allies afte ProceduresFinite Element Procedures Klaus-Jiirgen Bathe Professor of Mechanical Engineering Massachusetts Institute of Technology PRENTICE HALL, Upper Saddle River, New Jersey 07458Library of Congress Cataloging-in-Publication Data ‘Baru, KLAvs-JORGEN. Finite element procedures / Klaus-Jirgen Bathe. p.m, Revision of: Finite element procedures in engineering analysis. 1982. Includes bibliographical references and index. ISBN 0-13-301858-4 1 Finite element method. 2. Engineering mathematics. 1, Bathe, Klaus-Jirgen. Finite element procedures in engineering analysis. IL Tie, TA347.F5B36 1996 620'.001'51535—ae20 95-1231 cP ‘Acquisitions editor: William Stenquist Production editor: Merrill Peterson ‘Cover designer: Marianne Frasco Buyer: William Scazzero Eadtoria! assistant: Meg Weist ‘The author and publisher of this book have used their best efforts in preparing this book. These efforts include the development, research, and testing of the theories and educational computer programs given in the book. The author and publisher make no warranty of any kind, expressed or implied, with regard to ‘any text and the programs contained in this book. The author and publisher shall not be liable in any event for incidental or consequential damages in connection with, or arising out of, the furnishing, performance, or use of this text and these programs. BeBe © 1996 by Prentice-Hall, Inc. AE= Simon & Schuster / A Viacom Company Upper Saddle River, New Jersey 07458 All rights reserved. No part of this book may be reproduced, in any form or by any means, ‘without permission in writing from the publisher. Printed in the United States of America 10 976: 7 65 4 ISBN O-13-301458-4 PRENTICE-HALL INTERNATIONAL (UK) LiMiTED, London PRENTICE-HALL OF AUSTRALIA Pry. LIMITED, Sydney PRENTICE-HALL CANADA INC., Toronto PRENTICE-HALL HiSPANOAMERICANA, S.A., Mexico PRENTICE-HALL OF INDIA PRivaTE LIMITED, New Delhi PRENTICE-HALL OF JAPAN, INC., Tokyo SIMON a ScHuSTER Asta PTE. LrD., Singapore EDITORA PRENTICE-HALL DO BRASIL, LTDA., Rio de JaneiroTo my students . . . Progress in design of new structures seems to be unlimited. Last sentence of article: “The Use of the Electronic Computer in Structural Analysis,” by K. J. Bathe (undergraduate student), published in Impact, Journal of the University of Cape Town Engineering Society, pp. 57— 61, 1967.Contents Preface CHAPTER ONE An Introduction to the Use of Fi te Element Procedures 1 1.1 Introduction 1 1.2 Physical Problems, Mathematical Models, and the Finite Element Solution 2 1.3 Finite Element Analysis as an Integral Part of Computer-Aided Design 11 1.4 A Proposal on How to Study Finite Element Methods 14 CHAPTER TWO — Vectors, Matrices, and Tensors 7 2.1 Introduction 17 2.2 Introduction to Matrices. 18 2.3 Vector Spaces 34 2.4 Definition of Tensors 40 2.5 The Symmetric Eigenproblem Av = Av 51 2.6 The Rayleigh Quotient and the Minimax Characterization of Eigenvalues 60 2.7 Vector and Matrix Norms 66 28 Exercises 72vi Contents CHAPTER THREE Some Basic Concepts of Engineering Analysis and an Introduction to the Finite Element Method 7 3. Introduction 77 3.2 Solution of Discrete-System Mathematical Models 78 3.2.1 Steady-State Problems, 78 3.2.2. Propagation Problems, 87 3.2.3 Eigenvalue Problems, 90 3.2.4 On the Nature of Solutions, 96 3.2.5 Exercises, 101 3.3. Solution of Continuous-System Mathematical Models 105 3.3.1 Differential Formulation, 105 3.3.2. Variational Formulations, 110 3.3.3 Weighted Residual Methods; Ritz Method, 116 3.3.4 An Overview: The Differential and Galerkin Formulations, the Principle of Virtual Displacements, and an Introduction to the Finite Element Solution, 124 3.3.5 Finite Difference Differential and Energy Methods, 129 3.3.6 Exercises, 138 3.4 Imposition of Constraints 143 3.4.1 An Introduction to Lagrange Multiplier and Penalty Methods, 143 3.4.2 Exercises, 146 CHAPTER FOUR Formulation of the Element Method—Linear Analysis in Sol and Structural Mechanics 148 41 Introduction 148 4.2 Formulation of the Displacement-Based Finite Element Method 149 4.2.1 General Derivation of Finite Element Equilibrium Equations, 153 Imposition of Displacement Boundary Conditions, 187 Generalized Coordinate Models for Specific Problems, 193 Lumping of Structure Properties and Loads, 212 Exercises, 214 4.3 Convergence of Analysis Results 225 The Model Problem and a Definition of Convergence, 225 Criteria for Monotonic Convergence, 229 The Monotonically Convergent Finite Element Solution: A Ritz Solution, 234 Properties of the Finite Element Solution, 236 Rate of Convergence, 244 Calculation of Stresses and the Assessment of Error, 254 Exercises, 259 4.4 Incompatible and Mixed Finite Element Models 261 Incompatible Displacement-Based Models, 262 Mixed Formulations, 268 Mixed Interpolation—Displacement/Pressure Formulations for Incompressible Analysis, 276 4.4.4 Exercises, 296Contents vil 4.5 The Inf-Sup Condition for Analysis of Incompressible Media and Structural Problems 300 4.5.1 The Inf-Sup Condition Derived from Convergence Considerations, 301 4.5.2 The Inf-Sup Condition Derived from the Matrix Equations, 312 4.5.3 The Constant (Physical) Pressure Mode, 315 4.5.4 Spurious Pressure Modes—The Case of Total Incompressibility, 316 4.5.5. Spurious Pressure Modes—The Case of Near Incompressibility, 318 4.5.6 The Inf-Sup Test, 322 4.5.7. An Application to Structural Elements: The Isoparametric Beam Elements, 330 458 Exercises, 335 CHAPTER FIVE Formulation and Calculation of Isoparametric Fi ae Aes 5.3 5.4 5.5 es te Element Matrices 338 Introduction 338 Isoparametric Derivation of Bar Element Stiffness Matrix 339 Formulation of Continuum Elements 341 5.3.1 Quadrilateral Elements, 342 5.3.2 Triangular Elements, 363 5.3.3 Convergence Considerations, 376 5.3.4 Element Matrices in Global Coordinate System, 386 5.3.5 Displacement/Pressure Based Elements for Incompressible Media, 388 5.3.6 Exercises, 389 Formulation of Structural Elements 397 5.4.1 Beam and Axisymmetric Shell Elements, 399 5.4.2 Plate and General Shell Elements, 420 5.4.3 Exercises, 450 Numerical Integration 455 5.5.1 Interpolation Using a Polynomial, 456 5.5.2 The Newton-Cotes Formulas (One-Dimensional Integration), 457 5.5.3 The Gauss Formulas (One-Dimensional Integration), 461 5.5.4 Integrations in Two and Three Dimensions, 464 5.5.5 Appropriate Order of Numerical Integration, 465 5.5.6 Reduced and Selective Integration, 476 5.5.7 Exercises, 478 Computer Program Implementation of Isoparametric Finite Elements 480 CHAPTER SIX —___ Finite Element Nonlinear Analysis in Solid and Structural Mechanics 485 an 6.2 Introduction to Nonlinear Analysis 485 Formulation of the Continuum Mechanics Incremental Equations of Motion 497 6.2.1 The Basic Problem, 498 6.2.2 The Deformation Gradient, Strain, and Stress Tensors, 502vii 6.3 6.4 65 6.6 6.7 6.8 CHAPTER SEVEN Finite Element Analy: Contents 6.2.3 Continuum Mechanics Incremental Total and Updated Lagrangian Formulations, Materially-Nonlinear-Only Analysis, 522 6.2.4 Exercises, 529 Displacement-Based Isoparametric Continuum Finite Elements 538 6.3.1 Linearization of the Principle of Virtual Work with Respect to Finite Element Variables, 538 6.3.2 General Matrix Equations of Displacement-Based Continuum Elements, 540 6.3.3 Truss and Cable Elements, 543 6.3.4 Two-Dimensional Axisymmetric, Plane Strain, and Plane Stress Elements, 549 6.3.5 Three-Dimensional Solid Elements, 555 6.3.6 Exercises, 557 Displacement/Pressure Formulations for Large Deformations 561 64.1 Total Lagrangian Formulation, 561 6.4.2. Updated Lagrangian Formulation, 565 6.43 Exercises, 566 Structural Elements 568 6.5.1 Beam and Axisymmetric Shell Elements, 568 6.5.2 Plate and General Shell Elements, 575 65.3 Exercises, 578 Use of Constitutive Relations 581 6.6.1 Elastic Material Behavior—Generalization of Hooke’s Law, 583 6.6.2 Rubberlike Material Behavior, 592 6.6.3 Inelastic Material Behavior; Elastoplasticity, Creep, and Viscoplasticity, 595 6.64 Large Strain Elastoplasticity, 612 6.6.5 Exercises, 617 Contact Conditions 622 6.7.1 Continuum Mechanics Equations, 622 6.7.2 A Solution Approach for Contact Problems: The Constraint Function Method, 626 6.7.3 Exercises, 628 Some Practical Considerations 628 6.8.1 The General Approach to Nonlinear Analysis, 629 6.8.2 Collapse and Buckling Analyses, 630 6.8.3 The Effects of Element Distortions, 636 6.8.4 The Effects of Order of Numerical Integration, 637 6.8.5 Exercises, 640 of Heat Transfer, Field Problems, and incompressible Fluid Flows 642 7A 72 Introduction 642 Heat Transfer Analysis 642 7.2.1 Governing Heat Transfer Equations, 642 7.2.2 Incremental Equations, 646 7.2.3 Finite Element Discretization of Heat Transfer Equations, 651 7.24 Exercises, 659Contents x 73 TA Analysis of Field Problems 661 7.3.1 Seepage, 662 7.3.2 Incompressible Inviscid Flow, 663 7.3.3 Torsion, 664 7.3.4 Acoustic Fluid, 666 7.3.5. Exercises, 670 Analysis of Viscous Incompressible Fluid Flows 671 7.4.1 Continuum Mechanics Equations, 675 74.2. Finite Element Governing Equations, 677 74.3 High Reynolds and High Peclet Number Flows, 682 74.4 Exercises, 691 CHAPTER EIGHT — Solution of Equilibrium Equations in Static Analysis 695 8.1 8.2 83 84 CHAPTER NINE Introduction 695 Direct Solutions Using Algorithms Based on Gauss Elimination 696 8.2.1 Introduction to Gauss Elimination, 697 8.2.2. The LDL’ Solution, 705 8.2.3 Computer Implementation of Gauss Elimination—The Active Column Solution, 708 82.4 Cholesky Factorization, Static Condensation, Substructures, and Frontal Solution, 717 82.5 Positive Definiteness, Positive Semidefiniteness, and the Sturm Sequence Property, 726 8.2.6 Solution Errors, 734 8.2.7 Exercises, 741 Iterative Solution Methods 745 83.1 The Gauss-Seidel Method, 747 83.2 Conjugate Gradient Method with Preconditioning, 749 83.3 Exercises, 752 Solution of Nonlinear Equations 754 84.1 Newton-Raphson Schemes, 755 84.2 The BFGS Method, 759 84.3 Load-Displacement-Constraint Methods, 761 84.4 Convergence Criteria, 764 84.5 Exercises, 765 Solution of Equilibrium Equations in Dynamic Analysis 768 9.1 9.2 Introduction 768 Direct Integration Methods 769 9.2.1 The Central Difference Method, 770 9.2.2 The Houbolt Method, 774 9.2.3, The Wilson @ Method, 777Contents 9.2.4 The Newmark Method, 780 9.2.5 The Coupling of Different Integration Operators, 782 9.2.6 Exercises, 784 9.3 Mode Superposition 785 9.3.1 Change of Basis to Modal Generalized Displacements, 785 9.3.2. Analysis with Damping Neglected, 789 9.3.3 Analysis with Damping Included, 796 9.3.4 Exercises, 801 94 Analysis of Direct Integration Methods 801 9.4.1 Direct Integration Approximation and Load Operators, 803 9.4.2 Stability Analysis, 806 9.4.3 Accuracy Analysis, 810 9.4.4 Some Practical Considerations, 813 9.4.5 Exercises, 822 9.5 Solution of Nonlinear Equations in Dynamic Analysis 824 9.5.1 Explicit Integration, 824 9.5.2 Implicit Integration, 826 9.5.3 Solution Using Mode Superposition, 828 9.5.4 Exercises, 829 9.6 Solution of Nonstructural Problems; Heat Transfer and Fluid Flows 830 9.6.1 The a-Method of Time Integration, 830 9.6.2 Exercises, 836 CHAPTER TEN Preliminaries to the Solution of Eigenproblems 838 10.1 Introduction 838 10.2 Fundamental Facts Used in the Solution of Eigensystems 840 10.2.1 Properties of the Eigenvectors, 841 10.2.2 The Characteristic Polynomials of the Eigenproblem Keb = AM@ and of Its Associated Constraint Problems, 845 Shifting, 851 Effect of Zero Mass, 852 Transformation of the Generalized Eigenproblem Ke = AM¢ to a Standard Form, 