . . . . . . 501 9.4 Concluding Summary . . . . . . . . . . . . . . . . . . . . . . 512 9.5 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 512
© 2005 by Taylor & Francis Group, LLC
xv
CONTENTS
10 Nonlinear Dynamic Models
10.1 Examples of Nonlinear Vibration . . . . . . . . . . . . 10.2 Fundamental Nonlinear Equations . . . . . . . . . . . 10.3 Statistical Equivalent Linearization . . . . . . . . . . . 10.3.1 Equivalent Nonlinearization . . . . . . . . . . . 10.4 Perturbation Methods . . . . . . . . . . . . . . . . . . 10.4.1 LindstedtPoincaré Method . . . . . . . . . . . 10.4.2 Forced Oscillations of Quasiharmonic Systems . 10.4.3 Jump Phenomenon . . . . . . . . . . . . . . . . 10.4.4 Periodic Solutions of Nonautonomous Systems 10.4.5 Random Duffing Oscillator . . . . . . . . . . . 10.5 The van der Pol Equation . . . . . . . . . . . . . . . . 10.5.1 Limit Cycles . . . . . . . . . . . . . . . . . . . 10.5.2 The Forced van der Pol Equation . . . . . . . . 10.6 Markov ProcessBased Models . . . . . . . . . . . . . . 10.6.1 Probability Background . . . . . . . . . . . . . 10.6.2 The FokkerPlanck Equation . . . . . . . . . . 10.7 Concluding Summary . . . . . . . . . . . . . . . . . . 10.8 Problems . . . . . . . . . . . . . . . . . . . . . . . . .
11 Nonstationary Models
11.1 Some Applications . . . . . . . . . . . . . . . . . 11.2 Envelope Function Model . . . . . . . . . . . . . 11.2.1 Transient Response . . . . . . . . . . . . . 11.2.2 MS Nonstationary Response . . . . . . . . 11.3 Nonstationary Generalizations . . . . . . . . . . . 11.3.1 Discrete Model . . . . . . . . . . . . . . . 11.3.2 ComplexValued Stochastic Processes . . 11.3.3 Continuous Model . . . . . . . . . . . . . 11.4 Priestley’s Model . . . . . . . . . . . . . . . . . . 11.4.1 The Stieltjes Integral: An Aside . . . . . . 11.4.2 Priestley’s Model . . . . . . . . . . . . . . 11.5 SDoF Oscillator Response . . . . . . . . . . . . . 11.5.1 Stationary Case . . . . . . . . . . . . . . . 11.5.2 Nonstationary Case . . . . . . . . . . . . 11.5.3 Undamped Oscillator . . . . . . . . . . . 11.5.4 Underdamped Oscillator . . . . . . . . . . 11.6 Multi DoF Oscillator Response . . . . . . . . . . 11.6.1 Input Characterization . . . . . . . . . . . 11.6.2 Response Characterization . . . . . . . . 11.7 Nonstationary and Nonlinear Oscillator . . . . . 11.7.1 The Nonstationary and Nonlinear Duffing
© 2005 by Taylor & Francis Group, LLC
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
515
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. 516 . 519 . 521 . 530 . 532 . 534 . 539 . 543 . 544 . 552 . 555 . 556 . 557 . 561 . 561 . 565 . 587 . 587
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
. 592 . 598 . 599 . 605 . 607 . 608 . 610 . 610 . 612 . 612 . 614 . 615 . 615 . 616 . 618 . 619 . 622 . 622 . 624 . 625 . 627
591
xvi
CONTENTS
11.8 Concluding Summary . . . . . . . . . . . . . . . . . . . . . . 629 11.9 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 629
12 The Monte Carlo Method
631
13 FluidInduced Vibration
679
Index
727
12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 631 12.2 RandomNumber Generation . . . . . . . . . . . . . . . . . . 635 12.2.1 Standard Uniform Random Numbers . . . . . . . . . . 635 12.2.2 Generation of Nonuniform Random Variates . . . . . . 637 12.2.3 Composition Method . . . . . . . . . . . . . . . . . . . 648 12.2.4 Von Neumann’s RejectionAcceptance Method . . . . 651 12.3 Joint Random Numbers . . . . . . . . . . . . . . . . . . . . . 659 12.3.1 Inverse Transform Method . . . . . . . . . . . . . . . . 660 12.3.2 Linear Transform Method . . . . . . . . . . . . . . . . 661 12.4 Error Estimates . . . . . . . . . . . . . . . . . . . . . . . . . 664 12.5 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 670 12.5.1 Evaluation of FiniteDimensional Integrals . . . . . . . 670 12.5.2 Generating a Time History for a Stationary Random Process Defined by a Power Spectral Density . . . . . 673 12.6 Concluding Summary . . . . . . . . . . . . . . . . . . . . . . 676 12.7 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 676 13.1 Ocean Currents and Waves . . . . . . . . . . . . . . . . . . . 679 13.1.1 Spectral Density . . . . . . . . . . . . . . . . . . . . . 680 13.1.2 Ocean Wave Spectral Densities . . . . . . . . . . . . . 684 13.1.3 Approximation of Spectral Density from Time Series . 689 13.1.4 Generation of Time Series from a Spectral Density . . 691 13.1.5 ShortTerm Statistics . . . . . . . . . . . . . . . . . . 692 13.1.6 LongTerm Statistics . . . . . . . . . . . . . . . . . . . 698 13.1.7 Wave Velocities via Linear Wave Theory . . . . . . . . 700 13.2 Fluid Forces in General . . . . . . . . . . . . . . . . . . . . . 702 13.2.1 Wave Force Regime . . . . . . . . . . . . . . . . . . . 703 13.2.2 Wave Forces on Small Structures — Morison Equation 705 13.2.3 VortexInduced Vibration . . . . . . . . . . . . . . . . 711 13.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 713 13.3.1 Static Configuration of a Towing Cable . . . . . . . . 713 13.3.2 Fluid Forces on an Articulated Tower . . . . . . . . . 717 13.3.3 Weibull and Gumbel Wave Height Distributions . . . 720 13.3.4 Reconstructing Time Series for a Given Significant Wave Height . . . . . . . . . . . . . . . . . . . . . . . 722 13.4 Available Numerical Codes . . . . . . . . . . . . . . . . . . . 723
© 2005 by Taylor & Francis Group, LLC
Chapter 1
Introduction Probability and random processes are the quantifications of uncertainties. There is a steep learning curve to be covered before it becomes possible to consider any significant problem that includes a probabilistic component. We must learn a new way of thinking with uncertainty as our paradigm.1 The probabilistic paradigm is not the most comfortable for engineers since we are raised to believe that, given enough experimental data and theoretical development, any problem is solvable exactly, or at least to within measurement tolerances. Reality tells us that certainty exists only in idealized models, not in the actual physical systems that must be understood and designed. While an exact quantity exists only in our imagination, sometimes uncertainties can be ignored for particular applications. In our studies here, we begin to take account of uncertainty and learn some basic concepts in probability. First, let us consider some motivating examples.
1.1
Applications
To demonstrate the importance of uncertainty modeling in mechanical systems, a number of examples are chosen for brief and qualitative discussion. 1 A paradigm, pronounced “para—dime,” is a way of thinking. It may be viewed as the beliefs, values, and techniques shared by a particular group of people. Therefore, a new paradigm in a technical area implies a completely new way of thinking about that area. A recent example of a paradigm shift is the development of the field of chaos in nonlinear dynamics.
© 2005 by Taylor & Francis Group, LLC
1
CHAPTER 1. INTRODUCTION
2 1.1.1
Random Vibration
The discipline of random vibration of structures was borne of the need to understand how structures respond to dynamic loads that are too complex to model deterministically. Examples include aerodynamic loading on aircraft and earthquake loading of structures. Essentially, the question that must be answered is: Given the statistics (read: uncertainties) of the load
ing, what are the statistics (read: most likely values with bounds) of the response ? Generally, for engineering applications the statistics of greatest
concern are the mean, or average value, and the variance, or scatter. These concepts are discussed in detail subsequently. There are a number of texts on random vibration2 where the subject is explored fully. A significant portion of this book also deals with this important subject. But it is important to note that the methodologies introduced and developed in this text are applicable to a wide variety of problems in engineering and the physical sciences. Random vibration is a wonderful way to study linear systems, that is linear equations of all sorts. Suppose that we are aircraft designers currently working on the analysis and design of a wing for a new airplane. As engineers, we are very familiar with the mechanics of solids, and can size the wing for static loads. Also, we have vibration experience and can evaluate the response of the wing to a harmonic or impulsive forcing. But this wing needs to provide lift to an airplane flying through a turbulent atmosphere. The fluid dynamicists in the design group know that turbulence is a very complicated physical process. In fact, the fluid (air) motion is so complicated that probabilistic models are required to model the behavior. Here, a plausibly deterministic but very complicated dynamic process is taken to be random for purposes of modeling. Wing design requires force data resulting from the interaction
between fluid and structure. Such data can be shown as the time history in Figure 1.1. The challenge is to make sense of such intricate fluctuations. The analyst and designer must run scale model tests. A wing section is set up in the wind tunnel and representative aerodynamic forces are generated. Data on wind forces and structural response are gathered and analyzed. With additional data analysis, it is possible to estimate the force magnitudes. Estimates of mean values of these forces can be calculated, as well as of the range of 2 A very useful one that includes a broad spectrum of theory and application is by I. Elishakoff, Probabilistic Methods in the Theory of Structures, WileyInterscience, 1983, now in a Dover edition. Two exceptionally clear early books on random vibration are worth reading. The first is Random Vibration in Mechanical Systems by S.H. Crandall and W.D. Mark, Academic Press, 1963. The other is An Introduction to Random Vibration by J.D. Robson, Elsevier, 1964. Another introduction to the subject is An Introduction to Random Vibrations and Spectral Analysis by D.E. Newland, Longman, 1975, now in its third edition.
© 2005 by Taylor & Francis Group, LLC
1.1. APPLICATIONS
3
Figure 1.1: Time history of turbulent force. possible forces. With these estimates, it is possible to study the complex physical problem of the behavior of the wing under a variety of realistic loading scenarios using probabilistic and statistical methods. Probability provides the vehicle for quantifying uncertainties. This text introduces the use of probabilistic information in mechanical systems analysis and design, primarily structural and dynamic systems. However, we emphasize that these tools are applicable to all the sciences and engineering, even though the focus in this text is the mechanical sciences and engineering. 1.1.2
Fatigue Life
The fatigue life of mechanical components and structures3 depends on many factors such as material properties, temperature, corrosion environment, and also on vibration history. A first step in estimating fatigue life involves the characterization of the static and dynamic loading cycles the structure has experienced. Are there many cycles, what are the amplitude ranges, and is the loading harmonic or of broad frequency band? The estimates of fatigue life estimates are extremely important for the proper operation of a modern industrial society. Such estimates are intimately linked to the reliability of machines and structures. They determine the frequency with which components need to be replaced, the economics of the operation of the machine, and hence the insurance costs. 3
A very useful book with which to begin the study of fatigue is by V.V. Bolotin, ASME Press, 1989.
Prediction of Service Life for Machines and Structures ,
© 2005 by Taylor & Francis Group, LLC
CHAPTER 1. INTRODUCTION
4
Anyone studying fatigue life data will be immediately struck by the significant scatter. Components normally considered to be identical can have a wide range of lives. As engineers, we are concerned about having a rigorous basis for estimating the fatigue lives of nominally identical manufactured components. Eventually, it is necessary to relate the life estimate of the structure to that of its components. This is generally a difficult task, one that requires the ability to evaluate structural and machine response to random forces. Chapter 9 is devoted to reliability.
Example 1.1 Miner’s Rule for Fatigue Damage One of the most important early works on the estimation of fatigue life is by Miner,4 who was a strength test engineer with the Douglas Aircraft Company. Miner’s rule is a deterministic way to deal with the uncertainties of structural damage and fatigue. The phenomenon of cumulative damage under repeated loads is assumed to be related to the total work absorbed by a test specimen. The number of loading cycles applied, expressed as a fraction of the number of cycles to failure at a given stress level, is taken to be proportional to the useful structural life expended. When the total damage reaches 1, the fatigue test specimen is assumed to fail. Miner presented validation of his theory with data from experiments on Aluminum sheets. At a certain stress level for a specific material and geometry,5 this rule estimates the number of cycles to failure. Mathematically, this can be written as n = 1, (1.1) N
where n equals the number of cycles undergone by the structure at a specific stress level, and N is the experimentally known number of cycles to failure at that stress level. Since most structures undergo a mixture of loading cycles at different stress levels, Equation 1.1 must be written for each stress level i as follows,
n = 1, i
i
(1.2)
Ni
where each fraction represents the percentage of life used up at each stress level. Suppose the structure is loaded at two stress levels, corresponding cycles to failure
N1
Equation 1.2, the following relation,
= 100
and
n2 + 100 50 = 1,
N2
= 50.
i
= 1, 2, with
According to
n1
M.A. Miner, “Cumulative Damage in Fatigue,” pp. A159A164, Journal of Applied September 1945. 5 Corners and discontinuities result in high stress concentrations and lower fatigue life. 4
Mechanics,
© 2005 by Taylor & Francis Group, LLC
1.1. APPLICATIONS
5
holds between the number of possible cycles n1 and n2 for each stress level. There are numerous combinations that lead to failure. For exam
(
) = (50 25) (
) = (100 0) (
) = (0 50)
ple: n1 , n2 , , n1 , n 2 , , n1 , n 2 , , with others not hard to find. Miner realized that these summations were only approxi
mations. His experiments showed that sometimes a component failed before
1
the sum totaled , and, at other times, did not fail until the sum was greater than
1.
Furthermore, failure by this rule is independent of the ordering of
the stress cycles. This means that fatigue life is the same whether high stress cycles precede or follow lower stress cycles. We know, however, that stress history affects fatigue life. In the last 50 years since Miner’s paper, despite the vast amount of work that has been done to build on his and other work to better understand fatigue, Miner’s rule and its variants remain widely utilized practical methods.
⊛
1.1.3
OceanWave Forces
There are similarities between approaches used to model oceanwave forces on structures and those for wind forces. The differences are primarily due to the added mass6 of the water, and the different structural types de
signed for use in the ocean. The calculation of waveinduced forces is a very important aspect of ocean engineering.7 As might be expected, many
engineering disciplines are utilized in ocean engineering. The estimation of wave forces on offshore oildrilling platforms, ships, and other ocean and hydraulic structures, such as water channel spillways and dams, is the basis for the design of such structures. Without these estimates, there is no way to design or analyze the structure. The estimation of loads is always first on the list of tasks for an engineer. Figure 1.2 is a schematic showing a structural cylinder in the ocean that is subjected to random and harmonic waves, superimposed on the still water line. 6 The force required to accelerate a rigid body with mass m in a fluid medium that is normally at rest is
m + added mass of the surrounding fluid) × y¨.
(
The additional force is required to accelerate the surrounding fluid to y¨. If the fluid medium is air, the force required to accelerate the surrounding air is often neglected. 7 Some useful books, among hundreds of available volumes, are the following: J.F. Wilson, Dynamics of Offshore Structures, John Wiley & Sons, 1984. This book is out of print, but there is a second edition. O.M. Faltinsen, Sea Loads on Ships and Offshore Structures, Cambridge University Press, 1990. S. Gran, A Course in Ocean Engineering, Elsevier Science, 1992. This is a very thorough book. © 2005 by Taylor & Francis Group, LLC
CHAPTER 1. INTRODUCTION
6
Figure 1.2: Schematic of structural cylinder in the ocean subjected to random and harmonic waves.
Example 1.2 Wave Forces on an OilDrilling Platform The need to drill for oil in the oceans has driven our ability to design ocean structures for sites of everincreasing depths. Today’s fixedbottom ocean structures, when taken with their foundations, are taller than our tallest skyscrapers. As might be expected, the dynamic response of these towers to ocean waves and currents is significant and must be understood and analyzed. Consider the oceanwave force on an ocean platform such as the one shown in Figure 1.3.
The single most important paper8 on the force exerted by ocean waves
on fixed structures, even though it was written half a century ago, derived what became to be universally known as the Morison equation.
In this
paper, after much experimental work, Morison and his colleagues came to the conclusion that the force exerted by unbroken surface waves on a circular cylindrical column that extends from the bottom of the ocean upward above the wave crest is made up of two components:
•
a drag force proportional to the square of the water particle velocity, with proportionality represented by a drag coefficient having substantially the same value as for steady flow, and
8 J.R. Morison, M.P. O’Brien, J.W. Johnson, and S.A. Schaff, “The Force Exerted by Surface Waves on Piles,” pp. 149154, Petroleum Transactions, Vol. 189, 1950.
© 2005 by Taylor & Francis Group, LLC
1.1. APPLICATIONS
7
Figure 1.3: Offshore platform.
•
an inertia force proportional to the horizontal component of the inertia force exerted on the mass of water displaced by the column, with proportionality represented by an inertia coefficient.
The drag force on an element of length
dFD where
CD
of water,
dx
is given by
= CD ρD u2u dx,
(1.3)
is the experimentally determined drag coefficient,
D
is the diameter of the cylinder, and
horizontal water particle velocity. The term
uu
of the force is in the direction of the flow, where
u
ρ is the density
is the instantaneous
ensures that the direction
u
is the absolute value of
the particle velocity. The inertia force on an element of length by
dFI
2
= CI ρ πD4 u˙ dx,
dx
is given
(1.4)
u˙ ≡ ∂u/∂t is the instantaneous horizontal water particle acceleration CI is the inertia coefficient. The dimensionless drag and inertia co
where and
efficients are functions of flow characteristics, cylinder diameter, and fluid density. Depending on the application, the analyst may assume them to be effectively constant, or may need to account for the scatter in their values.
© 2005 by Taylor & Francis Group, LLC
CHAPTER 1. INTRODUCTION
8
The Morison force equation is the sum of the above drag and inertia components. Classical deterministic fluid mechanics is used to derive wave particle velocities and accelerations. Since many tall ocean structures oscillate appreciably, the relative velocity and acceleration between fluid and structure is commonly used in the Morison equation, where where
x˙
and
x¨
u is replaced by u − x˙ and u˙ is replaced by (u˙ − x¨),
are the structural velocity and acceleration, respectively.
Also, in order to better characterize the complexity of the wave motion, the fluid velocity, acceleration, and resulting force are modeled as random functions of time. We begin to explore the concept of random functions in detail in Chapter 5. More details on the vibration of structures in fluids
can be found in numerous books.9 Chapter 13 provides an introduction to
fluid induced vibration.
⊛
1.1.4
Wind Forces
Engineering structures such as cooling towers, aircraft, skyscrapers, rockets, and bridges are all exposed to wind and aerodynamic loads. Examples of these engineering structures are shown in Figures 1.41.7. Wind is the natural movement of the atmosphere due to temperature and pressure gradients. Aerodynamic loads are the atmospheric forces resulting from the interaction of wind and structure. While we know how to write an equation for a harmonic force, what does an equation for wind force look like? Due to the complexity of the fluid mechanics of wind, it is generally necessary to approximate the force due to wind. There are various levels of approximate relations, depending on the application. In all instances, the force relation includes at least one experimentally determined parameter or coefficient.
Such semiempirical
force equations are very valuable in wind engineering practice. These look very much like the Morison equation discussed in the last section. An excellent monograph is by Simiu and Scanlan.10
1.1.5
Material Properties
While the modeling of randomness in material properties is beyond our scope in this book, it is worthwhile to briefly mention this type of modeling 9 These two books are worth having a look at: R.D. Blevins, FlowInduced Vibration , van Nostrand Reinhold, 1977. (There is a second edition.) A.T. Ippen, Editor, Estuary and Coastline Hydrodynamics , McGrawHill, 1966. 10 Wind Effects on Structures: Fundamentals and Applications to Design , E. Simiu, R.H. Scanlan, Third Edition, 1996, WileyInterscience.
© 2005 by Taylor & Francis Group, LLC
1.1. APPLICATIONS
9
Figure 1.4: Cooling tower.
Figure 1.5: Vortex study in aircraft (courtesy of NASA, Dryer Flight Research Center).
© 2005 by Taylor & Francis Group, LLC
CHAPTER 1. INTRODUCTION
10
Figure 1.6: Saturn V Apollo 11.
Figure 1.7: Charleston Grace Memorial Bridge.
© 2005 by Taylor & Francis Group, LLC
1.1. APPLICATIONS
11
because of the importance of many new materials that have effective properties, that is, properties that are an average over a crosssection. These include various composites and tailored materials, modern materials designed for particular structural applications, especially where high strength and durability are needed concomitantly with light weight. Such designs require a complicated mix of fibers and substrates configured to obtain particular properties. Defining stressstrain relations and Young’s modulus for such components is not straightforward.
It is sometimes necessary that
properties be averaged or effective properties be defined. The soil is an example of a naturally occurring material that is extremely complex and cannot be modeled in a traditional manner. It is common that two nearby volumes of soil have very different mechanical properties. Therefore, in structural dynamics applications, such as earthquake engineering, the loading is effectively random because, in part, by the time it reaches the structure, the force has traversed a complex topology of the Earth.11
Data on the variability of material properties are tabulated in numerous
references.12 From Haugen, for example, hot rolled 1035 steel round bars
of diameters in the range 19 in have yield strengths of between 40,00060,000 psi, with an average yield of just under 50,000 psi. In addition, the variability can change appreciably depending on temperature. A titaniumaluminumlead alloy has an ultimate shear strength of between 88,000114,000 psi at 60,000 psi.
90o F but at 1000o F the strength drops to between 42,000
The obvious conclusion is that variability can be significant and is a function of different factors. Analysis and design require knowledge of the environment within which the structure will operate. While temperature and thermal effects are topics for specialized texts, these can be critical factors in many advanced aerospace and machine designs, and therefore the reader needs to be aware of the importance of these aspects.
1.1.6
Statistics and Probability
The previous examples of natural forces all have one factor in common. It is that they depend on experimentally determined parameters. Just as linearity of vibration depends on small oscillations, these semiempirical equations are valid only for a particular range in the data. While deterministic models also depend greatly on experimental data for their formulation and 11 The area of research known as earthquake engineering and the specific study of how energy propagates through complex materials such as soils is known as the study of waves in random media . 12 One can begin with the text by E.B. Haugen, Probabilistic Mechanical Design , WileyInterscience, 1980. There are interesting applications of probability to mechanical engineering, primarily based on the Gaussian distribution.
© 2005 by Taylor & Francis Group, LLC
CHAPTER 1. INTRODUCTION
12
ultimately their validity, random models are an attempt to explicitly deal with observed scatter in the data and with very intricate dynamic behavior. Random models also allow an estimate of how data scatter affects response scatter. Data are always our link to valid probabilistic models, their derivation and validation. While this is not our focus in this text, it is important that the reader is at least aware that this step precedes any valid probabilistic models.
Example 1.3 From Data to Model and Back to Data Modeling can be as much an art as a science. Engineers are generally handed a problem that needs to be solved, not an equation, not even a wellthoughtout description of the problem. For example: We need to go to the Moon in ten years ! Engineering is predicated on understanding how structures, machines, and materials behave under various operating conditions. This understanding is based on theory and data. Many experiments have been performed to get us to our current level of understanding and intuition. The experiments suggest cause and effect between variables. They provide us with parameter values. Finally, they are the basis for the equations we derive. We know that data has scatter, and the significance of the scatter to a particular problem determines whether it can be ignored. If it cannot be ignored, then the data is used to estimate the statistics of the randomness. The resulting probabilistic model is used to study the particular problem at hand, and the model’s validity is established by comparing its predictions with available data. Such comparisons help define the limits of model validity. In this way, a full circle is achieved. Data gives birth to understanding and parameter values, which lead to governing equations and their predictions, and finally validity is established by comparing model predictions with new data that is not part of the original set.
⊛
1.2
Units
All physical parameters have units that tie them to a particular system. There are primarily two systems of units, the English System and the SI System, where SI stands for Système International. The SI units are considered modern and more appropriate. In this book, both systems are used since both are used in practice in the United States. In Table 1.1, the English and SI system of units are shown for certain key physical parameters © 2005 by Taylor & Francis Group, LLC
1.3. ORGANIZATION OF THE TEXT Table 1.1: Parameter Force Mass
13
and SI Units for Key Physical Parameters English SI 1 lb 4.448 N (kg m/s2 ) 2 1 slug (lb s /ft) 14.59 kg (kilogram) 1 lbm 0.455 kg 1 ft 0.3048 m (meter) Length 1 ft/s2 0.3048 m/s2 Acceleration 1 lb/in 175.12 N/m Spring constant 0.1130 N m/rad Torsional spring constant 1 lb in/rad 1 lb s/in 175.12 N s/m Damping constant 0.1130 kg m2 Mass moment of inertia 1 lb in s2 1 deg 0.0175 rad Angle 1/6894.757 psi 1 N/m2 = 1 Pascal Pressure English
that we will encounter in this text. The SI system has taken hold in the scientific and engineering communities in Englishspeaking countries over the past several decades, but the English system continues to be popular.
1.3
Organization of the Text
There are many chapters in this book and many topics that, at first sight, may be difficult to organize in the reader’s mind. The major groupings are methods for “static” problems, the first four chapters, where time does not play a role, and “dynamic” problems where time is a critical factor, the remaining chapters. While this text contains many standard items and topics, it also includes sections on important material that is not exactly mainstream, as well as topics that break into more advanced disciplines. The first four chapters introduce basic concepts in probability and variables that are best described as random. The remaining chapters introduce time as a critical parameter in the randomness. An undergraduate course can be based on Chapters 15, and a graduate course based on Chapters 5 and subsequent.
1.4
Problems
1. Identify engineering and scientific applications where uncertainties can be ignored. Explain. 2. Identify engineering and scientific applications where uncertainties cannot be ignored. Explain. © 2005 by Taylor & Francis Group, LLC
CHAPTER 1. INTRODUCTION
14
3. Discuss how an engineer may ascertain whether uncertainties are important or can be ignored in an analysis and design. Use examples in your discussion. 4. If Miner’s rule for fatigue damage has to be extended to cases where the order of loading cycles is important, how can this be accomplished? Explain with or without using equations. 5. If Miner’s rule for fatigue damage has to be extended to cases where a cycle at stress nσ causes n times as much damage as a cycle at stress σ , how can this be accomplished? Explain with or without using equations. 6. Which variables or parameters in Equations 1.3 and 1.4 are better assumed to be random variables? Explain your choices.
© 2005 by Taylor & Francis Group, LLC
Chapter 2
Events and Probability 2.1
Sets
Set theory is a useful concept for understanding probability, and its building block is the event. Events are collections of outcomes on which probabilities have been assigned. Why is probability necessary? Suppose a set of experiments to measure the diameters of a group of rods are performed. Even though the rods are supposed to have the same dimensions, they do not because of manufacturing imperfections and measurement errors. Therefore, if the diameter for a design is to be specified, there is no single value, but many possible values. These possible values are events. As a way to organize this information, a
new language
that can accom
modate all the possibilities is needed. The language of probability is able to account for all the possibilities in a mathematical way so that computation is possible. A
set is a collection of things that share certain characteristics.
Included
are objects, numbers, colors, essentially any of the items and concepts we use to define our environment. A
sample space
is the set of all possibilities.
diameters is such a sample space.
take. Each of these possibilities is called a points occur, then a
realization
The group of possible
Or it can be the values a variable may
sample point.
If one of the sample
of that sample point has occurred. In this
case, the sample points are discretevalued, but in other applications they may be continuousvalued. A subset of the sample space is an
event
within
the sample space. We will examine the connection between events and probability, but first some examples of events are considered. In engineering reliability, the
15
© 2005 by Taylor & Francis Group, LLC
CHAPTER 2. EVENTS AND PROBABILITY
16
goal is to estimate the probability that the system will perform as designed. This means that it performs at or better than design specifications for a certain design life with a certain reliability level. The technical term for this characteristic of a structure or a component is its
safety.
The probability
that the system does not perform as designed is given by its
failure.
probability of
Generally, it is very difficult to estimate such a quantity due to
the complexity of engineered systems and the uncertainties associated with operating environments.
Example 2.1 Probabilities of Interest We have just mentioned safety and failure as two very important probabilities that define the robustness of a particular design. Other probabilities
(i) the live load1 is less than 10 N/m2 , (ii) the temper◦ ature is greater than 50 C, (iii) the frequency of vibration is in the range ω 1 < ω < ω 2 , and (iv) the flow rate is less than 3 ft3 /s. The safety of a of interest might be:
complex machine is a function of the reliability of all its components and how they are connected to each other. First, these individual safeties must be determined.
ω 1 < ω < ω 2 is a continuous sample space. A particular ω 1 < 33.01 < ω2 , is a sample point whose occurrence is a realization.
The range value,
⊛
2.1.1
Basic Events
Several events are needed for a complete modeling framework. They are:
1. The
impossible event, denoted by {} or φ, is the event with no sample null event is an empty
point. In the sample space, the impossible or set.
2. The
certain
or
universal event,
denoted by
S
or
Ω,
is the event that
contains all the sample points. The certain event is in fact the sample space itself.
3. The
complementary event
is denoted by the event symbol with an
E, the complementary event is E, meaning that it contains all sample points in S that are not in E. Therefore, E is not E.
overbar. For event
1 A live load is one that is due to transient effects, for example, a person walking across a floor. The counterpart is  you guessed it  the dead load. For example, the weight of the floor, which does not change.
© 2005 by Taylor & Francis Group, LLC
2.1. SETS
17
A useful visual representation of sets, spaces, and their manipulations is provided by the
Venn diagram.
represented by a rectangle. within the rectangle.
The sample space or the universe is
An event is represented by a closed region
The complement of this event will be the remaining
region. See Figure 2.1.
Figure 2.1: Venn diagram with event
E.
The concepts of impossible, universal, and complementary events are demonstrated by an example.
Example 2.2 Impossible, Universal, and Complementary Events Consider an experiment where a coin is tossed three times consecutively and the total number of heads counted.
The certain or universal event is
the set of all possible outcomes, the total number of heads counted,
S = {0, 1, 2, 3} . If
A is the event that the total
number of heads is greater than 3, then
A
is an impossible event, written as
A = {} Keep in mind that
{0}
or
A = φ.
is not an impossible event, but the event that no
heads appeared. Let
B
be the event that the total number of heads is 2. That is,
B = {2} . The complementary event of
B
is then
B¯ = {0, 1, 3} . © 2005 by Taylor & Francis Group, LLC
CHAPTER 2. EVENTS AND PROBABILITY
18
It should be noted that the impossible event is a complementary event of the universal event and vice versa,
S¯ = φ, φ¯ = S.
⊛ In many instances, it is necessary to consider more than one event and how the different events are coupled or affect each other.
For example, a
highstrength alloy has a range of yield stresses that depend on temperature. The two events are yield stresses and temperature.
It is important to be
able to make simultaneous or joint statements on both events. To do this, we need to operate with sets using the following rules on the combination of events. These rules can be easily seen utilizing the Venn diagrams in Figure 2.2.
union of two events E1 and E2 , denoted as E1 ∪ E2 , is defined as or E2 , or both. This means that either or both events may occur.
1. The
E1
2. The
intersection of events E1 and E2 , denoted as E1 ∩ E2 or E1 and E2 . This means that both events occur.
E1 E2 , is
defined as
Figure 2.2: The union and intersection of
E1
and
E2 .
The concept of the event can be used to signify certain regions of interest
S. Note the following: Ei ∪ S = S and Ei ∩ φ = φ. E1 , E2 are mutually exclusive or disjoint then E1 ∩ E2 =
within the sample space Also, if two events
φ. All these operations can be easily visualized using the Venn diagram.
Example 2.3 Union and Intersection I Conduct an experiment in which two dice are tossed at the same time. The experiment is repeated many times, and the numbers that show are then added.
S
is the universal event,
S = {2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12} . © 2005 by Taylor & Francis Group, LLC
2.1. SETS A
Let
19
be the event that the sum is greater than 7, and
B
the event that
the sum is less than 11. Then,
A = {8, 9, 10, 11, 12}
and
B = {2, 3, 4, 5, 6, 7, 8, 9, 10} .
Obtain the unions and intersections of
¯ A¯ and B.
Solution
¯ A and B, A¯ and B, A and B,
and
Using the definitions of union and intersection,
A ∪ B = {2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12} A¯ ∪ B = {2, 3, 4, 5, 6, 7, 8, 9, 10} A ∪ B¯ = {8, 9, 10, 11, 12} A¯ ∪ B¯ = {2, 3, 4, 5, 6, 7, 11, 12} , and
A ∩ B = {8, 9, 10} A¯ ∩ B = {2, 3, 4, 5, 6, 7} A ∩ B¯ = {11, 12} A¯ ∩ B¯ = φ.