854 10.26 Exercises, 860 10.3 Approximate Solution Techniques 861 10.3.1 Static Condensation, 861 10.3.2, Rayleigh-Ritz Analysis, 868 10.3.3 Component Mode Synthesis, 875 10.3.4 Exercises, 879 10.4 Solution Errors 880 10.4.1 Error Bounds, 880 10.4.2 Exercises, 886Contents xi CHAPTER ELEVEN Solution Methods for Wd 11.2 1L3 11.4 15 11.6 jenproblems 887 Introduction 887 Vector Iteration Methods 889 Inverse Iteration, 890 Forward Iteration, 897 Shifting in Vector Iteration, 899 Rayleigh Quotient Iteration, 904 Matrix Deflation and Gram-Schmidt Orthogonalization, 906 Some Practical Considerations Concerning Vector Iterations, 909 Exercises, 910 Transformation Methods 911 11.3.1 The Jacobi Method, 912 11.3.2. The Generalized Jacobi Method, 919 11.3.3 The Householder-QR- Inverse Iteration Solution, 927 11.3.4 Exercises, 937 Polynomial Iterations and Sturm Sequence Techniques 938 11.4.1 Explicit Polynomial Iteration, 938 11.4.2. Implicit Polynomial Iteration, 939 11.4.3 Iteration Based on the Sturm Sequence Property, 943 11.4.4 Exercises, 945 The Lanczos Iteration Method 945 11.5.1 The Lanczos Transformation, 946 41.5.2 Iteration with Lanczos Transformations, 950 11.5.3 Exercises, 953 The Subspace Iteration Method 954 11.6.1 Preliminary Considerations, 955 11.6.2 Subspace Iteration, 958 11.6.3 Starting Iteration Vectors, 960 11.6.4 Convergence, 963 11.6.5 Implementation of the Subspace Iteration Method, 964 11.6.6 Exercises, 978 CHAPTER TWELVE ———__ Implementation of the Finite Element Method 979 12.1 bes 12.3 Introduction 979 Computer Program Organization for Calculation of System Matrices 980 12.2.1 Nodal Point and Element Information Read-in, 981 12.2.2 Calculation of Element Stiffness, Mass, and Equivalent Nodal Loads, 983 12.2.3 Assemblage of Matrices, 983 Calculation of Element Stresses 987xii Contents 12.4 Example Program STAP 988 12.4.1 Data Input to Computer Program STAP, 988 12.42 Listing of Program STAP, 995 12.5 Exercises and Projects 1009 References 1013 Index 1029Preface Finite element procedures are now an important and frequently indispensable part of engineering analysis and design. Finite element computer programs are now widely used in practically all branches of engineering for the analysis of structures, solids, and fluids. My objective in writing this book was to provide a text for upper-level undergraduate and graduate courses on finite element analysis and to provide a book for self-study by engineers and scientists, With this objective in mind, I have developed this book from my earlier publication Finite Element Procedures in Engineering Analysis (Prentice-Hall, 1982). I have kept the same mode of presentation but have consolidated, updated, and strengthened the earlier writing to the current state of finite element developments. Also, I have added new sections, both to cover some important additional topics for completeness of the presentation and to facilitate (through exercises) the teaching of the material discussed in the book. This text does not present a survey of finite element methods. For such an endeavor, anumber of volumes would be needed. Instead, this book concentrates only on certain finite element procedures, namely, on techniques that I consider very useful in engineering practice and that will probably be employed for many years to come. Also, these methods are introduced in such a way that they can be taught effectively—and in an exciting manner—to students. An important aspect of a finite element procedure is its reliability, so that the method can be used in a confident manner in computer-aided design. This book emphasizes this point throughout the presentations and concentrates on finite element procedures that are general and reliable for engineering analysis. Hence, this book is clearly biased in that it presents only certain finite element procedures and in that it presents these procedures in a certain manner. In this regard, the book refiects my philosophy toward the teaching and the use of finite element methods. xiiixiv Preface While the basic topics of this book focus on mathematical methods, an exciting and thorough understanding of finite element procedures for engineering applications is achieved only if sufficient attention is given to both the physical and mathematical charac- teristics of the procedures. The combined physical and mathematical understanding greatly enriches our confident use and further development of finite element methods and is there- fore emphasized in this text. These thoughts also indicate that a collaboration between engineers and mathemati- cians to deepen our understanding of finite element methods and to further advance in the fields of research can be of great benefit. Indeed, I am thankful to the mathematician Franco Brezzi for our research collaboration in this spirit, and for his valuable suggestions regard- ing this book. Tconsider it one of the greatest achievements for an educator to write a valuable book. In these times, all fields of engineering are rapidly changing, and new books for students are needed in practically all areas of engineering. I am therefore grateful that the Mechanical Engineering Department of M.I.T. has provided me with an excellent environment in which to pursue my interests in teaching, research, and scholarly writing. While it required an immense effort on my part to write this book, I wanted to accomplish this task as a commitment to my past and future students, to any educators and researchers who might have an interest in the work, and, of course, to improve upon my teaching at M.I.T. 1 have been truly fortunate to work with many outstanding students at M.I.T., for which I am very thankful. It has been a great privilege to be their teacher and work with them. Of much value has also been that I have been intimately involved, at my company ADINA R & D, Inc., in the development of finite clement methods for industry. This involvement has been very beneficial in my teaching and research, and in my writing of this ‘book, A text of significant depth and breadth on a subject that came to life only a few decades ago and that has experienced tremendous advances, can be written only by an author who has had the benefit of interacting with many people in the field. I would like to thank all my students and friends who contributed—and will continue to contribute—to my knowledge and understanding of finite element methods. My interaction with them has given me great joy and satisfaction, I also would like to thank my secretary, Kristan Raymond, for her special efforts in typing the manuscript of this text. Finally, truly unbounded thanks are due to my wife, Zorka, and children, Ingrid and Mark, who, with their love and their understanding of my efforts, supported me in writing this book, K. J. BatheMl CHAPTER ONE Se An Introduction to the Use of Finite Element Procedures 1.1 INTRODUCTION Finite element procedures are at present very widely used in engineering analysis, and we can expect this use to increase significantly in the years to come. The procedures are employed extensively in the analysis of solids and structures and of heat transfer and fluids, and indeed, finite element methods are useful in virtually every field of engineering analysis. The development of finite element methods for the solution of practical engineering problems began with the advent of the digital computer. That is, the essence of a finite element solution of an engineering problem is that a set of governing algebraic equations is established and solved, and it was only through the use of the digital computer that this process could be rendered effective and given general applicability. These two properties— effectiveness and general applicability in engineering analysis—are inherent in the theory used and have been developed to a high degree for practical computations, so that finite element methods have found wide appeal in engineering practice. As is often the case with original developments, it is rather difficult to quote an exact “date of invention,” but the roots of the finite element method can be traced back to three separate research groups: applied mathematicians—see R. Courant [A]; physicists—see J. L. Synge [A]; and engineers—see J. H. Argyris and S. Kelsey [A]. Although in principle published already, the finite element method obtained its real impetus from the develop- ments of engineers. The original contributions appeared in the papers by J. H. Argyris and S. Kelsey [A]; M. J. Turner, R. W. Clough, H. C. Martin, and L. J. Topp [A]; and R. W. Clough [A]. The name “finite element” was coined in the paper by R. W. Clough [A]. Important early contributions were those of J. H. Argyris [A] and O. C. Zienkiewicz and Y. K. Cheung [A]. Since the early 1960s, a large amount of research has been devoted to the technique, and a very large number of publications on the finite element method is 12 An Introduction to the Use of Finite Element Procedures = Chap. 1 available (see, for example, the compilation of references by A. K. Noor [A] and the Finite Element Handbook edited by H. Kardestuncer and D. H. Norrie [A]). The finite element method in engineering was initially developed on a physical basis for the analysis of problems in structural mechanics. However, it was soon recognized that the technique could be applied equally well to the solution of many other classes of problems. The objective of this book is to present finite element procedures comprehen- sively and in a broad context for solids and structures, field problems (specifically heat transfer), and fluid flows. . To introduce the topics of this book we consider three important items in the following sections of this chapter. First, we discuss the important point that in any analysis we always select a mathematical model of a physical problem, and then we solve that model. The finite element method is employed to solve very complex mathematical models, but it is important to realize that the finite element solution can never give more information than that contained in the mathematical model. Then we discuss the importance of finite element analysis in the complete process of computer-aided design (CAD). This is where finite element analysis procedures have their greatest utility and where an engineer is most likely to encounter the use of finite element methods. In the last section of this chapter we address the question of how to study finite element methods. Since a voluminous amount of information has been published on these tech- niques, it can be rather difficult for an engineer to identify and concentrate on the most important principles and procedures. Our aim in this section is to give the reader some guidance in studying finite element analysis procedures and of course also in studying the various topics discussed in this book. 1.2 PHYSICAL PROBLEMS, MATHEMATICAL MODELS, AND THE FINITE ELEMENT SOLUTION The finite element method is used to solve physical problems in engineering analysis and design. Figure 1.1 summarizes the process of finite element analysis. The physical problem typically involves an actual structure or structural component subjected to certain loads. The idealization of the physical problem to a mathematical model requires certain assump- tions that together lead to differential equations governing the mathematical model (see Chapter 3). The finite element analysis solves this mathematical model. Since the finite element solution technique is a numerical procedure, it is