⊛ Example 2.4 Union and Intersection II Let
E1 , E2 , and E3 be three events.
set notation,
(i)
Express the following statements in
at least one event occurs,
(ii)
all three events occur,
(iii)
exactly one of the three events occurs.
Solution
(i) At least one event occurring is same as either E1 , E2 , or E3 occurring. That is,
(ii)
That is,
(iii)
E1 ∪ E2 ∪ E3 .
All three events occurring is same as all
E1 ∩ E2 ∩ E3 .
E1 , E2 , and E3
occurring.
There are three ways that exactly one of three events occurs. The
first is when
E1
occurs, and
E2
and
E3
do not.
occurs, and the others do not. The third is when do not. Call these events
The second is when E2 E3 occurs, and the others
A1 , A2 , and A3 , respectively, A1 = E1 ∩ E2 ∩ E3 A2 = E1 ∩ E2 ∩ E3 A3 = E1 ∩ E2 ∩ E3 .
© 2005 by Taylor & Francis Group, LLC
CHAPTER 2. EVENTS AND PROBABILITY
20
The event that exactly one of three mutually exclusive events occurs is the union of
A1 , A2 , and A3 , that is, A1 ∪ A2 ∪ A3 , or
E1 ∩ E2 ∩ E3 ∪ E1 ∩ E2 ∩ E3 ∪ E1 ∩ E2 ∩ E3 .
⊛
Figure 2.3: John Venn (18341923).
Venn John Venn came from a Low Church Evangelical background and when he entered Gonville and Caius College Cambridge in 1853 he had so slight an acquaintance with books of any kind that he may be said to have begun there his knowledge of literature. He graduated in 1857, was elected a Fellow in that year and two years later was ordained a priest. For a year he was curate at Mortlake. In 1862 he returned to Cambridge University as a lecturer in Moral Science, studying and teaching logic and probability theory. He developed Boole’s mathematical logic and is best known for his diagrammatic way of representing sets and their unions and intersections. Venn wrote Logic of Chance in 1866, which Keynes described as strikingly original and considerably influenced the development of the theory of statistics. Venn published Symbolic Logic in 1881 and The Principles of Empirical Logic in 1889. The second of these is rather less original but the first was described by Keynes as probably his most enduring work on logic. © 2005 by Taylor & Francis Group, LLC
2.1. SETS
21
In 1883 Venn was elected a Fellow of the Royal Society. About this time his career changed direction. He had already left the Church in 1870 but his interest now turned to history. He wrote a history of his college, publishing The Biographical History of Gonville and Caius College 13491897 in 1897. He then undertook the immense task of compiling a history of Cambridge University, the first volume of which was published in 1922. He was assisted by his son in this task, which was described by another historian in these terms: It is difficult for anyone who has not seen the work in its making to realize the immense amount of research involved in this great undertaking.
Venn had other skills and interests, too, including a rare skill in building machines. He used his skill to build a machine for bowling cricket balls, which was so good that when the Australian Cricket team visited Cambridge in 1909, Venn’s machine clean bowled one of its top stars four times. 2.1.2
Operational Rules
As with the mathematical operations of addition and multiplication, set operations have similar rules. last section.
We have started operating with sets in the
By doing so, it is possible to derive additional information.
For example, sets
A
and
B
are known.
information about a new set is found,
By performing the union
C = A ∪ B.
A ∪ B,
In engineering a similar expansion of knowledge is possible. The design life and reliability of an aircraft wing depends on our knowledge of the states of stress in the wing. We need to know how the stress event varies with time and location on the wing. Stress, however, depends on many other events. Two such events are the wind velocity and the surrounding temperature distribution. While it is possible to estimate the velocity and temperature, these must be related to the stress state in the structure. This is what is meant by operating with events. Such operations allow us to make statements on other related and more important events. Some of the more important properties of unions and intersections follow.
It is
worthwhile to use the Venn diagram symbolism to verify the properties developed next. The
commutative
property is a statement of whether the
union or intersection changes the outcome. two events
E1
and
E2 ,
ordering
of a
It does not, as shown for the
E1 ∪ E2 = E2 ∪ E1 E1 ∩ E2 = E2 ∩ E1 . The
associative
property is a statement of whether operations
© 2005 by Taylor & Francis Group, LLC
grouping
CHAPTER 2. EVENTS AND PROBABILITY
22
affects the result. It does not, as shown for events
E1 , E2 , and E3 ,
E1 ∪ (E2 ∪ E3 ) = (E1 ∪ E2 ) ∪ E3 E1 ∩ (E2 ∩ E3 ) = (E1 ∩ E2 ) ∩ E3 . The
distributive
property is a statement about mixed operations, for
example, intersections of sums and sums of intersections.
E2 , and E3 ,
For events
E1 ∩ (E2 ∪ E3 ) = (E1 ∩ E2 ) ∪ (E1 ∩ E3 ) E1 ∪ (E2 ∩ E3 ) = (E1 ∪ E2 ) ∩ (E1 ∪ E3 ).
E1 ,
(2.1) (2.2)
Using the shorthand notation for intersection, Equation 2.1 can be written as
∪
E1 (E2 ∪ E3 ) = E1 E2 ∪ E1 E3 . New rules can be obtained by interchanging ∩. De Morgan’s rules formalize this:
and
E1 ∪ E2 = E1 ∩ E2 E1 ∩ E2 = E1 ∪ E2 . These rules can be generalized to
n events.
(2.3)
(2.4)
The validity of Equation 2.3 is
shown in the Venn diagram of Figure 2.4. De Morgan’s rules demonstrate a duality relation: the complement of unions and intersections is equal to the intersections and unions of complements. This property can simplify a calculation since in numerous instances is it simpler to evaluate the complement of an event than the event itself.
Figure 2.4: De Morgan’s rule visualized using Venn diagrams.
© 2005 by Taylor & Francis Group, LLC
2.1. SETS
23
Example 2.5 Failure of Drive Train In a drive train, two drive shafts are connected via a joint that permits the transmission of power even if the shafts are naturally misaligned, as shown in Figure 2.5.
The drive shafts transmit a torque
the shafts fails, then the drive train is no longer operable.
T.
If one of
Formulate this
problem and find the failure event for the drive train in terms of the events that define the breakage of each shaft individually.
Figure 2.5: Drive train under torque.
Solution
Define the following events:
E1 = breakage of shaft 1
E2 = breakage of shaft 2
.
The failure of the drive train is defined as the failure of either of the two shafts, or of both shafts. Using set notation,
failure of drive train
= E1 ∪ E2 .
Therefore, the event of no failure is given by
no failure of drive train
= E1 ∪ E2.
Also, in another way of looking at the event of no failure, if the drive train is operational, this implies that both shafts are operational, or,
no failure of drive train
= E1 ∩ E2.
Since these two events must be equal to each other, we have an illustration of De Morgan’s rule,
⊛ © 2005 by Taylor & Francis Group, LLC
E1 ∪ E2 = E1 ∩ E2.
24
CHAPTER 2. EVENTS AND PROBABILITY
Figure 2.6: Augustus De Morgan (18061871).
De Morgan Augustus De Morgan, the fifth child of LieutenantColonel John De Morgan, was born while his father was stationed in India. While he was stationed there his fifth child Augustus was born. Augustus lost the sight of his right eye shortly after birth and, when seven months old, returned to England with the family. His father died when Augustus was 10 years old. De Morgan did not excel at school and, because of his physical disability he did not join in the sports of other boys. He was even made the victim of cruel practical jokes by some school fellows. In 1823 at the age of 16, De Morgan entered Trinity College Cambridge where he was taught by Peacock and Whewell; the three became lifelong friends. He received his B.A. but, because an M.A. required a theological test, something to which De Morgan strongly objected despite being a member of the Church of England, he could go no further at Cambridge, being not eligible for a Fellowship without his M.A. In 1826 he returned to his home in London and entered Lincoln’s Inn to study for the Bar. In 1827 (at the age of 21) he applied for the chair of mathematics in the newly founded University College London and, despite having no mathematical publications, he was appointed. In 1828 De Morgan became the first professor of mathematics at University College. He gave his inaugural lecture On the study of mathematics. De Morgan was to resign his chair, on a matter of principle, in 1831. He was appointed to the chair again in 1836 and held it until 1866 when he was to resign for a second time, again on a matter of principle. © 2005 by Taylor & Francis Group, LLC
2.1. SETS His book, Elements of Arithmetic (1830), was his second publication and was to see many editions. In 1838 he defined and introduced the term “mathematical induction,” putting a process that had been used without clarity on a rigorous basis. The term first appears in De Morgan’s article, Induction (Mathematics), in the Penny Cyclopedia. (Over the years he was to write 712 articles for the Penny Cyclopedia .) The Penny Cyclopedia was published by the Society for the Diffusion of Useful Knowledge, set up by the same reformers who founded London University, and that Society also published the famous work by De Morgan, The Differential and Integral Calculus. De Morgan was always interested in odd numerical facts and writing in 1864 he noted that he had the distinction of being x years old in the year x2 . In 1849 he published Trigonometry and Double Algebra in which he gave a geometric interpretation of complex numbers. He recognized the purely symbolic nature of algebra and he was aware of the existence of algebras other than ordinary algebra. He introduced De Morgan’s laws and his greatest contribution is as a reformer of mathematical logic. De Morgan corresponded with Charles Babbage and gave private tutoring to Lady Lovelace who, it is claimed, wrote the first computer program for Babbage. De Morgan also corresponded with Hamilton and, like Hamilton, attempted to extend double algebra to three dimensions. In 1866 he was a cofounder of the London Mathematical Society and became its first president. De Morgan’s son, George, a very able mathematician, became its first secretary. In the same year De Morgan was elected a Fellow of the Royal Astronomical Society. De Morgan was never a Fellow of the Royal Society as he refused to let his name be put forward. He also refused an honorary degree from the University of Edinburgh. He was described by Thomas Hirst thus: A dry dogmatic pedant I fear is Mr. De Morgan, notwithstanding his unquestioned ability .
Macfarlane remarks that ... De Morgan considered himself a Briton unattached, neither English, Scottish, Welsh or Irish. He also says: He disliked the country and while his family enjoyed the seaside, and men of science were having a good time at a meeting of the British Association in the country, he remained in the hot and dusty libraries of the metropolis. He had no ideas or sympathies in common with the physical philosopher. His attitude was doubtless due to his physical infirmity, which prevented him from being either an observer or an experimenter. He never voted in an election, and he never visited the House of Commons, or the Tower, or Westminster Abbey .
© 2005 by Taylor & Francis Group, LLC
25
CHAPTER 2. EVENTS AND PROBABILITY
26
2.2
Probability
Let us define probability. Believe it or not, much dispute has centered on the definition of probability, primarily pitting those who view it as a subjective quantity against others who believe that only with experimentation can a rigorously derived probability be possible. The former will counter that it is often not possible to perform enough experiments to arrive at that rigorous probability and judgment must be used. Fortunately, our purposes here do not require us to resolve this debate. We can assume that in some manner it is possible to obtain the probabilities necessary for our computations, usually based on data analysis.
2
Think of a randomly vibrating oscillator where random behavior implies unpredictable periods, amplitudes, and frequencies.
These all appear to
vary from one instant of time to the next, as shown in Figure 2.7.
As
another example, what do you suppose is the temperature of the flame in a fire? Clearly, the temperature at a particular point within the flame varies in a very complex way, fluctuating in much the same way as the amplitude of Figure 2.7.
Figure 2.7: Amplitude variation in time of a random oscillator.
How can we answer a question such as: What is the probability that the value of amplitude A is greater than a specific number A0 ? Begin by expressing the question using probability notation,
Pr(A > A0 ). The ques
tion can be restated as: for how much time is the oscillator at amplitudes greater than
A0
?
This implies a
fraction
or
frequency
interpretation for
probability. Looking at a longtime history of the oscillation, it is possible to estimate the amount of time the amplitude is greater than
A0
.
That
2 J.S. Bendat and A.G. Piersol, Random Data: Analysis and Measurement Procedures , Second Edition, John Wiley & Sons, 1986. This book provides an excellent development of the theory and techniques of data analysis.
© 2005 by Taylor & Francis Group, LLC
2.2. PROBABILITY excursion frequency
27
is the probability estimate,
of time A > A0 . Pr (A ≥ A0) ≃ amounttotal time
For example, if the oscillation time history is 350 hr long and for 37 of those
A > A0 A > A0
hours
, then
Pr (A > A0 ) ≃ 37/350 = 0.106; the probability that 10.6%. This is only an estimate since, if the test
is estimated to be
lasted 3500 hr instead of 350 hr, we expect some change in the probability, although hopefully slight so that there is confidence in the estimate.
The
key to a good estimate is that the test is long enough so the probability estimate has approximately converged. Figure 2.8 depicts this procedure of estimating probabilities by frequencies of occurrence for the case of a discrete process that can have one of 7 outcomes. The frequency of each outcome is counted and the ratio with respect to the total number of all outcomes provides an estimate of the occurrence probability. Figure 2.8 is known as a
histogram.
The frequency
interpretation for estimating probabilities is the most useful, and is the approach used in this text.
Figure 2.8: Relative frequency of occurrence.
Probabilities may or may not involve time. Probabilistic models of mechanical systems are a natural result of the observation that most physical variables may take on a range of possible values.
For example, if 100
machine shafts are manufactured, there will be 100 different diameters if enough significant figures are kept in the measurements. Figure 2.9 depicts a histogram of diameter data where three significant figures are kept.
As
expected, the diameters are very close in value, but not exactly the same.
© 2005 by Taylor & Francis Group, LLC
CHAPTER 2. EVENTS AND PROBABILITY
28
Note that the sum of all frequencies must add up to
1 since all of the possible
outcomes have been included.
Figure 2.9: Histogram of machine shaft diameters.
How is such a spread of values considered in measuring the strength of the shaft in torsion? What numbers should be substituted into the stressstrain relation?
Similarly, running ultimate tensilestrength tests a num
ber of times on “identical specimens” will show no two identical results. Small differences in dimensions, material properties, and boundary conditions make it impossible to exactly duplicate experimental results.
will always be some scatter.
There
How should this information be utilized?
Randomness is possible for constants as well as functions (of time or A constant with a scatter of possible values is called a random variable. A function of time with a scatter is generally called a random 3 process. Random variables are those that can only be prescribed to a certain space).
level of certainty. Important examples are material yield characteristics that define the transition from elastic to plastic behavior.
Random processes
are timedependent (or spacedependent) phenomena that, with repeated observation under essentially identical conditions, do not show the same time histories. For increasingly complex engineering demands, it is important to understand, and be able to model, uncertainties and qualitative information for
3
Or the Greek stochastic :
στ oκoς.
© 2005 by Taylor & Francis Group, LLC
2.2. PROBABILITY
29
analysis and design. An example of qualitative information is a verbal description of size or strength. Developing the ability to analyze uncertainties allows the engineer to decide for which applications they are insignificant and may be ignored. For many applications discussed in this text, scatter cannot be ignored because of the resulting variability of response.
2.2.1
Axioms of Probability
Mathematical probability is a ma jor branch of mathematics. Yet all of the vast understanding can be traced to the three
axioms of probability.
axiom is a stipulation of a property that is not proven.
An
It is postulated.
It is used as a foundation upon which a deductive framework is built. The veracity and usefulness of an axiom is linked to the veracity and usefulness of the framework that is created.
All of mathematics is based on various
axioms. The axioms of probability are the following:
1. For every event governed by
E in a sample space S , the probability of the event is
Pr(E ) ≥ 0.
Pr(S ) = 1. For two events E1 , E2 , that are mutually exclusive, that is, E1 ∩E2 = ∅,
2. The probability of the certain event 3.
S
is given by
the probability of the occurrence of either or both events is given by
Pr(E1 ∪ E2) = Pr(E1 ) + Pr(E2). Using these axioms along with the previous combination rules allows us to derive all the rules of probability.
Example 2.6 The Third Axiom of Probability Machine shafts are being manufactured. The shaft can be rejected if the diameter is less than 98% or greater than 102% of its nominal value. The “nominal” value is the desired or design value. The probability that a shaft is being rejected because the diameter is less than 98% of its nominal value is given as 0.02 and the probability that a shaft is being rejected because the diameter is greater than 102% of its nominal value is given as 0.015. What is the probability that a shaft will be rejected?
Solution
Let
E1
be the event that the diameter of the shaft is less
than 98% of its nominal value, and
E2
be the event that the diameter of
the shaft is greater than 102% of its nominal value. The probability of each event is
Pr (E1) = 0.02,
© 2005 by Taylor & Francis Group, LLC
and
Pr (E2) = 0.015.
CHAPTER 2. EVENTS AND PROBABILITY
30
A shaft will be rejected if either event
E1 or E2 occurs.
That is, the proba
E1 and E2 , that is, Pr (E1 ∪ E2 ) . Since the shaft diameter cannot be too small and too large at the same time, the events E1 and E2 are mutually exclusive, bility that a shaft will be rejected equals the probability of the union
and using the third axiom of probability,
Pr(E1 ∪ E2 ) = Pr (E1) + Pr (E2) = 0.035.
⊛ 2.2.2
Extensions from the Axioms
A probability is numerically in the range
0 ≤ Pr(E ) ≤ 1.
To show this,
write the following equivalent relations:
Pr(E ∪ E ) = Pr(E ) + Pr(E ) Pr(E ∪ E ) = Pr(S ) = 1
by axiom
3
by axiom
2.
Equate the two probability statements to find
Pr(E ) + Pr(E ) = 1
0 ≤ Pr(E ) ≤ 1.
and therefore
These two probability rules are universal and used often. axiom 3 for the case where
An extension of
E1 ∩ E2 = ∅ is very useful in applications since
in most instances events that are part of the same system are not disjoint. For overlapping events, as shown in the Venn diagram of Figure 2.10,
Pr(E1 ∪ E2) = Pr(E1) + Pr(E2 ) − Pr(E1 ∩ E2). It is clear from the Venn diagram that the event since it has been counted twice by adding
E1
(2.5)
E1 ∩ E2 must be subtracted E2 . Equation 2.5 can be
and
extended to any number of variables as shown here,
Pr(E1 ∪ E2 ∪ E3) = Pr([E1 ∪ E2] ∪ E3 ) = Pr(E1 ∪ E2 ) + Pr(E3) − Pr([E1 ∪ E2 ] ∩ E3 ) = Pr(E1) + Pr(E2 ) − Pr(E1 ∩ E2) + Pr(E3) − Pr([E1 ∩ E3 ] ∪ [E2 ∩ E3]) = Pr(E1) + Pr(E2 ) + Pr(E3) − Pr(E1 ∩ E2) − Pr(E1 ∩ E3) − Pr(E2 ∩ E3) + Pr(E1 ∩ E2 ∩ E3). Estimating probabilities such as
Pr(Ei ∩Ej ) and Pr(Ei ∩Ej ∩Ek ) requires
information, usually experimentally derived, on whether these events “depend” on each other. If they are disjoint,
© 2005 by Taylor & Francis Group, LLC
Ei ∩ Ej = φ and Ei ∩ Ej ∩ Ek = φ,
2.2. PROBABILITY
Figure 2.10: Union of
31
E1
and
E2
visualized using Venn diagrams.
then their probabilities equal zero. Using De Morgan’s rule, Equation 2.3, we also find that
Pr(E1 ∪ E2 ∪ E3) = 1 − Pr(E1 ∪ E2 ∪ E3) = 1 − Pr(E1 ∩ E2 ∩ E3).
Example 2.7 Extension of Probability Axioms Show that
Solution
Pr(E1 ∪ E2 ) ≤ Pr(E1) + Pr(E2). From the first axiom, we know that
Pr (E1 ∩ E2 ) ≥ 0.
Therefore,
Pr(E1) + Pr(E2) − Pr (E1 ∩ E2 ) ≤ Pr(E1) + Pr(E2).
(2.6)
Using Equation 2.5, the lefthand side of Equation 2.6 is identical to the probability
Pr (E1 ∪ E2) . Thus,
Pr(E1 ∪ E2) ≤ Pr(E1) + Pr(E2 ).
⊛ 2.2.3
Conditional Probability
Thinking about the implications of Equation 2.5 leads us to consider the modeling of
causality
probabilistically. It is important to be able to estimate
the probability that one event occurs
© 2005 by Taylor & Francis Group, LLC
given
that another event
has already
CHAPTER 2. EVENTS AND PROBABILITY
32
occurred. of
E1
Formally, the
conditional probability
is defined as the probability
E2 has occurred, and is given by E1 ∩ E2 ) . Pr(E1 E2) = Pr(Pr( E2 )
occurring given that
(2.7)
This equation will sometimes be useful in its equivalent form:
Pr(E1 ∩ E2) = Pr(E1E2) Pr(E2). Note that Pr(E1 ∩ E2 ) also equals Pr(E2 E1 ) Pr(E1 ) since E1 ∩ E2 = E2 ∩ E1 . Interpret Equation 2.7 in terms of the sample space frequency interpretation of probability.
S
and the relative
Recall that in this interpretation,
the ratio of desired outcomes to the complete space
S
is a measure of the
probability of realization of that outcome. Since the conditional probability
E2 has occurred, the complete space is reduced from S to E1 occurs given that E2 has occurred is given by the intersection E1 ∩ E2 and the probability equals the ratio of the probability of the intersection to the probability of E2 . states that event
E2 .
The event that
Given what we have just said about the conditional probability makes it clear that we have been actually using conditional probabilities when describing probabilities, except that the condition was on the whole space
S. That is, with E2 = S,
Pr(E1 ∩ S ) Pr(S ) Pr( = E1 ) 1 = Pr(E1 ),
Pr(E1S ) =
so that the probability of an event to the space
S.
E1 , as given by Pr(E1 ), is with respect
Example 2.8 Conditional Probability Light bulbs are tested to establish lifetimes. Suppose a light bulb can fail prematurely for a variety of reasons, one reason being a defective filament. The probability that the part will fail for any reason is 0.01.
If the bulb
happens to have a defective filament, then the probability that it will fail prematurely is 0.1. If the bulb fails, then the probability that the cause of failure is the defective filament is 0.05. filament is defective?
Solution
Let
What is the probability that the
A be the event that the bulb fails prematurely and B
the event that the filament is defective. The given probabilities are
Pr (A) = 0.01, Pr (AB) = 0.1, © 2005 by Taylor & Francis Group, LLC
and
Pr (BA) = 0.05.
2.2. PROBABILITY Find
33
Pr (B) .
Using the definition of conditional probability in Equation 2.7, we find that
Pr (B ∩ A) = Pr (BA) Pr (A) = 0.05 × 0.01 = 0.0005. The probability that the bulb fails and also has a defective filament is 0.0005. Again using the definition of conditional probability,
∩ B) Pr (B) = PrPr(A (AB) 0 .0005 = 0 .1 = 0.005. The probability that any given bulb has a defective filament is 0.005.
⊛
2.2.4
Statistical Independence
Two events are said to be event is
not dependent on,
statistically independent if the occurrence of one and does not affect the occurrence of, the other
event. From the definition of conditional probability, for statistically independent events
E1
and
E2 ,
Pr(E1E2) = Pr(E1) Pr(E2E1) = Pr(E2), and therefore,
Pr(E1 ∩ E2) = Pr(E1 E2) Pr(E2) = Pr(E1) Pr(E2) Pr(E2 ∩ E1) = Pr(E2 E1) Pr(E1) = Pr(E2) Pr(E1). Statistical independence can then be defined as
Pr(E1 ∩ E2) = Pr(E1) Pr(E2 ).
(2.9)
Statistical independence relates how the probability of the joint event
E1 ∩ E2
is related to the probabilities of the individual events
E1
and
E2 .
On the other hand, mutual exclusiveness is a statement of whether the joint event
E1 ∩ E2 exists.
Mutually exclusive events cannot both happen at the
same time. That is, if For nontrivial events,
E1 and E2 are mutually exclusive, Pr (E1 ∩ E2 ) = 0. Pr (E1) = 0 and Pr (E2 ) = 0, the righthand side
of Equation 2.9 cannot be zero, and in this case mutually exclusive events cannot be statistically independent.
© 2005 by Taylor & Francis Group, LLC
CHAPTER 2. EVENTS AND PROBABILITY
34
Example 2.9 Statistically Independent versus Mutually Exclusive Events Toss a coin once.
E1 is the event that the result is heads, and E2 is the
event that the result is tails
Pr(E1) = 0.5 Pr(E2) = 0.5. These two events are mutually exclusive since both cannot happen at the
E1 ∩ E2 ) = 0. These events are not statistically independent E2 cannot occur. Now toss two fair coins. Let E1 be the event that the result of the first coin toss is heads, and E2 be the event that the result of the second coin toss same time, Pr(
since if
E1
occurs,
is heads. These two events are statistically independent since the outcome of one does not affect the outcome of the other.
These events can also
happen at the same time. That is, these events are not mutually exclusive, but are statistically independent.
⊛
Example 2.10 Statistical Independence A die and a coin are tossed.
Assume that they are thrown such that
they do not interfere with each other. What is the probability of rolling a 2 provided that the coin comes up heads. Also, what is the probability that the die shows an even number and the coin comes up heads?
Solution
Let
A be the event that the die will show 2 and B the event
that the coin shows heads. Assuming a fair coin and die, the probability of each event is
Pr (A) = 1/6 and Pr (B) = 1/2. The two events can be assumed to be independent.
That is, the outcome
of one event does not affect the outcome of the other. Therefore, the probability that the die will show 2 is not affected by the outcome of the coin. The answer is 1/6
. We just showed
Pr (AB) = Pr (A) . The probability that the die shows 2 and the coin shows heads can be written as Pr
⊛
(A ∩ B) . The probability is 1/6 × 1/2 = 1/12.
© 2005 by Taylor & Francis Group, LLC
2.2. PROBABILITY 2.2.5
35
Total Probability
Most applications that engineers encounter, and must design for, contain many components and processes.
For each of these many events we can
define their operating characteristics and reliability. event
E0
Suppose there is an
that affects the design, but it cannot be measured or determined
directly. Its occurrence, however, is always accompanied by the occurrence of one or several other events. For example, perhaps the pressure is needed but cannot be measured, whereas the temperature that can be related to the pressure is measurable. Let these other possible events be denoted by the mutually exclusive events
E1 , E2 , · · · , En , as shown in Figure 2.11.
Figure 2.11: Constructing the theorem of total probability.
E0 can overlap any of the other events in the Venn diagram.
Conditional
probabilities can be used to construct the equivalent statements,
Pr(E0 ) = Pr(E0E1) + Pr(E0E2) + · · · + Pr(E0En) = Pr(E0E1) Pr(E1) + Pr(E0 E2) Pr(E2)+ · · · + Pr(E0En) Pr(En). This equation is called the
theorem of total probability,
(2.10)
which is very useful
in deducing probabilities of an event in terms of its dependence on other events.
© 2005 by Taylor & Francis Group, LLC
CHAPTER 2. EVENTS AND PROBABILITY
36
Example 2.11 Inspection of a Manufactured Product via Total Probability The process of manufacturing an item requires that two of the components be welded, after which a visual inspection is performed. The item fails the inspection if either the weld has gaps or the components are misaligned. The manufacturing history allows us to estimate the following probabilities. It is estimated that 1% of the welds have gaps and that 4% of the components are misaligned.
It is also known that if a weld is misaligned that it
is 30% more likely to have gaps than if it were not misaligned. We need to estimate the probability that a given weld will pass inspection.
Figure 2.12: Schematic of regions:
Solution
G, M , GM , and G ∪ M.
First find the probability of failure,
and then use this
G = gap in weld, M = misaligned component. We are given the probabilities Pr(G) = 0.01 and Pr(M ) = 0.04 as shown schematically on the Venn diagram of Figure to answer the question.
Define the following events:
2.12. We are also given a relation between misalignment and gaps, that is,
Pr(GM ) = 1.3 Pr(GM ). Since failure occurs when either or both events G and
M
occur, we need to evaluate
Pr(G ∪ M ) = Pr(G) + Pr(M ) − Pr(GM ), where
Pr(GM ) = Pr(GM ) Pr(M ). Pr(GM ) is unknown. Therefore, another equation is needed since
But
there are two unknowns. The theorem of total probability is that equation,
© 2005 by Taylor & Francis Group, LLC
2.2. PROBABILITY
37
stating that a gap may occur with or without a mismatch,
Pr(G) = Pr(GM ) Pr(M ) + Pr(GM ) Pr(M ) 0.01 = Pr(GM )0.04 + 11.3 Pr(GM )(1 − 0.04) 0.01 = 0.778 Pr(GM ) 0.0128 = Pr(GM ). This is the value of the needed probability, and
Pr(G ∪ M ) = Pr(G) + Pr(M ) − Pr(GM ) Pr(M ) = 0.01 + 0.04 − 0.0128(0.04) = 0.0495. This is the probability that the weld will fail inspection.
The probability
that it will pass inspection is
Pr(G ∪ M ) = 1 − 0.0495 = 0.9505, or
95.05%.
If the designer or the end user believes this to be too small
a reliability, then the previous set of calculations provide guidance where manufacturing improvements need to be made. In applications, such precise numbers may not be obtainable. one needs to make additional assumptions.
For example, if
G
and
Then
M
are
Pr(GM ) = Pr(G) Pr(M ). Performing the same calculations as above leads to the probability Pr(G ∪ M ) = 0.0496. If on the other hand the unrealistic assumption is made that G and M are mutually exclusive (they cannot both occur), then Pr(GM ) = 0, and Pr(G ∪ M ) = 0.0500. For this example these assumptions do not
assumed to be statistically independent, then
significantly alter the results. Other problems may not be so forgiving.
⊛
2.2.6
Bayes’ Theorem
It is possible to combine the expressions for conditional probability and the theorem of total probability into one that provides a mechanism for updat
ing probabilities as new information becomes available. Since Pr(E1 ∩ E2 ) = Pr(E2 ∩ E1), then Pr(E1 ∩ E2 ) = Pr(E1E2) Pr(E2) = Pr(E2E1 ) Pr(E1). For an arbitrary event
Ea ,
Pr(EaEi ) Pr(Ei ) = Pr(EiEa) Pr(Ea) =⇒ Pr(EiEa) = Pr(EaEi ) Pr(Ei ) , Pr(Ea) © 2005 by Taylor & Francis Group, LLC
(2.11)
CHAPTER 2. EVENTS AND PROBABILITY
38
where
i = 1, ..., n,
known as
and
n = number of possible events.
Bayes’ theorem.
This equation is
The theorem of total probability can replace
Pr(Ea ) in the denominator, resulting in the generalized Bayes’ theorem, Ea Ei ) Pr(Ei ) . Pr(Ei Ea) = nPr(Pr( Ea Ej ) Pr(Ej ) j =1
The meaning of this equation is not obvious. We first explain it qualitatively, and then fix these ideas with examples. Rewrite Equation 2.11 for the case
Ei = E0 :
E0) Pr(E0) . Pr(E0Ea) = Pr(EaPr( E)
(2.12)
a
E0
The initial probability estimate that
occurs is
Pr(E0 ). To improve this
estimate, a test is performed on the system resulting in event
Ea .
Figure
2.11 provides a hint at where we are heading. In the theorem of total probability,
Pr(E0) is found by evaluating the intersections of E0
other events in the space.
these other events provides additional information about ilar way here, knowing that
Ea
Ea
Pr(E0). In a simPr(Ea ) provides
occurred with probability
additional information. The estimate tional knowledge that
with all the
The occurrence, or nonoccurrence, of any of
Pr(E0 ) can be updated with the addiPr(Ea ), that is, Pr(E0Ea).
has taken place with
Pr(E E )
To signify this updating, rewrite Equation 2.12 in the following way:
Pr(E0 Ea) =
a
0
Pr(Ea)
Pr(E0),
where the term in the square brackets is sometimes called a
tion
or
transfer function
on the initial estimate
likelihood func
and represents the effect of the new information
Pr(E0). The following example helps to demonstrate
this procedure.
Example 2.12 Bayes’ Rule in Manufacturing Cameras are being manufactured by two factories, A and B. It is known that the probability that a defective unit will still pass the final inspection is 0.01. If the unit is defective, the probability that it came from factory A is 0.3. If the unit is not defective, the probability that it came from factory B is 0.8. Suppose that a consumer can find out where the camera came from. This additional information can help a customer to decide which camera he or she will buy.
In order to decide which camera to buy, we need to find
the probability that the camera is defective if it came from factory A, and the probability that the camera is defective if it came from factory B.