necessary to assess the solution accuracy, If the accuracy criteria are not met, the numerical (i.e., finite element) solution has to be repeated with refined solution parameters (such as finer meshes) until a sufficient accuracy is reached. Tt is clear that the finite element solution will solve only the selected mathematical model and that all assumptions in this model will be reflected in the predicted response. We cannot expect any more information in the prediction of physical phenomena than the information contained in the mathematical model. Hence the choice of an appropriate mathematical model is crucial and completely determines the insight into the actual physical problem that we can obtain by the analysis,Sec. 1.2 Physical Problems, Mathematical Models, the Finite Element Solution 3 ntial equations Assumptions on Geometry + Kinematic + Meterial law + Loading + Boundary conditions Ete. Finite element solution Choice of elements: + Mesh density Finite + Solution parameters element Representation of solution + Loading of + Boundery conditions mathematicel * Etc, model Figure 1.1 The process of finite element analysis, Let us emphasize that, by our analysis, we can of course only obtain insight into the physical problem considered: we cannot predict the response of the physical problem exactly because it is impossible to reproduce even in the most refined mathematical model all the information that is present in nature and therefore contained in the physical problem. Once a mathematical model has been solved accurately and the results have been interpreted, we may well decide to consider next a refined mathematical model in order to increase our insight into the response of the physical problem. Furthermore, a change in the physical problem may be necessary, and this in turn will also lead to additional mathemat- ical models and finite element solutions (see Fig. 1.1). The key step in engineering analysis is therefore choosing appropriate mathematical models. These models will clearly be selected depending on what phenomena are to be4 An Introduction to the Use of Finite Element Procedures Chap. 1 predicted, and it is most important to select mathematical models that are reliable and effective in predicting the quantities sought. To define the reliability and effectiveness of a chosen model we think of a very- comprehensive mathematical model of the physical problem and measure the response of our chosen model against the response of the comprehensive model. In general, the very- comprehensive mathematical model is a fully three-dimensional description that also in- cludes nonlinear effects. Effectiveness of a mathematical model The most effective mathematical model for the analysis is surely that one which yields the required response to a sufficient accuracy and at least cost. Reliability of a mathematical model The chosen mathematical model is reliable if the required response is known to be predicted within a selected level of accuracy measured on the response of the very- comprehensive mathematical model. Hence to assess the results obtained by the solution of a chosen mathematical model, it may be necessary to also solve higher-order mathematical models, and we may well think of (but of course not necessarily solve) a sequence of mathematical models that include increasingly more complex effects. For example, a beam structure (using engineering termi- nology) may first be analyzed using Bernoulli beam theory, then Timoshenko beam theory, then two-dimensional plane stress theory, and finally using a fully three-dimensional continuum model, and in each case nonlinear effects may be included. Such a sequence of models is referred to as a hierarchy of models (see K. J. Bathe, N. S. Lee, and M. L. Bucalem [A}. Clearly, with these hierarchical models the analysis will include ever more complex response effects but will also lead to increasingly more costly solutions. As is well known, a fully three-dimensional analysis is about an order of magnitude more expensive (in computer resources and engineering time used) than a two-dimensional solution. Let us consider a simple example to illustrate these ideas. Figure 1.2(a) shows a bracket used to support a vertical load. For the analysis, we need to choose a mathematical model. This choice must clearly depend on what phenomena are to be predicted and on the geometry, material properties, loading, and support conditions of the bracket. We have indicated in Fig. 1.2(a) that the bracket is fastened to a very thick steel column. The description “very thick” is of course relative to the thickness f and the height hi of the bracket. We translate this statement into the assumption that the bracket is fastened to a (practically) rigid column, Hence we can focus our attention on the bracket by applying a “rigid column boundary condition” to it. (Of course, at a later time, an analysis of the column may be required, and then the loads carried by the two bolts, as a consequence of the load W, will need to be applied to the column.) We also assume that the load W is applied very slowly. The condition of time “very slowly” is relative to the largest natural period of the bracket; that is, the time span over which the load W is increased from zero to its full value is much longer than the fundamen- tal period of the bracket. We translate this statement into requiring a static analysis (and not, a dynamic analysis).Sec. 1.2 Physical Problems, Mathematice! Models, the Finite Element Solution 5 With these preliminary considerations we can now establish’an appropriate mathe- matical model for the analysis of the bracket—depending on what phenomena are to be predicted. Let us assume, in the first instance, that only the total bending moment at section AA in the bracket and the deflection at the load application are sought. To predict these quantities, we consider a beam mathematical model including shear deformations [see Fig. 1.2(b)] and obtain M = WL = 27,500 Nm an Slomaw = LE + nd? , WL + m9) oe! 3AG a2) = 0.053 cm Uniform thickness t We 1000 N Two bolts tiasen ty= 050m E = 2x 107 Nlcm? v =03 h =6.0cm t =0.40m Pin (a) Physical problem of steel bracket 1A | Me ry 0.50m L+ y= 280m {b) Beam model Figure 1.2 Bracket to be analyzed and two mathematical models6 An Introduction to the Use of Finite Element Procedures Chap. 1 ‘Areas with imposed zero displecements u, v nv Load applied at point B 2w xu Equilibrium equations (see Example 4.2) Sts 4 My ax Gy Be, Ay Se 4 oy ax ay Tan = 0, Tm = 0 on surfaces except at point B and at imposed zero displacements Stress-strain relation (see Table 4.3): lop 0 Ex vi 0 & 0 0 (— »)/2\L Ye. E = Young's modulus, v = Poisson's ratio 0 in domain of bracket Strain-displacement relations (see Section 4.2): =, 5 =H, ae f= 5,0 Sn = yi Y= ay oe (c) Plene stress model Figure 1.2 (continued) where L and ry ate given in Fig. 1.2(a), Eis the Young's modulus of the steel used, G is the shear modulus, / is the moment of inertia of the bracket arm (I = 7h), A is the cross- sectional area (A = ht), and the factor } is a shear correction factor (see Section 5.4.1). Of course, the relations in (1.1) and (1.2) assume linear elastic infinitesimal displace- ment conditions, and hence the load must not be so large as to cause yielding of the material and/or large displacements. Let us now ask whether the mathematical model used in Fig. 1.2(b) was reliable and effective. To answer this question, strictly, we should consider a very-comprehensive math- ematical model, which in this case would be a fully three-dimensional representation of theSec. 1.2 Physical Problems, Mathematical Models, the Finite Element Solution 7 full bracket. This model should include the two bolts fastening the bracket to the (assumed rigid) column as well as the pin through which the load Wis applied. The three-dimensional solution of this model using the appropriate geometry and material data would give the numbers against which we would compare the answers given in (1.1) and (1.2). Note that this three-dimensional mathematical model contains contact conditions (the contact is between the bolts, the bracket, and the column, and between the pin carrying the load and the bracket) and stress concentrations in the fillets and at the holes. Also, if the stresses are high, nonlinear material conditions should be included in the model. Of course, an analytical solution of this mathematical model is not available, and all we can obtain is a numerical solution. We describe in this book how such solutions can be calculated using finite element procedures, but we may note here already that the solution would be rela- tively expensive in terms of computer resources and engineering time used. Since the three-dimensional comprehensive mathematical model is very likely too comprehensive a model (for the analysis questions we have posed), we instead may consider a linear elastic two-dimensional plane stress model as shown in Fig. 1.2(c). This mathemat- ical model represents the geometry of the bracket more accurately than the beam model and assumes a two-dimensional stress situation in the bracket (see Section 4.2). The bending moment at section AA and deflection under the load calculated with this model can be expected to be quite close to those calculated with the very-comprehensive three- dimensional model, and certainly this two-dimensional model represents a higher-order model against which we can measure the adequacy of the results given in (1.1) and (1.2). Of course, an analytical solution of the model is not available, and a numerical solution must be sought. Figures 1.3(a) to (¢) show the geometry and the finite element discretization used in the solution of the plane stress mathematical model and some stress and displacement results obtained with this discretization. Let us note the various assumptions of this math- ematical model when compared to the more comprehensive three-dimensional model dis- cussed earlier. Since a plane stress condition is assumed, the only nonzero stresses are Trx, ‘yy. and Try. Hence we assume that the stresses 7.2, Tye, and 7,.are zero. Also, the actual bolt fastening and contact conditions between the steel column and the bracket are not included (a) Geometry of bracket as obtained from a CAD program Figure 1.3 Plane stress analysis of bracket in Fig. 1.2. AutoCAD was used to create the geometry, and ADINA was used for the finite element analysis.(b) Mesh of nine-node elements used in finite element dis- cretization (c) Deflected shape. Deflections are drawn with a magnifi- cation factor of 100 together with the original configura- tion {d) Maximum principal stress near notch. Un- (e) Maximum principal stress near notch. smoothed stress results are shown. The small ‘Smoothed stress results. (The averages of breaks in the bands indicate that a reasonably nodal point stresses are taken and interpo- accurate solution of the mathematical mode! lated over the elements.) has been obtained (see Section 4.3.6) re 13 (continued)Sec. 1.2 Physical Problems, Mathematics! Mod , the Finite Element Solution 9 in the model, and the pin carrying the load into the bracket is not modeled. However, since our objective is only to predict the bending moment at section AA and the deflection at point B, these assumptions are deemed reasonable and of relatively little influence. Let us assume that the results obtained in the finite element solution of the mathernat- ical model are sufficiently accurate that we can refer to the solution given in Fig. 1.3. as the solution of the plane stress mathematical model. Figure 1.3(C) shows the calculated deformed configuration. The deflection at the point of load application B as predicted in the plane stress solution is Blacinaw = 0.064 cm (1.3) Also, the total bending moment predicted at section AA is M|s-o = 27,500 N cm (1a) Whereas the same magnitude of bending moment at section AA is predicted by the beam model and the plane stress model,' the deflection of the beam model is considerably Jess than that predicted by the plane stress model [because of the assumption that the beam in Fig. 1.2(b) is supported rigidly at its left end, which neglects any deformation between the beam end and the bolts]. Considering these results, we can say that the beam mathematical model in Fig. 1.2(b) is reliable if the required bending moment is to be predicted within 1 percent and the deflection is to be predicted only within 20 precent accuracy. The beam model is of course also effective because the calculations are performed with very little effort. On the other hand, if we next ask for the maximum stress in the bracket, then the simple mathematical beam model in Fig. 1.2(b) will not yield a sufficiently accurate answer. Specifically, the beam model totally neglects the stress increase due to the fillets.? Hence a plane stress solution including the fillets is necessary. The important points to note here are the following. 1, The selection of the mathematical model must depend on the response to be predicted (ie., on the questions asked of nature). 2, The most effective mathematical model is that one which delivers the answers to the questions in a reliable manner (i.e., within an acceptable error) with the least amount of effort. 3. A finite element solution can solve accurately only the chosen mathematical model (e-g., the beam model or the plane stress model in Fig. 1.2) and cannot predict any more information than that contained in the model. 4. The notion of reliability of the mathematical model hinges upon an accuracy assess- ment of the results obtained with the chosen mathematical model (in response to the questions asked) against the results obtained with the very-comprehensive mathemat- ical model. However, in practice the very-comprehensive mathematical model is “The bending moment at section AA in the plane stress mode! is calculated here from the finite element nodal point forces, and for this statically determinate analysis problem the internal resisting moment must be equal to the externally applied moment (see Example 4.9). 20f course, the effect of the filets could be estimated by the use of stress concentration factors that have been established from plane stress solutions.10 An Introduction to the Use of Finite Element Procedures Chap. 1 usually not solved, and instead engineering experience is used, or a more refined ‘mathematical model is solved, to judge whether the mathematical model used was adequate (i.e., reliable) for the response to be predicted. Finally, there is one further important general point. The chosen mathematical model may contain extremely high stresses because of sharp corners, concentrated loads, or other effects. These high stresses may be due solely to the simplifications used in the model when compared with the very-comprehensive mathematical model (or with nature). For example, the concentrated load in the plane stress model in Fig. 1.2(c) is an idealization of a pressure load over a small area. (This pressure would in nature be transmitted by the pin carrying the load into the bracket.) The exact solution of the mathematical model in Fig. 1.2(c) gives an infinite stress at the point of load application, and we must therefore expect a very large stress at point B as the finite element mesh is refined. Of course, this very large stress is an artifice of the chosen model, and the concentrated load should be replaced by a pressure load over a small area when a very fine discretization is used (see further discussion). Furthermore, if the model then still predicts a very high stress, a nonlinear mathematical model would be appropriate. Note that the concentrated load in the beam model in Fig. 1.2(b) does not introduce any solution difficulties. Also, the right-angled sharp corners at the support of the beam model, of course, do not introduce any solution difficulties, whereas such corners ina plane stress model would introduce infinite stresses. Hence, for the plane stress model, the corners have to be rounded to more accurately represent the geometry of the actual physical bracket. We thus realize that the solution of a mathematical model may result in artificial difficulties that are easily removed by an appropriate change in the mathematical model to more closely represent the actual physical situation. Furthermore, the choice of a more encompassing mathematical model may result, in such cases, in a decrease in the required solution effort. While these observations are of a general nature, let us consider once again, specifically, the use of concentrated loads. This idealization of load application is exten- sively used in engineering practice. We now realize that in many mathematical models (and therefore also in the finite element solutions of these models), such loads create stresses of infinite value. Hence, we may ask under what conditions in engineering practice solution difficulties may arise. We find that in practice solution difficulties usually arise only when the finite element discretization is very fine, and for this reason the matter of infinite stresses under concentrated loads is frequently ignored. As an example, Fig. 1.4 gives finite element results obtained in the analysis of a cantilever, modeled as a plane stress problem. The cantilever is subjected to a concentrated tip load. In practice, the 6 X 1 mesh is usually considered sufficiently fine, and clearly, a much finer discretization would have to be used to accurately show the effects of the stress singularities at the point of load application and at the support. As already pointed out, if such a solution is pursued, it is necessary to change the mathematical model to more accurately represent the actual physical situation of the structure. This change in the mathematical model may be important in self-adaptive finite element analyses because in such analyses new meshes are generated automatically and artificial stress singularities cause —artificially—extremely fine discretizations. We refer to these considerations in Section 4.3.4 when we state the general elasticity problem considered for finite element solution.Sec, 1.3 Finite Element Analysis as an Integral Part of Computer-Aided Design 11 Weao.t 4 Y 20 E = 200,000 v= 0.30 10 Thickness = 0.1 (a) Geometry, boundary condition: Bernoulli beam theory resulte: 5= (b) Typical finite element discretization, 6 x 1 mesh of 9-node elements; results are: 8 = 0.16, tmex= 116 Figure 1.4 Analysis of a cantilever as a plane stress problem In summary, we should keep firmly in mind that the crucial step in any finite element analysis is always choosing an appropriate mathematical model since a finite element solution solves only this model. Furthermore, the mathematical model must depend on the analysis questions asked and should be reliable and effective (as defined earlier). In the process of analysis, the engineer has to judge whether the chosen mathematical model has been solved to a sufficient accuracy and whether the chosen mathematical model was appropriate (i.e., reliable) for the questions asked. Choosing the mathematical model, solving the model by appropriate finite element procedures, and judging the results are the fundamental ingredients of an engineering analysis using finite element methods. 1.3 FINITE ELEMENT ANALYSIS AS AN INTEGRAL PART OF COMPUTER-AIDED DESIGN Although a most exciting field of activity, engineering analysis is clearly only a support activity in the larger field of engineering design. The analysis process helps to identify good new designs and can be used to improve a design with respect to performance and cost. In the early use of finite element methods, only specific structures were analyzed, mainly in the aerospace and civil engineering industries. However, once the full potential of finite element methods was realized and the use of computers increased in engineering design environments, emphasis in research and development was placed upon making the use of finite element methods an integral part of the design process in mechanical, civil, and aeronautical engineering.12 An Introduction to the Use of Finite Element Procedures Chap. 1 Figure 1.5 The field of CAD/CAM viewed schematically Figure 1.5 gives an overview of the steps in a typical computer-aided design process. Finite element analysis is only a small part of the complete process, but it is an important art. P ‘We note that the first step in Figure 1.5 is the creation of a geometric representation of the design part. Many different computer programs can be employed (.g.,a typical and popular program is AutoCAD). In this step, the material properties, the applied loading and boundary conditions on the geometry also need to be defined. Given this information, a finite element analysis may proceed. Since the geometry and other data of the actual physical part may be quite complex, it is usually necessary to simplify the geometry and loading in order to reach a tractable mathematical model. Of course, the mathematical model should be reliable and effective for the analysis questions posed, as discussed in the preceding section. The finite element analysis solves the chosen mathematical model, which may be changed and evolve depending on the purpose of the analysis (see Fig. 1.1). Considering this process—which generally is and should be performed by engineer- ing designers and not only specialists in analysis—we recognize that the finite element methods must be very reliable and robust. By reliability of the finite element methods we now’ mean that in the solution of a well-posed mathematical model, the finite element procedures should always for a reasonable finite element mesh give a reasonable solution, Note that this meaning of “reliability of finite element methods” is different from that of “reliability of a mathematical model" defined in the previous section.Sec. 1.3 Finite Element Analysis as an Integral Part of Computer-Aided Design 13 and if the mesh is reasonably fine, an accurate solution should always be obtained. By robustness of the finite element methods we mean that the performance of the finite element procedures should not be unduly sensitive to the material data, the boundary conditions, and the loading conditions used. Therefore, finite element procedures that are not robust will also not be reliable. For example, assume that in the plane stress solution of the mathematical model in Fig. 1.2), any reasonable finite element discretization using a certain element type is employed. Then the solution obtained from any such analysis should not be hugely in error, that is, an order of magnitude larger (or smaller) than the exact solution. Using an unreliable finite element for the discretization would typically lead to good solutions for some mesh topologies, whereas with other mesh topologies it would lead to bad solutions. Elements based on reduced integration with spurious zero energy modes can show this unreliable behavior (see Section 5.5.6). Similarly, assume that a certain