© 2005 by Taylor & Francis Group, LLC
2.2. PROBABILITY B
Solution
Let
39
F
be the event that a camera is defective, and
A and
the events that a camera came from factory A and B, respectively.
are given
Pr (F ) = 0.01, Pr (AF ) = 0.3,
and
We
Pr BF¯ = 0.8.
Also, if the camera is not manufactured by factory A, then it must be manufactured by factory B,
¯ A¯ = B and A = B. We are asked to find
Pr (F A) and Pr (F B) .
It is given that 30% of the defective items are from factory A, which leaves 70% for factory B. Based on this figure, the consumer may be tempted to choose a camera from factory A. However, if factory B manufactures substantially more products than factory A, it may safer to choose the product from factory B. Using Bayes’ rule, the following conditional probabilities are known,
Pr (AF ) Pr(F ) Pr (AF ) Pr (F ) + Pr AF¯ Pr F¯ Pr (BF ) Pr(F ) . Pr (F B) = Pr (BF ) Pr (F ) + Pr BF¯ Pr F¯ Pr (F A) =
A and B
Pr AF¯ = 1 − Pr BF¯ = 0.2, ¯ = 0.99, and the probability, the probability Pr F Since
are complementary,
Pr (BF ) = 1 − Pr (AF ) = 0.7. The probability that the camera from factory A is defective is
0.3 × 0.01 Pr (F A) = 0.3 × 0.01 + 0.2 × 0.99 = 0.0149, and the probability that the camera from factory B is defective is
0.7 × 0.01 Pr (F B) = 0.7 × 0.01 + 0.8 × 0.99 = 0.00876. Therefore, a consumer is better off with a camera from factory B.
© 2005 by Taylor & Francis Group, LLC
CHAPTER 2. EVENTS AND PROBABILITY
40
Table 2.1: Probabilities
1
Die 1 Die 2 Die 3
2
3
4
5
6
1/12 1/6 1/12 1/3 1/6 1/6 1/6 1/6 1/6 1/12 1/12 1/3 1/3 1/6 1/6 1/6 1/12 1/12
When the origin of the product is not known, the consumer is faced with a probability of 0.01 that the product is defective. However, with the new piece of information, buying the camera from factory B reduced the probability that the product is defective to 0.00876. In this way, additional information can be used to refine and make more precise the probability estimates, thus updating them.
⊛ Example 2.13 Bayes’ Rule and the Loaded Dice Table 2.1 shows the probabilities of all the realizations for three dice. The dice are all loaded, meaning that they are unbalanced in some way leading to an unequal probability of any number coming up.
Note that
summing each row still must yield 1, since all the numbers can appear.
(i) Find the probability of rolling a 6 if one die is selected at random. Use the theorem of total probability. Let
A1, A2 , A3
B
equal the event of rolling a 6;
equal the events of selecting die 1, 2, or 3, respectively.
Since
Pr(A1) = Pr(A2) = Pr(A3) = 1/3. From the table, Pr(B A1 ) = 1/6, Pr(B A2 ) = 1/3, Pr(B A3 ) = 1/12. Substitute the die is selected at random,
these into the total probability theorem to obtain the probability of rolling a 6,
Pr(B) = Pr(BA1) Pr(A1) + Pr(BA2 ) Pr(A2 ) + Pr(BA3) Pr(A3) 1 1 · = 1 · 13 + 13 · 13 + 12 3 6 7 = = 0.19444. 36 If all the die were fair, we would have found Pr(B ) = 3 × (1/18) = 1/6 = 0.16667. (ii) Determine the probability that die 2 was chosen if a 6 was rolled with the randomly selected die. This question looks at the problem in a reverse way than did the last one. The required probability is
Pr(A2B) = Pr(BA2) Pr(A2) , Pr(B ) © 2005 by Taylor & Francis Group, LLC
Pr(A2B ), given by
2.2. PROBABILITY using Bayes’ rule.
41
Note that
Pr(B)
is required in order to evaluate this
expression, and, therefore, even if the previous question was not asked, we would have had to apply the total probability theorem. We have then,
Pr(A2 B) =
1 3
·
7 36
= 47 = 0.57143.
1 3
This result can also be obtained by considering the reduced space of probabilities for a roll of 6. From the table, the probabilities under the 6 are the reduced space, thus,
Pr(A2B) =
1 6
1 3 1 3
+ + 121
= 0.57143.
Applying Bayes’ rule, we can similarly find the following,
Pr(A1B) =
Pr(A3B) =
1 6
·
1 3
7 36 1 1 12 · 3 7 36
These results can be verified by adding
= 0.28571
= 0.14286.
Pr(A1B)+Pr(A2B )+Pr(A3 B) = 1,
as they should, signifying that the realizations have been tracked accurately.
⊛
Example 2.14 Liquid Sloshing in a Container4 The sloshing of a liquid in a container is a very important problem in engineering design.
A number of key engineering systems contain an
enclosed fluid, for example, the water tower on top of a building, fuel in aircraft wings and fuselage, fuel in spacecraft flying in low or microgravity, liquid flow in pipes.
Each of these is an example of liquid sloshing in a
container. A common characteristic is that the amount of liquid varies with time, the sloshing (moving) liquid has inertia, and the amount of liquid affects the vibration characteristics of the enclosing structure. The purpose of this problem is to consider the simplified problem of a water tower in a seismic zone. Studies have been performed on the strength of the water tower to the shock waves of various earthquakes. There are two primary uncertainties in this problem, the strength and arrival time of the earthquake, and the amount of water in the tower at the seismic event. The arrival time uncertainty and the quantity of water in the tower are really
4 Problem based on A.HS. Ang and W.H. Tang, Probability Concepts in Engineering and Planning, Vol.1 Basic Principles, John Wiley & Sons, 1975.
© 2005 by Taylor & Francis Group, LLC
CHAPTER 2. EVENTS AND PROBABILITY
42
the same uncertainty, because if we knew the arrival time of the quake, we could extrapolate the amount of water in the tower. From test data on the tower and historical data on the seismicity of the region, the following information is available.
When an earthquake
occurs, the probability that the tower fails depends on the magnitude of the earthquake and on the amount of water in the tower at that time.
Since
the earthquake magnitude and the fullness of the tower can be any of a broad range of numbers, we simplify these by assuming that the tower is either full or halffull with a relative likelihood of 1:3. The lowest magnitude earthquake that is of significance is called a weak earthquake, and the one of highest magnitude is called a strong earthquake. The relative likelihood is, respectively, 8:2. It is determined that if a strong earthquake hits the tower, it collapses regardless of the quantity of water in the tank.
We also know
that if a weak earthquake hits the tower and it is at most halffull, then it will definitely survive the event.
5
However, if during a weak seismic event
the tower is full, then it has only a 1 in 2 chance of survival.
(This
50%
chance of survival implies that within the group of weak earthquakes there are characteristic differences that can be substantial and can sometimes lead to tower failure.) The two questions of interest about this system and its response to its environment are:
(i)
what is the probability of tower collapse, and
(ii)
if
the tower collapsed during a seismic event, what is the probability that the tank was full at the time of the quake?
Solution
The first step is to define all possible events along with
their respective probabilities. Then the theorem of total probability is used along with Bayes’ rule. Define the following events:
F = full tower, H = halffull tower, C = tower collapse, S = strong seismic event, W = weak seismic event. Using the relative likelihood information given in the problem statement, the following probabilities can be deduced:
Pr(F ) = 0.25, Pr(H ) = 0.75 Pr(S ) = 0.20, Pr(W ) = 0.80 Pr(C SF ) = Pr(C SH ) = 1 Pr(C W F ) = Pr(C W F ) = 0.5 Pr(C W H ) = 0. 5 This bit of information suggests to a designer that the tower be sensored so that in the event of an earthquake, enough water be released into the sewers so that the tower is less than half full.
© 2005 by Taylor & Francis Group, LLC
2.2. PROBABILITY
43
To calculate the probability of collapse, use the total probability theorem, summing all the possible ways collapse can occur,
Pr(C ) = Pr(CSH ) + Pr(CSF ) + Pr(CWH ) + Pr(CW F ) = Pr(C SH ) Pr(SH ) + Pr(C SF ) Pr(SF ) + Pr(C W H ) Pr(W H ) + Pr(C W F ) Pr(W F ). At this point a physical judgement is required about the joint event (seismic event, quantity of water in tank). It is reasonable to assume that these two events are statistically independent. Thus,
Pr(C ) = 1 × 0.20 × 0.75 + 1 × 0.20 × 0.25 + 0 + 0.5 × 0.80 × 0.25 = 0.3. The probability that the tower is full, assuming a failure event, is given by
Pr(F C ) = Pr(F S C ) + Pr(F W C ) Pr(C F S ) Pr(F S ) + Pr(C FW ) Pr(F W ) = Pr(C ) Pr(C ) 0 .5 × 0.25 × 0.80 1 × 0.25 × 0.20 + = 0 .3 0.3 = 0.5. While this is the answer to the problem, a designer would look at the results and think about how the probability of collapse can be reduced. The general equation for
Pr(C ) shows how each component probability on the right hand
side adds up to the total probability. It suggests weaknesses in the structure. Therefore, the above procedure can be used iteratively as a design tool until
Pr(C ) is sufficiently small.
⊛
Bayes Thomas Bayes was ordained a Nonconformist minister like his father, who was one of the six Nonconformist ministers to be ordained in England. Thomas was educated privately, something that appears to have been necessary for the son of a Nonconformist minister at that time. Nothing is known of his tutors, but Barnard points out the intriguing possibility that he could have been tutored by de Moivre, who was certainly giving private tuition in London at this time. At first Bayes assisted his father in Holborn. In the late 1720s he became minister of the Presbyterian Chapel in Tunbridge Wells, 35 miles southeast of London. On August 24, 1746 William Wiston describes having breakfast with Bayes, who he says is ... a dissenting Minister at Tunbridge Wells, and a Successor, though not immediate, to Mr. Humphrey Ditton, and like him a very good mathematician.
© 2005 by Taylor & Francis Group, LLC
44
CHAPTER 2. EVENTS AND PROBABILITY
Figure 2.13: Thomas Bayes (17021761).
Bayes apparently tried to retire from the ministry in 1749 but remained minister at Tunbridge Wells until retirement in 1752, but continued to live in Tunbridge Wells. Bayes set out his theory of probability in Essay towards solving a problem in the doctrine of chances published in the Philosophical Transactions of the Royal Society of London in 1764. The paper was sent to the Royal Society by Richard Price, a friend of Bayes, who wrote: I now send you an essay which I have found among the papers of our deceased friend Mr. Bayes, and which, in my opinion, has great merit. In an introduction which he has writ to this Essay, he says, that his design at first in thinking on the subject of it was, to find out a method by which we might judge concerning the probability that an event has to happen, in given circumstances, upon supposition that we know nothing concerning it but that, under the same circumstances, it has happened a certain number of times, and failed a certain other number of times.
Bayes’ conclusions were accepted by Laplace in a 1781 memoir, rediscovered by Condorcet (as Laplace mentions), and remained unchallenged until Boole questioned them in the Laws of Thought. Since then Bayes’ techniques have been subject to controversy. Bayes also wrote an article An Introduction to the Doctrine of Fluxions, and a Defence of the Mathematicians Against the Objections of the Author of The Analyst (1736) attacking Berkeley for his attack on the logical foundations of the calculus. Bayes writes that Berkeley ...represents the disputes and controversies among mathematicians as disparaging the
© 2005 by Taylor & Francis Group, LLC
2.3. CONCLUDING SUMMARY
45
evidence of their methods and ... he represents Logics and Metaphysics as proper to open their eyes, and extricate them from their difficulties. If the disputes of the professors of any science disparage the science itself, Logics and Metaphysics are much more disparaged than Mathematics, why, therefore, if I am half blind, must I take for my guide one that cannot see at all?
Bayes was elected a Fellow of the Royal Society in 1742 despite the fact that at that time he had no published works on mathematics, indeed none were published in his lifetime under his own name. The article on fluxions referred to above was published anonymously. Another mathematical publication on asymptotic series appeared after his death .
2.3
Concluding Summary
This chapter has introduced the underlying basis for probability theory, sets and operations with sets.
These and the axioms of probability permit the
development of the subsequent rules for working with uncertain parameters and variables. The interpretation of probability as a frequency of occurrence is provided as a useful practical definition. In addition, some basic probabilistic concepts have been introduced, in particular, conditional probability and independence, very important concepts that are used in the following chapters.
2.4
Problems
Section 2.1: Sets 1. Give 3 examples of impossible events from everyday life. 2. Give 3 examples of impossible events from fluids engineering. 3. Give 3 examples of impossible events from materials engineering. 4. Give 3 examples of impossible events from strength of materials. 5. Give 3 examples of impossible events from mechanical vibration. 6. Give 3 examples of impossible events from thermal engineering. 7. Give 3 examples of certain events from everyday life. 8. Give 3 examples of certain events from fluids engineering. 9. Give 3 examples of certain events from materials engineering.
© 2005 by Taylor & Francis Group, LLC
CHAPTER 2. EVENTS AND PROBABILITY
46
10. Give 3 examples of certain events from strength of materials. 11. Give 3 examples of certain events from mechanical vibration. 12. Give 3 examples of certain events from thermal engineering. 13. Give an example of complementary events from everyday life. 14. Give an example of complementary events from fluids engineering. 15. Give an example of complementary events from materials engineering. 16. Give an example of complementary events from strength of materials. 17. Give an example of complementary events from mechanical vibration. 18. Give an example of complementary events from thermal engineering. 19. Given the three events:
X = Y = Z =
{odd numbers} {even numbers} {negative numbers} ,
obtain the following:
(i) X ∪ Y (iii) X (v) Z
(ii) X ∩ Y (iv) Y (vi) Y ∩ Z.
20. Extend Example 2.5 to the case where there are three shafts connecting two rotors (instead of two shafts connected to one rotor). shafts are numbered from left to right as
The
1, 2, 3. The failure of the
drive train is defined as the failure of either of the three shafts, with events defined by
E1, E2,
and
E3,
respectively.
Find the following
events:
(i)
failure of the drive train,
(ii) (iii)
no failure of the drive trains, and show an illustration of de Morgan’s rule.
Section 2.2: Probability 21. In your own words, explain the essential ideas of the Theorem of Total Probability, and briefly discuss its importance.
© 2005 by Taylor & Francis Group, LLC
2.4. PROBLEMS
47
22. Consider Figure 2.7 where
dav = 50
mm, and suppose 50 shafts are
manufactured. From the measurements we observe that
25 have the diameter dav 10 have the diameter 1.01 dav 6 have the diameter 1.02 dav 5 have the diameter 0.99 dav 4 have the diameter 0.98 dav . Sketch the frequency diagram showing appropriate numbers along the axes. Using the frequency interpretation for probability, calculate the probability of occurrence for each shaft size and verify that the sum of these probabilities equals 1. 23. Suppose
(i)
If
E1
Pr(E1 ) = 0.20, and Pr(E2) = 0.30. and
E2
relate to a particular process, are any events not
accounted for here? Why?
(ii)
If
Pr(E1 ∪ E2) = 0.90
are these processes mutually exclusive?
Why?
Pr(E1 ∪ E2 ) = 0.50, then what is the value of Pr(E1E2)? Suppose Pr(A) = 0.5, Pr(AB ) = 0.3, and Pr(B A) = 0.1, calculate Pr(B ). Suppose Pr(A) = 0.5, Pr(AB ) = 0.5, and Pr(B A) = 0.1, calculate Pr(B ). What can be concluded about the statistical relationship, if (iii)
24.
25.
If
any, between
A
and
B.
26. Continuing Example 2.13, let
G
equal the event of rolling a 3;
A1, A2, A3
equal the events of selecting die 1, 2, or 3, respectively.
(i)
Find the probability of rolling a 3 if one die is selected at random.
Use the theorem of total probability.
(ii)
Determine the probability that die 2 was chosen if a 3 was rolled
with the randomly selected die. 27. Consider the drive train Example 2.5 from a different perspective. The drive train consists of a rotor
R
and turbine blades
are manufactured in high precision.
B.
The components
How well the system operates
depends on the precision of the manufactured components. Testing of the individual components by the manufacturer yields the following information:
0.1% of R have imperfections 0.01% of B fail. © 2005 by Taylor & Francis Group, LLC
CHAPTER 2. EVENTS AND PROBABILITY
48
Also, it is determined that if are
R
has imperfections, then the blades
B
50% more likely to fail due to the additional vibration forces that
result. Determine the probability that the system will pass inspection.
© 2005 by Taylor & Francis Group, LLC
Chapter 3
Random Variable Models In our previous discussions, a language based on set theory was introduced to work with probability and uncertainties.
Several key equations were
used to relate probabilities with each other.
But these probabilities are
about “extreme” events. For example, a component fails or not, with some probability. What we would like to be able to do is to provide a complete probabilistic description of a parameter, not just the probability of a particular realization. This complete probabilistic description is in the form of a function over the whole range of realizations. Begin by exploring the properties of random variables and the functions that define them. Probability affords a framework for defining and utilizing such variables in the models developed for engineering analysis and design. Mathematical models of physical phenomena are essentially relationships between variables. Where some of these variables have associated uncertainties, there are a multiplicity of possible values for each random variable. An example is the set of possible values of Young’s modulus determined from a series of experiments on “identical” test specimens. This multiplicity is represented by the probability functions introduced next. A random variable may be discrete, continuous, or mixed. Once a parameter is taken to be random, its probability functions provide a complete description of its variability. The importance of these probability functions is that they can be used in mathematical operations involving randomness. 49
© 2005 by Taylor & Francis Group, LLC
CHAPTER 3. RANDOM VARIABLE MODELS
50
3.1
Probability Distribution Function
The likelihood of a random variable taking on a particular range of values 1 is defined by its probability distribution function, defined as
where
Pr(X ≤ x)
FX (x) = Pr (X ≤ x) , is the probability that random variable
or equal to the number the particular value shown that
FX (x)
(3.1)
x.
x.
X
is less than
This probability is, of course, a function of 2 Based on the axioms of probability, it can be
is an increasing function of
x,
and is bound by 0 and
1. The impossible event has a zero probability, and the certain event has a probability of one. In particular,
since
Pr(X < −∞) = 0;
lim FX (x) = 0 x→−∞ all realizations of the random variable must be
greater than negative infinity.
A realization is one of the many possible
values of a random variable. Similarly,
x→lim +∞ FX (x) = 1
since
Pr(X < +∞) = 1; all realizations of the random variable must be less FX (x) are 0 ≤ FX (x) ≤ 1, and for
than positive infinity. Thus, bounds on
x1 ≤ x2,
FX (x1) ≤ FX (x2), since Pr(X ≤ x1 ) ≤ Pr(X ≤ x2 ). See Figure 3.1.
Note that the probabil
ity distribution function is nondecreasing. Where the context is clear, the
subscript
X
is omitted.
While the above definitions apply to all types of random variables, if the
Pr(X = xi) x ≤x pX (xi),
variable is discrete, Equation 3.1 becomes
FX (x) =
=
al l
al l 1
F (x)
is sometimes called the
accumulated as
2 An
x b ecomes
i
(3.2)
xi ≤x
cumulative distribution function ,
since probability is
larger.
excellent b ook on the basics of probabilistic modeling is by A. Pap oulis,
ability, Random Variables, and Stochastic Processes ,
M cGraw—Hill.
Prob
There are several
editions. We find the first edition most readable. We would also encourage the reader to lo ok up some of the other fine texts by Pap oulis on probability and stochastic processes; they are among the b est. A different approach to explaining probability is offered by C. Ash in The Probability Tutoring Book: An Intuitive Course for Engineers and Scientists (and everyone else!) , IEEE Press, 1993. It offers an intro duction through problem solving.
© 2005 by Taylor & Francis Group, LLC
3.1. PROBABILITY DISTRIBUTION FUNCTION
51
Figure 3.1: A continuous probability distribution function.
where
pX (xi) are the individual probabilities of each realization.
A hint of
the cumulative distribution function can be found in the histogram of Figure 2.9. Each frequency is a probability,
pX (xi ), and the sum from left to right
is the cumulative distribution function. The histogram is a practical way to build a distribution function.
Example 3.1 From Histogram to Cumulative Distribution Function The cumulative distribution function corresponding to the histogram of
F (0.98dAV ) = n0.98 , F (0.99dAV ) = n0.98 + n0.99 , ..., and F (1.02dAV ) = n0.98 + ... + n1.02 = 1. See Figure 3.2.
Figure 2.9 can be drawn using Equation 3.2 . For example,
⊛
Example 3.2 Discrete and Continuous Random Variables A packaging machine fills containers with bolts. If the machine does not count the number of bolts it places in the containers, then that number is approximate.
The number of bolts in each package is a discrete random
variable because only integer values are allowed. On the other hand, the diameter of each bolt is only approximately the same. The diameter is a continuous random variable because the diameter can take on any value over a range.
⊛
© 2005 by Taylor & Francis Group, LLC
CHAPTER 3. RANDOM VARIABLE MODELS
52
Figure 3.2: The discrete cumulative distribution function.
3.2 The
Probability Density Function
probability density function presents the same information contained in
the probability distribution function, but in a more useful form. Assuming 3 continuity of the distribution, the density function is defined as
fX (x)
fX (x) =
dFX (x) . dx
Alternatively, by integrating both sides and rearranging, the distribution can be related to the density,
FX (x) = Pr (X ≤ x) =
x
−∞
fX (ξ)dξ.
(3.3)
fX (x) is analogous to the individual probpX (xi) of a discrete random variable.
The probability density function abilities
Equation 3.3 provides a useful interpretation of the density function: the probability that a continuous random variable X has a value less than or equal to the number x is equal to the area under the density function for 3 The
distribution function does not have to b e a continuous function.
In many in
stances it may have discrete jumps where a finite probability exists for a certain realization. It is just easier to work initially with a continuous function.
© 2005 by Taylor & Francis Group, LLC
3.2. PROBABILITY DENSITY FUNCTION
53
values less than x. Similarly, for arbitrary x1 and x2 , the probability that x1 < X ≤ x2 is given by the integral, x2 Pr(x1 < X ≤ x2 ) = f (x)dx, (3.4) x1 as shown in Figure 3.3, and for small dx = x2 − x1 , the probability can be
Figure 3.3: The probability that
x1 < X ≤ x2
is given by the shaded area
under the probability density function. estimated as
Pr(x1 < X ≤ x2) ≃ fX (x) dx. The probability of any specific
realization equals zero, that is,
Pr(X = λ) = Note the important
λ
normalization
∞
−∞
λ
f (x)dx = 0.
property of the probability density,
f (x)dx = 1,
(3.5)
signifying that the density function is representative of all possible outcomes or realizations of the random variable. The area under the density function is
normalized
to 1. Since probability is numerically in the range 0 to 1, the 4 function: . It is
density function must be a positive semidefinite
f (x) ≥ 0
important to recall that the random variable has a static property. That is, the shape of the density function does not change with time. the density function is timedependent, the variable is called a
stochastic process.
Where
random
or
This more advanced topic will be discussed in the second
part of this book, beginning with Chapter 5. 4 A positive definite function is one that has all values greater than zero. semidefinite, then it may also b e equal to zero.
© 2005 by Taylor & Francis Group, LLC
If it is p ositive
CHAPTER 3. RANDOM VARIABLE MODELS
54
a and b, where b > a, b a Pr(a < X ≤ b) = f (x)dx − f (x)dx −∞ −∞ = F (b) − F (a).
For two values
X may assume discrete values X is then both discrete and continuous.
In some special cases, the random variable as well as all values in some interval.
Pr (a < X ≤ b) can then be written as b Pr (a < X ≤ b) = f (x)dx + p (xi) . a a 10) = 0.025.
⊛
© 2005 by Taylor & Francis Group, LLC
10
hr is given by
we asked “greater than,” the
CHAPTER 3. RANDOM VARIABLE MODELS
66
Which density functions are of use in engineering applications ?
To be
able to design a product such as a structure or a machine, one needs to be able to understand the behavior of materials, the characteristics of a vibrating system, and the external forces. Usually the largest uncertainties occur with load definition. Even so, in practice, we expect that the graphs of probability densities will have most of their area about the mean value, that is, with a small variance. Sometimes in engineering, we only know the able.
high/low values of a vari
In this instance all intermediate values are equally probable.
This
property leads to the uniform probability density. Other times our experience tells us that parameter values significantly different from the mean can happen, even if these are unlikely. This characteristic implies the Gaussian density, studied in Section 3.4.3. Data from testing and design experience helps considerably in the choice of the most physically realistic probability density function. Section 3.4 provides the details.
3.4
Useful Probability Densities
It turns out that a handful of density functions are sufficient for probabilistic modeling in most engineering applications. Here, five of these are discussed: the uniform, exponential, normal or Gaussian, lognormal, and the Rayleigh density functions.
3.4.1
The Uniform Density
uniform density is a good model for a variable with known upper and equally likely values within the range. From Figure 3.12 it can be seen that for any range ∆x, the area under the density curve (a horizontal line ) is the same. The lefthand side of Equation 3.4 must be The
lower bounds, and
a constant so that there is equal probability for the variable to be in any range. Suppose then that value in the interval
X
is a continuous random variable that can have any
[a, b], where both a and b are finite.
1( − )
If the probability
density function is given by
f (x) = then
X
is
/b
0,
a,
a≤x≤b
otherwise,
uniformly distributed. The probability distribution function for a
© 2005 by Taylor & Francis Group, LLC
3.4. USEFUL PROBABILITY DENSITIES
67
Figure 3.12: Three examples of a uniform density function.
uniformly distributed random variable is
F (x) = Pr (X ≤ x) = 0,
x
−∞
f (s)ds
= (x − a)/ (b − a) , 1,
x 20 N is given by
the mean is value,
Pr (P > 20) =
25
20
1 dp = 1 . 3 15
The variance is evaluated as
σ2P = E {P 2} − µ2P = 18.75 N2, and the coefficient of variation is then
δ=
σP = 4.3 = 0.25, µP 17.5
or 25%. This is a relatively large scatter about the mean value. In engineering applications, coefficients of variation greater than 15% or, 0.15, imply a need for further data gathering.
⊛ Example 3.11 Quadratic Density P is disfP (p) = αp2, also with 10 ≤ P ≤ 25. straightforward to find α = 0.00021,
For comparison, suppose that instead of a uniform density, tributed according to a quadratic law, Following the above procedure, it is
µP = 19.98 N, and
Pr (P > 20) = © 2005 by Taylor & Francis Group, LLC
25
20
0.00021p2dp = 0.53,
3.4. USEFUL PROBABILITY DENSITIES
69
which makes sense since much more of the area under the density function is located near the upper end of the range. Here, the variance is
σ2P = 6.76 N2,
a much smaller value than for the
uniform density, and the coefficient of variation is
δ = 0.13 or 13%, again
signifying that the spread of values is much smaller for the quadratic than the uniform.
Figure 3.13 is a sketch of the quadratic superimposed on
the uniform, providing confirmation that for the quadratic density a large fraction of the area is clustered at the extreme values of
⊛
Figure 3.13:
3.4.2
fP (p) = 0.00021p2 and fP (p) =
1 15 for
P.
10 ≤ p ≤ 25.
The Exponential Density
For mechanical reliability,
6
the exponential function is most commonly used
to estimate failure times. The failure density is given by
f (t) = λe−λt , where
λ > 0, t ≥ 0,
(3.13)
λ is a constant (failure) rate per unit time, and 1/λ is the mean (time λ = 1 in Figure 3.14.
to failure). A sample exponential density is plotted for
6
A good starting point for studying reliability is the book by B.S. Dhillon, Mechanical
Reliability: Theory, Models and Applications, AIAA, 1988. © 2005 by Taylor & Francis Group, LLC
CHAPTER 3. RANDOM VARIABLE MODELS
70
Figure 3.14: Exponential distribution with λ = 1.
Example 3.12 Time to Failure A pump is known to fail according to the exponential density with a mean of 1000 hr. Then λ = 1/1000. Suppose that a critical mission requires the pump to operate 200 hr. Calculate the failure probability. Solution The probability density function f (t) tells us how the failure time is distributed. As long as the failure time is greater than 200 hr, the pump’s performance is satisfactory. The failure probability is then the probability that the failure time is less than or equal to 200 hr,
Pr (failure) = Pr (t ≤ 200) or F (200) , where F (t) is the cumulative distribution function corresponding to the density f (t) . For an exponential density, the probability distribution function is
F (t) = 1 − e−λt .
For this example,
F (200) = 1 − e−200/1000 = 0.1813. The probability that the pump will fail during the first 200 hr is 0.1813 or 18.13%. Knowing this value will help in making the decision whether a backup pump needs to be on hand.
⊛
© 2005 by Taylor & Francis Group, LLC
3.4. USEFUL PROBABILITY DENSITIES 3.4.3
71
The Normal (Gaussian) Density
Many physical variables are assumed to be governed by the normal or Gaussian density. See Figure 3.15. There are two reasons for this: the broad applicability of the central limit theorem,7 and the Gaussian is mathematically tractable and tabulated.
Figure 3.15: The normal or Gaussian density function. A random variable governed by the Gaussian density has the probability density function,
1 x−µ 1 f (x) = √ exp − 2 σ σ 2π
2
,
− ∞ < x < ∞,
where the meaning of parameters µ and σ are found by taking the expected 7 The central limit theorem states that under very general conditions, as the number of variables in a sum becomes large, the density of the sum of random variables will approach the Gaussian regardless of the individual densities. Examples of variables that arise as the sum of a number of random effects, where no one effect dominates, are noise generated by falling rain, the effects of a turbulent boundary layer, and the response of linear structures to a turbulent environment. Many naturally occurring physical processes approach a Gaussian density.
© 2005 by Taylor & Francis Group, LLC
CHAPTER 3. RANDOM VARIABLE MODELS
72
values8 of X and X 2 , respectively, ∞ 2 1 (σy + µ)e−y /2 dy = µ E {X } = √ 2π −∞ ∞ 2 1 2 (σy + µ)2 e−y /2 dy = µ2 + σ 2 . E {X } = √ 2π −∞ Thus, the mean value is is
σ.
µ and, using Equation 3.10, the standard deviation
Note that the Gaussian density extends from
−∞ to ∞, and therefore,
cannot represent any physical variable except approximately. Since there are no physical parameters that can take on all possible values on the real
number line, we may rightly wonder how good a model is the Gaussian for physical processes. But the approximation, in many instances, turns out to be very good. For example, consider a positivedefinite random variable
X
δ = 0.20, = 5σ. How significant is the area under the density function in the negative X region? Integrating numerically for x < 0, one finds an area of approximately 24 10−8 , a negligible probability for most purposes. The or
that is modeled as a Gaussian with coefficient of variation
µ
×
suitability of the Gaussian model depends on the application, and how much the tails extend into physically forbidden regions. When it is not possible to accept any negative values,
9
the analyst some
times resorts to the truncated Gaussian,
f (x) =
σ
and zero elsewhere.
A
√
2π
(If
exp
x1
− x0 )2 − 2σ 2 , (x
0
≤ x1 ≤ x ≤ x2,
→ −∞ and x2 → ∞, then A → 1, and X E {X } = x0 and V ar(X ) = σ 2 .
becomes a Gaussian random variable with See Figure 3.16.)
For ease in applications, the Gaussian variable formed so that the resulting variable
S=
S
X
is sometimes trans
is zero mean with unit variance,
X − µX , σX
resulting in the standard normal density,
fS (s) =
√1 e−s2/2. 2π
=( − )
8 Make
=
use of the transformation of variables: y x µ /σ and note that dx σdy . is especially true in reliability calculations where the probabilities of failure may be very small, even on the order of −8 , and extra care must be taken to ensure that the density function is suitable. 9 This
© 2005 by Taylor & Francis Group, LLC
10
3.4. USEFUL PROBABILITY DENSITIES
73
Figure 3.16: The truncated Gaussian for
The probability distribution is then
FS (s) = Pr (S ≤ s) = where
s
1.5 ≤ X ≤ 4.5.
√1 e−s2/2ds,
−∞ 2π
FS (s) is often denoted as Φ (s) and is tabulated. See Table 4.7. X is within k standard deviations
Finally, compute the probability that
of its mean. This probability is expressed as
Pr (µX − kσX < X ≤ µX + kσX ) = Pr −k < X − µX ≤ k σX = Pr (−k < S ≤ k) ,
which is equal to the probability that the standard normal random variable
S
is between
−k and k. Note that the probability depends only on k and
σX . In Figure 3.17, the shaded areas under the µX and σX are different. It will be shown in Example 3.14 that the probabilities that X is within 1σ X , 2σ X , and 3σ X
is independent of
µX
and
curves are equal even though
are 0.6827, 0.9545, and 0.9973, respectively, and that 95% of the area under
the normal distribution lies within
1.96σX
of the mean. See Figure 3.18.
Gaussian Table When utilizing the Gaussian model, it has become the practice to utilize tables of values of the
Standard Normal Cumulative Distribution Function.
© 2005 by Taylor & Francis Group, LLC
74
CHAPTER 3. RANDOM VARIABLE MODELS
Figure 3.17: Equal shaded areas.
Figure 3.18: Area under the normal density.
© 2005 by Taylor & Francis Group, LLC
3.4. USEFUL PROBABILITY DENSITIES
75
Utilizing the table is generally easier than integrating the Gaussian function over a range of values. The tabulated probability values are for the standard normal function. Suppose the normal random variable of interest mean
µ
and a standard deviation
σ,
in shorthand
transformed into the standard normal variable
S=
X −µ . σ
N (µ, σ).
X
has a
This can be
S by using the relation,
X ≤ x is probabilistically equivalent to the event ≤ (x − µ)/σ, for which values are tabulated. A common notation used for this is (3.14) Pr S ≤ x − µ = Φ x − µ , Therefore, the event that
that
S
where
Φ(−a) = 1 − Φ(a).
σ
The argument of
Φ
σ
is the number of standard
deviations above or below the mean value. The values of
Φ(s) are tabulated
in Section 4.7. This is called the Gaussian table or the standard normal table. Note that the numbers
0 to 9 across the table are the second digits
after the decimal.
Example 3.13 Reading from the Standard Normal Table Using the Standard Normal Table in Section 4.7, obtain probabilities
Pr(S ≤ −2.03), (ii) Pr (S ≤ 1.76) , and (iii) Pr (S ≥ −1.58) .
(i)
Solution
(i) From the standard normal table, Pr (S ≤ −2.03) = 0.0212. (ii) Pr (S ≤ 1.76) can be found using the fact that
Pr (S ≤ 1.76) = 1 − Pr (S < −1.76) . = 1 − 0.0392 = 0.9608. This result can be confirmed by looking up Table.
(iii)
Φ (1.76) in the Standard Normal
Pr (S ≥ −1.58) can be written as Pr (S ≥ −1.58) = 1 − Pr (S < −1.58) .
Pr (S ≤ −1.58) = 0.0570 from the Standard Normal Table. Pr (S ≥ −1.58) = 1 − 0.0570 = 0.9430.
⊛ © 2005 by Taylor & Francis Group, LLC
Therefore,
CHAPTER 3. RANDOM VARIABLE MODELS
76
Example 3.14 Probabilities Within k Standard Deviations Find the probabilities that
X is within 1σX , 2σX , and 3σX , respectively. σX
Also, show that 95% of the area under the normal density lies within 1.96 of the mean.
Solution
The probability that
lows,
X is within 1σX
is calculated as fol
Pr (µX − σX < X ≤ µX + σX ) = = = = =
Pr (−1 < S ≤ 1) Pr (S ≤ 1) − Pr (S < −1) Φ (1) − Φ (−1) 2Φ (1) − 1 0.6826. Similarly, the probabilities that X is within 2σ X and 3σ X are Pr (µX − 2σX < X ≤ µX + 2σX ) = 2Φ (2) − 1 = 0.9544, and
Pr (µX − 2σX < X ≤ µX + 2σX ) = 2Φ (3) − 1 = 0.9974. It can be shown that 95% of the area under the normal distribution lies within
1.96σX
of the mean,
Pr (µX − kσX < X ≤ µX + kσX ) = 0.95 2Φ (k) − 1 = 0.95 Φ (k) = 0.975. From the Gaussian table, k = 1.96, and thus about 96% of the area under the normal density lies within 1.96 σ X of the mean value.
⊛
Example 3.15 Cable Failure The cable of a suspension bridge is taken from a lot delivered by the manufacturer who guarantees that, statistically, the lot has the following
N (50 ksi, 5 ). If a load of 40 ksi acts on one cable, what is the probability of failure?
properties. The cables have a yield stress that is normal with ksi
Solution
Let
Y
be the yield event. The transformation is given as
S= © 2005 by Taylor & Francis Group, LLC
Y
− 50 . 5
3.4. USEFUL PROBABILITY DENSITIES
77
The probability of failure under the given load can be calculated in the following way,
Pr (Y < 40) = Pr S < 40 −5 50 = Φ(−2.0) = 0.0228.
This is the probability that the yield stress of the given cable is less than 40 ksi, that is, 2.28%.
⊛
Example 3.16 Rocket Ship Traffic Volume It is the year 2150. Permanent habitats, really small cities, have been growing on the Moon after the first groundbreaking in 2019. On Mars there are now 3 cities, the oldest one of which dates back to 2028. Populations are growing and people are being born on both the Moon and Mars. Interplanetary commerce has become a significant fraction of total human commerce. There is even talk about sending settlers to permanent sites on one of the Jovian Moons, but really, that sounds incredible. We are being asked to perform a traffic study for Lunar Spaceport Alpha. As the oldest and largest spaceport on the Moon, it needs to be expanded. Here is what we know, and what we need to estimate. On Lunar Spaceport Alpha, the current volume and blastoffs) during the peak hour is
N (50, 10).
V
(rocket ship landings
The present capacity is
for a total of 80 operations per hour. If there are more than 80 operations, there is a backlog of blastoffs and rockets entering the lunar space must orbit until clearance is granted.
(i) Find the current probability of congestion.
Since the volume is a
normal random variable, we have the option of utilizing the normal density function, or using the standard normal table.
Let us do both.
For the
normal density function, the required probability is given by
2 ∞ exp − 1 V σ− µ dV Pr (V > 80) = √1 2 σ 2π 80 ∞ V − 50 2 1 1 √ dV exp − = 10 2 10 2π 80 = 1.3499 × 10−3 . This is an extremely small number. If we are to use the standard normal
© 2005 by Taylor & Francis Group, LLC
CHAPTER 3. RANDOM VARIABLE MODELS
78
table, we need to transform as follows,
S=
V
− µ = V − 50 .
σ
10
Then,
Pr (V > 80) = 1 − Pr (V < 80) − 50 = 1 − Pr S < 80 10 = 1 − Pr (S < 3.0) = 1 − Φ (3.0) , where from the standard normal tables
Φ(3.0) = 0.9987. This is approxi
mately the same number we obtained above. Note that the argument of the function
Φ is in reality the number of standard deviations above or below
the mean value. Therefore, currently congestion is not a problem. But wait!
(ii) It is estimated that the mean volume will increase by 5% of the current volume per year while the coefficient of variation will stay the same. Perform the same computations as above to estimate the probability of congestion in 10 years time assuming that capacity is not increased.
δ = σ/µ = 10/50 = 0.2. The mean volume 50 + (0.05 × 50) × 10 = 75.0. The standard deviation will be σ 10 = δµ10 = 0.2 × 75 = 15.0, that is, V10 = N (75, 15). Now we can The coefficient of variation is
in 10 years time is
repeat the calculations of part one. First, use the density function,
∞ V − 75 2 1 1 dV = 0.36944, exp − Pr(V10 > 80) = √ 15 2 15 2π 80 and then using the standard normal tables,
Pr (V10 > 80) = 1 − Pr (V10 < 80) 80 − 75 = 1 − Pr S10 < 15 = 1 − Φ (0.33) = 1 − 0.6293 = 0.3707, which is approximately the same, the difference being due to the roundoff in the table lookup.
(iii) Finally, since such a congestion probability is too large, find the required capacity in 10 years time such that the present service conditions are maintained. Let C10 be the necessary capacity after ten years. It is © 2005 by Taylor & Francis Group, LLC
3.4. USEFUL PROBABILITY DENSITIES
79
required that
Pr (V10 > C10 ) = 1.3499 × 10−3 C − 75 = 1 − Φ (3.0) 1 − Φ 1015 C10 − 75 = 3 .0 15 C10 = 120. Therefore, capacity must be increased to 120 landing and blastoff operations per hour in order to maintain the same low congestion rate.
⊛
Figure 3.19: Carl Friedrich Gauss (17771855).
Gauss At the age of seven, Carl Friedrich Gauss started elementary school, and his potential was noticed almost immediately. His teacher, Büttner, and his assistant, Martin Bartels, were amazed when Gauss summed the integers from 1 to 100 instantly, spotting that the sum was 50 pairs of numbers each pair summing to 101. In 1788 Gauss began his education at the Gymnasium with the help of Büttner and Bartels, where he learned High German and Latin. After receiving a stipend from the Duke of BrunswickWolfenbüttel, Gauss entered Brunswick Collegium Carolinum in 1792. At the academy Gauss independently discovered Bode’s law, the binomial theorem, and the arithmeticgeometric mean, as well as the law of quadratic reciprocity and the prime number theorem. © 2005 by Taylor & Francis Group, LLC
80
CHAPTER 3. RANDOM VARIABLE MODELS In 1795 Gauss left Brunswick to study at Göttingen University. Gauss’ teacher there was Kaestner, whom Gauss often ridiculed. His only known friend among the students was Farkas Bolyai. They met in 1799 and corresponded with each other for many years. Gauss left Göttingen in 1798 without a diploma, but by this time he had made one of his most important discoveries, the construction of a regular 17gon by ruler and compasses. This was the most major advance in this field since the time of Greek mathematics and was published as Section VII of Gauss’ famous work, Disquisitiones Arithmeticae. Gauss returned to Brunswick where he received a degree in 1799. After the Duke of Brunswick had agreed to continue Gauss’ stipend, he requested that Gauss submit a doctoral dissertation to the University of Helmstedt. He already knew Pfaff, who was chosen to be his advisor. Gauss’ dissertation was a discussion of the fundamental theorem of algebra. With his stipend to support him, Gauss did not need to find a job so devoted himself to research. He published the book Disquisitiones Arithmeticae in the summer of 1801. There were seven sections, all but the last section, referred to above, being devoted to number theory. In June 1801, Zach, an astronomer whom Gauss had come to know two or three years previously, published the orbital positions of Ceres, a new “small planet,” which was discovered by G. Piazzi, an Italian astronomer on January 1, 1801. Unfortunately, Piazzi had only been able to observe 9 degrees of its orbit before it disappeared behind the Sun. Zach published several predictions of its position, including one by Gauss, which differed greatly from the others. When Ceres was rediscovered by Zach on December 7, 1801 it was almost exactly where Gauss had predicted. Although he did not disclose his methods at the time, Gauss had used his least squares approximation method. In June 1802 Gauss visited Olbers who had discovered Pallas in March of that year and Gauss investigated its orbit. Olbers requested that Gauss be made director of the proposed new observatory in Göttingen, but no action was taken. Gauss began corresponding with Bessel, whom he did not meet until 1825, and with Sophie Germain. Gauss married Johanna Ostoff on October 9, 1805. Despite Gauss having a happy personal life for the first time, his benefactor, the Duke of Brunswick, was killed fighting for the Prussian army. In 1807 Gauss left Brunswick to take up the position of director of the Göttingen observatory. Gauss arrived in Göttingen in late 1807. In 1808 his father died, and a year later Gauss’ wife Johanna died after giving birth to their second son, who was to die soon after her. Gauss was shattered and wrote to Olbers asking him to give him a home for a few weeks, to gather new strength in the arms of your friendship  strength for a life which is only valuable
© 2005 by Taylor & Francis Group, LLC
3.4. USEFUL PROBABILITY DENSITIES because it belongs to my three small children. Gauss was married for a second time the next year, to Minna, the best friend of Johanna, and although they had three children, this marriage seemed to be one of convenience for Gauss. Gauss’ work never seemed to suffer from his personal tragedy. He published his second book, Theoria Motus Corporum Coelestium in Sectionibus Conicis Solem Ambientium , in 1809, a major twovolume treatise on the motion of celestial bodies. In the first volume he discussed differential equations, conic sections, and elliptic orbits, while in the second volume, the main part of the work, he showed how to estimate and then to refine the estimation of a planet’s orbit. Gauss’ contributions to theoretical astronomy stopped after 1817, although he went on making observations until the age of 70. Much of Gauss’ time was spent on a new observatory, completed in 1816, but he still found the time to work on other subjects. His publications during this time include Disquisitiones generales circa seriem infinitam, a rigorous treatment of series and an introduction of the hypergeometric function, Methodus nova integralium valores per approximationem inveniendi, a practical essay on approximate integration, Bestimmung der Genauigkeit der Beobachtungen, a discussion of statistical estimators, and Theoria attractionis corporum sphaeroidicorum ellipticorum homogeneorum methodus nova tractata. The latter work was inspired by geodesic problems and was principally concerned with potential theory. In fact, Gauss found himself more and more interested in geodesy in the 1820s. Gauss had been asked in 1818 to carry out a geodesic survey of the state of Hanover to link up with the existing Danish grid. Gauss was pleased to accept and took personal charge of the survey, making measurements during the day and reducing them at night, using his extraordinary mental capacity for calculations. He regularly wrote to Schumacher, Olbers, and Bessel, reporting on his progress and discussing problems. Because of the survey, Gauss invented the heliotrope, which worked by reflecting the Sun’s rays using a design of mirrors and a small telescope. However, inaccurate base lines were used for the survey and an unsatisfactory network of triangles. Gauss often wondered if he would have been better advised to have pursued some other occupation but he published over 70 papers between 1820 and 1830. In 1822 Gauss won the Copenhagen University Prize with Theoria Attractionis ... together with the idea of mapping one surface onto another so that the two are similar in their smallest parts . This paper was published in 1825 and led to the much later publication of Untersuchungen über Gegenstände der Höheren Geodäsie (1843 and 1846). The paper Theoria combinationis observationum erroribus minimis obnoxiae (1823), with its © 2005 by Taylor & Francis Group, LLC
81
82
CHAPTER 3. RANDOM VARIABLE MODELS supplement (1828), was devoted to mathematical statistics, in particular to the leastsquares method. From the early 1800s Gauss had an interest in the question of the possible existence of a nonEuclidean geometry. He discussed this topic at length with Farkas Bolyai and in his correspondence with Gerling and Schumacher. In a book review in 1816 he discussed proofs that deduced the axiom of parallels from the other Euclidean axioms, suggesting that he believed in the existence of nonEuclidean geometry, although he was rather vague. Gauss confided in Schumacher, telling him that he believed his reputation would suffer if he admitted in public that he believed in the existence of such a geometry. In 1831 Farkas Bolyai sent to Gauss his son János Bolyai’s work on the subject. Gauss replied to praise it would mean to praise myself. Again, a decade later, when he was informed of Lobachevsky’s work on the subject, he praised its “genuinely geometric” character, while in a letter to Schumacher in 1846, states that he had the same convictions for 54 years indicating that he had known of the existence of a nonEuclidean geometry since he was 15 years of age (this seems unlikely). Gauss had a major interest in differential geometry, and published many papers on the subject. Disquisitiones generales circa superficies curva (1828) was his most renowned work in this field. In fact, this paper rose from his geodesic interests, but it contained such geometrical ideas as Gaussian curvature. The paper also includes Gauss’ famous theorema egregrium : If an area in E3 can be developed (i.e., mapped isometrically) into another area of E3, the values of the Gaussian curvatures are identical in corresponding points. The period 18171832 was a particularly distressing time for Gauss. He took in his sick mother in 1817, who stayed until her death in 1839, while he was arguing with his wife and her family about whether they should go to Berlin. He had been offered a position at Berlin University and Minna and her family were keen to move there. Gauss, however, never liked change and decided to stay in Göttingen. In 1831 Gauss’ second wife died after a long illness. In 1831, Wilhelm Weber arrived in Göttingen as physics professor filling Tobias Mayer’s chair. Gauss had known Weber since 1828 and supported his appointment. Gauss had worked on physics before 1831, publishing Über ein neues allgemeines Grundgesetz der Mechanik, which contained the principle of least constraint, and Principia generalia theoriae figurae fluidorum in statu aequilibrii, which discussed forces of attraction. These papers were based on Gauss’ potential theory, which proved of great importance in his work on physics. He later came to believe his potential theory and his method of least squares provided vital links between science and
© 2005 by Taylor & Francis Group, LLC
3.4. USEFUL PROBABILITY DENSITIES nature. In 1832, Gauss and Weber began investigating the theory of terrestrial magnetism after Alexander von Humboldt attempted to obtain Gauss’ assistance in making a grid of magnetic observation points around the Earth. Gauss was excited by this prospect and by 1840 he had written three important papers on the subject: Intensitas vis magneticae terrestris ad mensuram absolutam revocata (1832), Allgemeine Theorie des Erdmagnetismus (1839), and Allgemeine Lehrsätze in Beziehung auf die im verkehrten Verhältnisse des Quadrats der Entfernung wirkenden Anziehungs und Abstossungskräfte (1840). These papers all dealt with the current theories on terrestrial magnetism, including Poisson’s ideas, absolute measure for magnetic force and an empirical definition of terrestrial magnetism. Dirichlet’s principle was mentioned without proof. Allgemeine Theorie showed that there can only be two poles in the globe and went on to prove an important theorem, which concerned the determination of the intensity of the horizontal component of the magnetic force along with the angle of inclination. Gauss used the Laplace equation to aid him with his calculations, and ended up specifying a location for the magnetic South Pole. Humboldt had devised a calendar for observations of magnetic declination. However, once Gauss’ new magnetic observatory (completed in 1833, free of all magnetic metals) had been built, he proceeded to alter many of Humboldt’s procedures, not pleasing Humboldt greatly. However, Gauss’ changes obtained more accurate results with less effort. Gauss and Weber achieved much in their six years together. They discovered Kirchhoff’s laws, as well as building a primitive telegraph device that could send messages over a distance of 5000 ft. However, this was just an enjoyable pastime for Gauss. He was more interested in the task of establishing a worldwide net of magnetic observation points. This occupation produced many concrete results. The Magnetischer Verein and its journal were founded, and the atlas of geomagnetism was published, while Gauss and Weber’s own journal in which their results were published ran from 1836 to 1841. In 1837, Weber was forced to leave Göttingen when he became involved in a political dispute and, from this time, Gauss’ activity gradually decreased. He still produced letters in response to fellow scientists’ discoveries, usually remarking that he had known the methods for years but had never felt the need to publish. Sometimes he seemed extremely pleased with advances made by other mathematicians, particularly that of Eisenstein and Lobachevsky. Gauss spent the years from 1845 to 1851 updating the Göttingen University widow’s fund. This work gave him practical experience in financial © 2005 by Taylor & Francis Group, LLC
83
CHAPTER 3. RANDOM VARIABLE MODELS
84
matters, and he went on to make his fortune through shrewd investments in bonds issued by private companies. Two of Gauss’ last doctoral students were Moritz Cantor and Dedekind. Dedekind wrote a fine description of his supervisor: ... usually he sat in a comfortable attitude, looking down, slightly stooped, with hands folded above his lap. He spoke quite freely, very clearly, simply and plainly: but when he wanted to emphasize a new viewpoint ... then he lifted his head, turned to one of those sitting next to him, and gazed at him with his beautiful, penetrating blue eyes during the emphatic speech. ... If he proceeded from an explanation of principles to the development of mathematical formulas, then he got up, and in a stately very upright posture he wrote on a blackboard beside him in his peculiarly beautiful handwriting: he always succeeded through economy and deliberate arrangement in making do with a rather small space. For numerical examples, on whose careful completion he placed special value, he brought along the requisite data on little slips of paper. Gauss presented his golden jubilee lecture in 1849, fifty years after his diploma had been granted by Hemstedt University. It was appropriately a variation on his dissertation of 1799. From the mathematical community only Jacobi and Dirichlet were present, but Gauss received many messages and honors. From 1850 onwards Gauss’ work was again of nearly all of a practical nature although he did approve Riemann’s doctoral thesis and heard his probationary lecture. His last known scientific exchange was with Gerling. He discussed a modified Foucalt pendulum in 1854. He was also able to attend the opening of the new railway link between Hanover and Göttingen, but this proved to be his last outing. His health deteriorated slowly, and Gauss died in his sleep early in the morning of February 23, 1855. 3.4.4
The Lognormal Density
Sometimes it is important to strictly limit possible values of a parameter to the positive range.
In such instances, a likely choice is the lognormal
density, depicted in Figure 3.20.
Applications for the lognormal include
material strength, fatigue life, loading intensity, time to the occurrence of an event, and volumes and areas. A random variable probability density function if
fX (x) = 10
√ 1
xζ
π
2
exp
ln
X
− 1
ln
2
x−λ ζ
The probability density function of X
ample 4.1.
© 2005 by Taylor & Francis Group, LLC
X
has a lognormal 10 that is,
is normally distributed,
ln
2
,
0
< x < ∞,
(3.15)
is derived from the normal density in Ex
3.4. USEFUL PROBABILITY DENSITIES λ = E {ln X } is the mean value of ln X standard deviation of ln X .
where
85
and
ζ=
V ar {ln X } is the
Figure 3.20: The lognormal density. It is of interest to relate λ and ζ to the mean and standard deviation of X. Define Y = ln X. From the definition of lognormal, Y is a normal random variable with mean λ and standard deviation ζ, that is, Y = N (λ, ζ ). Solve for X by taking the exponential of both sides; X = exp(Y ). Define the following,
µX = E {X } V ar{X } = E {X 2 } − E 2 {X }. Therefore, using the property that
Y
is normal,
∞ y−λ 2 1 1 dy µX = E {exp(Y )} = exp(y) √ exp − 2 ζ ζ 2π −∞ ∞ y−λ 2 1 1 √ exp y − 2 ζ dy = −∞ ζ 2π
= exp(λ + 1 ζ 2), 2
and solving for
λ,
© 2005 by Taylor & Francis Group, LLC
1 2
λ = ln µX − ζ 2 .
CHAPTER 3. RANDOM VARIABLE MODELS
86
E{X 2 } = exp 2 λ + ζ 2 , and the variance is given by V ar{X } =
µ2X exp ζ 2 − 1 . Then, σX = + V ar{X }
= µ2X exp ζ 2 − 1 σ2 , ζ 2 = ln 1 + X µ2X
1 2 2 and λ = ln µX − 2 ln 1 + σ X /µX . For most engineering applications, the 2 2 ratio σ X /µX is on the order of 0.1. Given this, 1 ≫ σ X /µX , and using the 11 expansion for ln(1 + x), we have 2 2 ln 1 + σµX2 ≃ σµX2 ; X X σX , ζ≃ µX Similarly,
which is the coefficient of variation of
X;
the mean value of
mately equal to the coefficient of variation of
X.
Y
is approxi
Example 3.17 Cable Failure Revisited The cable of a suspension bridge studied in Example 3.15, demonstrating the use of the Gaussian density, is now assumed to have a yield stress that is lognormal with mean
50 ksi and standard deviation 5 ksi. If a load of 40
ksi acts on one cable, what is the probability of failure? ksi.
Solution
12
Let
N (λ, ζ ).
Let
Y = ln X,
X
be the yield stress with
where
Y
µX = 50
x = 40
ksi equals the probability that the yield stress is less
than the applied stress. Therefore, to find
Pr (X < 40) evaluate
1 ln x − λ exp − Pr (X < 40) = ζ 2 −∞ xζ 2π λ
σX = 5
From the figure it is clear that the area under the lognormal and
to the left of
where
ksi and
is normally distributed with a distribution
and
ζ
40
√1
2
are mean and the standard variation of
dx,
ln X
or
(3.16)
Y.
Rather
than integrating Equation 3.16, the Gaussian table can be used to evaluate the integral. The probability that that
Y
is less than
1 1 The 1 2 ksi
X
is less than 40 equals the probability
ln 40, Pr (X < 40) = Pr (Y < ln 40) . x x − x2 /2 + x3 /3 − ..., −1 < x ≤ 1.
expansion is ln(1 + ) = is shorthand for kips per square inch, and kip is one thousand.
© 2005 by Taylor & Francis Group, LLC
3.4. USEFUL PROBABILITY DENSITIES To evaluate mation,
Pr (Y
87
< ln 40) using the Gaussian tables, use the transforS = Y −ζ λ ,
where
λ = ln µX − 12 ζ 2
ζ = ln 1 + δ2X . Substitute the values µX = 50 ksi and σX = 5 ksi, and find that
ln(1 + (5/50)2) = 9.9751 × 10−2 ksi,
ζ = and
λ = ln 50 − 0.5 × 9.9751 × 10−22 = 3.907 ksi.
Then, the probability of failure under the given load can be calculated from the Gaussian tables in the following way:
Pr (Y
40 − 3.907 < ln 40) = Pr S < 9ln.9751 × 10−2 = Φ(−2.18) = 0.0146.
Compare this with 0.0228 where the variable was assumed normal. Note that the probability is also equivalent to the area under the probability density function. Figure 3.21 shows the probability density function, and the shaded area is the probability that X is less than 40.
⊛
3.4.5
The Rayleigh Density
The Rayleigh density, like the lognormal, is also limited to strictly positivevalued random variables,
f (x) =
© 2005 by Taylor & Francis Group, LLC
x exp σ2
x2 − 2σ 2 ,
x ≥ 0.
CHAPTER 3. RANDOM VARIABLE MODELS
88
Figure 3.21: The lognormal density for The shaded area equals
Pr(X < 40).
ζ = 9. 9751 × 10−2
and
λ = 3.907.
The first and secondorder statistics can be derived to be
E {X } =
π
2
σ
E {X 2 } = 2σ 2 σX =
4 − π σ ≃ 0. 655σ. 2
Figure 3.22 shows the Rayleigh density with three different values of
σ.
As an example of where the Rayleigh density is a good model, consider a random oscillation, or any time dependent process that is governed by the Gaussian. This oscillation has peaks that are distributed randomly as well. The Rayleigh density is a good model for the distribution of these peak values.
Lord Rayleigh
John William Strutt, better known as Lord Rayleigh, suffered from poor health so his schooling at Eton and Harrow was disrupted and for four years he had a private tutor. He entered Trinity College, Cambridge, in 1861, graduating in 1864. His first paper in 1865 was on Maxwell’s electromagnetic theory. He worked on propagation of sound and, while on an excursion to Egypt taken for health reasons, Strutt wrote Treatise on Sound (18701871). In 1879 he wrote a paper on travelling waves. This theory has now developed into
© 2005 by Taylor & Francis Group, LLC
3.4. USEFUL PROBABILITY DENSITIES
Figure 3.22: The Rayleigh probability density function.
Figure 3.23: John William Strutt, Lord Rayleigh (18421919).
© 2005 by Taylor & Francis Group, LLC
89
CHAPTER 3. RANDOM VARIABLE MODELS
90
the theory of solitons. His theory of scattering (1871) was the first correct explanation of why the sky is blue. In 1873 he succeeded to the title of Baron Rayleigh. From 1879 to 1884 he was the second Cavendish professor of experimental physics at Cambridge, succeeding Maxwell. Then in 1884 he became secretary of the Royal Society. Rayleigh discovered the inert gas argon in 1895, work that earned him a Nobel Prize in 1904. He was awarded the De Morgan Medal of the London Mathematical Society in 1890 and was president of the Royal Society between 1905 and 1908. He became chancellor of Cambridge University in 1908.
3.4.6
Probability Density Functions of a Discrete Random Variable
So far we have discussed the probability density functions of continuous random variables.
Here, we consider such functions for discrete random
variables.
The Binomial Distribution ¯ The A or A. A occurs is po . If this experiment is repeated n times, what will be the probability that event A occurs exactly k times? Note that each repetition is assumed statistically independent and that probability po is constant between experiments. If X is the binomial variable, based on n 13 repetitions, the probability density function of X is given by
Perform an experiment where the outcome can be either event probability that event
pX (x = k) = where
n k
=
n k
pko (1 − po )n−k ,
(3.17)
n! = n (n − 1) · · · (n − k + 1) . k! k! (n − k)!
1 3 The reason why this distribution is called binomial is that the k th term in the summation is the same as the kth term in the binomial expansion of (po + (1 po ))n .
−
© 2005 by Taylor & Francis Group, LLC
3.4. USEFUL PROBABILITY DENSITIES For
91
n = 5 and po = 0.3, the probabilities are given by 5!
pX (x = 0) =
0!5! 5!
pX (x = 1) =
1!4! 5!
pX (x = 2) =
2!3! 5!
pX (x = 3) =
3!2! 5!
pX (x = 4) =
pX (x = 5) = The reader can verify that binomial distribution is
µX
=
=
4!1! 5!
5!0!
5
× 0.30 × (1 − 0.3)5 = 0.16807
× 0.31 × (1 − 0.3)4 = 0.36015
× 0.32 × (1 − 0.3)3 = 0.30870
× 0.33 × (1 − 0.3)2 = 0.13230
× 0.34 × (1 − 0.3)1 = 0.02835
× 0.35 × (1 − 0.3)0 = 0.00243.
k=0 pX (x = k) = 1. The expected value of the
n
kpX (x = k)
k=0 n
k
k=0
n! pk (1 − po )n−k . k! (n − k)! o
Since the first term equals zero,
µX Replace
k
with
µX
n
k
k=1
can be written as
n! pk (1 − po )n−k . k! (n − k)! o
l + 1,
=
=
Further replace
=
µX
n −1 l=0 npo
(l + 1)
n −1 l=0
n!
(l + 1)! (n
l+1 (1 − po )n−1−l
− 1 − l)! po
(n − 1)! pl (1 − po )n−1−l . l! (n − 1 − l)! o
n − 1 with m,
m
µX
=
npo
µX
=
npo ,
l=0
m! pl (1 − po )m−l l! (m − l)! o
.
because the summation of all the individual probabilities add to 1
© 2005 by Taylor & Francis Group, LLC
CHAPTER 3. RANDOM VARIABLE MODELS
92
Similarly, the variance is found to be aged to derive this value.
npo (1 − po ) . The reader is encour
Example 3.18 Binomial Distribution Suppose that 300 cars are being assembled every day in an assembly line. After they are assembled, each vehicle is tested. On average, 2% of the tested vehicles are rejected. The rejected vehicles are then taken apart and reassembled. Find the expected value and standard deviation of the number of vehicles that will pass the test. Also, what is the probability that all 300 vehicles will pass the test?
Solution
X be the number of vehicles that will pass the test. X
Let
has a binomial distribution. The expected value and the standard deviation are given by
µX σX and the probability that
npo = 294 npo (1 − po ) = 2.425,
= =
k
vehicles will pass the test is
pX (x = k) =
300
k
k
300 0.98 0.02
−k .
The probability that all 300 vehicles will pass the test is
pX (x = 300)
= =
300 300
2.3325
0.98
300
0.02
0
× 10−3,
a very small value. It is highly unlikely that all 300 vehicles will pass the test on any given day.
⊛
The Poisson Distribution Consider experiments similar to those that led to the binomial distribution. ¯ The probability A or A. po . If this experiment is repeated an infinite n times, what will be the probability that the
The outcome of the experiment can be either event that the event
A
occurs is
number of times instead of event
A
occurs exactly
k
times?
Each repetition is assumed statistically
po is constant between experiments. n → ∞ such that the expected value npo
independent and outcome probability Further, assume that
po
→
stays constant and equal to
© 2005 by Taylor & Francis Group, LLC
0
as
λ.
3.4. USEFUL PROBABILITY DENSITIES Replace
po
with
93
λ/n in Equation 3.17,
pbinomial X (x = k)
k λ n−k λ 1− = n (n − 1) · · · (n − k + 1) n k! n n−k k−1 1 1 k 1− λ (λ) · · · 1− = 1· 1− n k! n n k λ n λn k−1 1 1 · · · 1 − n k! n − λ 1 − n . = 1· 1− n 1
Let
n → ∞, pbinomial X (x = k) =
n→∞ n −λ . where limn→∞ (1 − λ/n) = e lim
pPoisson X (x = k) =
e−λ λk , k!
This is called the
e−λ λk , k!
Poisson distribution,
k = 0, 1, 2, 3, · · · ,
with mean and variance both given by
µX
=
σ2X
=
λ.
This shows that, in the limit, the Poisson distribution converges to the binomial distribution as
n → ∞, po
→
0
, and npo
=
λ.
This interpretation
of the Poisson distribution is useful when we approximate the value of the binomial distribution for a large value of
n.
may be in order. Sometimes,
λ
is written as
νt,
where
ν
n! in the binomial n, and approximations
The term
distribution can be quite large for a moderate value of
equals the average number of
occurrences per unit time or over a physical space, and such as time or physical space.
t
is a parameter
The Poisson distribution is particularly
useful in reliability theory. The subject of reliability is explored in Chapter 9.