finite element discretization of a mathematical model gives accurate results for one set of material parameters and that a small change in the parameters corresponds to a small change in the exact solution of the mathematical model. Then the same finite element discretization should also give accurate results for the math- ematical model with the small change in material parameters and not yield results that are very much in error. These considerations regarding effective finite element discretizations are very im- portant and are discussed in the presentation of finite element discretizations and their stability and convergence properties (see Chapters 4 to 7). For use in engineering design, it is of utmost importance that the finite element methods be reliable, robust, and of course efficient. Reliability and robustness are important because a designer has relatively little time for the process of analysis and must be able to obtain an accurate solution of the chosen mathematical model quickly and without “trial and error.” The use of unreliable finite element methods is simply unacceptable in engineering practice. ‘An important ingredient of a finite element analysis is the calculation of error esti- mates, that is, estimates of how closely the finite element solution approximates the exact solution of the mathematical model (see Section 4.3.6). These estimates indicate whether a specific finite element discretization has indeed yielded an accurate response prediction, and a designer can then rationally decide whether the given results should be used. In the case that unacceptable results have been obtained using unreliable finite element methods, the difficulty is of course how to obtain accurate results. Finally, we venture to comment on the future of finite element methods in computer- aided design. Surely, many engineering designers do not have time to study finite element methods in depth or finite element procedures in general. Their sole objective is to use these techniques to enhance the design product. Hence the integrated use of finite element meth- ods in CAD in the future will probably involve to an increasingly smaller degree the scrutiny of finite element meshes during the analysis process. Instead, we expect that in linear elastic static analysis the finite element solution process will be automatized such that, given a mathematical model to be solved, the finite element meshes will be automatically created, the solution will be calculated with error estimates, and depending on the estimated errors and the desired solution accuracy, the finite element discretization will be automatically refined (without the analyst or the designer ever seeing a finite element mesh) until the required solution accuracy has been attained. In this automatic analysis process—in which4 An Introduction to the Use of Finite Element Procedures Chap. 1 of course the engineer must still choose the appropriate mathematical model for analysis— the engineer can concentrate on the design aspects while using analysis tools with great efficiency and benefit. While this design and analysis environment will be commonly available for linear elastic static analysis, the dynamic or nonlinear analysis of structures and fluids will probably still require from the engineer a greater degree of involvement and expertise in finite element methods for many years to come. With these remarks we do not wish to suggest overconfidence but to express a realistic outlook with respect to the valuable and exciting future use of finite element procedures. For some remarks on overconfidence in regard to finite element methods (which are still pertinent after almost two decades), see the article “A Commentary on Computational Mechanics” by J. T. Oden and K. J. Bathe [A]. 1.4 A PROPOSAL ON HOW TO STUDY FINITE ELEMENT METHODS With a voluminous number of publications available on the use and development of finite element procedures, it may be rather difficult for a student or teacher to identify an effective plan of study. Of course, such a plan must depend on the objectives of the study and the time available. In very general terms, there are two different objectives: 1. To learn the proper use of finite element methods for the solution of complex prob- lems, and 2. To understand finite element methods in depth so as to be able to pursue research on finite element methods. This book has been written to serve students with either objective, recognizing that the population of students with objective 1 is the much larger one and that frequently a student may first study finite element methods with objective | and then develop an increas- ing interest in the methods and also pursue objective 2. While a student with objective 2 will need to delve still much deeper into the subject matter than we do in this book, it is hoped that this book will provide a strong basis for such further study and that the “philosophy” of this text (see Section 1.3) with respect to the use of and research into finite element methods will be valued. Since this book has not been written to provide a broad survey of finite element methods—indeed, for a survey, a much more comprehensive volume would be necessary— it is clearly biased toward a particular way of introducing finite element procedures and toward only certain finite element procedures. The finite element methods that we concentrate on in this text are deemed to be effective techniques that can be used (and indeed are abundantly used) in engineering practice. The methods are also considered basic and important and will probably be em- ployed for a long time to come. The issue of what methods should be used, and can be best used in engineering practice, is important in that only reliable techniques should be employed (see Section 1.3), and this book concentrates on the discussion of only such methods.Sec. 1.4 A Proposal on How to Study Finite Element Methods 15 In introducing finite element procedures we endeavor to explain the basic concepts and equations with emphasis on physical explanations. Also, short examples are given to demonstrate the basic concepts used, and exercises are provided for hand calculations and the use of a finite element computer program. The preceding thoughts lead to our proposal for the study of finite element procedures. If objective 1 is being pursued, the reader will be mostly interested in the basic formulations, the properties of the finite elements and solution algorithms, and issues of convergence and efficiency. ‘An important point to keep in mind is that numerical finite element procedures are used to solve a mathematical model (with some reasonably small solution errors) (see Section 1.2). Hence, it is important for the user of a finite element computer program to always be able to judge the quality of the finite element results obtained in a solution. We demonstrate in this book how such judging is properly performed. However, if objective 2 is being pursued, much more depth in the formulations and numerical algorithms must be acquired. This text provides a broad basis for such study (but of course does not give all details of derivations and implementations). In cither case, we believe that for a study of finite element methods, it is effective to use a finite element computer program while learning about the theory and numerical procedures. Then, at the same time as the theoretical concepts are being studied on paper, it will be possible to test and demonstrate these concepts on a computer. Based on this book, various courses can be taught depending on the aim of the instruction. A first course would be “Introduction to Finite Element Analysis,” which could be based on Sections 1.1 to 1.4, 2.1 t0 2.3, 3.1 to 3.4, 4.1 to 4.3, 5.1 to 5.3, and 8.1 to 8.2.2. ‘A more advanced course would be “Finite Element Procedures,” based on Sec- tions 1.1. to 1.4, 3.1, 3.3, 4.1 to 4.4, 5.1 to 5.6, and 8.1 to 8.3. ‘A course on finite element methods for dynamic analysis would be “Computer Meth- ods in Dynamics” and could be based on Sections 1.1 to 1.4, 2.1 to 2.7, 3.1, 3.3, 4.1, 4.2, 5.1 to 5.3, 8.1 to 8.2.2, 9.1 to 9.4, 10.1, 10.2, 11.1, 11.2.1, 11.3.1, and 11.6. A course emphasizing the continuum mechanics principles for linear and nonlinear finite element analysis would be “Theory and Practice of Continuum Mechanics,” which could be based on Sections 1.1 to 1.4, 3.1, 3.3,4.1, 4.2.1, 4.2.2, 5.1, 5.2, 5.3.1, 5.3.3, 5.3.5, 6.1, 6.2, 6.3.1, 6.3.2, 6.4.1, 6.6, 7.1, and 7.4, A course on the analysis of field problems and fluid flows would be “Finite Element Analysis of Heat Transfer, Field Problems, and Fluid Flows,” based on Sections 1.1 to 1.4, 3.1, 3.3, 7.1 to 7.4, 5.3, 5.5, and 4.5.1 to 4.5.6. Note that the presentation in this course would first provide the finite element formulations (Sections 7.1 to 7.4) and then numerical procedures and mathematical results. A course on nonlinear finite element analysis of solids and structures would be “Nonlinear Analysis of Solids and Structures” and could be based on Sections 1.1 to 1.4, 6.1 to 6.8, 8.1, and 8.4. ‘A course on the numerical methods used in finite element analysis would be “Numerical Methods in Finite Element Analysis,” which could be based on Sections 1.1 to 1.4, 2.1 0 2.7, 4.1, 4.2.1, 5.1, 5.3, 5.5, 8.1 t0 8.4, 9.1 t0 9.6, 10.1 to 10.3, and 11.1 to 11.6. ‘And, finally, a course on more theoretical aspects of finite element analysis could be offered, entitled “Theoretical Concepts of Finite Element Analysis,” based on Sections 1.1 to 1.4, 4.1 to 4.5, and 5.1 t0 5.5.16 An Introduction to the Use of Finite Element Procedures Chap. 1 In addition, in most of these courses, the material in Chapter 12 would be useful. In general, such courses are best taught with homework that includes problem solu- tions, as suggested in the exercise sections of this book, as well as the completion of a project in which a finite element computer program is used. Various projects are proposed in the exercise sections of this book. One type of project requires the student to use the program STAP and to modify this program for certain new capabilities as suggested in Section 12.5. Such a project would be of value to students interested in learning about the actual detailed implementation of finite element methods. The other type of project is based on the use of a standard (commercial) finite element program—as it might be employed in engineering practice—to analyze an interesting physical problem. A valuable way to then proceed is to first solve a mathematical model for which an analytical exact solution is known and then introduce complicating features in the mathematical model that require a numerical solution. In this way, studies using different finite element discretizations and solution procedures in which the results are evaluated against a known exact solution can first be performed, which establishes confidence in the use of the methods. The required expertise and confidence then become valuable assets in the finite element solution of more complicated mathematical models. It is of particular importance during the study of finite element procedures that good judgment be developed and exercised concerning the quality of any finite element solution and of any mathematical model considered. Of course, exercising such judgments requires that the analyst be sufficiently familiar with possible mathematical models and available finite element procedures, and this aspect stimulates learning about