SiméonDenis Poisson was born (June 21, 1781) in Pithiviers, France, and died (April 25, 1840) in Sceaux near Paris, France. Poisson’s parents were not from the nobility and it was becoming increasingly difficult to distinguish between the nobility and the bourgeoisie in France in the years prior to the Revolution. Nevertheless the French class system still had a major influence on his early years. The main reason for this was that the army was one of the few occupations where the nobility enjoyed significant institutional privileges and Poisson’s father had been a Poisson
© 2005 by Taylor & Francis Group, LLC
94
CHAPTER 3. RANDOM VARIABLE MODELS
Figure 3.24: SiméonDenis Poisson (17811840).
soldier. Certainly Poisson’s father was discriminated against by the nobility in the upper ranks of the army and this made a large impression on him. After retiring from active service he was appointed to a lowly administrative post that he held at the time that his son SiméonDenis was born. There is no doubt that SiméonDenis’ family put a great deal of energy into helping him to a good start in life. Now SiméonDenis was not the first of his parents’ children but several of his older brothers and sisters had failed to survive. Indeed his health was also very fragile as a child and he was fortunate to pull through. This may have been because his mother, fearing that her young child would die, entrusted him to the care of a nurse to bring him through the critical period. His father had a large influence on his young son, devoting time to teach him to read and write. SiméonDenis was eight years old when the Parisian insurrection of July 14, 1789, heralded the start of the French Revolution. As might be expected of someone who had suffered discrimination at the hands of the nobility, Poisson senior was enthusiastic about the political turn of events. One immediate consequence for his support of the Revolution was the fact that he became president of the district of Pithiviers, which is in central France, about 80 km south of Paris. From this position he was able to influence the future career of his son. Poisson’s father decided that the medical profession would provide a secure future for his son. An uncle of Poisson’s was a surgeon in Fontainebleau
© 2005 by Taylor & Francis Group, LLC
3.4. USEFUL PROBABILITY DENSITIES and Poisson was sent there to become an apprentice surgeon. However, Poisson found that he was ill suited to be a surgeon. First, he lacked coordination to quite a large degree, which meant that he completely failed to master the delicate movements required. Second, it was quickly evident that, although he was a talented child, he had no interest in the medical profession. Poisson returned home from Fontainebleau having essentially failed to make the grade in his apprenticeship and his father had to think again to find a career for him. Times were changing quite quickly in France, which was by this time a republic. No longer were certain professions controlled by the nobility as they had been and there had been moves toward making education available to everyone. In 1796 Poisson was sent back to Fontainebleau by his father, this time to enroll in the École Centrale. On the one hand he had shown a great lack of manual dexterity, but he now showed that he had great talents for learning, especially mathematics. His teachers at the École Centrale were extremely impressed and encouraged him to sit for the entrance examinations for the École Polytechnique in Paris. He proved his teachers right, for although he had far less formal education than most of the young men taking the examinations he achieved the top place. Few people can have achieved academic success as quickly as Poisson did. When he began to study mathematics in 1798 at the École Polytechnique, he was therefore in a strong position to cope with the rigors of a hard course, yet overcome the deficiencies of his early education. There were certainly problems for him to overcome for he had little experience in the social or academic environment into which he was suddenly thrust. It was therefore to his credit that he was able to undertake his academic studies with great enthusiasm and diligence, yet find time to enjoy the theater and other social activities in Paris. His only weakness was the lack of coordination that had made a career as a surgeon impossible. This was still a drawback to him in some respects, for drawing mathematical diagrams was quite beyond him. His teachers Laplace and Lagrange quickly saw his mathematical talents. They were to become friends for life with their extremely able young student and they gave him strong support in a variety of ways. A memoir on finite differences, written when Poisson was 18, attracted the attention of Legendre. However, Poisson found that descriptive geometry, an important topic at the École Polytechnique because of Monge, was impossible for him to succeed with because of his inability to draw diagrams. This would have been an insurmountable problem had he been going into public service, but those aiming at a career in pure science could be excused the drawing requirements, and Poisson was not held back. In his final year of study he wrote a paper on the theory of equations
© 2005 by Taylor & Francis Group, LLC
95
96
CHAPTER 3. RANDOM VARIABLE MODELS and Bezout’s theorem, and this was of such quality that he was allowed to graduate in 1800 without taking the final examination. He proceeded immediately to the position of répétiteur in the École Polytechnique, mainly on the strong recommendation of Laplace. It was quite unusual for anyone to gain a first appointment in Paris, most of the top mathematicians having to serve in the provinces before returning to Paris. Poisson was named deputy professor at the École Polytechnique in 1802, a position he held until 1806 when he was appointed to the professorship at the École Polytechnique that Fourier had vacated when he had been sent by Napoleon to Grenoble. In fact Poisson had little time for politics; his energies were directed to support mathematics, science, education and the École Polytechnique. When the students at the École had been about to publish an attack on Napoleon’s ideas for the Grand Empire in 1804, Poisson had managed to stop them, not because he supported Napoleon’s views but rather because he saw that the students would damage the École Polytechnique by their actions. Poisson’s motives were not understood by Napoleon’s administration, however, and they saw Poisson as a supporter, which did his career no harm at all. During this period Poisson studied problems relating to ordinary differential equations and partial differential equations. In particular he studied applications to a number of physical problems such as the pendulum in a resisting medium and the theory of sound. His studies were purely theoretical, however, for as we mentioned above, he was extremely clumsy with his hands: Poisson was content to remain totally unfamiliar with the
vicissitudes of experimental research. It is quite unlikely that he ever attempted an experimental measurement, nor did he try his hand at drafting experimental designs.
His first attempt to be elected to the Institute was in 1806 when he was backed by Laplace, Lagrange, Lacroix, Legendre, and Biot for a place in the Mathematics Section. Bossut was 76 years old at the time and, had he died, Poisson would have gained a place. However Bossut lived for another seven years so there was no route into the mathematics section for Poisson. He did, however, gain further prestigious posts. In addition to his professorship at the École Polytechnique, in 1808 Poisson became an astronomer at Bureau des Longitudes. In 1809 he added another appointment, namely that of the chair of mechanics in the newly opened Faculté des Sciences. In 1808 and 1809 Poisson published three important papers with the Academy of Sciences. In the first, Sur les inégalités des moyens mouvement des planètes, he looked at the mathematical problems that Laplace and Lagrange had raised about perturbations of the planets. His approach to these problems was to use series expansions to derive approximate solutions. This was typical of the type of problem that he found interesting. Libri
© 2005 by Taylor & Francis Group, LLC
3.4. USEFUL PROBABILITY DENSITIES ... he especially liked unresolved questions that had been treated by others or areas in which there was still work to be done. In 1809 he published two papers, the first, Sur le mouvement de rotation de la terre, and the second, Sur la variation des constantes arbitraires dans les questions de méchanique, were a direct consequence of developments
wrote:
in Lagrange’s method of variation of arbitrary constants, which had been inspired by Poisson’s 1808 paper. In addition he published a new edition of Clairaut’s Théorie de la Figure de la Terre in 1808. The work had been first published by Clairaut in 1743 and it confirmed the NewtonHuygens belief that the Earth was flattened at the poles. In 1811 Poisson published his twovolume treatise, Traité de Mécanique, which was an exceptionally clear treatment based on his course notes at the École Polytechnique. Malus was known to have a terminal illness by 1811 and his death would leave a vacancy in the physics section of the Institute. The mathematicians, aiming to have Poisson fill that vacancy when it occurred, set the topic for the Grand Prix on electricity so as to maximize Poisson’s chances. The topic for the prize was as follows: To determine by calculation and to confirm by experiment the manner in which electricity is distributed at the surface of electrical bodies considered either in isolation or in the presence of each other, for example, at the surface of two electrified spheres in the presence of each other. In order to simplify the problem, one needed to look only at an examination of cases where the electricity spread on each surface remains always of the same kind. Poisson had made considerable progress with the problem before Malus died on February 24, 1812. Poisson submitted the first part of his solution to the Academy on 9 March entitled Sur la distribution de l’électricité à la surface des corps conducteurs. As the mathematicians had intended, this was the deciding factor in Poisson being elected to the physics section of the Institute to replace Malus. It also marked a move away from experimental research toward theoretical research in what was considered to constitute physics, and in this the Institute was following the lead given by Laplace. Poisson continued to add various responsibilities to his already busy life. In 1815 he became examiner for the École Militaire and in the following year he became an examiner for the final examinations at the École Polytechnique. It is remarkable how much work Poisson put into his research, his teaching, and into playing an ever increasingly important role in the organization of mathematics in France. When he married Nancy de Bardi in 1817 he found that family life put yet another pressure on him, yet somehow he survived the pressures, continuing to take on further duties. His research contributions covered a wide range of applied mathematics topics. Although he devised no innovative new theories, he made major contributions to fur
© 2005 by Taylor & Francis Group, LLC
97
98
CHAPTER 3. RANDOM VARIABLE MODELS ther develop the theories of others, often being the first to exhibit their real significance. We mention now just a few of the topics he studied after his election to the Academy. In 1813 Poisson studied the potential in the interior of attracting masses, producing results that would find application in electrostatics. He produced major work on electricity and magnetism, followed by work on elastic surfaces. Papers followed on the velocity of sound in gasses, on the propagation of heat, and on elastic vibrations. In 1815 he published a work on heat, which annoyed Fourier, who wrote: Poisson has too much talent to apply
it to the work of others. To use it to discover what is already known is to waste it ... . Fourier went on to make valid objections to Poisson’s arguments, which he corrected in later memoirs of 1820 and 1821.
In 1823 Poisson published on heat, producing results that influenced Sadi Carnot. Much of Poisson’s work was motivated by results of Laplace, in particular his work on the relative velocity of sound and his work on attractive forces. This latter work was not only influenced by Laplace’s work but also by the earlier contributions of Ivory. Poisson’s work on attractive forces was itself a major influence on Green’s major paper of 1828 although Poisson never seems to have discovered that Green was inspired by his formulations. In Recherchés sur la probabilité des jugements en matière criminelle et matière civile, an important work on probability published in 1837, the
Poisson distribution first appears. The Poisson distribution describes the probability that a random event will occur in a time or space interval under the conditions that the probability of the event occurring is very small, but the number of trials is very large so that the event actually occurs a few times. He also introduced the expression “law of large numbers.” Although we now rate this work as of great importance, it found little favor at the time, the exception being in Russia where Chebyshev developed his ideas. It is interesting that Poisson did not exhibit the chauvinistic attitude of many scientists of his day. Lagrange and Laplace recognized Fermat as the inventor of the differential and integral calculus; he was French after all while neither Leibniz nor Newton were! Poisson, however, wrote in 1831:
This [differential and integral] calculus consists in a collection of rules ... rather than in the use of infinitely small quantities ... and in this regard its creation does not predate Leibniz, the author of the algorithm and of the notation that has generally prevailed.
He published between 300 and 400 mathematical works in all. Despite this exceptionally large output, he worked on one topic at a time. Libri writes: Poisson never wished to occupy himself with two things at the same
time; when, in the course of his labors, a research project crossed his mind that did not form any immediate connection with what he was doing at © 2005 by Taylor & Francis Group, LLC
3.4. USEFUL PROBABILITY DENSITIES
99
the time, he contented himself with writing a few words in his little wallet. The persons to whom he used to communicate his scientific ideas know that as soon as he had finished one memoir, he passed without interruption to another subject, and that he customarily selected from his wallet the questions with which he should occupy himself. To foresee beforehand in this manner the problems that offer some chance of success, and to be able to wait before applying oneself to them, is to show proof of a mind both penetrating and methodical. Poisson’s name is attached to a wide variety of ideas, for example: Poisson’s integral, Poisson’s equation in potential theory, Poisson brackets in differential equations, Poisson’s ratio in elasticity, and Poisson’s constant in electricity. However, he was not highly regarded by other French mathematicians either during his lifetime or after his death. His reputation was guaranteed by the esteem that he was held in by foreign mathematicians who seemed more able than his own colleagues to recognize the importance of his ideas. Poisson himself was completely dedicated to mathematics. Arago reported that Poisson frequently said: Life is good for only two
things: to study mathematics and to teach it.
3.4.7 The
MomentGenerating Functions
momentgenerating function MX (t) of random variable X is very useful
because all the moments of
X
can be derived directly from this function. It
is defined as
MX (t)
= =
where
fX (x)
E {exp (t X )} ∞ exp (t x) fX (x) dx,
(3.18)
−∞
is the probability density function of
X,
and
MX (t)
is its
Laplace transform. Write the exponential function in its power series form,
exp (tx) =
∞
1 (tx)k , k=0 k!
and substitute into Equation 3.18 to find that
∞ 1 1 1 + t x + (t x)2 + (t x)3 + · · · fX (x) dx 3! 2! −∞ 1 = 1 + t E {X } + t2E X 2 + 1 t3E X 3 + 4!1 t4E X 4 · · · . 3! 2!
MX (t) =
© 2005 by Taylor & Francis Group, LLC
CHAPTER 3. RANDOM VARIABLE MODELS
100
All the moments can be evaluated using the appropriate derivative of the moment generating function,
k dk E X = k MX (t) dt
t=0
.
(3.19)
It is interesting to note that if experimental data can be used to estimate the moments of
in principle,
X, then Equation 3.19 can be used to find MX (t) and finally, by an inverse Laplace transform fX (x) can be recovered.
Example 3.19 MomentGenerating Function For
X
governed by the standard normal density,
E {X } = 0 E {X 2 } = 1, with all higherorder moments related to the first two moments by
E {X n} = 1 · 3 · · · · (n − 1) σn and zero for
for
n even,
n odd, the moment generating function is given by
1
1
MZ (t) = 1 + t2 σ2 + t4(3σ4 ) · · · 4! 2! 3 1 4 2 = 1+ t + t ··· 4! 1 2 = exp t . 2
2!
⊛
Characteristic Function A related function is the
characteristic function,
defined as
ΦX (t) = E {exp (itX )} , the Fourier transform of fX (x) . The characteristic function exists for all
probability density functions, whereas this is not true for the momentgenerating function. The moments can be derived using this function, as shown in the example below.
© 2005 by Taylor & Francis Group, LLC
3.5. TWO RANDOM VARIABLES
101
Example 3.20 Characteristic Function Using the definition of characteristic function,
ΦX (t) = 1 + (it) E {X } + 2!1 (it)2 E {X 2} + · · · .
Then,
dn 1 n E {X } = n n ΦX (t) . i dt t=0
For the Gaussian,
ΦX (t) = 1 + 2!1 (it)2 σ2 + 4!1 (it)4 (3σ4) · · · = exp − 12 t2 .
Therefore, for the (standard) Gaussian,
1 d exp − 1 t2 E {X } = 2 t=0 i dt d exp − 12 t2 = 1i dt t=0 1 1 2 = i −t exp − 2 t t=0 = 0. and
⊛ 3.5
1 d2 exp − 1 t2 2 t=0 i2 dt2 2 = − d 2 exp − 12 t2 dt t=0 1 12 2 = exp − t − t exp − 2 t2 2 t=0 = 1.
E {X 2} =
Two Random Variables
We have already introduced concepts that relate two random variables. These are the concepts of statistical independence and conditional probability. Earlier concepts and definitions need to be generalized so that it is
© 2005 by Taylor & Francis Group, LLC
CHAPTER 3. RANDOM VARIABLE MODELS
102
possible to address questions such as:
variable
How do the possible values of random
X affect the possible values of related random variable Y ? General
izations for two random variables are utilized to answer these questions since these permit exploration of the necessary generalizations without needlessly crowding the concepts with too much algebra. Consider the two random variables
distribution function is denoted by
X
and
Y.
The
joint cumulative
completely defines their probabilistic properties, and
FXY (x, y) = Pr (X ≤ x, Y
≤ y) .
This function defines the probability that random variable
X is less than or
equal to
The definition is
x and random variable Y
is less than or equal to
y.
valid for discrete and continuous random variables. Based on this definition, the following general statements can be made about
FXY (−∞, −∞) = 0 FXY (−∞, y) = 0 FXY (∞, y) = FY (y) As for a single random variable,
0.
FXY (x, y),
FXY (∞, ∞) = 1 FXY (x, −∞) = 0 FXY (x, ∞) = FX (x).
FXY (x, y) is nondecreasing and FXY (x, y) ≥
Generalizing from our study of one random variable, and of joint discrete variables, the
joint density function
variables is for small
for a continuous pair of random
dx and dy, approximately defined by
fXY (x, y)dx dy ≃ Pr (x < X ≤ x + dx, y < Y
≤ y + dy) ,
and exactly equal to
Pr (a < X ≤ b, c < Y ≤ d) =
d b
c a
fXY (x, y)dx dy.
(3.20)
The joint distribution function is then
FXY (x, y) =
y x
f (u, v)du dv. −∞ −∞ XY
Conversely,
fXY (x, y) =
∂2FXY (x, y) . ∂x ∂y
Equation 3.20 is the probability that two random variables are
neously
within a certain range, specifically, that
simulta
a < X ≤ b and c < Y
≤ d.
Such information is useful when considering problems that have two or more variables where it is necessary to establish how the values of one variable
affect those of the other variables. This topic is considered in Section 3.5.1.
© 2005 by Taylor & Francis Group, LLC
3.5. TWO RANDOM VARIABLES
103
For a process with two random variables, the volume under the density function must equal Suppose
X
and
1.
Y
are discrete random variables. There then exists a
joint probability mass function
pXY (x, y), defined as
pXY (x, y) = Pr(X = x, Y = y). The joint cumulative distribution function for these discrete random variables is given by the summation,
FXY (x, y) =
xi ≤x, yi ≤y
A conditional probability mass function Equation 2.7, as
pXY (xi , yi ) .
pX Y (xy)
is defined, referring to
pX Y (xy) = Pr (X = xY = y) p (x, y) . = XY pY (y) A
marginal probability mass function
total probability,
can be derived using the theorem of
Pr(A) = ni=1 Pr(ABi) Pr(Bi),
pX (x) =
= =
yi
yi
yi
Pr (X = xY = yi) Pr (Y = yi) Pr (X = x, Y = yi) pXY (x, yi ) .
If X and Y are statistically pY X (yx) = pY (y) , and
independent, then
pXY (x, y) = pX (x) pY (y) for all possible combinations of
pX Y (xy) = pX (x) , (3.21)
x and y.
Example 3.21 Discrete Marginal and Total Probabilities from Probability Mass Data Suppose data are obtained from a series of experiments on two parameters,
X and Y. For example, in the flow of atmosphere over a wing, X could
© 2005 by Taylor & Francis Group, LLC
CHAPTER 3. RANDOM VARIABLE MODELS
104
be the normalized flow velocity and
Y
the turbulence fluctuation length
scale. The data is tabulated as follows:
y \x 1 .0
25 7/30 0 1 .5 1/30 2 .0 0 2 .5 pX (x) [8/30]
27 1/30 5/30 6/30 1/30 [13/30]
29 0 1/30 1/30 7/30 [9/30]
X takes on values x = 25, 27, and 29, and Y
pY (y)
[8/30] [6/30] [8/30] [8 /30] =1
takes on the values
y = 1.0,
1.5, 2.0, and 2.5. Some sample joint probabilities are
1 Pr(X = 25, Y = 2.0) = pXY (25, 2.0) = 30 7. Pr(X = 29, Y = 2.5) = pXY (29, 2.5) = 30 The marginal probability pX (27) is found by summing the joint probabili
ties,
pX (27) = Pr(X = 27) = Pr(X = 27, Y = 1.0) + Pr(X = 27, Y = 1.5) + Pr(X = 27, Y = 2.0) + Pr(X = 27, Y = 2.5)
= 13 , 30
which is obtained by summing down the column under
pX (25) = 8/30
and
pX (29) = 9/30.
X = 27.
Similarly,
Note that the marginal probability
masses sum to 1, as they should since the complete probability space is covered. One can check whether
X
and
Y
are statistically independent by
applying Equation 3.21 to all the respective probability mass values, for example, from the table,
pXY (25, 1.0) = pX (25) pY (1.0) , and therefore the random variables are not statistically independent. Note that for statistical independence to hold, it must hold for
pXY (xi , yj ) = pX (xi ) pY (yj ) , for all (i, j ).
⊛ © 2005 by Taylor & Francis Group, LLC
3.5. TWO RANDOM VARIABLES
105
Example 3.22 Wind Speed Amplification due to Large Structures We have all experienced high wind speeds around large structures such as skyscrapers.
This amplification depends on the size and shape of the
structure as well as on the surrounding terrain, and atmospheric conditions such as temperature (gradients).
The possible speeds are modeled by a
probability density, in this case a joint density in the plane of ground level. Suppose the component speeds in two orthogonal directions are the random variables
U
V, and the joint probability density function is a decaying fUV (u, v) = C exp (−3 [u + v ]) , u ≥ 0, v ≥ u. This density and
and
exponential,
the range on the variables has been derived for a specific site.
Find the
C. Calculate the probability Pr(V < 3U ), and the probability that the maximum speed in both directions is less than 5. Solution The value of C is obtained by requiring the joint density to satisfy the relation u v fUV (u, v) dv du = 1. For this problem, ∞ ∞ C exp (−3 [u + v]) dv du = 1 value of
0
u
1 = 18. u exp (−3 [u + v]) dv du 0 To calculate the probability Pr(V < 3U ), the limits of the integration are needed, and can be found by sketching the lines u = v and v = 3u. See C = ∞∞
Figure 3.25. Integrate between these lines,
∞ 3u
1 18 exp (−3 [u + v]) dv du = . 2 0 u Similarly, the probability that both speeds are less than 5 is Pr(V < 3U ) =
Pr(U, V < 5) =
5 0
5
u
18 exp (−3 [u + v]) dv du
= exp (−30) − 2 exp (−15) + 1
≃ 1.0. ⊛
Continuing the study of joint random variables, define and understand
conditional densities. Recall
Pr(AB ) = Pr(AB)/ Pr(B). The conditional Pr(B). The conditional
probability is taken on a reduced sample space, here
density function is needed to find probabilities such as
Pr(a < X ≤ bY = y). © 2005 by Taylor & Francis Group, LLC
106
CHAPTER 3. RANDOM VARIABLE MODELS
Figure 3.25: Domain of integration bounded by
For continuous random variables
V = 3U
and
V = U.
Pr(Y = y) = 0, therefore, define the con
ditional density function to be
Pr(a < X ≤ bY = y) =
b
a
f (xy)dx,
where f (xy ) is sometimes written as fX Y and is the conditional density of X given Y. Given that the area under a region of the density function is a measure of the respective probability, then
f (xy )dx ≃ Pr(x < X ≤ dxY = y ) Pr(x < X ≤ dx, y < Y ≤ y + dy) = dylim →0 Pr( y < Y ≤ y + dy) f (x, y )dx dy f (x, y)dx = dylim →0 f (y)dy = f (y) f (x, y ) . f (xy ) = f (y ) Similarly,
(3.22)
f (y x) = f (x, y )/f (x). These expressions lead to the relation that
defines statistical independence in the following way.
f (x, y) = f (xy)f (y). If X and Y f (xy ) = f (x) and f (x, y ) = f (x)f (y ).
3.22 as
© 2005 by Taylor & Francis Group, LLC
are
Rewrite Equation
statistically independent,
then
3.5. TWO RANDOM VARIABLES
107
Another version of the total probability theorem can be obtained if the variable
B
is taken to be continuous. The summation becomes an integral,
found by rewriting Equations 2.9 and 2.10, leading to
fXY (x, y ) = fX Y (xy)fY (y ) ∞ fX (x) = fXY (x, y )dy,
−∞
fX (x) is called the marginal probability density function. ∞ fXY (x, y)dx. fY (y ) = −∞
where
(3.23)
Similarly,
Example 3.23 Marginal Densities Y
Given that the joint density of
X
and
Y
is uniform over the domain
≥ X 2 − 1, Y ≤ 0, as shown in Figure 3.26, find the density function and
both marginal densities.
Figure 3.26: Domain of integration bounded by
Solution
Y
≥ X 2 − 1 and Y ≤ 0.
Since the density function is uniform, it equals the inverse
of the area shown in the sketch. That is,
Area
and the joint density
=
fXY (x, y)
© 2005 by Taylor & Francis Group, LLC
1
−1 =
4 x2 − 1 dx = , 3
1/Area = 3/4. The marginal densities
CHAPTER 3. RANDOM VARIABLE MODELS
108
can now be evaluated,
fX (x) =
0
x2 −1
fXY (x, y ) dy
3 dy x− 4 = 34 − 34 x2, √y+1 fY (y) = √ fXY (x, y ) dx − y+1 √y+1 3 = √ 4 dx − y+1 3 = 2 (y + 1). =
X fXY (x, y) = fX (x)fY (y ). The conclusion is that
and
0
2 1
Y
are not statistically independent, that is,
⊛ Example 3.24 Two Geometrical Parameters Data taken from a manufacturing site concludes that two geometrical parameters primarily affect the design life. These two parameters,
W, are jointly distributed according to 2w2e−w(1+2h) , fWH (w, h) = 0,
Find the probability
w, h > 0 otherwise
H
and
.
Pr(W ≤ 0.1, H ≤ 2.0). Derive the marginal densities. Pr(H ≤ 2.0  0.1 < W ≤ 0.2).
Evaluate the probability
Solution
The joint density is plotted in Figure 3.27. Then,
Pr(W ≤ 0.1, H ≤ 2.0) =
0.1
= 1.0707 × 10−3.
0
2 0
2w2e−w(1+2h) dh dw
The marginal densities are obtained using Equation 3.23 for each variable,
fW (w) =
0
∞ ∞
fWH (w, h) dh
2w2e−w(1+2h) dh = we−w .
=
© 2005 by Taylor & Francis Group, LLC
0
3.5. TWO RANDOM VARIABLES
fWH (w, h) = 2w2 e−w(1+2h) ,
Figure 3.27:
The joint density function for random variables
w, h > 0.
Similarly,
fH (h) =
0
∞ ∞
109
w
fWH (w, h) dw
2w2e−w(1+2h) dw 4 . = (1 + 2h)3
=
0
Finally,
Pr(H ≤ 2.0  0.1 < W ≤ 0.2) = Pr(H ≤ 2.0 , 0.1 < W ≤ 0.2) Pr(0.1 < W ≤ 0.2) 2 0.2 f (w, h) dw dh = 0.02 0.1∞ WH 0.1 0 fWH (w, h) dh dw 2 0.2 2 −w(1+2h) 2w e dw dh = 0.02 0.1∞ 2 −w(1+2h) dh dw 0.1 0 2w e − 3 5.8828 × 10 = 0.45801. = 1.2844 × 10−2 © 2005 by Taylor & Francis Group, LLC
and
h:
CHAPTER 3. RANDOM VARIABLE MODELS
110
where Equation 3.23 is used.
Note that
W
and
H
are not statistically
independent.
⊛
3.5.1
Covariance and Correlation
Proceeding from models with one random variable to those with two requires the introduction of the concept of
correlation coefficient.
covariance, and the related parameter, the linear the relationship is
These are measures of how
between the two random variables. Consider the two random variables and
Y
E {XY } = If
X
X
with second joint moment,
and
Y
∞∞ xyf (x, y )dx dy. −∞ −∞ XY
(3.24)
are statistically independent, then the joint density function
can be separated into the product of the respective marginal densities,
fXY (x, y) = fX (x)fY (y),
and by Equation 3.24, the integrals can be sepa
rated, resulting in the relation
E {XY }
=
E {X }E {Y }.
This is a oneway
relation; the product of the densities implies the product of the expectations
E {XY } = E {X }E {Y } X and Y are not statistically independent. Suppose Z = X + Y, where the mean values and standard deviations of X and Y are given. Find the mean value and standard deviation of the derived variable Z by taking the expectation of the equation, but not the other way around. It is possible to have
even when
E {Z } = E {X + Y } = E {X } + E {Y } µZ = µX + µY . This is true regardless of whether To find
σZ
X
and
Y
are statistically independent.
requires a bit more algebra,
V ar {Z } = E {[Z − µZ ]2 } 2 = E [X + Y − µX − µY ]
E {X 2 } − µ2X + E {Y 2 } − µ2Y + 2E {XY } − 2µX µY = V ar {X } + V ar {Y } + 2E {XY } − 2µX µY .
=
Z, it is necessary to know the joint E {XY }. If X and Y are statistically independent, then E {XY } = E {X }E {Y } and To evaluate the variance of the sum, statistical properties embodied in
V ar {Z } = V ar {X } + V ar {Y } σ2Z = σ 2X + σ 2Y . © 2005 by Taylor & Francis Group, LLC
(3.25)
3.5. TWO RANDOM VARIABLES
µX
The
covariance
and
µY ,
111
is defined as the second joint moment about mean values
Cov {X, Y } = E {(X − µX )(Y
− µY )} = E {XY } − µX µY . (3.26) Note that if the variables are statistically independent, Cov {X, Y } = 0. The
correlation coefficient
is defined as the normalized (dimensionless) co
variance,
ρ=
Cov {X, Y } . σ X σY
(3.27)
X and Y aY , where a is a constant. Then,
To better understand the correlation coefficient, assume that
X
are linearly related by the equation
=
the variance is calculated to be
V ar {X } = E X 2 − E 2 {X } 2 2 − E 2 {aY } =E a Y
2 2 =a E Y − E 2 {Y } 2 = a V ar {Y } . Taking the square root of both sides, the standard deviations of are related by
σX
=
X
and
Y
a  σ Y .
Similarly, the covariance is calculated,
Cov {X, Y } = E {XY } − E {X } E {Y }
2 − E 2 {Y } =a E Y 2 = a V ar {Y } = aσ Y . Therefore, Equation 3.27 becomes
ρ=
aσ 2Y σX σ Y
=
aσY σX
=
a
a  .
If a > 0, the correlation coefficient ρ = +1. The random variables X and Y are in perfect positive correlation. If a < 0, the correlation coefficient ρ = −1, or X and Y are in perfect negative correlation.14
ρ may indicate strong correlation, but X and Y may be correlated by virtue of being related to some third variable. Also, if X and Y are indep endent, ρ = 0, as we see in the following example. The converse is not necessarily true. ρ = 0 indicates the absence of a linear relationship; a random or a nonlinear functional relationship between X and Y is still a 1 4 It is important to realize that a high value for
not direct
cause
and
effect
since
possibility.
© 2005 by Taylor & Francis Group, LLC
CHAPTER 3. RANDOM VARIABLE MODELS
112
Even though it is not shown here, it can be shown that these values,
ρ = +1
and
−1, are the extreme values of ρ. That is, −1 ≤ ρ ≤ 1. Figure
3.28 depicts representative correlations between data points for nonlinear and perfect linear relationships. Note that if the sample points for nonlinear relationships happen to be symmetrical around either the correlation coefficient is zero.
X = µX
or
Y
=
µY ,
Figure 3.28: Discrete data distributions for various correlation coefficients.
Example 3.25 Jointly Distributed Variables X and Y are jointly distributed according to the fXY (x, y) = 12 e−y , y > x , −∞ < x < ∞. See Figure 3.29.
Two random variables joint density
Compute the marginal densities and the covariance.
© 2005 by Taylor & Francis Group, LLC
3.5. TWO RANDOM VARIABLES
Figure 3.29: Joint probability density function,
x and −∞ < x < ∞. Solution
113
fXY (x, y) = 12 e−y
y>
The marginal densities are defined and evaluated:
∞
∞ 1
1 e−y dy = e−x, 2 2 x −∞ ∞ y 1 e−y dx = ye−y , fY (y) = fXY (x, y)dx = −∞ −y 2
fX (x) =
for
fXY (x, y)dy =
−∞ < x < ∞ y > 0.
Since the joint density is not equal to the product of the two marginal densities, the variables are not statistically independent. Covariance Equation 3.26 requires an evaluation of the second joint moment,
∞ y
1 xye−y dx dy −y 2 0 ∞ y ye−y xdx dy = 0, =1 2 0 −y
E {XY } =
and the respective mean values,
E {X } = and similarly,
∞
−∞
xfX (x)dx =
1 xex dx + ∞ 1 xe−x dx = 0, 0 2 −∞ 2
0
E {Y } = 2. Therefore, Cov {X, Y } = 0, not because the two E {XY } = E {X } E {Y } due to the
variables are independent, but because
© 2005 by Taylor & Francis Group, LLC
CHAPTER 3. RANDOM VARIABLE MODELS
114
particular numbers in the problem.
E {XY } can be equal to E {X } E {Y }
even when the two variables are not statistically independent.
⊛
Example 3.26 Statistics of Support Reaction of Simply Supported Beam
Figure 3.30: Simply supported beam loaded by random forces
F1 and F2.
Consider the simply supported beam in Figure 3.30. There are two loads
F1 at a distance of d1 from the left end and F2 at a d3 from the right end. The length of the beam is L = d1 + d2 + d3.
acting on the beam, distance of
Given are the following loading statistics:
µ1 = E {F1}
E {F12} − µ21 σ1 = µ2 = E {F2}
σ2 =
E {F22} − µ22.
Derive the statistics of the reactions at
A
and
that the forces are statistically independent.
Solution
B, RA
and
RB ,
assuming
From a static analysis of the free body of the beam as
suming the forces to be exactly known, the functional relation between the forces, the dimensions and the reactions are
F1 (d3 + d2) + F2d3 L L ( F1 + F2 ) d1 + F2d2 F2 (d1 + d2) + F1d1 . = RB = L L RA =
(F1 + F2 ) d3 + F1d2
=
While the forces are given as statistically independent, the reactions are dependent because they are both functions of the same parameters
© 2005 by Taylor & Francis Group, LLC
F1
3.5. TWO RANDOM VARIABLES and
115
F2. This can be shown by deriving the correlation coefficient. For this di are known exactly. Then,
problem it is also assumed that the distances
the mean values of the reactions are straightforwardly given by
µA =
(µ1 + µ2) d3 + µ1 d2 L
(µ + µ ) d + µ d µB = 1 2 1 2 2 . L The variances can be calculated for these sums of independent random variables using Equation 3.25,
1 (d + d )2 σ2 + d2σ2 L2 3 2 1 3 2 1 σ2B = 2 d21σ21 + (d1 + d2)2 σ22 . L σ2A =
The standard deviations are the positive square roots of the respective variances.
E {RARB },
To find the correlation coefficient first find the expression for
E {RARB } = E
[F1 (d3 + d2) + F2d3] [F2 (d1 + d2) + F1 d1]
L2
= L12 E 2d1d3 + d2d3 + d1d2 + d22 F1 F2 + d1 (d3 + d2 ) F12 + d3 (d1 + d2) F22 1
= 2 2d1d3 + d2 d3 + d1d2 + d22 E {F1F2} L + d1 (d3 + d2 ) E {F12} + d3 (d1 + d2) E {F22} = a1E {F1F2} + a2E {F12} + a3E {F22 },
ai are constant coefficients introduced to simplify the expression. We E {F1 F2} = E {F1 }E {F2}. (If F1 and F2 are not statistically independent, their correlation coefficient ρF1 F2 is needed.) Also,
where
make use of independence,
E {F12 } = σ21 + µ21 E {F22 } = σ22 + µ22. Then,
E {RARB } = a1µ1µ2 + a2 σ21 + µ21 + a3 σ22 + µ22 .
The covariance is given by
Cov {RA , RB } = E {RA RB } − µAµB , © 2005 by Taylor & Francis Group, LLC
CHAPTER 3. RANDOM VARIABLE MODELS
116
and the correlation coefficient can now be expressed as
ρRARB =
Cov {RA, RB } , σA σB
which can be used for numerical evaluation. Consider the case where
σ2 = σ ,
σ1 =
σ 2 2d2 + 2 d3 d2 + d2 1 2 2 3 2 2 2 σA = 2 (d3 + d2 ) σ + d3 σ = 2 L L σ 2 2d2 + 2 d1 d2 + d2 1 1 2 , σ2B = 2 d21 σ2 + (d1 + d2)2 σ2 = L2 L 2
and
L2 a1µ1 µ2 + a2 σ2 + µ21 + a3 σ2 + µ22 − µAµB . ρRARB = σ2 (2d23 + 2 d3 d2 + d22 ) (2d21 + 2 d1d2 + d22)
⊛ Example 3.27 Correlation Coefficient and Reliability
A number of examples have been drawn from the discipline of reliability, in part because it is very important in all aspects of engineering. Reliability is also a function of vibration characteristics, which depend on uncertainties in loading and material parameter values. Suppose that we have been tasked with the job of estimating the reliability of a component that is to be used in a particular stress environment. Assume that tests are performed on both component and loading to gather data. How do we proceed?
Solution
Define the strength of the component as
stress it experiences due to loading as
Y
X
psi and the
psi. The strength may be a yield
stress or an ultimate stress. Test a sufficient number of “identical” components in order to establish its strength probability density function, Similarly, loading data leads us to a loading or stress density function
fX (x). fY (y).
The strength is designed to exceed the stress for all but the most rare of cases, as one would expect. This is shown schematically in Figure 3.31. The shaded region in the figure represents the realizations where the
loading stress is greater than the component strength. This is a failure of
Pr ([X − Y ] ≤ 0). Define Pr (Z ≤ 0). The reliability of the component is then defined as R = 1 − Pr (Z ≤ 0), which is designed to be the component. The probability of failure equals
Z = X−Y
with the goal of finding
a very large number.
From Equations 3.26 and 3.27 we have
E {XY } = ρσX σY + µX µY , © 2005 by Taylor & Francis Group, LLC
3.6. CONCLUDING SUMMARY
117
Figure 3.31: Reliability schematic where shaded region is a measure of the probability of failure.
Z is V ar{Z } = V ar{X } + V ar{Y } − 2ρσX σY . Since strength and loading stress are uncorrelated, ρ = 0, and the variance of Z equals the variance of X plus the variance of Y , or
and using Equation 3.9, the variance of the new variable
σZ = σ2X + σ2Y . The mean value of
Z
is
µZ = µX − µY .
These are the mean and variance of the probability of failure. Given the density function
fZ (z ), the reliability of the component is R=1−
0
−∞
fZ (z)dz.
(3.28)
X and Y are Gaussian. Then it is known, and not proven Z is also Gaussian. In this instance, fZ (z ) is fully defined given µZ and σZ , and Equation 3.28 can then be evaluated.
Suppose that both here, that
⊛
3.6
Concluding Summary
In this chapter, the functions used to define random variables, their distribution functions, are defined. The probability distribution function or the
© 2005 by Taylor & Francis Group, LLC
CHAPTER 3. RANDOM VARIABLE MODELS
118
probability density function fully describe the probabilistic properties of the random variable. The mathematical expectation is defined, a function that is utilized to derive the moments of the random variable. Various specific densities are defined and worked with, showing how the expectation function can be utilized. Additional descriptors are introduced in order to be able to work with models of more than one random variable. It becomes necessary to account for how the variables interact with each other.
To
address such questions, the concept of correlation is introduced.
3.7
Problems
Section 3.1: Probability Distribution Function 1. Which of the following Figures 3.32
(a), (b), (c)
represent possible
probability distribution functions?
Figure 3.32: Possible distributions.
2. Show that the expected value of a constant equals that constant. 3. Suppose
X
is distributed as indicated in Figure 3.33.
All lines are
1 < x < 3. FX (x) algebraically and then find the following probabilities: Pr(X = 1/3) Pr(X = 1) Pr(X < 1/3) Pr(X ∼ 1/3) Pr(X < 1) Pr(X ≤ 1) Pr(1 < X ≤ 2) Pr(1 ≤ X ≤ 2) Pr(X = 1 or 1/4 < X < 1/2).
straight except for the exponential curve in the range Express
Section 3.2: Probability Density Function 4. For each of the following functions, state why they are valid or invalid probability density functions. If arbitrary constants are present, evaluate them. Sketch the probability density functions. Then, find the cumulative distribution functions and sketch.
© 2005 by Taylor & Francis Group, LLC
3.7. PROBLEMS
119
Figure 3.33: Distribution
FX (x).
fX (x) = 1/10, 0 ≤ x ≤ 10, (ii) fX (x) = c exp(−λx), x ≥ 0, (iii) fX (x) = c + 1/20, 0 ≤ x ≤ 10. (i)
Section 3.3: Mathematical Expectation 5. Find the expectation and variance of the following functions and data sets:
X, fX (x) = 1/3, 1 ≤ x ≤ 4, 2 (ii) X , fX (x) = 1/3, 1 ≤ x ≤ 4, (iii) 2, 2.5, 2.5, 4, 4.3, 4.9, 7, 10, 10, 11
(i)
6. Derive Equation 3.9. 7. For the random variable density function,
X,
we are given the following probability
f (x) =
c , x ≥ 10. x2
Derive the cumulative distribution function, functions. Find the expected value of of
X
F (x),
and sketch both
and the standard deviation
X. What is the coefficient of variation?
Is this a large number?
8. Two different probability density functions are recommended as models for a random variable that is representative of a particular material’s strength, in nondimensional units:
f (x) = c1 exp(−x), 0 ≤ x ≤ 15, f (x) = c2x, 0 ≤ x ≤ 15. © 2005 by Taylor & Francis Group, LLC
CHAPTER 3. RANDOM VARIABLE MODELS
120
Discuss how these models are different. Refer to their mean values, standard deviations, coefficients of variation.
Plot both functions.
Provide your recommendations as to which you think is a better physical model, and justify. 9. Materials and structures age with time. The rate of aging is variable and depends on many factors beyond the control of the designer. It is reasonable to say that this aging process is a random function of
How does an engineer define the aging of a material or a structure such as an aircraft? Second,
time. First, state a definition of aging:
can one use any of the probability functions we have studied to date, such as the cumulative distribution function or the probability density function, to model such an aging process? 10. The time duration
T
of a force acting on a structure is a random
variable having the probability density function,
2 at fT (t) = b 0
0 ≤ t ≤ 12 12 ≤ t ≤ 16
elsewhere.
(i) Determine the appropriate values for constants a and b. (ii) Calculate the mean value and variance of T. (iii) Calculate Pr(T ≥ 6). 11. A strong wind at a particular site may come from any direction be
θ = 0◦ and due North at θ = 90◦. The wind speed V can vary between 0 and 150 mph. (i) Sketch the sample space for wind speed and direction in one figure. (ii) Let A = {V ≥ 20 mph } B = {12 mph ≤ V ≤ 30 mph } C = {θ ≤ 30◦}. tween due East at
Identify the events
A, B, C, A in the sample space sketched in part (i).
(iii) Use new sketches to identify the following events: D = A∩C E = A∪B F = A ∩ B ∩ C. (iv) Are the events D
and
C
and
mutually exclusive?
© 2005 by Taylor & Francis Group, LLC
E
mutually exclusive? Are the events
A
3.7. PROBLEMS
121
12. A fiberoptic cable must be manufactured to the highest tolerances. An important parameter is cable diameter
D.
Testing of the manu
factured product finds the diameter to be normally distributed with
0 .1
a mean diameter of
0.02.
m with a coefficient of variation of
cable is considered unacceptable if its diameter is value, that is, more than
3%
A
off the mean
3% above or below the mean value.
What is
the probability of a cable being unacceptable? Sketch the probability density function and the unacceptable region. 13. Two cables are used to lift load carrying the load; cable
B
W.
Normally, only cable
is slightly longer than
participate in carrying the load. But if cable have to carry the full load until
A
0.30.
A B Pr(A) Pr(BA)
B
A
will be
so it does not
A breaks,
is replaced.
following information: The probability that cable The probability that
A,
then
B
will
We are given the
A will break is 0.02.
will fail if it has to carry the load by itself is
= cable A breaks = cable B breaks = 0.02 = 0.30 B would break only if A already broke.
(i) What is the probability that both cables will fail? (ii) If the load remains lifted, what is the probability that none of the cables have failed?
Section 3.4: Useful Probability Densities 14. Verify that the variance of a random variable governed by the binomial density equals
npo (1 − po ) .
15. In Example 3.12, another design offers two pumps in parallel in the pipe. As in the example, each has a mean time to failure of 1000 hr, governed by an exponential density. State any assumptions made, and calculate the probability that both pumps will survive the first 200 hr of operation.
Section 3.5: Two Random Variables 16. Consider again Example 3.26 of the simply supported beam, with
µ1 = 30, σ1 = 10
and
µ2 = 50, σ2 = 5.
Calculate the statistics
of the two reactions and the correlation coefficient between the two reactions.
© 2005 by Taylor & Francis Group, LLC
CHAPTER 3. RANDOM VARIABLE MODELS
122
17. Consider again Example 3.26.
Suppose the external forces are due
to related causes, and therefore are statistically dependent. Explain what differences occur in the derivation. 18. For the statistically dependent version of Example 3.26, derived in the previous problem, calculate the statistics of the two reactions and their correlation coefficient for and
E {F1F2 } = 75.
19. Random variables
X and Y
µ1 = 30, σ1 = 10 and µ2 = 50, σ2 = 5,
have the following joint density function:
f (x, y) = K exp[−(x + y)], Sketch the domain of
(i) find the
value of
x ≥ 0, y ≥ x.
X, Y on the xy plane, and do the following: K that makes this function a valid probability
density function,
(ii) find Pr(Y < 2X ), and (iii) find the probability that the maximum of X or Y is ≤ 4. In parts (ii) and (iii) sketch the domains of integration in the xy plane. 20. The random variables
X
and
Y
have the joint uniform probability
density function over the region bounded by the
x axis and the curve
1 − x2. Sketch this domain in the xy plane. Find the joint probability density function f (x, y ), and the marginal densities: f (x) and f (y ). Sketch the domains of integration. 21. Find the marginal density of
X, where
fXY (x, y) = 22. The random variables
X, Y
X and the conditional density of Y x + y < 1, x ≥ 0, y > 0
2 0
otherwise.
are jointly distributed according to
fXY (x, y) = a cos x,
(i) (ii) (iii) (iv) (v) (vi)
0 ≤ x ≤ π/2 0 ≤ y ≤ 1.
sketch this joint density function find the marginal density functions find find are
Pr(X ≤ π/4, Y ≤ 1/2) Pr(X > π/4, Y > 1/2)
X, Y
statistically independent?
what is the covariance
© 2005 by Taylor & Francis Group, LLC
Cov(X, Y )?
fX
and
fY
given
3.7. PROBLEMS
123
23. In Example 3.26, rederive the statistics assuming that the forces are
Then reduce these equations for the cases ρ12 = 1 and ρ12 = −1. Compare results and discuss the importance of the correla
correlated.
tion coefficient to the final results.
24. For Example 3.26, evaluate the statistics for the parameter values: Parameter
F1 F2 d1 d2 d3
Value
10 20 1 1 1
Do this for the uncorrelated statistics. Then for the correlated statistics of the previous problem. Discuss the importance of correlation to the final results.
© 2005 by Taylor & Francis Group, LLC
Chapter 4
Functions of Random Variables Most applications require engineers to be able to relate the values of one variable to the values of another. Sometimes there are many variables involved in this relation called a function. When one or more of the variables are randomly distributed according to particular densities, there is the need to derive the densities of the derived functions in terms of the densities that are given. For the simplest case of Y = g(X ), given fX (x) and the functional relationship g(X ) between the variables X and Y, find the density fY (y ) , or at least the mean value and standard deviation of Y. How much we are able to derive depends on the complexity of the function g(X ). In some instances, it is not possible to find the density fX (x) , but rather the statistical measures µX and σX . In such instances, the goal is to estimate µY and σY . In this chapter, general techniques for the exact and approximate evaluation of derived densities and derived approximate statistical measures are developed. 4.1
Exact Functions of One Variable
Given fX (x) and g(X ), where Y = g(X ), there is an interest in finding fY (y ) . A general relation can be derived based on the assumption that g(X ) is simple enough to allow for a calculation of the inverse X = g −1 (Y ) . The approach is best motivated using the graphical representation shown in Figure 4.1. Using the figure as a reference, the strategy is to find the probability Pr(y < Y ≤ y + dy) which, for sufficiently small dy, is approximately equal to fY (y) dy. In deriving fY (y) it is useful to find an equivalent statement 125
© 2005 by Taylor & Francis Group, LLC
126
CHAPTER 4. FUNCTIONS OF RANDOM VARIABLES
Figure 4.1: Transformation procedure schematic for a function of a random variable using the equation y = g(x).
© 2005 by Taylor & Francis Group, LLC
127
4.1. EXACT FUNCTIONS OF ONE VARIABLE
to Pr(y < Y ≤ y + dy). From the figure it is observed that there is an equivalence between the following events, {y < Y ≤ y + dy} = {x1 < X ≤ x1 + dx1} + {x2 + dx2 < X ≤ x2} + {x3 < X ≤ x3 + dx3} ,
where dx1 > 0, dx2 < 0, and dx3 > 0. Then,
Pr(y < Y ≤ y + dy) = Pr(x1 < X ≤ x1 + dx1 ) + Pr(x2 + dx2 < X ≤ x2) + Pr(x3 < X ≤ x3 + dx3),
where the righthand side equals the sum of the shaded areas in the figure, thus, fY (y) dy ≃ fX (x1 ) dx1 + fX (x2 ) dx2  + fX (x3 ) dx3 (4.1) is an approximation of the equality. To complete this derivation it is necessary to relate the increments dxi to the functional relation between X and Y. From the graph,
and, therefore,
g ′ (X ) ≡
dg dX
dy , ≡ dX
g ′ (xi ) dX X =xi = dy, since all the values of dy are identical. Let dxi ≡ dX X=xi dxi by dxi  in Equation 4.1 for generality. Then,
and replace each
f (x ) f (x ) f (x ) fY (y ) dy = X′ 1 dy + X′ 2 dy + X′ 3 dy. g (x3) g (x2) g (x1)
In general, for any number n of roots xi, the relation becomes fY (y ) =
n fX′ (xi) . i=1 g (xi )
Again, the use of this equation requires that g−1 (X ) exist. The roots xi are implicitly real since for application to probability, only the real roots have any significance. In addition, the physical statement of the problem sometimes rules out certain roots as being not physically feasible. The following are examples of this procedure for some special cases of g(X ) and fX (x) . Example 4.1 Lognormal Density
The lognormal density in Section 3.4.4 is revisited. Two random variables Y and X are related by Y = eX . The random variable X is normally distributed with mean λ and standard deviation ζ. Find fY (y) . © 2005 by Taylor & Francis Group, LLC
128
CHAPTER 4. FUNCTIONS OF RANDOM VARIABLES
Solution
The probability density function of X is given by
.
1 x−λ 1 fX (x) = √ exp − ζ 2 ζ 2π
2
The transformation function is g (X ) = eX . Solving for x that corresponds to y, there is only one root, x1 = ln y. The derivative of g (X ) evaluated at x1 is given by
dg = ex1 = y. dX X =x1 The probability density function of Y is then given by
1 f (x ) fY (y ) = X′ 1 = √ exp g (x1) yζ 2π
1 ln y − λ , − 2
2
ζ
which is identical to Equation 3.15 with x and y reversed.
⊛
Sums of Random Variables Let the relation between X and Y be given by Y = aX + b, where a and b are constants. Here g (X ) = aX + b and g ′ (X ) = a. Solving for X we find one root, x1 = (y − b)/a for every y. Therefore, for any density fX (x), the density of Y is then
y−b 1 f (x ) . fY (y ) = X′ 1 = fX a g (x1 ) a Now substitute the density function of X to find the density function of Y. As a simple example, suppose that fX (x) = 1/ (m − n) . Note that for the uniform density, the argument makes no difference since the density is a constant value. Then fX ((y − b) /a) = 1/ (m − n) , and
fY (y ) =
1
1
a  m − n ,
n
0. This relation also has only one root, x1 = a/y for © 2005 by Taylor & Francis Group, LLC
4.1. EXACT FUNCTIONS OF ONE VARIABLE
Figure 4.2: Density function for Y
fY (y ) = 1/y2 exp (−1/y ) .
every y. Since g (X ) Therefore,
= a/X,
= a/X,
129
(a) fX (x)
the derivative is g ′ (X )
a f ( x ) a  , fY (y ) = X′ 1 = 2 fX y g (x1) y
= exp (−x) ,
(b)
= −a/X 2 = −Y 2/a. y > 0.
If the density function of X is an exponential, fX (x) = λ exp (−λx) , then fY (y) = a /y2 λ exp (− (a/y) λ) . For numerical comparison, let a = 1 and λ = 1. The respective densities are shown in Figures 4.2(a) and (b).
Parabolic Transformation of Random Variables In this case, the random variables X and Y are related by the parabolic equation Y = aX 2 , a > 0. Since only the real roots are needed, and there are no real solutions if Y < 0, then fY (y) = 0 for this domain. If Y ≥ 0, there are two solutions,
y
x1 = +
a
y
x2 = −
a
.
The functional relation √ is g(X ) = aX 2 , with its derivative g′ (X ) = 2aX = 2a Y/a = 2 aY . Therefore, the general transformation is given © 2005 by Taylor & Francis Group, LLC
CHAPTER 4. FUNCTIONS OF RANDOM VARIABLES
130 by
fX′ (xi) = fY (y ) = 2
i=1 g (xi )
1
2√ay
f y + f − y , X
X
a
y ≥ 0.
a
For X normally distributed, the probability density for X is given by
µX )2 , exp − (x − fX (x) = 2σ2X σX 2π
1 √
(4.2)
−∞ < x < ∞,
y/a − µ X 1 exp − fY (y) = √ , 2 aσ σ X 2πay X
and the density for Y is
2
2
For the case σX = 1, µX = 0 and a given in Figures 4.3 (a) and (b).
Figure 4.3: Density function for Y (b) fY (y ) = √21πy exp {−y/2} .
y > 0.
(4.3)
= 1, the plots of these densities are
= aX 2, (a) fX (x) = √12π exp −x2/2 ,
The parabolic transformation has important applications, such as the drag force due to flow around a body. This drag force is proportional to the square of the velocity F ∼ V 2 . The proportionality constant is the drag coefficient CD , that is, F = CD V 2 . Given the density of V, then using the © 2005 by Taylor & Francis Group, LLC
4.1. EXACT FUNCTIONS OF ONE VARIABLE
131
above discussion the density of F can be derived. With this force density function the analyst can calculate the probability that the force is in a certain range and base the design of the structure to a certain probability of failure. This subject of structural reliability is examined in more depth in Chapter 9.
Example 4.2 Parabolic Transformation Consider a case where X is uniformly distributed in
fX (x) = 1/d,
(0, d) . That is,
0 < x < d.
The random variables Y and X are related by Y = aX 2 . The probability density function for Y is then given by Equation 4.2,
y + f − y . X a a y/a = 1/d. fX − y/a = 0 fX 1 f fY (y ) = √ 2 ay X
However,
Note that defined only between
0 and d. Therefore, 1 0 ≤ y ≤ ad2. fY (y ) = √ , 2 ayd
since fX (X ) is
(4.4)
⊛ Harmonic Transformation of Random Variables In this case, the transformation leads to an infinity of roots due to the periodic nature of the harmonic function. Consider the relation,
Y = a sin (X + θ) ,
a > 0.
(4.5)
There are roots only when Y  < a. These roots are given by
xi = arcsin
y − θ, a
i = · · · − 1, 0, 1, · · · ,
with g′ (xi ) = a cos (xi + θ ) . Then,
fY (y ) =
a 1− y i ∞−∞ fX (xi) , 2
2
=
y < a.
(4.6)
Use the fact that g 2 + (g ′ ) = a2 . If y  > a, there are no real solutions and fY (y ) = 0. Note that fY (±a) = ∞. Since there are no probability masses 2
© 2005 by Taylor & Francis Group, LLC
132
CHAPTER 4. FUNCTIONS OF RANDOM VARIABLES
at ±a, this implies that Pr(y = ±a) = 0. Even though it appears that in Equation 4.5 Y can equal ±a, we must be careful to distinguish between a deterministic and a random equation. The example that follows shows how the physics of the problem helps limit the number of roots.
Example 4.3 Harmonic Transformation An interesting application of this transformation is the problem of the path of a projectile. Suppose a particle is ejected at an initial velocity v and an angle of θ to the horizontal, from a point in space that we label as the origin. It is assumed that v is a constant, but θ is uniformly distributed between (0, π/2). Find the density of the horizontal distance Y the projectile will fly before it returns to the horizontal, and find the probability that this distance is less than or equal to a particular value y . Basic physics yields the relation,
v2 sin 2θ. (4.7) g This can be written more simply as Y = a sin φ, where a = v2 /g, and φ = 2θ, where φ is uniformly distributed between (0, π). Note that the maximum value of Y is a. Using Equation 4.6, we find Y =
fY (y ) =
a 1− y π1 + π1 , 2
2
0 < y < a.
The probability Pr(Y ≤ y) can now be found by integrating the area under the density function,
y Pr(Y ≤ y) = FY (y) = fY (yo ) dyo 0 y 2 1 dyo = a2 − yo2 π 0 0 < y < a. = 2 arcsin y ,
π
a
For a = 1, fY (y ) and FY (y ) are plotted in Figures 4.4 (a) and (b).
⊛
4.2
Functions of Two or More RVs
The technique developed above can, in principle, be extended to two or more variables. However, as a practical matter, only simple functions are
© 2005 by Taylor & Francis Group, LLC
4.2. FUNCTIONS OF TWO OR MORE RVS
133
Figure 4.4: Density function for projectile Equation 4.7, (a) fY (y ) 2/(π 1 − y2), (b) FY (y) = (2/π) sin−1 y.
=
Table 4.1: Table of Discrete Probabilities, pXY (xi , yi )
Y X X = 0 Y = 0 0.01 Y = 1 0.02 Y = 2 0.02 Y = 3 0.02 Sum
0.07
X=1 0.03 0.04 0.05 0.05 0.17
X =2 0.08 0.08 0.08 0.07 0.31
X =3 0.13 0.12 0.10 0.10 0.45
Sum 0.25 0.26 0.25 0.24 1
amenable to inversion. The procedure is developed for any functional relation Z = g(X, Y ). It is useful, however, to have some function in mind, for example, Z = aX + bY, (4.8) where X, Y, and Z are random variables, and a and b are given constants. Given the joint density function fXY (x, y), find fZ (z ) . Start with an example with discrete variables. Given pXY (xi , yi ), find pZ (zi ) . Suppose that a = b = 1 and (X, Y ) has the discrete distribution given in Table 4.1. Z = X + Y can take numbers from 0 to 6. For instance, z = 0 when (X, Y ) = (0, 0), z = 1 when (X, Y ) = (0, 1) or (1, 0) , z = 2 when (X, Y ) = (0, 2) , (1, 1) , or (2, 1) , and so on. In general, for Z = Zj these set of points are the points along the line Y = Zj − X or the diagonal lines in Figure 4.5.
© 2005 by Taylor & Francis Group, LLC
CHAPTER 4. FUNCTIONS OF RANDOM VARIABLES
134
Figure 4.5: The line Y
= Zj − X.
The individual probabilities for Zj can be obtained by adding the individual probabilities of each point along the line Y = Zj − X. The individual probabilities are
pZ (z = 0) = 0.01, pZ (z = 1) = 0.05, pZ (z = 2) = 0.14, pZ (z = 3) = 0.28, pZ (z = 4) = 0.25, pZ (z = 5) = 0.17, pZ (z = 6) = 0.10. The equation used to obtain these values is
pXY (xi, yi)y z −x x pXY (xi, zj − xi) . =
pZ (zj ) =
3
i =0
i= j
i
3
xi =0
For example,
pZ (z = 3) =
pXY (xi, 3 − xi) 3
xi =0 = pXY (0, 3) + pXY (1, 2) + pXY (2, 1) + pXY (3, 0) .
© 2005 by Taylor & Francis Group, LLC
4.2. FUNCTIONS OF TWO OR MORE RVS
135
Note that xi can eliminated instead of yi so that
pXY (xi, yi)x z −y pZ (zj ) = y pXY (zj − yi, yi) . = 3
i= j
i =0
i
3
yi =0
To demonstrate the procedure for continuous random variables, assume the transformation function g(X, Y ) = aX + bY. Begin with the respective cumulative distribution,
FZ (z ) = Pr(Z ≤ z ) = Pr(g (X, Y ) ≤ z)
=
g(x,y)≤z
fXY (x, y ) dx dy.
The procedure developed for discrete variables was based on rewriting the required probability on Z, that is, Pr(Z ≤ z ), in terms of its equivalent, Pr(g (X, Y ) ≤ z), for which fXY (x, y) is known. Figure 4.6 shows the region of integration, ax + by ≤ z.
Figure 4.6: Region of integration g (x, y ) ≤ z. The equation of the boundary line is z = ax + by. The limits on x can be found by solving for x in terms of y and z. Solving for x in Equation 4.8, we have x = (z − by )/a. Therefore, x is integrated © 2005 by Taylor & Francis Group, LLC
CHAPTER 4. FUNCTIONS OF RANDOM VARIABLES
136 from
−∞ to this value of x, FZ (z ) =
∞ z−by /a f ˜)
(
−∞ −∞
XY (˜x, y˜) dx˜ dy˜,
(4.9)
where x ˜ and y˜ are dummy variables. The goal is to derive an expression for fZ (z ) . Therefore, replace dx˜ by its equivalent in (˜ y , z˜). The dummy variables are also related by x˜ = (˜ z − by˜)/a. Therefore,
∂ x˜ dz˜ ∂ z˜ = ∂∂z˜ z˜ −a by˜ dz˜
dx˜ =
= a1 dz˜,
which is interpreted as 1/a so that the probability functions are strictly positive regardless of the value of a. We do all this so that the result after integration is only a function of z. The limits of integration on x˜ are transformed to
z−by /a dx˜ → z dz˜. (
˜)
0
Then,
FZ (z ) =
0
∞z
1
−∞ −∞ a
fXY
z˜ − by˜, y˜ dz˜ dy˜, a
and
dFZ (z ) dz
∞ 1 z − by , y dy. f =
fZ (z ) =
−∞ a
XY
(4.10)
a
The same result can be obtained by taking the derivative of Equation 4.9 with respect to z. Leibniz’s rule is useful here. In general, it is given by the relation,
b(z) ∂f (x, z) da (z ) db (z ) d b(z) dx. f (a (z ) , z )+ f (b (z ) , z )− f (x, z ) dx = ∂z dz dz dz a(z) a(z) © 2005 by Taylor & Francis Group, LLC
4.2. FUNCTIONS OF TWO OR MORE RVS
137
Using Leibniz’s rule, the probability density function fZ (z ) is given by
dFZ (z ) dz d ∞ (z−by)/a fXY (x, y ) dx dy = dz −∞ −∞ ∞ 1 z − by , y dy, fXY = a  a  −∞
fZ (z ) =
(4.11)
where the absolute value of a is used to ensure that fZ (z ) remains positive. This is identical to Equation 4.10. In general terms, following the above procedure, the relation is x = h1 (y, z ) , and
∞ x h y,z F (z ) = f = 1(
Z
−∞ −∞
with
dx =
)
XY (x, y) dx dy,
∂h1 (y, z ) dz. ∂z
The density function is then found to be
dFZ (z ) dz
∂h (y, z ) XY [h (y, z ) , y] ∂z dy. −∞
∞f =
fZ (z ) =
1
1
(4.12)
Solving for y instead of x above leads to the following equivalent relations,
∞ f [x, h (x, z)] ∂h (x, z) dx XY ∂z −∞ ∞ 1 z − ax dx, x, f =
fZ (z ) =
2
2
−∞ b
XY
b
(4.13)
(4.14)
where Equation 4.14 is the counterpart to Equation 4.11. Equations 4.12 and 4.13 are completely general, but they require that g(X, Y ) be invertible.