mathematical models and methods of solution. The value of using a finite element computer code—the program STAP or a commercial finite element program as previously mentioned—is clearly in stimulating the process of learning and in reinforcing the understanding of what has been learned in class and from homework problem solutions. Using a powerful analysis program, in particular, also adds to the excitement of actually solving complicated analysis problems that heretofore could not be tackled.MI CHAPTER TWO Se Vectors, Matrices, and Tensors 2.1 INTRODUCTION The use of vectors, matrices, and tensors is of fundamental importance in engineering analysis because it is only with the use of these quantities that the complete solution process can be expressed in a compact and elegant manner. The objective of this chapter is to present the fundamentals of matrices and tensors, with emphasis on those aspects that are important in finite element analysis. From a simplistic point of view, matrices can simply be taken as ordered arrays of numbers that are subjected to specific rules of addition, multiplication, and so on. It is of course important to be thoroughly familiar with these rules, and we review them in this chapter. However, by far more interesting aspects of matrices and matrix algebra are recog- nized when we study how the elements of matrices are derived in the analysis of a physical problem and why the rules of matrix algebra are actually applicable. In this context, the use of tensors and their matrix representations are important and provide a most interesting subject of study. Of course, only a rather limited discussion of matrices and tensors is given here, but we hope that the focused practical treatment will provide a strong basis for understanding the finite element formulations given later. W718 Vectors, Matrices, and Tensors Chap. 2 2.2 INTRODUCTION TO MATRICES ‘The effectiveness of using matrices in practical calculations is readily realized by consider- ing the solution of a set of linear simultaneous equations such as. Sx > 4a + x3 =0 ~4n +6m- 4+ mol (2.1) a1 = daz + 625 — dre = 0 wa dey + Sze = where the unknowns are x1, x2, x3, and x,. Using matrix notation, this set of equations is written as 5-4 1 fx -4 6-4 UI 1-4 6 ~4} Jas o 1-4 5} Lx where it is noted that, rather logically, the coefficients of the unknowns (5, —4, 1, etc.) are grouped together in one array; the left-hand-side unknowns (x1, x2, x3, and x.) and the righthand-side known quantities are each grouped together in additional arrays. Although written differently, the relation (2.2) still reads the same way as (2.1). However, using matrix symbols to represent the arrays in (2.2), we can now write the set of simultaneous equations as (2.2) con] Ax=b (2.3) where A is the matrix of the coefficients in the set of linear equations, x is the matrix of unknowns, and b is the matrix of known quantities; i.e.,- $74 1 gl x 0 6-4 if xl, hi ~4 6 -a) FA[n bb RIy (24) o 1-4 5 x 0 The following formal definition of a matrix now seems apparent. Definition: A matrix is an array of ordered numbers. A general matrix consists of mn numbers arranged in m rows and n columns, giving the following array: On ans de @y an ++ dn aslo a (2.5) Gm om We say that this matrix has order m X n (m by n). When we have only one row (m = 1) or one column (n = 1), we also call A a vector. Matrices are represented in this book by boldface letters, usually uppercase letters when they are not vectors. On the other hand, vectors can be uppercase or lowercase boldface.Sec. 2.2 Introduction to Matrices 19 We therefore see that the following are matrices: [I 2 where the first and the last matrices are also column and row vectors, respectively. A typical element in the ith row and jth column of A is identified as ay; e.g., in the first matrix in (2.4), a, = 5 and az = —4. Considering the elements ay in (2.5), we note that the subscript i runs from 1 to m and the subscript j runs from 1 to n. A comma between subscripts will be used when there is any risk of confusion, €.g., ai+,,j+s Of to denote differentiation (see Chapter 6). In general, the utility of matrices in practice arises from the fact that we can identify and manipulate an array of many numbers by use of a single symbol. We shall use matrices in this way extensively in this book. 1 4 5.3), [i a } [61 22 3] (2.6) Symmetric, Diagonal, and Banded Matrices; A Storage Scheme Whenever the elements of a matrix obey a certain law, we can consider the matrix to be of special form. A real matrix is a matrix whose elements are all real, A complex matrix has elements that may be complex. We shall deal only with real matrices. In addition, the matrix will often be symmetric. Definition: The transpose of the m X n matrix A, written as AT, is obtained by interchanging the rows and columns in A. If A = A, it follows that the number of rows and columns in A are equal and that ay = ay. Because m = n, we say that A is a square matrix of order n, and because dy = ay, we Say that A is a symmetric matrix. Note that symmetry implies that A is square, but not vice versa; ie., a square matrix need not be symmetric. For example, the coefficient matrix A in (2.2) is a symmetric matrix of order 4. We can verify that Av = A by simply checking that ax = a, for é, ‘Another special matrix is the identity (or unit) matrix I,, Which is a square matrix of order n with only zero elements except for its diagonal entries, which are unity. For example, the identity matrix of order 3 is 10 0 h=]0 10 (2.7) oo. In practical calculations the order of an identity matrix is often implied and the subscript is not written. In analogy with the identity matrix, we also use identity (or unit) vectors of order n, defined as e,, where the subscript i indicates that the vector is the ith column of an identity matrix. We shall work abundantly with symmetric banded matrices. Bandedness means that all elements beyond the bandwidth of the matrix are zero. Because A is symmetric, we can state this condition as ay=0 forj>itm (2.8)20 Vectors, Matrices, and Tensors = Chap. 2 where 2m, + 1 is the bandwidth of A. As an example, the following matrix is a symmetric banded matrix of order 5. The half-bandwidth ma is 2: 32100 a A=]1 4561 (2.9) Ohgud 00143 If the half-bandwidth of a matrix is zero, we have nonzero elements only on the diagonal of the matrix and denote it as a diagonal matrix. For example, the identity matrix is a diagonal matrix. In computer calculations with matrices, we need to use a scheme of storing the elements of the matrices in high-speed storage. An obvious way of storing the elements of a matrix A of order m X n is simply to dimension in the FORTRAN program an array A(M, N), where M = m and N = n, and store each matrix element ay in the storage location A(I, J). However, in many calculations we store in this way unnecessarily many zero elements of A, which are never needed in the calculations..Also, if A is symmetric, we should probably take advantage of it and store only the upper half of the matrix, including the diagonal elements. In general, only a restricted number of high-speed storage locations are available, and it is necessary to use an effective storage scheme in order to be able to take into high-speed core the maximum matrix size possible. If the matrix is too large to be contained in high-speed storage, the solution process will involve reading and writing on secondary storage, which can add significantly to the solution cost. Fortunately, in finite element analysis, the system matrices are symmetric and banded. Therefore, with an effec- tive storage scheme, rather large-order matrices can be kept in high-speed core. Let us denote by A(I) the Ith element in the one-dimensional storage array A. A diagonal matrix of order n would simply be stored as shown in Fig. 2.1(a): AM =a 1 am (2.10) Consider a banded matrix as shown in Fig. 2.1(b). We will see later that zero elements within the “skyline” of the matrix may be changed to nonzero elements in the solution process; for example, ass may be a zero element but becomes nonzero during the solution process (see Section 8.2.3). Therefore we allocate storage locations to zero elements within the skyline but do not need to store zero elements that are outside the skyline. The storage scheme that will be used in the finite element solution process is indicated in Fig. 2.1 and is explained further in Chapter 12. Matrix Equality, Addition, and Multiplication by a Scalar We have defined matrices to be ordered arrays of numbers and identified them by single symbols. In order to be able to deal with them as we deal with ordinary numbers, it is necessary to define rules corresponding to those which govern equality, addition, subtrac- tion, multiplication, and division of ordinary numbers. We shall simply state the matrix rules and not provide motivation for them. The rationale for these rules will appear later, as it will turn out that these are precisely the rules that are needed to use matrices in theSec. 2.2 Introduction to Matrices 21 an Elements not a2 shown are zeros, 833 au AU) = ayy, A(2) = 822, A(3) = 239 . ALA) = Btu oor ALND = Bon nn (a) Diagonal matrix mat iad ec matiat 1 Skyline A(t) = ayy, AQ) = a2, ALB) = 29, A(4) = 223, AUS) = 13, A(6) = B44, ALT) = Ba, A(B) = 855, AI9) = 245, A(10) = a5, anes — ou = 75, A(12) = 256, {b) Banded matrix, ma=3 Figure 2.1 Storage of matrix A in a one-dimensional array solution of practical problems. For matrix equality, matrix addition, and matrix multiplica- tion by a scalar, we provide the following definitions. Definition: The matrices A and B are equal if and only if 1. A and B have the same number of rows and columns, 2. All corresponding elements are equal; i.e. ay = by for all i and j. Definition: Two matrices A and B can be added if and only if they have the same number of rows and columns. The addition of the matrices is performed by adding all corresponding elements; i.e., if ay and by denote general elements of A and B, respectively, then cy = ay + by denotes a general element of C, where C = A + B, It follows that C has the same number of rows and columns as A and B. EXAMPLE 2.1: Calculate C = A + B, where _f2 147, Efa.1 2 a=[2 3 bf melee | Here we have C=A+B [ss ai22 Vectors, Matrices, and Tensors Chap. 2 It should be noted that the order in which the matrices are added is not important. The subtraction of matrices is defined in an analogous way. EXAMPLE 2.2: Calculate,C = A — B, where A and B are given in Example 2.1. Here we have ay Ot cxa-e=[ 7), =I a] From the definition of the subtraction of matrices, it follows that the subtraction of a matrix from itself results in a matrix with zero elements only. Such a matrix is defined to be a null matrix 0. We turn now to the multiplication of a matrix by a scalar. Definition: A matrix is multiplied by a scalar by multiplying each matrix element by the scalar; ie, C = kA means that cy = kay. The following example demonstrates this definition. EXAMPLE 2.3: Calculate C = kA, where 2a a=[os 3 a a 422 We have c=ra=[! 