Leibniz Gottfried Leibniz was the son of Friedrich Leibniz, a professor of moral philosophy at Leipzig. Friedrich Leibniz ...was evidently
a competent though not original scholar, who devoted his time to his offices and to his family as a pious, Christian father. Leibniz’s mother was
Catharina Schmuck, the daughter of a lawyer and Friedrich Leibniz’s third
© 2005 by Taylor & Francis Group, LLC
138
CHAPTER 4. FUNCTIONS OF RANDOM VARIABLES
Figure 4.7: Gottfried Leibniz (16461716) wife. However, Friedrich Leibniz died when Leibniz was only six years old, and he was brought up by his mother. Certainly, Leibniz learned his moral and religious values from her, which would play an important role in his life and philosophy. At the age of seven, Leibniz entered the Nicolai School in Leipzig. Although he was taught Latin at school, Leibniz had taught himself far more advanced Latin and some Greek by the age of 12. He seems to have been motivated by wanting to read his father’s books. In particular, he read metaphysics books and theology books from both Catholic and Protestant writers. As he progressed through school, he was taught Aristotle’s logic and theory of categorizing knowledge. Leibniz was clearly not satisfied with Aristotle’s system and began to develop his own ideas on how to improve on it. In later life, Leibniz recalled that, at this time, he was trying to find ordering on logical truths that were the ideas behind rigorous mathematical proofs. In 1661, at the age of fourteen, Leibniz entered the University of Leipzig. It may appear that this was a truly exceptionally early age for anyone to enter university, but it is fair to say that by the standards of the time, although he was quite young, there would be others of a similar age. He studied philosophy, which was well taught at the University of Leipzig, and mathematics, which was very poorly taught. Among the other topics that were included in this two year general degree course were rhetoric, Latin, © 2005 by Taylor & Francis Group, LLC
4.2. FUNCTIONS OF TWO OR MORE RVS Greek, and Hebrew. He graduated with a bachelors degree in 1663, with the thesis De Principio Individui (On the Principle of the Individual ) which ... emphasized the existential value of the individual, who is not to be explained either by matter alone or by form alone but rather by his whole being. In
this work one finds the beginning of his notion of “monad.” Leibniz then went to Jena to spend the summer term of 1663. At Jena, the professor of mathematics was Erhard Weigel, who was also a philosopher. Through him Leibniz began to understand the importance of the method of mathematical proof for subjects such as logic and philosophy. Weigel believed that the number was the fundamental concept of the universe, and his ideas were to have considerable influence on Leibniz. After being awarded a bachelor’s degree in law, Leibniz worked on his habilitation in philosophy. His work was to be published in 1666 as Dissertatio de Arte Combinatoria (Dissertation on the Combinatorial Art ). In this work, Leibniz aimed to reduce all reasoning and discovery to a combination of basic elements such as numbers, letters, sounds, and colors. He was awarded his Master’s Degree in philosophy for a dissertation that combined aspects of philosophy and law, studying the relations between these subjects and mathematical ideas that he had learned from Weigel. A few days after Leibniz presented his dissertation, his mother died. By October 1663, Leibniz was back in Leipzig starting his studies towards a doctorate in law. Despite his growing reputation and acknowledged scholarship, Leibniz was refused the doctorate in law at Leipzig. It is unclear why this happened. He served as secretary to the Nuremberg alchemical society for a while. Then, he met Baron Johann Christian von Boineburg. By November 1667, Leibniz was living in Frankfurt, employed by Boineburg. During the next few years Leibniz undertook a variety of different projects, scientific, literary, and political. He also continued his law career taking up residence at the courts of Mainz before 1670. One of his tasks there, undertaken for the Elector of Mainz, was to improve the Roman civil law code for Mainz but Leibniz was also occupied by turns as Boineburg’s secretary, assistant, librarian, lawyer and advisor, while at the same time a personal friend of the Baron and his family .
Boineburg was a Catholic, while Leibniz was a Lutheran, but Leibniz had as one of his lifelong aims the reunification of the Christian Churches, and ... with Boineburg’s encouragement, he drafted a number of mono
graphs on religious topics, mostly to do with points at issue between the churches....
Another of Leibniz’s lifelong aims was to collate all human knowledge. Certainly, he saw his work on Roman civil law as part of this effort. Leibniz also tried to bring the work of the learned societies together to coordinate © 2005 by Taylor & Francis Group, LLC
139
140
CHAPTER 4. FUNCTIONS OF RANDOM VARIABLES research. Leibniz began to study motion, and although he had in mind the problem of explaining the results of Wren and Huygens on elastic collisions, he began with abstract ideas of motion. In 1671, he published Hypothesis Physica Nova (New Physical Hypothesis ). In this work he claimed, as had Kepler, that movement depends on the action of a spirit. He communicated with Oldenburg, the secretary of the Royal Society of London, and dedicated some of his scientific works to the Royal Society and the Paris Academy. Leibniz was also in contact with Carcavi, the Royal Librarian in Paris. As Ross explains ... Although Leibniz’s interests were clearly developing in a scientific direction, he still hankered after a literary career. All his life he prided himself on his poetry (mostly Latin), and boasted that he could recite the bulk of Virgil’s “Aeneid” by heart. During this time with Boineburg, he would have passed for a typical late Renaissance humanist.
Leibniz’s work followed parallel paths of scientific and nonscientific inquiry. He wished to visit Paris to make more scientific contacts and had started the construction of a calculating machine that he hoped would be of interest. In 1672, Leibniz went to Paris on behalf of Boineburg to try to use his plan to divert Louis XIV from attacking German areas. He formed a political plan to try to persuade the French to attack Egypt and this provided the reason for visiting Paris. His first objective in Paris was to make contact with the French government but, while waiting for such an opportunity, Leibniz made contact with mathematicians and philosophers there, in particular Arnauld and Malebranche, discussing with Arnauld a variety of topics, but particularly church reunification. In Paris, beginning in the autumn of 1672, Leibniz studied mathematics and physics under Christiaan Huygens. On Huygens’ advice, Leibniz read SaintVincent’s work on summing series and made some discoveries of his own in this area. Also in the autumn of 1672, Boineburg’s son was sent to Paris to study under Leibniz, which meant that his financial support was secure. Accompanying Boineburg’s son was Boineburg’s nephew on a diplomatic mission to try to persuade Louis XIV to set up a peace congress. Boineburg died on December 15, but Leibniz continued to be supported by the Boineburg family. In January, 1673, Leibniz and Boineburg’s nephew went to England to try the same peace mission, the French one having failed. Leibniz visited the Royal Society, and demonstrated his incomplete calculating machine. He also talked with Hooke, Boyle, and Pell. While explaining his results on series to Pell, he was told that these were to be found in a book by Mouton. The next day he consulted Mouton’s book and found that Pell was correct. At the meeting of the Royal Society on February 15, which Leibniz did not attend, Hooke made some unfavorable comments on Leibniz’s calculating
© 2005 by Taylor & Francis Group, LLC
4.2. FUNCTIONS OF TWO OR MORE RVS machine. Leibniz returned to Paris on hearing that the Elector of Mainz had died. Leibniz realized that his knowledge of mathematics was less than he would have liked so he redoubled his efforts on the subject. The Royal Society of London elected Leibniz a fellow on April 19, 1673. Leibniz met Ozanam and solved one of his problems. He also met again with Huygens who gave him a reading list including works by Pascal, Fabri, Gregory, SaintVincent, Descartes, and Sluze. He began to study the geometry of infinitesimals and wrote to Oldenburg at the Royal Society in 1674. Oldenburg replied that Newton and Gregory had found general methods. Leibniz was, however, not on best terms with the Royal Society, since he had not kept his promise of finishing his mechanical calculating machine. Nor was Oldenburg to know that Leibniz had changed from the rather ordinary mathematician who visited London, into a creative mathematical genius. In August, 1675, Tschirnhaus arrived in Paris, forming a close friendship with Leibniz that proved mathematically beneficial to both. It was during this period in Paris that Leibniz developed the basic features of his version of the calculus. In 1673, he was still struggling to develop a good notation for his calculus and his first calculations were clumsy. On November 21, 1675, he wrote a manuscript using the f (x)dx notation for the first time. In the same manuscript, the product rule for differentiation is given. By autumn 1676, Leibniz discovered the familiar d(xn ) = nx(n−1) dx for both integral and fractional n. Newton wrote a letter to Leibniz, through Oldenburg, which took some time to reach him. The letter listed many of Newton’s results but it did not describe his methods. Leibniz replied immediately, but Newton, not realizing that his letter had taken a long time to reach Leibniz, thought he had six weeks to work on his reply. Certainly one of the consequences of Newton’s letter was that Leibniz realized that he must quickly publish a fuller account of his own methods. Newton wrote a second letter to Leibniz on October 24, 1676, that did not reach Leibniz until June, 1677, by which time Leibniz was in Hanover. This second letter, although polite in tone, was clearly written by Newton believing that Leibniz had stolen his methods. In his reply, Leibniz gave some details of the principles of his differential calculus, including the rule for differentiating a function of a function. Newton was to claim, with justification, that ... not a single previously unsolved problem was solved ... by Leibniz’s approach. But the formalism was to prove vital in the latter development of the calculus. Leibniz never thought of the derivative as a limit. This concept does not appear until the work of d’Alembert. Leibniz would have liked to have remained in Paris in the Academy © 2005 by Taylor & Francis Group, LLC
141
142
CHAPTER 4. FUNCTIONS OF RANDOM VARIABLES of Sciences, but it was believed locally that there were already enough foreigners there and so no invitation came. Reluctantly, Leibniz accepted the position of librarian and Court Councillor at Hanover from the Duke of Hanover, Johann Friedrich. He left Paris in October, 1676, making the journey to Hanover via London and Holland. The rest of Leibniz’s life was spent at Hanover, except for the many travels that he made. His duties at Hanover ... as librarian were onerous, but fairly mundane:
general administration, purchase of new books and secondhand libraries, and conventional cataloging.
He undertook a whole collection of other projects, however. For example, one major project started in 16781679 involved draining water from the mines in the Harz mountains. His idea was to use wind and water power to operate pumps. He designed many different types of windmills, pumps, and gears, but ... every one of these projects ended in failure. Leibniz himself believed that this was because of deliberate obstruction by administrators and technicians, and the worker’s fear that technological progress would cost them their jobs.
In 1680, Duke Johann Friedrich died and his brother Ernst August became the new Duke. The Harz project had always been difficult, and it failed by 1684. However, Leibniz had achieved important scientific results, becoming one of the first people to study geology through the observations he compiled for the Harz project. During this work he formed the hypothesis that the Earth was at first molten. Another of Leibniz’s great achievements in mathematics was his development of the binary system of arithmetic. He perfected his system by 1679, but he did not publish anything until 1701, when he sent the paper, Essay d’une Nouvelle Science des Nombres, to the Paris Academy to mark his election to the Academy. His major work on determinants arose from his methods to solve systems of linear equations. Although he never published this work in his lifetime, he developed many different approaches to the topic with many different notations being tried out to find the one that was most useful. An unpublished paper dated January 22, 1684, contains his notation and results. Leibniz continued to perfect his metaphysical system in the 1680s, attempting to reduce reasoning to an algebra of thought. Leibniz published Meditationes de Cognitione, Veritate et Ideis (Reflections on Knowledge, Truth, and Ideas ), which clarified his theory of knowledge. In February, 1686, Leibniz wrote his Discours de Métaphysique. Another major project which Leibniz undertook, this time for Duke Ernst August, was writing the history of the Guelf family, of which the House of Brunswick was a part. He made a lengthy trip to search archives for material on which to base this history, visiting Bavaria, Austria, and © 2005 by Taylor & Francis Group, LLC
4.2. FUNCTIONS OF TWO OR MORE RVS Italy between November, 1687 and June, 1690. As always, Leibniz took the opportunity to meet with scholars of many different subjects on these journeys. In Florence, for example, he discussed mathematics with Viviani, who had been Galileo’s last pupil. Although Leibniz published nine large volumes of archival material on the history of the Guelf family, he never wrote the work that was commissioned. In 1684, Leibniz published details of his differential calculus, Nova Methodus pro Maximis et Minimis, Itemque Tangentibus, in Acta Eruditorum , a journal established in Leipzig two years earlier. The paper contained the familiar d notation, and the rules for computing the derivatives of powers, products, and quotients. However, it contained no proofs, and Jacob Bernoulli called it an enigma rather than an explanation. In 1686, Leibniz published, in Acta Eruditorum , a paper dealing with the integral calculus where the notation first appeared. Newton’s Principia appeared the following year. Newton’s “method of fluxions” was written in 1671, but Newton failed to get it published, and it did not appear in print until John Colson produced an English translation in 1736. This time delay in the publication of Newton’s work resulted in a dispute with Leibniz. Another important piece of mathematical work undertaken by Leibniz was his work on dynamics. He criticized Descartes’ ideas of mechanics and examined what are effectively kinetic energy, potential energy, and momentum. This work was started in 1676, but he returned to it at various times, in particular, while he was in Rome in 1689. It is clear that while he was in Rome, in addition to working in the Vatican library, Leibniz worked with members of the Accademia, to which he was elected a member at this time. Also while in Rome he read Newton’s Principia. His two part treatise, Dynamica, studied abstract dynamics and practical dynamics, and is written in a style somewhat similar to Newton’s Principia. Ross writes ... although Leibniz was ahead of his time in aiming at a genuine dynamics, it was this very ambition that prevented him from matching the achievement of his rival Newton. ... It was only by simplifying the issues... that Newton succeeded in reducing them to manageable proportions.
Leibniz put much energy into promoting scientific societies. He was involved in moves to set up academies in Berlin, Dresden, Vienna, and St. Petersburg. He began a campaign for an academy in Berlin in 1695, and visited Berlin in 1698 as part of his efforts. In 1700, he finally persuaded Friedrich to found the Brandenburg Society of Sciences, on July 11. Leibniz was appointed its first president, this being an appointment for life. However, the Academy was not particularly successful, and only one volume of the proceedings were ever published. It did lead to the creation of the © 2005 by Taylor & Francis Group, LLC
143
144
CHAPTER 4. FUNCTIONS OF RANDOM VARIABLES Berlin Academy some years later. Other attempts by Leibniz to found academies were less successful. He was appointed as Director of a proposed Vienna Academy in 1712, but Leibniz died before the Academy was created. Similarly, he did much of the work to prompt the setting up of the St. Petersburg Academy, but again it did not come into existence until after his death. It is no exaggeration to say that Leibniz corresponded with most of the scholars in Europe. He had over 600 correspondents. Among the mathematicians with whom he corresponded was Grandi. The correspondence started in 1703, and later concerned the results obtained by putting x = 1 into 1/(1 + x) = 1 − x + x2−x3 + .... Leibniz also corresponded with Varignon on this paradox. Leibniz discussed logarithms of negative numbers with Johann Bernoulli. In 1710, Leibniz published Théodicée, a philosophical work intended to tackle the problem of evil in a world created by a good God. Leibniz claims that the universe had to be imperfect, otherwise it would not be distinct from God. He then claims that the universe is the best possible without being perfect. Leibniz is aware that this argument looks unlikely — surely a universe in which nobody is killed by floods is better than the present one, but still not perfect. His argument here is that the elimination of natural disasters, for example, would involve such changes to the laws of science that the world would be worse. In 1714, Leibniz wrote Monadologia, which synthesized the philosophy of his earlier work, Théodicée. Much of the mathematical activity of Leibniz’s last years involved the priority dispute over the invention of the calculus. In 1711, he read the paper by Keill in the Transactions of the Royal Society of London that accused Leibniz of plagiarism. Leibniz demanded a retraction, saying that he had never heard of the calculus of fluxions until he had read the works of Wallis. Keill replied to Leibniz, saying that ... the two letters from Newton, sent through Oldenburg, had given ... pretty plain indications... whence Leibniz derived the principles of that calculus, or at least could have derived them .
Leibniz wrote again to the Royal Society asking them to correct the wrong done to him by Keill’s claims. In response to this letter, the Royal Society set up a committee to pronounce on the priority dispute. It was totally biased, not asking Leibniz to give his version of the events. The report of the committee, finding in favor of Newton, was written by Newton himself, and published as Commercium Epistolicum near the beginning of 1713, but not seen by Leibniz until the autumn of 1714. He learned of its contents in 1713, in a letter from Johann Bernoulli, reporting on the copy of the work brought from Paris by his nephew Nicolaus(I) Bernoulli. Leibniz then published an anonymous pamphlet, Charta Volans, setting
© 2005 by Taylor & Francis Group, LLC
4.2. FUNCTIONS OF TWO OR MORE RVS
145
out his side, in which a mistake by Newton in his understanding of second and higher derivatives, spotted by Johann Bernoulli, is used as evidence of Leibniz’s case. The argument continued with Keill, who published a reply to Charta Volans. Leibniz refused to carry on the argument with Keill, saying that he could not reply to an idiot. However, when Newton wrote to him directly, Leibniz did reply and gave a detailed description of his discovery of the differential calculus. From 1715 up until his death, Leibniz corresponded with Samuel Clarke, a supporter of Newton, on time, space, freewill, gravitational attraction across a void, and other topics. Leibniz is described as follows: Leibniz was a man of medium height with a stoop, broadshouldered but bandylegged, as capable of thinking for several days sitting in the same chair as of travelling the roads of Europe summer and winter. He was an indefatigable worker, a universal letter writer (he had more than 600 correspondents), a patriot and cosmopolitan, a great scientist, and one of the most powerful spirits of Western civilization.
Ross points out that Leibniz’s legacy may have not been quite what he had hoped for: It is ironical that one so devoted to the cause of mutual un
derstanding should have succeeded only in adding to intellectual chauvinism and dogmatism. There is a similar irony in the fact that he was one of the last great polymaths — not in the frivolous sense of having a wide general knowledge, but in the deeper sense of one who is a citizen of the whole world of intellectual inquiry. He deliberately ignored boundaries between disciplines, and lack of qualifications never deterred him from contributing fresh insights to established specialisms. Indeed, one of the reasons why he was so hostile to universities as institutions was because their faculty structure prevented the crossfertilization of ideas which he saw as essential to the advance of knowledge and of wisdom. The irony is that he was himself instrumental in bringing about an era of far greater intellectual and scientific specialism, as technical advances pushed more and more disciplines out of the reach of the intelligent layman and amateur .
Example 4.4 Kinetic Energy Density
A particle of mass m moving in the xy plane has a kinetic energy T where Z is the resultant velocity that is related
to2 the component velocities (speed) in each coordinate direction by Z = X˙ + Y˙ 2. Suppose that X˙ and Y˙ are statistically independent (for convenience) and each is distributed as a standard normal random variable, that is, zero mean and unitary standard deviation. Derive the density of the kinetic energy.
= mZ 2 /2
© 2005 by Taylor & Francis Group, LLC
146
CHAPTER 4. FUNCTIONS OF RANDOM VARIABLES
Solution Use the result from Section 4.1 where a parabolic transformation of densities was derived. To do this, write the kinetic energy as T
=
1
1
mZ 2 = m X˙ 2 + Y˙ 2
2 = U + V.
2
The2 first step is to apply Equation 4.3 to transform between X˙ 2 and U, and Y˙ and V. The resulting densities are fU (u) =
fV (v) =
v
u 1 √πmu exp − m 1 √πmv exp − m
,
u ≥ 0,
,
v ≥ 0.
(4.15) (4.16)
Now following the procedure of this section, solve for either u or v, say,
Figure 4.8: Domain of integration for Example 4.4. v = t − u, where t − u ≥ 0, or t ≥ u. This relation is useful in setting the integration limits. The domain of integration is shown in Figure 4.8. © 2005 by Taylor & Francis Group, LLC
147
4.2. FUNCTIONS OF TWO OR MORE RVS
Applying Equation 4.12, the kinetic energy density is given by fT (t) = =
t
t 0
0
fU (u) fV (t − u) du
t−u u 1
1 √πmu exp − exp − m m πm (t − u)
du
(4.17) which was integrated directly using MAPLE, or could be integrated by transforming according to r = u/t and du = tdr, resulting in a beta function. Equations 4.15—4.17 are chisquared densities with one and two degrees of freedom. =
1
m
exp (
−t/m) ,
1
⊛
Example 4.5 Products and Quotients Two other important transformations are for the functions Z = XY and Z = X/Y. Products and quotients are everywhere in mathematics. For the product functional relations, the following procedure is developed,
x = z/y ≡ h1 (y, z )
= y1 z ∞ 1 , y dy f fZ (z ) = y XY y −∞ ∞ 1 z fXY x, dx. or = x x ∂h1 ∂z
 
−∞ 

The last equality was obtained by solving for y
∂h2 /∂z
= 1/x.
1
fT (t) = =
=
=
1
πm 1
πm 1
πm 1
m
exp
exp
exp
exp
© 2005 by Taylor & Francis Group, LLC
− mt
− mt
t
−m
− mt
.
0
= z/x = h2 (x, z) and using
t
u−1/2 (t
− u)−1/2 du
1
r−1/2 (1
− r)−1/2 dr
0
1 1 2 2
B( , )
148
CHAPTER 4. FUNCTIONS OF RANDOM VARIABLES = X/Y, follow the procedure, x = zy ≡ h1 (y, z )
Considering next the quotient Z
∂h1 =y ∂z
∞ y fXY (zy, y) dy −∞ ∞ x f (x, zx) dx. = −∞ XY
fZ (z ) =
Suppose that Z = XY , X, and Y are statistically independent, and their probability density functions are given by
x2 x fX (x) = 2 exp − 2 , x ≥ 0 2σ σ 1 ln b , 0 < a ≤ y ≤ b. fY (y ) = y a The cumulative distribution FZ (z ) is then
FZ (z ) =
xy≤z
fX (x) fY (y ) dx dy.
The domain of integration is defined by
x ≥ 0, a ≤ y ≤ b, and xy ≤ z, and is shown in Figure 4.9. Therefore, FZ (z ) can be written as FZ (z ) =
b z/y a
0
fX (x) fY (y ) dx dy.
Differentiating FZ (z ) with respect to z, obtain
fZ (z ) =
d F (z ) dz Z
b 1 z f (y) dy fX y Y a y  z2 1 b b 1 z exp − 2σ2y2 y ln a dy. = a y yσ2
=
Since y is always positive, the absolute value sign can be omitted. The resulting probability density is
1 ln b b z2 exp − z2 dy 2σ2y2 z a a y 3 σ2 2 2 = 1 ln b exp − 2σz2 b2 − exp − 2σz2a2 . z a
fZ (z ) =
© 2005 by Taylor & Francis Group, LLC
4.2. FUNCTIONS OF TWO OR MORE RVS
149
Figure 4.9: Domain of integration for Example 4.5. With the density function for Z, moments of the variable can be calculated.
⊛
The examples above are only meant to demonstrate the basic procedure, which is generalizable to any number of random variables, and can by way of a sequence of transformations be applied to very complex functional relationships. For example, a complex relation can be built in the following way, where the arrows indicate the direction of the density transformation,
Y1 ←− X12 Y2 ←− Y1 + X2 Y3 ←− Y22 Y4 ←− X3 X4 Y5 ←− Y3 + Y4
=⇒ Y5 = X12 + X2 2 + X3X4.
4.2.1
General Case
Consider the twodimensional continuous random variable, (X, Y ), with joint probability density function fXY (x, y ) . Z and W are twodimensional continuous random variables that are functions of X and Y such that
Z = H1 (X, Y ) and W = H2 (X, Y ) . Then, how do we find the probability density function fZW (z, w)? In order to answer this question, a similar procedure is followed as that for the
© 2005 by Taylor & Francis Group, LLC
150
CHAPTER 4. FUNCTIONS OF RANDOM VARIABLES
function of one variable, demonstrated in Section 4.1.
Figure 4.10: Transformation of variables. Assume that there is a onetoone relationship between (X, Y ) and (Z, W ) such that the inverse transform exists,
X = g1 (Z, W ) and Y = g2 (Z, W ) .
(4.19)
Consider the probability,
Pr(x < X ≤ x + dx
and y < Y
≤ y + dy).
For sufficiently small dx and dy, the probability can be approximated by fXY (x, y ) dx dy . The same probability can be written as
Pr(z < Z ≤ z + dz
and w < W
≤ w + dw)
and approximated by fZW (z, w) dz dw. That is,
fXY (x, y ) dx dy = fZW (z, w) dz dw ,
(4.20)
which is equivalent to Equation 4.1 for probability density functions of one variable only. The next step is to find dx dy in terms of dz dw. This is not as simple as in the previous case. The term dz dw is the incremental area R shown in Figure 4.10 (a). Figure 4.10 (b) shows the same incremental area in xy plane after it is transformed using the inverse transformation defined in Equation 4.19. The actual transformation, denoted as F, results in the
© 2005 by Taylor & Francis Group, LLC
4.2. FUNCTIONS OF TWO OR MORE RVS
151
image S, defined by the vertices ABCD. The corners of S are
A B C D
: : : :
[g1 (z, w) , g2 (z, w)] [g1 (z + dz, w) , g2 (z + dz, w)] [g1 (z + dz, w + dw) , g2 (z + dz, w + dw)] [g1 (z, w + dw) , g2 (z, w + dw)] .
Using the Taylor series about (z, w), we can approximate
∂g ∂g1 dz + 1 dw ∂w ∂z ∂g1 dw g1 (z, w + dw) ≃ g1 (z, w) + ∂w ∂g g1 (z + dz, w) ≃ g1 (z, w) + 1 dz. ∂z The same expansion can be performed for g2 . The area S is approximated by the parallelogram with vertices A′ B ′ C ′ D′ . The corners of the parallelogram g1 (z + dz, w + dw)
are
≃
g1 (z, w) +
A′ = A : (g1 (z, w) , g2 (z, w)) ∂g2 ∂g1 ′ dz dz, g2 (z, w) + B : g1 (z, w) + ∂z ∂z ∂g2 ∂g2 ∂g1 ∂g1 ′ dw dz + dw, g2 (z, w) + dz + C : g1 (z, w) + ∂w ∂z ∂w ∂z ∂g ∂g D′ : g1 (z, w) + 1 dw, g2 (z, w) + 2 dw . ∂w ∂w
Figure 4.11: Area of the parallelogram.
The area of the parallelogram, A′ B ′ C ′ D′ can be obtained by performing the operations (A1 + A2 ) − (A3 + A4 ) , as shown in Figure 4.11. The areas © 2005 by Taylor & Francis Group, LLC
CHAPTER 4. FUNCTIONS OF RANDOM VARIABLES
152
A1 , A2 , A3 , and A4 are given by
1 ∂g1 dw ∂g2 dw 2 ∂w ∂w ∂g2 ∂g2 ∂g2 ∂g1 1 = 2 ∂z dz ∂w dw + ∂z dz + ∂w dw ∂g 1 dz 2 dz = 12 ∂g ∂z ∂z 1 ∂g1 dw ∂g2 dz + ∂g2 dz + ∂g2 dw . = 2 ∂w ∂w ∂z ∂z
A1 =
A2
A3
A4
Therefore, the area of the parallelogram is given by
=
1 ∂g1 ∂g2 dw2 + ∂g1 ∂g2 dz dw + 1 ∂g1 ∂g2 dz 2 2 ∂z ∂z ∂z ∂w 2 ∂w ∂w 1 ∂g1 ∂g2 dz 2 − ∂g1 ∂g2 dz dw − 1 ∂g1 ∂g2 dw2 − 2 ∂z ∂z 2 ∂w ∂w ∂w ∂z ∂g1 ∂g2 ∂z ∂w
1 ∂g2 − ∂g ∂w ∂z
dz dw.
That is, the area dz dw has increased by a factor of
∂g
∂g2 ∂z ∂w 1
1 ∂g2 − ∂g ∂w ∂z
under the transformation. This factor is the Jacobian J, or the determinant of the matrix
∂g1 /∂z ∂g1 /∂w , ∂g2 /∂z ∂g2 /∂w and the incremental area S in the xy plane is dx dy. Therefore,
∂g1 /∂z ∂g1 /∂w dx dy = det ∂g2 /∂z ∂g2 /∂w = Jdz dw.
dz dw
Rewriting Equation 4.20, we obtain
fZW (z, w) = fXY (g1 (z, w) , g2 (z, w)) J  ,
(4.21)
where J  is the absolute value of the Jacobian. The use of Equation 4.21 is demonstrated in the following example.
© 2005 by Taylor & Francis Group, LLC
4.2. FUNCTIONS OF TWO OR MORE RVS
153
Example 4.6 Functions of a TwoDimensional Random Variable Suppose (X, Y ) is a random variable with probability density function fXY (x, y ) defined in a circle with radius 1 and center at (0, 0) . The joint probability density function is given by
2 2 , (x, y) inside the circle 1 /2πσ2 exp − 12 x σ+2y fXY (x, y) = 0,
(x, y)
elsewhere.
Suppose that it is of interest to express the probability density function in terms of the distance from the center, R, and the angle from the positive axis, Θ. The random variables R and Θ are related to X and Y by
X2 + Y 2 Y . Θ = tan−1 X Find the probability density function fRΘ (r, θ) . Solution Solving for X and Y, we obtain R=
X = R cos Θ Y = R sin Θ. The Jacobian is given by
J
= det = det =
∂x/∂r ∂x/∂θ ∂y/∂r ∂y/∂θ cos θ −r sin θ sin θ r cos θ
r.
Then the joint probability density function fRΘ (r, θ) is given by
fRΘ (r, θ) = r fXY (r cos θ, r sin θ ) 1 r2 r exp − = 2 σ2 2πσ 2
for 0 ≤ r < 1, 0 ≤ θ < 2π.
⊛ Example 4.7 Distribution of a Sum of Random Variables Suppose that (X, Y ) is a random variable with density fXY (x, y ). We seek to find the probability density fZ (z ) where Z is defined as follows,
Z = aX + bY. © 2005 by Taylor & Francis Group, LLC
CHAPTER 4. FUNCTIONS OF RANDOM VARIABLES
154
Let fXY (x, y ) be defined on 0 < X < 1 and 0 < Y < 1. Solution A similar problem was solved previously where fXY (x, y) was defined on the infinite domain, and the results shown in Equations 4.10 and 4.14. In order to use the general method, we define another variable
W,
W
= X.
The choice of W is arbitrary, and so a convenient function is chosen. Solve for X and Y ,
X=W Z − aW . Y = b The Jacobian is evaluated as
J
= det
0 1/b
1
−a/b
=
− 1b ,
and the joint probability density is then
z − aw 1 . fZW (z, w) = − fXY w, b b
If fXY (x, y) is defined on the infinite domain, −∞ < X < ∞ and −∞ < Y < ∞, then the corresponding domain in W and Z is also the infinite domain,
−∞ < X < ∞ and − ∞ < Y < ∞) → (−∞ < W < ∞, −∞ < Z < ∞) .
(
The marginal density fZ (z ) is then
fZ (z ) =
∞ 1 z − aw − f dw. w, f (z, w) dw = b −∞ b XY −∞ ZW
∞
This is an alternate method to obtain the result in Equation 4.14. If fXY (x, y ) is defined on the rectangle defined by 0 < X < 1 and 0 < Y < 1, obtaining the corresponding domain in W and Z and the marginal density fZ (z ) requires more work. Start with the domain of W and Z. Since X = W, we know 0 < W < 1. The range for Z can be derived from that of Y. It is given that 0 < Y < 1. Writing Y in terms of Z and W, results in 0
0, ω ω
where m4 (rad/s)4 4 B = 0.0175 (rad/s) for the wind velocity at 19.5 m above the still water level, taken to be 25 m/s. The wave elevation at x = 0 (zero horizontal location) is then given by N A
=
η (t) =
0.7795
2
n=1
N
σ cos (ω n t − ϕn ) ,
where it was previously found that σ2 = A/4B. ωn are chosen according to o (ω) /σ2 , the probability density f (ω) = Sηη f (ω ) =
4B exp ω5
− ωB4
The cumulative distribution function is F (ω) = exp
π
21 2
2 2 M.
is excluded since
rad is the same as
, 0
ω
ω
≥ 0.
≥ 0.
rad.
Shinozuka, “M onte Carlo Solution of Structure Dynamics,”
tures 2, ,
2π
− ωB4
,
855874, 1972.
© 2005 by Taylor & Francis Group, LLC
Computers and Struc
12.5. APPLICATIONS
675
Table 12.9: Random Frequencies and Phase Angles for Ocean Wave Height x1 0.9501
0.4860
0.4565
0.2311
0.8913
0.0185
0.6068
0.7621
0.8214
ϕ = 2πx1 5.970 3.054 2.868 1.452 5.600 0.1162 3.813 4.788 5.161
x2 0.4447
0.9218
0.4057
0.6154
0.7382
0.9355
0.7919
0.1763
0.9169
ω = F −1 (x2 ) = (−B/ ln(x2 )) 0.3834 0.6809 0.3732 0.4357 0.4900 0.7158 0.5233 0.3169 0.6702
0.25
The random frequencies and the random phase angles are generated from standard uniform random numbers x1 and x2 and tabulated in Table 12.9. Figure 12.10 shows the time history of a sample wave derived using the above procedure for wind velocity V19.5 = 25 m/s and N = 10 at x = 0.
⊛
Figure 12.10: Sample wave profile derived by Shinozuka’s method applied to the PiersonMoskowitz wave height spectrum.
© 2005 by Taylor & Francis Group, LLC
676 12.6
CHAPTER 12. THE MONTE CARLO METHOD Concluding Summary
This chapter introduced the fundamental ideas underlying the numerical methods grouped by the term Monte Carlo methods. Such methods are widely used in both probabilistic and deterministic problems. The methods are based on the generation of numbers according to specific probability density functions. These are introduced and discussed. In addition, as a numerical method, it is necessary to establish error estimates. All these are introduced briefly. With this background the reader can move to more advanced studies. 12.7
Problems
Section 12.2: Random Number Generation
1. Generate uniform random numbers, using available software such as MATLAB, either using the linear congruential generator or the program’s builtin random number generator. Plot the average value and the standard deviation and compare them with the theoretical values for the uniform distribution. 2. Show that random numbers that are distributed according to the lognormal density function with mean value exp (µ) , and standard deviation exp (σ) , can be obtained by using
y = exp σ
−2 ln x1 cos 2πx2 + µ
,
where x1 and x2 are uniform random numbers between 0 and 1. 3. Verify that the random numbers produced in the previous problem are indeed distributed according to the lognormal density. Plot the mean value and the standard deviation as functions of the number of random numbers used. 4. Show that random numbers that are distributed according to the lognormal density function with mean value exp (µ) , and standard deviation exp (σ) , can be obtained by using
y = exp σ
−2 ln x1 cos 2πx2 + µ
,
where x1 and x2 are uniform random numbers between 0 and 1. 5. Verify that the random numbers produced in the previous problem are indeed distributed according to the lognormal density. Plot the mean value and the standard deviation as functions of the number of random numbers used. © 2005 by Taylor & Francis Group, LLC
12.7. PROBLEMS
677
6. Use the inverse numerical transformation method to generate normal random numbers with µ = 0 and σ = 1. Use ten equally spaced intervals between y = −3σ and y = 3σ. 7. Use the composition method to generate random numbers distributed according to fY (y ) =
1
4
sin y +
9
32π 3
y2 ,
0
< y < 2π.
8. Use Von Neumann’s rejectionacceptance method to generate normally distributed random numbers. How does the method compare to the BoxMuller method? Section 12.3: Random Numbers for Joint Probability Densities
9. Use the inverse transform method to generate random numbers for the joint probability density function, fY1 Y2 (y1 , y2 ) = y1 y2 ,
0
< y1 , y2 < 1.
10. Use the linear transform method to generate random numbers y1 and y2 for the joint probability density function given in the previous problem.
© 2005 by Taylor & Francis Group, LLC
Chapter 13
FluidInduced Vibration In modeling offshore structures, one needs to account for the forces exerted by the surrounding fluid.
The vibrational characteristics of a structure
can be significantly altered when it is surrounded by water. For example, damping by the fluid lowers the natural frequency of vibration. The purpose of this chapter is to show how the fluid forces on an offshore structure due to current and random waves are modeled. In Section 13.1, we discuss how the fluid forces in an ocean can be described. In Section 13.2, we find how the fluid force on an offshore structure can be estimated. In Section 13.3, some example problems are presented to show how the first two sections can be applied. Readers who are interested in more detailed studies 1 2 3 4 should refer to texts by Chakrabarti, Faltinsen, Kinsman, Sarpkaya, and 5 Wilson. This chapter is meant to be selfcontained and therefore repeats small parts of earlier chapters for the ease of the reader.
13.1
Ocean Currents and Waves
When one considers the dynamics of an offshore structure, one must also consider the forces due to the surrounding fluid. The two important sources of fluid motion are ocean waves and ocean currents. 1
Hydrodynamics of Offshore Structures,
Publications, 1987. 2
S. Chakrabarti, Computational Mechanics
Sea Loads on Ships and Offshore Structures,
Press, 1993. 3 4
Wind Waves, B. Kinsman, PrenticeHall, 1965, now available in a Dover edition. Mechanics of Wave Forces on Offshore Structures, T. Sarpkaya, M. Isaacson, Van
Nostrand Reinhold, 1981. 5
O. Faltinsen, Cambridge University
Dynamics of Offshore Structures,
2003.
J. Wilson, John Wiley & Sons, Second Edition,
679
© 2005 by Taylor & Francis Group, LLC
CHAPTER 13. FLUIDINDUCED VIBRATION
680
Most steady large currents are generated by the drag of the wind passing over the surface of the water, with the currents confined to a region near the ocean surface. Tidal currents are generated by the gravitational attraction of the Sun and the Moon, and they are most significant near coasts. The ultimate source of the ocean circulation is the uneven radiative heating of the Earth by the Sun. 6 Isaacson suggested an empirical formula for the current velocity in the horizontal direction as a function of depth,
1/7 x − d + do , + Udrif t (d) Uc (x) = (Utide(d) + Ucirculation(d)) xd do
Udrif t is the windinduced drift current, Utide is the tidal current, Ucirculation is the lowfrequency longterm circulation, x is the vertical distance measured from the ocean bottom, d is the depth of the water, and do is the smaller of the depth of the thermocline or 50 m. The value of Utide is obtained from tide tables, and Udrif t is about 3% of the 10 minute mean where
wind velocity at 10 m above sea level. It should be noted that these currents evolve slowly compared to the time scales of engineering interests. Therefore, they can be treated as quasisteady phenomena. Waves, on the other hand, cannot be treated as steady phenomena.
The underlying physics that govern wave dynamics are too
complex, and therefore waves must be modeled stochastically. In the subsequent section, we discuss the concept of the spectral density, the available ocean wave spectral densities, a method to obtain the spectral density from wave timehistories, methods to obtain a sample time history from a spectral density, shortterm and longterm statistics, and a method to obtain fluid velocities and accelerations from wave elevation using linear wave theory.
13.1.1
Spectral Density
Here, only surface gravity waves are considered. A regular wave is examined first in order to become familiar with the terms that are used to describe a wave. The wave surface elevation is denoted as as
η (x, t)
=
A cos (kx − ωt) ,
where
k
η (x, t) and can be written ω is the
is the wave number and
angular frequency. Figure 13.1(a) shows the surface elevation at two time instances (t
= 0 and t = τ ) and (b) the surface elevation at a fixed location A is the amplitude, H is the wave height or the distance between the maximum and minimum wave elevation or twice the amplitude, and T is the period, given by T = 2π/ω. (x
= 0).
6 “Wave
and Current Forces on Fixed Offshore Structures,” M. Isaacson, Canadian
Journal of Civil Engineering, 15:937947, 1988.
© 2005 by Taylor & Francis Group, LLC
13.1. OCEAN CURRENTS AND WAVES
Figure 13.1: Defining parameters of a regular wave.
© 2005 by Taylor & Francis Group, LLC
681
CHAPTER 13. FLUIDINDUCED VIBRATION
682
In practice, waves are not regular. Figure 13.2 shows a schematic time history of an irregular wave surface elevation.
The wave height and fre
quency are not easily defined. Therefore, the wave height spectral density is utilized for a statistical description of the wave elevation.
Figure 13.2: Time history of a random wave. The random surface elevation
η (t)
can be thought of as a summation
of regular waves with different frequencies. The surface elevation therefore be related to its Fourier transform
η (t) =
1
X (ω) by
∞ X (ω ) exp (−iωt) dω. −∞
2π
Suppose that the energy of the system is proportional to energy becomes
η (t)
can
η2 (t) so that the
1
E = Cη2 (t) , where
C
2
is the proportionality constant. The expected value of the energy
is then given by
where
E η 2 (t)
1 E {E} = CE η 2 (t) , 2
is the meansquare value of
process, the meansquare value of
η (t)
average over a long period of time,
E η2 (t) =
1
η (t) .
If
η (t)
Ts /2
2
→∞ Ts −Ts/2 η (t) dt ∞ 1 1 X (ω )2 dω, = lim Ts →∞ Ts 2π −∞ lim
Ts
where Parseval’s theorem has been utilized,
∞ ∞ 2 1 X (ω )2 dω, η (t) dt = 2π −∞ −∞
© 2005 by Taylor & Francis Group, LLC
is an ergodic
can be approximated by the time
13.1. OCEAN CURRENTS AND WAVES
683
where
X (ω)2 = X (ω ) X ∗ (ω )
X (ω ) = X ∗ (ω ) =
∞ η (t) exp (−iωt) dt −∞ ∞
−∞
η (t) exp (iωt) dt.
The power spectral density (or simply the spectrum) is defined as
Sηη (ω) ≡ so that
E η2 (t) is given by
1
2πTs
X (ω )2 ,
(13.1)
∞ S (ω ) dω. −∞ ηη
E η2 (t) = For a zeromean process,
E η2 (t)
2 2 density has units of η t, or m s.
(13.2)
equals the variance
σ 2η .
The spectral
Sηη (ω ) is related to the autocorrelation function, Rηη (τ ), by the Wiener
Khinchine relations,
7 ,8
Sηη (ω )
=
Rηη (τ )
=
∞ Rηη (τ ) exp (−iωτ ) dτ 2π −∞ ∞ 1
−∞
Sηη (ω ) exp (iωτ ) dω.
It should be noted that in some textbooks, the factor
(13.3)
(13.4)
1/2π appears in the
second equation instead of the first. There are a few properties of the spectral density with which the reader should become familiar.
The first property is that the spectral density
function of a realvalued stationary process is both real and symmetric.
Sηη (ω ) = Sηη (−ω ) (Equation 13.1). Secondly, the area under the E η2 (t) (Equation 13.2) and also equal to Rηη (0) = σ2η − µ2η , where σ 2η is the variance and µη is the mean of η (t) . In That is,
spectral density is equal to
most cases, we only consider a zero mean process so that the area under the spectral density equals
σ2η .
If the process does not have a zero mean, the
mean value can be subtracted from the process so that it now has a zero mean value. 7 “Generalized
Harmonic Analysis,” N. Wiener, Acta Math., 55:117258, 1930. Theorie der Stationaren Stochastischen Prozesse,” A. Khinchine, Math. Ann., 109:604615, 1934. 8 “Korrelations
© 2005 by Taylor & Francis Group, LLC
CHAPTER 13. FLUIDINDUCED VIBRATION
684
For ocean applications, a onesided spectrum in terms of cycles per sec
Hertz (Hz) is often used. The onesided spectrum is given a supero, and can be obtained from the twosided spectrum by the relation,
ond or script
o Sηη (ω ) = 2Sηη (ω ) ,
ω ≥ 0.
ω can be transformed to the spectrum ω = 2πf ) by the relation,
The twosided spectrum in terms of in terms of
f
(where
Sηη (f ) = 2πSηη (ω ) . Then, the twosided spectrum in terms of
ω
can be transformed to the
onesided spectrum in units of Hz by the relation,
o Sηη (f ) = 4πSηη (ω ) ,
for
f, ω > 0.
It should be noted that the spectral density that we have defined in this text is the
amplitude halfspectrum.
For an amplitude halfspectrum
S (ω ),
the amplitude, height, and height double spectra are related by
S A (ω) = 2S (ω) S H (ω) = 8S (ω) S 2H (ω) = 16S (ω ) .
13.1.2
Ocean Wave Spectral Densities
In this section, spectral density models for a random sea are introduced. An excellent review of existing spectral density models is given in Chapter 4 of Chakrabarti. The ocean wave spectrum models are semiempirical formulas.
That
is, they are derived mathematically, but the formulation requires one or more experimentally determined parameters. The accuracy of the spectrum depends significantly on the choice of these parameters. In formulating spectral densities, the parameters that influence the spectrum are
current,
fetch limitations, decaying versus developing seas, water depth, swell. The fetch is the distance over which a wind blows in
and
a wavegenerating phase. Fetch limitation refers to the limitation on the distance due to some physical boundaries so that full wave development is prohibited. In a developing sea, wave development has not yet reached its stationary state under a stationary wind. In contrast, the wind has blown a sufficient time in a fully developed sea, and the sea has reached its stationary state. In a decaying sea, the wind has dropped off from its stationary value. Swell is the wave motion caused by a distant storm that persists even after the storm has died down or moved away.
© 2005 by Taylor & Francis Group, LLC
13.1. OCEAN CURRENTS AND WAVES
685
9 The PiersonMoskowitz (PM) spectrum is the most extensively used
spectrum for representing a fully developed sea. It is a oneparameter model in which the sea severity can be specified in terms of the wind velocity. The PM spectrum is given by
S
ω
g . × 10−3g2 exp −0.74 Uw,19.5 ω5
81 o ηη ( ) =
g
4
ω−4 ,
U
where is the gravitational constant, and w, 19.5 is the wind speed at a height of 19.5 m above the still water. The PiersonMoskowitz spectrum is also called the windspeed spectrum because it requires wind data. It can
ωm as well, ω m 4 8.1 × 10−3 g2 o . exp − 1 . 25 Sηη (ω ) = ω ω5
be written in terms of the modal frequency
Note that the modal frequency is the frequency at which the spectrum is the maximum. In some cases, it may be more convenient to express the spectrum in terms of significant wave height rather than the wind speed or modal fre10 quency. For a narrowband Gaussian process, the significant wave height is related to the standard deviation by
Hs = 4ση . The standard
∞ o deviation 2is 0 Sηη (ω ) dω = σ η .
the square root of the area under the spectral density, Then, the spectrum can be written as
0.0324g2 −4 8.1 × 10−3 g 2 o ω , exp − Sηη (ω ) = 5 Hs2 ω
and the peak frequency and the significant wave height are related by
ωm = 0.4 g/Hs. The PiersonMoskowitz spectrum is applicable for deep water, unidirectional seas, fully developed and localwindgenerated with unlimited fetch, and was developed for the North Atlantic.
The effect of swell is not ac
counted for in this spectrum. It is found that even though it is derived for the North Atlantic, the spectrum is valid for other locations. However, the limitation that the sea is fully developed may be too restrictive because it cannot model the effect of waves generated at a distance. Therefore, in such instances a twoparameter spectrum, such as the Bretschneider spectrum, is used to model a nonfully developed sea as well as a fully developed sea. 9 “A Proposed Spectral Form for Fully Developed Wind Seas Based on the Similarity Theory of S.A. Kitaigorodskii,” W. Pierson, L. Moskowitz, Journal of Geophysical Research , 69(24):51815203, 1964. 1 0 See Section 13.1.5 for details.
© 2005 by Taylor & Francis Group, LLC
CHAPTER 13. FLUIDINDUCED VIBRATION
686
11 , 12 The Bretschneider spectrum is a twoparameter spectrum in which
both the sea severity and the state of development can be specified. The Bretschneider spectrum is given by
S
ω
.
ω H 2 exp −0.675 ωs 4 , ω ω s
4 s o ( ) = 0 169 ηη 5
ωs = 2π/Ts and Ts is the significant period. The sea severity can be Hs and the state of development can be specified √ by ωs. It can be shown that for ω s = 1.167ω m (equivalently ω s = 1.46/ Hs ) the Bretschneiwhere
specified by
der spectrum and the PiersonMoskowitz spectrum are equivalent. Figure
Hs = 4 m and various values of ωs. ωs = 0.731 rad/s, the PiersonMoskowitz and the Bretschneider spectra
13.3 plots the Bretschneider spectra for For
are identical. It should be noted that the developing sea will have a slightly
√
higher modal frequency than the fully developed sea and can be described by
ωs greater than 1.461/ Hs .
Figure 13.3: Bretschneider spectra for various values of
ωs .
Other twoparameter spectral densities that are often used are the ISSC (International Ship Structures Congress) and the ITTC (International Towing Tank Conference) spectra. The ISSC spectrum is written in terms of the 1 1 Wave Variability and Wave Spectra for WindGenerated Gravity Waves , C. Bretschneider, Technical Memorandum No. 118, Beach Erosion Board, U.S. Army Corps of Engineers, Washington, D.C., 1959. 1 2 “Wave Forcasting,” in Handbook of Ocean and Underwater Engineering, John Meyers, Editor, McGrawHill, 1969.
© 2005 by Taylor & Francis Group, LLC
13.1. OCEAN CURRENTS AND WAVES
687
−
Table 13.1: The parameters for the generalized twoparameter spectrum, 4 2 4 o ˜ /ω 5 exp A (ω/ω ˜) (ω ) = A Sηη 4 Hs ω
−
A 0.675 0.318 0.4427
Model
Bretschneider ITTC ISSC
ω ˜ ωs ωz ω ¯
significant wave height and the mean frequency, where the mean frequency is given by
∞ 0 ωS (ω ) dω ω ¯= ∞ S (ω )
0
= 1.30 ω m .
The ISSC spectrum is given by
S
o ηη
ω ¯4 (ω) = 0.111 5 Hs2 exp ω
−0.444
ω¯ 4 ω
.
The ITTC spectrum is based on the significant wave height and the zero crossing frequency and is given by o Sηη (ω ) = 0.0795
ω4z 2 H exp ω5 s
where the zero crossing frequency
ωz
−0.318
ω
,
is given by
∞ ω 2 S (ω) dω 0
ωz = ∞ 0
ω z 4
= 1.41ωm .
S (ω )
The Bretschneider, ITTC, and ISSC spectra are called twoparameter spectra. A generalized equation for such twoparameter spectra is
S with
A
and
ω ˜
o ηη
(ω ) =
A 4
˜ 2ω
Hs
4
ω5
exp
−A
4 ω ˜ ω
,
given in Table 13.1 for the examples discussed here.
The spectra that we have discussed so far do not permit spectra with two peaks to represent local or distant storms, nor do they allow specification of 13 the sharpness of the peaks. The OchiHubble spectrum is a sixparameter 1 3 “Six
Parameter Wave Spectra,” M. Ochi, E. Hubble,
Engineering Conference, Honolulu, 301328, 1976. © 2005 by Taylor & Francis Group, LLC
Proc. Fifteenth ASCE Coastal
CHAPTER 13. FLUIDINDUCED VIBRATION
688
spectrum that provides such modeling flexibility. It has the form,
S
o ηη
(ω ) =
1
4
2
4λi +1
ω4mi Γ (λi )
λi
4
i=1
2 Hsi
ω4λi +1
exp
−
4λi + 1
4
ω mi 4 , ω
where Γ (λi ) is the gamma function, Hs1 , ω m1 , and λ1 are the significant wave height, modal frequency, and shape factor for the lower frequency
Hs2 , ωm2 , and λ2 are those for the higher Assuming that the entire spectrum is that of a
components, respectively, and frequency components.
narrow band, the equivalent significant wave height is given by
Hs =
Hs21 + Hs22 .
λ1 = 1 and λ2 = 0, the spectrum reduces to the PiersonMoskowitz spectrum. With the assumption that the entire spectrum is narrow band,
For
λ1 is much larger than λ2 . The OchiHubble spectrum represents unidirectional seas with unlimited fetch. The sea severity and the
the value of
state of development can be specified by dition,
λi
Hsi
and
ω mi ,
respectively. In ad
can be selected appropriately to control the frequency width of
the spectrum. For example, a small developing sea and a large
λi
λi
(wider frequency range) describes a
(narrower frequency range) describes a swell
condition. Figure 13.4 shows the OchiHubble spectrum with
ω m1 = 0.626 rad/s, Hs1 = 3.35 Hs2 = 2.19 m.
m,
λ2 = 2.72, ω m2 = 1.25
λ1 = 2.72, rad/s, and
Figure 13.4: OchiHubble spectrum. Finally, another spectrum that is commonly used is the JONSWAP
© 2005 by Taylor & Francis Group, LLC
13.1. OCEAN CURRENTS AND WAVES
689
(Joint North Sea Wave Project) spectrum developed by Hasselmann et al.
14
It is a fetchlimited spectrum because the growth over only a limited fetch is taken into account. In addition, the attenuation in shallow water is taken into account. The JONSWAP spectrum is o Sηη (ω ) =
where
αg 2 exp ω5
−1.25
ω m 4 ω
2 2 2 γ exp(−(ω−ωm ) /2τ ωm ) ,
γ = peakedness parameter, and τ = shape parameter. The peakedγ is the ratio of the maximum spectral energy to the max
ness parameter
imum spectral energy of the corresponding PiersonMoskowitz spectrum. That is, when
γ = 7, the peak spectral energy is 7 times that of the Pierson
Moskowitz spectrum. The parameters are valued as follows:
7.0 for very peaked data γ = 3.3 for mean of selected JONSWAP data 1.0 for PM spectrum for ω ≤ ω m τ = 00..07 09 for ω > ω m α = 0.076 X¯ −0.22 or 0.0081 if fetch independent X¯ = gX/Uw2 X = fetch length (nautical miles) Uw = wind speed (knots) ωm = 2π · 3.5 · (g/Uw )X¯ −0.33. Figure 13.5 depicts the JONSWAP spectrum for
0.626 rad/s, for threepeakedness parameters.
13.1.3
α = 0.0081 and ωm =
Approximation of Spectral Density from Time Series
From the time history of the wave elevation, the spectral density function can be obtained by two methods. The first method is to use the autocorrelation function
Rηη (τ ) = E {η (t) η (t + τ )} , which is related to the spectral Sηη (ω) by the WienerKhinchine relations (Equations 13.3
density function and 13.4).
1 4 Measurement of WindWave Growth and Swell Decay During the Joint North Sea Wave Project (JONSWAP) , K. Hasselmann, T.P. Barnett, E. Bouws, H. Carlson, D.E.
Cartwright, K. Enke, J.A. Ewing, H. Gienapp, D.E Hasselmann, P. Kruseman, A. Meerburg, P. Muller, D.J. Olbers, K. Richter, W. Sell, and H. Walden, Deutschen Hydrographischen Zeitschrift, Technical Report 13A, 1973.
© 2005 by Taylor & Francis Group, LLC
CHAPTER 13. FLUIDINDUCED VIBRATION
690
Figure 13.5: JONSWAP spectrum for
γ = 1.0, 3.3, and 7.0.
Assuming that the process is ergodic, the autocorrelation function for a given time history of length
1 Rˆηη (τ ) = T lim s →∞ Ts − τ where the notation
Ts can be approximated by Ts −τ 0
η (t) η (t + τ ) dt,
for
0
< τ < Ts ,
ˆ is used to emphasize that the variable is an approx
imation based on a sample time history of length
Ts .
Then, the spectral
density is obtained by taking the Fourier cosine transform of
1
Sˆηη (ω) = π
Ts 0
Rˆηη (τ ) cos ωτdτ.
Rˆηη (τ ) ,
(13.5)
The second method for obtaining the spectral density function is to use the relationship between the spectral density and the Fourier transform of the time series. They are related by
Sˆηη (ω) = T lim →∞ s
where
1
2πTs
Xˆ (ω) Xˆ ∗ (ω ) ,
Xˆ (ω) is given by Xˆ (ω) =
© 2005 by Taylor & Francis Group, LLC
Ts 0
η (t) exp (−iωt) dt,
(13.6)
13.1. OCEAN CURRENTS AND WAVES and
691
Xˆ ∗ (ω) is the complex conjugate, given by Xˆ ∗ (ω) =
Ts 0
η (t) exp (iωt) dt.
In order to obtain the Fourier transforms of the time series, the Discrete Fourier Transform (DFT) or the Fast Fourier Transform (FFT) procedure can be used. For detailed descriptions of how this is done, refer to Appendix 15 1 in Tucker. Spectral analysis is almost always carried out via FFTs because it is easier to use and faster than the formal method via correlation function. It should be noted that the length of the sample time history only needs to be long enough so that the limits converge. Taking a longer sample will not improve the accuracy of the estimate. Instead, one should take many samples or break one long sample into many parts.
For
n
samples, the
spectral densities are obtained for each sample time history using either Equation 13.5 or 13.6, and they are averaged to give the estimate. The determination of the spectral density from wave records depends on the details of the procedure, such as the length of the record, sampling interval, degree and type of filtering and smoothing, and time discretization.
13.1.4
Generation of Time Series from a Spectral Density
In a nonlinear analysis, the structural response is found by a numerical integration in time. Therefore, one needs to convert the wave elevation spectrum into an equivalent time history. The wave elevation can be represented as a sum of many sinusoidal functions with different angular frequencies and
η (t) as N
η (t) = cos(ω i t − ϕi ) 2Sηη (ωi ) ∆ω i ,
random phase angles. That is, write
(13.7)
i=1
where
ϕi
is a uniformly distributed random number between
are discrete sampling frequencies,
∆ωi =
ωi − ω i−1 , and N
0 and 2π,
ωi
is the number of
partitions. Recall that the area under the spectrum is equal to the variance,
σ2η . The incremental area under the spectrum, Sηη (ωi ) ∆ω i , can be denoted 2 as σ i such that the sum of all the incremental areas equals the variance of N 2 2 the wave elevation or σ η = σ . The time history can then be written i=1 i as
η (t) =
N
i=1
15
cos(ω i t
√ − ϕ ) 2σ . i
i
Waves in Ocean Engineering: Measurement, Analysis, Interpretation,
Ellis Horwood Limited, England, 1991.
© 2005 by Taylor & Francis Group, LLC
M. Tucker,
CHAPTER 13. FLUIDINDUCED VIBRATION
692
The sampling frequencies, ω i , can be chosen at equal intervals such that ω i = iω 1 . However, the time history will then have the lowest frequency of ω 1 and will have a period of T = 2π/ω1 . In order to avoid this unwanted 16 periodicity, Borgman suggested that the frequencies be chosen so that the
area under the spectrum curve for each interval is equal or
σ 2i = σ2 = σ 2η /N.
The time history is written as
η (t) =
2
N
ση
N
ωi t cos(¯
i=1
− ϕ ),
(13.8)
i
ω¯ i = (ω i + ω i−1 ) /2. The discrete frequencies ωi are chosen such that 0 < ω < ω i is equal to i/N of the total area under the curve between the interval 0 < ω < ω N or ωi i ωN Sηη (ω) dω for i = 1, ..., N, Sηη (ω ) dω = N 0 0 where
the area between the interval
where it is assumed that the area under the spectrum beyond ligible. If
η (t)
ωN
is neg
is a narrowband Gaussian process, the standard deviation
can be replaced by
ση
Hs /4, and the time history can be written as N Hs 2 cos(¯ ω i t − ϕi ). η(t) = N i=1 4 =
17 Shinozuka proposed that the sampling frequencies
ω¯ i
in Equation
f (ω ) ≡ o Sηη (ω) /σ2η . This is equivalent to performing an integration using the Monte Carlo method. The random frequencies ω distributed according to f (ω ) can −1 (x) , be obtained from uniformly distributed random numbers x by ω = F where F (ω ) is the cumulative distribution of f (ω ) . 13.8 should be randomly chosen according to the density function
The random frequencies obtained in this way are used in Equation 13.8 to generate a sample time series. It should be noted that many sample time histories should be obtained and averaged to synthesize a time history for use in numerical simulations.
13.1.5
ShortTerm Statistics
In discussing wave statistics, we often use the term, significant wave, to describe an irregular sea surface. The significant wave is not a physical wave 1 6 “Ocean
Wave Simulation for Engineering Design,” L. Borgman,
terways and Harbors Division, ASCE, 95:557583, 1969. 1 7 “Monte
Journal of the Wa
Carlo Solution of Structural Dynamics,” M. Shinozuka,
Structures, 2:855874, 1972.
© 2005 by Taylor & Francis Group, LLC
Computers and
13.1. OCEAN CURRENTS AND WAVES
693
that can be seen, but a statistical description of random waves. The concept 18 of significant wave height was first introduced by Sverdrup and Munk as the average height of the highest onethird of all waves. Usually ships cooperate in programs to find sea statistics by reporting a rough estimate of the storm severity in terms of an observed wave height. It is found that this observed wave height is consistently very close to the significant wave height. Two assumptions that are made in describing shortterm wave statistics are stationarity and ergodicity. These assumptions are valid only for “short” time intervals, on the order of a couple of hours or the duration of a storm, but not for weeks or years. The wave elevation is assumed to be weakly stationary so that its autocorrelation is a function of time lag only. As a result, the mean and the variance are constant, and the spectral density is invariant with time. Therefore, the significant wave height and the significant wave period are constant when considering short term statistics. In this case, the individual wave height and wave period are the stochastic variables. Consider a sample time history of a zero mean random process, as shown in Figure 13.6.
The questions that we ask are:
level (for example, distributed?
Z
how often is a certain
in the figure) exceeded, and how are the maxima
Equivalently, we can ask when can we expect to see that a
certain level is exceeded for the first time, and what are the values of the peaks of a random process. The first question is important for determining when a structure may fail due to a onetime excessive load, and the second question is important for establishing when a structure may fail due to cyclic loads.
Figure 13.6: A sample time history with maxima highlighted. 18 Wind,
Sea, and Swell: Theory of Relations for Forecasting , H. Sverdrup, W. Munk,
Technical Report 601, U.S. Navy Hydrographic Office, 1947.
© 2005 by Taylor & Francis Group, LLC
CHAPTER 13. FLUIDINDUCED VIBRATION
694
X (t) crosses the Z with a positive slope can be calculated using the relation,
It is found that the rate at which a random process random variable
ν z+ where
=
∞ 0
vfX X˙ (z, v) dv,
fX X˙ (x, x˙ ) is the joint probability density function of X
and
˙ The X.
expected time of the first upcrossing is then the inverse of the crossing rate, or
E {T } = 1/ν z + . The probability density function of the maxima,
A, can be calculated from
0
−∞ −wfX X˙ X¨ (a, 0, w) dw , 0 −∞ wfX˙ X¨ (0, w) dw
fA (a) = where
¨ X.
If
˙ and fX X˙ X¨ (x, x, ˙ x ¨) is the joint probability density function of X, X,
X (t) is a Gaussian process, the joint probability density functions are fX X˙ (x, x˙ )
−∞
=
1
2πσX σX˙
< x < ∞,
2 2 x 1 , − 12 σx˙ exp − 2 σ X
X˙
− ∞ < x˙ < ∞,
and
fX X˙ X¨ (x, x, ˙ x ¨) =
1
(2π)
3/2
M 
1 − 1 T exp − ({x} − {µX }) [M ] ({x} − {µX }) , 1/2 2
where
[M ]
{x} − {µX } © 2005 by Taylor & Francis Group, LLC
=
=
σ 2X 0
σ 2X˙
0
σ2X˙ 0
σ 2X˙ 0
σ 2X¨
x − µX x˙ − µX˙ . x¨ − µX¨
13.1. OCEAN CURRENTS AND WAVES
695
Then, for a stationary Gaussian process, the upcrossing rate is given by
ν+ z
= =
=
∞
fX X˙ (Z, x˙ ) x˙ dx. ˙
2 ∞ 2 x˙ 1 Z 1 x˙ dx˙ exp − exp − σ 2 2 σX 2πσ X σX˙ 0 X˙ 2 σX˙ Z 1 , exp − 2 σ 2πσ 0
1
X
X
and the probability density function of maxima is given by the Rice density 19 function,
√ −a2 1 − α2 √ exp fA (a) = 2σ2η (1 − α2 ) 2 2πση −a , aα α 2 exp +a 2Φ 2σ2η ση ση (α − 1)
where
Φ (x) is the cumulative distribution function of the standard normal
random variable,
Φ (x) =
x
1
√
−∞
2π
exp
−z2/2 dz,
α is the irregularity factor, equivalent to the ratio of the number of zero η (t) crosses zero with a positive slope) to the number of peaks. α ranges from 0 to 1, and it is also equal to and
upcrossings (number of times that
α= If
X (t)
σ2η˙ . σ η σ η¨
α
is a broadband process,
= 0 and the Rice distribution is
reduced to the Gaussian probability density function,
fA (a) = If
X (t) is
whenever
√1
2πση
exp
− a2 2σ 2η
for
− ∞ < a < ∞.
a narrowband process, it is guaranteed that it will have a peak
η (t) crosses its mean.
In this case, the irregularity factor is close
to unity, and the Rice distribution is reduced to the Rayleigh probability density function given by
fA (a) = 1 9 Mathematical
a σ 2η
exp
2
− 12 σa2
η
for
0
< a < ∞.
Analysis of Random Noise , S. Rice, Dover Publications, 1954.
© 2005 by Taylor & Francis Group, LLC
CHAPTER 13. FLUIDINDUCED VIBRATION
696
In other words, the amplitudes of a
narrowband stationary Gaussian process
are distributed according to the Rayleigh distribution.
Figure 13.7: Rice distribution for maxima.
Figure 13.7 shows the Rice distribution for various values of
α. Note that
the Rice distribution includes both positive and negative maxima except when
α = 1,
in which case all the maxima are positive.
maxima are the local maxima that occur above the mean of
The positive
X (t),
and the
negative maxima are the local maxima that occur below the mean, as shown in Figure 13.6. In some cases, the negative maxima may not have physical meaning. In those cases, we can use the truncated Rice distribution, where
fA (a) is used. fA (a) is normalized by the area 20 ,21 under the probability density for positive maxima,
only the positive portion of
fAtrunc (a) =
∞fA (a) 0
fA (a) da
a ≥ 0.
,
The truncated Rice distribution is shown in Figure 13.8. If
X (t)
is the wave elevation, its maxima,
2 0 “On
A,
are the amplitudes of the
the Statistical Distribution of the Height of Sea Waves,” M. LonguetHiggins, 11(3):245266, 1952. 2 1 “On Prediction of Extreme Values,” M. Ochi, Journal of Ship Research , 17:2937, 1973.
Journal of Marine Research ,
© 2005 by Taylor & Francis Group, LLC
13.1. OCEAN CURRENTS AND WAVES
697
Figure 13.8: Truncated Rice distributions for various
H = 2A,
wave elevation. The wave height,
dA dH
fH (h) = fA (H/2)
h exp 4σ2η
=
2
− 12 4hσ2
η
α.
is then distributed according to
for
0