2 3 Ie should be noted that so far all definitions are completely analogous to those used in the calculation with ordinary numbers, Furthermore, to add (or subtract) two general matrices of order n X m requires nm addition (subtraction) operations, and to multiply a general matrix of order n X m by a scalar requires nm multiplications. Therefore, when the matrices are of special form, such as symmetric and banded, we should take advantage of the situation by evaluating only the elements below the skyline of the matrix C because all other elements are zero. Multiplication of Matrices ‘We consider two matrices A and B and want to find the matrix product C = AB. Definition: Two matrices A and B can be multiplied to obtain C = AB if and only if the number of columns in A is equal to the number of rows in B. Assume that A is of order p X m and B is of order m X q. Then for each element in C we have > aby Qn) %y where C is of order p X q; i.e., the indices i and j in (2.11) vary from \ to p and \ 10 4, respectively. Therefore, to calculate the (i, j)th element in C, we multiply the elements in the ith row of A by the elements in the jth column of B and add all individual products, By taking the product of each row in A and each column in B, it follows that C must be of order p X 4.Sec. 2.2 Introduction to Matrices 23 EXAMPLE 2.4: Calculate the matrix product C = AB, where 531 ng =| 4 6 2[; =|24 10 3 4 32 We have. en = (5)(1) + (3)(2) + (1)(3) = 14 en = (4)() + (2) + (2)G) = 22 ea = (10)(1) + (3)(2) + (4)(3) = 28 ete. 14 39 Hence we obtain C=|22 48 28 70 As can readily be verified, the number of multiplications required in this matrix multiplication is p X q X m. When we deal with matrices in practice, however, we can often reduce the number of operations by taking advantage of zero elements within the matrices. EXAMPLE 2.5: Calculate the matrix product ¢ = Ab, where 2) O10) 2 oF me 2-1)’ ae 1 3 Here we can take advantage of the fact that the bandwidth of A is 3; ie., ma = 1. Thus, taking into account only the elements within the band of A, we have ex = (2)(4) + (-1I)) = 7 2 = (-1)(4) + Q)(1) + (-1)Q) = 3 = (14) + QQ) + (1G) = ca = (~1)Q) + (1G) = 1 ‘As is well known, the multiplication of ordinary numbers is commutative; ie., ab = ba. We need to investigate if the same holds for matrix multiplication, If we consider the matrices A= [3h B=B 4] (2.12) we have AB = ie al BA = [11] (2.13)2 Vectors, Matricas, and Tensors Chap. 2 Therefore, the products AB and BA are not the same, and it follows that matrix multiplica- tion is not commutative. Indeed, depending on the orders of A and B, the orders of the two product matrices AB and BA can be different, and the product AB may be defined, whereas the product BA may not be calculable. To distinguish the order of multiplication of matrices, we say that in the product AB, the matrix A premultiplies B, or the matrix B postmultiplies A. Although AB # BA in general, it may happen that AB = BA for special A and B, in which case we say that A and B commute. Although the commutative law does not hold in matrix multiplication, the distributive law and associative law are both valid. The distributive law states that E=(A+B)C =AC+BC (2.14) In other words, we may first add A and B and then multiply by C, or we may first multiply A and B by C and then do the addition. Note that considering the number of operations, the evaluation of E by adding A and B first is much more economical, which is important to remember in the design of an analysis program. The distributive law is proved using (2.11); that is, using ey = Z (ay + bude (2.15) we obtain ey =D ayey + D bycy (2.16) The associative law states that G = (AB)C = A(BC) = ABC (2.17) in other words, that the order of multiplication is immaterial. The proof is carried out by using the definition of matrix multiplication in (2.11) and calculating in either way a general element of G. Since the associative law holds, in practice, a string of matrix multiplications can be carried out in an arbitrary sequence, and by a clever choice of the sequence, many opera- tions can frequently be saved. The only point that must be remembered when manipulating the matrices is that brackets can be removed or inserted and that powers can be combined, but that the order of multiplication must be preserved. Consider the following examples to demonstrate the use of the associative and distribu- tive laws in order to simplify a string of matrix multiplications. EXAMPLE 2.6: Calculate A‘, where alia] One way of evaluating A* is to simply calculate ee 7 be = - le alk 4} E Al S$ s][2 1] _Jis 20 2 = ATA = 7 ae cL. i alk 4} B 2Sec. 2.2 Introduction to Metrices a3 Ar iS 20][2 1] _ [50 75 o3 [20 alli atl as| Alternatively, we may use 5 S)[5 5 50. 75 42 tat = are [3 walls wal [ts ual and save one matrix multiplication. EXAMPLE 2.7: Evaluate the product v"Av, where 3.20 A=/]2 4 2) v= ne = The formal procedure would be to calculate x = Av; i. 3.2107 1 x=Av=|]2 4 2] 2] = 12 6jL-1 and then calculate 7x to obtain 6 Wav =[1 2 -f t]-» -1 However, it is more effective to calculate the required product in the following way. First, we write A=U+D+U" where U is a lower triangular matrix and D is a diagonal matrix, 0 0 0 3.0 0 U=|2 00); D=|040 12 0 00 6 Hence we have vay = v(U + D + Uy VAv = vUy + v'Dv + v'U"y However, v"Uv is a single number and hence v'U"y = v"Uy, and it follows that viAy = 2y"Uv + v'Dy @ The higher efficiency in the matrix multiplication is obtained by taking advantage of the fact that U is a lower triangular and D is a diagonal matrix. Let x = Uy; then we have n=O wa = (2)(1) = 2 3 = (1)(1) + (2)Q) = 526 Vectors, Matrices, and Tensdrs Chap. 2 Next, we obtain WUv = vx = (2)(2) + (-1)(5) = -1 Also ¥Dv = (1)(1)3) + (2)(2)4) + (-(- 6) = 25 Hence using (a) we have v"Av = 23, as before, Apart from the commutative law, which in general does not hold in matrix multipli- cations, the cancellation of matrices in matrix equations also cannot be performed, in general, as the cancellation of ordinary numbers. In particular, if AB = CB, it does not necessarily follow that A = C. This is easily demonstrated considering a specific case: 2 afi] _ [4 oft (i dll) f0 IE] allel 21 40 but Fal +[s 3] 2.19) However, it must be noted that A = C if the equation AB = CB holds for all possible B. Namely, in that case, we simply select B to be the identity matrix I, and hence A = C. Tt should also be noted that included in this observation is the fact that if AB = 0, it does not follow that*either A or B is a null matrix. A specific case demonstrates this observation: eB eB hg om Some special rules concerning the use of transposed matrices in matrix multiplica- tions need to be pointed out. It is noted that the transpose of the product of two matrices A and B is equal to the product of the transposed matrices in reverse order; i.e.. (AB)’ = BTA” (2.21) The proof that (2.21) does hold is obtained using the definition for the evaluation of a matrix product given in (2.11). Considering the matrix products in (2.21), it should be noted that although A and B may be symmetric, ABis, in general, not symmetric, However, if A is symmetric, the matrix B’AB is always symmetric. The proof follows using (2.21): (BTAB)’ = (AB)"(B")" (2.22) = BAB (2.23) But, because A’ = A, we have (BTAB)’ = BTAB (2.24) and hence B7AB is symmetric.Sec, 2.2 Introduction to Metrices 2 The Inverse Matrix ‘We have seen that matrix addition and subtraction are carried out using essentially the same laws as those used in the manipulation of ordinary numbers. However, matrix multiplication is quite different, and we have to get used to special rules. With regard to matrix division, it strictly does not exist. Instead, an inverse matrix is defined. We shall define and use the inverse of square matrices only. Definition: The inverse of a matrix A is denoted by A~'. Assume that the inverse exists; then the elements of A~' are such that A“'A = Land AA“ = I. A matrix that possesses an inverse is said to be nonsingular. A matrix without an inverse is a singular matrix. As mentioned previously, the inverse of a matrix does not need to exist. A trivial example is the null matrix. Assume that the inverse of A exists. Then we still want to show that either of the conditions A"'A = or AA“! = Limplies the other. Assume that we have evaluated the elements of the matrices Ay! and A;' such that A7'A = I and AA;' = L Then we have Art = Ar(AA?}) = (Art A)AT! = Art (2.25) and hence Aj’ = A;'. EXAMPLE 2.8: Evaluate the inverse of the matrix A, where Ae i 3 For the inverse of A we need AA™' = I. By trial and error (or otherwise) we find that ‘We check that AA! = Tand A“‘A = I: o-L TE | eH Db To calculate the inverse of a product AB, we proceed as follows. Let G = (AB)"', where A and B are both square matrices. Then GAB =I (2.26) and postmultiplying by B-' and A~!, we obtain GA -B" (2.27) G=B IAT (2.28) Therefore, (AB)" = BAT (2.29) ‘We note that the same law of matrix reversal was shown to apply when the transpose of a matrix product is calculated.28 Vectors, Matrices, and Tensors Chap. 2 EXAMPLE 2.9: For the matrices A and B given, check that (AB)“' = B-'A™', _ ee a-[a A s=[5 ‘] ‘The inverse of A was used in Example 2.8. The inverse of B is easy to obtain: {0 4 04 = B-'A™', we need to evaluate C = AB: 2- | 3.0 = 6 -4 3jLo 4. 3 12. Assume that C~' = B~'A™!. Then we would have Sa -0 2 -[} + a (ili J-G e . ‘To check that the matrix given in (a) is indeed the inverse of C, we evaluate C~'C and find that 4 47 6 -4 agaft 3 - cre= [2 als al“! But since C-' is unique and only the correct C~' satisfies the relation C~'C = I, we indeed have found in (a) the inverse of C, and the relation (AB)~' = B-'A™* is satisfied. c In Examples 2.8 and 2.9, the inverse of A and B could be found by trial and error. However, to obtain the inverse of a general matrix, we need to have a general algorithm. One way of calculating the inverse of a matrix A of order n is to solve the n systems of equations AX=1 (2.30) where I is the identity matrix of order n and we have X = A”'. For the solution of each system of equations in (2.30), we can use the algorithms presented in Section 8.2. These considerations show that a system of equations could be solved by calculating the inverse of the coefficient matrix; i.e., if we have Ay=e (2.31) where A is of order n X n and y and c are of order n X 1, then y=A'e (2.32) However, the inversion of A is very costly, and it is much more effective to only solve the equations in (2.31) without inverting A (see Chapter 8). Indeed, although we may write symbolically that y = A”'e, to evaluate y we actually only solve the equations. Partitioning of Matrices To facilitate matrix manipulations and to take advantage of the special form of matrices, it may be useful to partition ‘a matrix into submatrices. A submatrix is a matrix that is obtained from the original matrix by including only the elements of certain rows and columns. TheSec. 2.2 Introduction to Matrices 29 idea is demonstrated using a specific case in which the dashed lines are the lines of partitioning: au; an ay ay A=|an i an an ay art (2.33) It should be noted that each of the partitioning lines must run completely across the original matrix. Using the partitioning, matrix A is written as Au An Aw a ie An ‘a _ where Au= [eh Au= [ee ae al etc. (2.35) an an G2 au. The right-hand side of (2.34) could again be partitioned, such as _ [An | An Aw A [i 1 An x] Ce and we may write A as -aoa. oa elAn. ge [Ae Ae A=[A. Aj}; AL [| Ay [ae al (2.37) ‘The partitioning of matrices can be of advantage in saving computer storage; namely, if submatrices repeat, it is necessary to store the submatrix only once. The same applies in arithmetic, Using submatrices, we may identify a typical operation that is repeated many times, We then carry out this operation only once and use the result whenever it is needed. The rules to be used in calculations with partitioned matrices follow from the definition of matrix addition, subtraction, and multiplication. Using partitioned matrices we can add, subtract, or multiply as if the submatrices were ordinary matrix elements, provided the original matrices have been partitioned in such a way that it is permissible to perform the individual submatrix additions, subtractions, or multiplications. These rules are easily justified and remembered if we keep in mind that the partition- ing of the original matrices is only a device to facilitate matrix manipulations and does not change any results. EXAMPLE 2.10: Evaluate the matrix product C = AB in Example 2.4 by using the following partitioning: -[te A} w= [B] aor a lie An! B= [py ‘AnB, + ABz Deve AB TAB + er o30 Vectors, Matricas, and Tensors Chap. 2 “tal acts al aat=([!]o a=[2 3] AaB, = [10 al, §]= 06 62) But AnB = [4][3 2] = [12 8] ‘Then substituting into (a) we have 14 39) AB =| 22 48 28 70. EXAMPLE 2.11: Taking advantage of partitioning, evaluate ¢ = Ab, where The only products that we need to evaluate are [4 af = 3] 13 olla] ~ Ly. oe = 7 2 ili 3 and = [ ‘We can now construct ¢: ea ftw 201 + 2wo 17 — 2 of, substituting, oe 24 The Trace and Determinant of a Matrix The trace and determinant of a matrix are defined only if the matrix is square. Both quantities are single numbers, which are evaluated from the elements of the matrix and are therefore functions of the matrix elements. Definition: The trace of the matrix A is denoted as tr(A) and is equal to D}-1 ay, where n is the order of A.Sec. 2.2 Introduction to Matrices 31 EXAMPLE 2.12: Calculate the trace of the matrix A given in Example 2.11. Here we have (A) = 44648 + 12=30 ‘The determinant of a matrix A can be defined in terms of the determinants of subma- trices of A and by noting that the determinant of a matrix of order 1 is simply the element of the matrix; i.e., if A = (a), then det A = ay. Definition: The determinant of an n X n matrix A is denoted as det A and is defined by the recurrence relation det A = > (=1)!ay; det Ay, (2.38) = where Ar; is the (n — 1) X (n ~ 1) matrix obtained by eliminating the Ist row and jth column from the matrix A. EXAMPLE 2.13: Evaluate the determinant of A, where a an a an. Using the relation in (2.38), we obtain det A = (~1)'an, det An + (—1)an2 det Arr But det Aun = az; det Ar. = an A Hence det A = gua — aan This relation is the general formula for the determinant of a2 X 2 matrix. Itcan be shown that to evaluate the determinant of a matrix we may use the recurrence relation given in (2.38) along any row or column, as indicated in Example 2.14. EXAMPLE 2.14: Evaluate the determinant of the matrix A, where 210 A=|1 31 012 Using the recurrence relation in (2.38), we obtain 301 = (-19 det A (-0@ ae |? 4 + (-1)() det iF 5 + (=1)"(0) det [3 I32 Vectors, Matrices, and Tensors Chap. 2 ‘We now employ the formula for the determinant of a2 X 2 matrix given in Example 2.13 and have det A = (2){(3)(2) — ()()} — {€)2) — OM} + 0 Hence det A=8 Let us check that the same result is obtained by using (2.38) along the second row instead of the first row, In this case we have, changing the 1 to 2 in (2.38), det A = (-1)%(1) det ki 2 + nee ae |? no 7 + cnrayae [> | Again using the formula given in Example 2.13, we have det A = —{(1)(2) — (O)(1)} + BM(2)(2) — (0))} = {(2)(1) - (O)} or, as before, det A = Finally, using (2.38) along the third column, we have det A = (~1)4(0) det [3 ] +.(=1)(1) det i ] + (—1)2) det i? 4] and, as before, obtain det A = 8. Many theorems are associated with the use of determinants. Typically, the solution of a set of simultaneous equations can be obtained by a series of determinant evaluations (see, for example, B. Noble [A]). However, from a modern viewpoint, most of the results that are obtained using determinants can be obtained much more effectively. For example, the solution of simultaneous equations using determinants is very inefficient. As we shall see later, a primary value of using determinants lies in the convenient shorthand notation we can use in the discussion of certain questions, such as the existence of an inverse of a matrix. ‘We shall use determinants in particular in the solution of eigenvalue problems. In evaluating the determinant of a matrix, it may be effective to first factorize the matrix into a product of matrices and then use the following result: det (BC - - - F) = (det By(det C) - - - (det F) (2.39)‘Sec. 2.2 Introduction to Matrices 33 Relation (2.39) states that the determinant of the product of a number of matrices is equal to the product of the determinants of each matrix. The proof of this result is rather lengthy and clumsy [it is obtained using the determinant definition in (2.38)], and therefore we shall not include it here. We shall use the result in (2.39) often in eigenvalue calculations when the determinant of a matrix, say matrix A, is required. The specific decomposition used is A= LDL’, where L is a lower unit triangular matrix and D is a diagonal matrix (see Section 8.2.2). In that case, det A = det L det D det L™ (2.40) and because det L = 1, we have det A= Th dy 41) EXAMPLE 2.15: Using the LDL” decomposition, evaluate the determinant of A, where A is given in Example 2.14. ‘The procedure to obtain the LDL” decomposition of A is presented in Section 8.2. Here we simply give L and D, and it can be verified that LDL? = A: 100 20 0 71-0, D=|o ¢ o31 00 det A = (2)()@) = ‘This is also the value obtained in Example 2.14. we Using (2.41), we obtain ‘The determinant and the trace of a matrix are functions of the matrix elements. However, it is important to observe that the off-diagonal elements do not affect the trace of a matrix, whereas the determinant is a function of all the elements in the matrix. Although we can conclude that a large determinant or a large trace means that some matrix elements are large, we cannot conclude that a small determinant or a small trace means that all matrix elements are small. EXAMPLE 2.16; Calculate the trace and determinant of A, where nef 2] wo 2 Here we have wr(A)=3 and det A = (1)(2) — (10~4)(10,000) ie, det A= 1 Hence both the trace and tHe determinant of A are small in relation to the off-diagonal ele- ment ay2.34 Vectors, Matrices, and Tensors Chap. 2 2.3 VECTOR SPACES In the previous section we defined a vector of order n to be an array of n numbers written in matrix form. We now want to associate a geometric interpretation with the elements of a vector. Consider as an example a column vector of order 3 such as a 2 =]al= (2.42) Xs 3 We know from elementary geometry that x represents a geometric vector in a chosen coordinate system in three-dimensional space. Figure 2.2 shows assumed coordinate axes and the vector corresponding to (2.42) in this system. We should note that the geometric representation of x depends completely on the coordinate system chosen; in other words, if (2.42) would give the components of a vector in a different coordinate system, then the geometric representation of x would be different from the one in Fig. 2.2. Therefore, the coordinates (or components of a vector) alone do not define the actual geometric quantity, but they need to be given together with the specific coordinate system in which they are measured. Figure 2.2 Geometric representation of ‘vector x The concepts of three-dimensional geometry generalize to a vector of any finite order n. If n > 3, we can no longer obtain a plot of the vector; however, we shall see that mathematically all concepts that pertain to vectors are independent of n. As before, when we considered the specific case n = 3, the vector of order n represents a quantity in a specific coordinate system of an n-dimensional space. Assume that we are dealing with a number of vectors all of order n, which are defined in a fixed coordinate system. Some fundamental concepts that we shall find extremely powerful in the later chapters are summarized in the following definitions and facts.Sec.2.3 Vector Spaces 35 Definition: A collection of vectors x1, X2,. . . , X: is said to be linearly dependent if there exist numbers 01, 02, ...» Oy which are not all zero, such that Ox) + ax, t+ + aK = 0 (2.43) If the vectors are not linearly dependent, they are called linearly independent vectors. We consider the following examples to clarify the meaning of this definition. EXAMPLE 2.17: Let n = 3 and determine if the vectors e,, i = 1, 2, 3, are linearly dependent or independent. According to the definition of linear dependency, we need to check if there are constants a, a, and ars, not all zero, that satisfy the equation 1 0 07 «fo a) 0} + om) 1) +a3}0)=]0 (a) 0. 0 1 0 But the equations in (a) read a] fo a|=|0 al Lo which is satisfied only if a; = 0, i = 1, 2, 3; therefore, the vectors e, are linearly independent. EXAMPLE 2.18; With n = 4, investigate whether the following vectors are linearly dependent or independent. 1 -1 oO welt we] Oh gn }Os ee edt 0.5) 0 oe We need to consider the system of equations 1 -1 0 0 1 0 -0.5 0 ao | a] *%]-05 | “lo 0.5, 0. 10,25) i) or, considering each row, a — a a 05 ay a: - 05 a, = 0 0.5 — 0.25a, = 0 where we note that the equations are satisfied for a = 1, a2 = 1, and as = 2. Therefore, the vectors are linearly dependent.36 Vectors, Matrices, and Tensors Chap. 2 In the preceding examples, the solution for a, a2, and a3 could be obtained by inspection. We shall later develop a systematic procedure of checking whether a number of vectors are linearly dependent or independent. Another way of looking at the problem, which may be more appealing, is to say that the vectors are linearly dependent if any one of them can be expressed in terms of the others. That is, if not all of the a; in (2.43) are zero, say a; # 0, then we can write (2.44) Geometrically, when n = 3, we could plot the vectors and if they are linearly dependent, we would be able to plot one vector in terms of multiples of the other vectors. For example, plotting the vectors used in Example 2.17, we immediately observe that none of them can be expressed in terms of multiples of the remaining ones; hence the vectors are linearly independent, ‘Assume that we are given q vectors of order n, n = g, which are linearly dependent, but that we only consider any (q — 1) of them. These (q — 1) vectors may still be linearly dependent. However, by continuing to decrease the number of vectors under consideration, we would arrive at p vectors, which are linearly independent, where, in general, p = q. The other (q — p) vectors can be expressed in terms of the p vectors. We are thus led to the following definition. Definition: Assume that we have p linearly independent vectors of order n, where n= p. These p vectors form a basis for a p-dimensional vector space. We talk about a vector space of dimension p because any vector in the space can be expressed as a linear combination of the p base vectors. We should note that the base vectors for the specific space considered are not unique; linear combinations of them can give another basis for the same space. Specifically, if p = n, then a basis for the space considered ise, i= 1,...,, from which it also follows that p cannot be larger than n. Definition: q vectors, of which p vectors are linearly independent, are said to span a -dimensional vector space. ‘We therefore realize that all the importance lies in the base vectors since they are the smallest number of vectors that span the space considered. All q vectors can be expressed in terms of the base vectors, however large q may be (and indeed q could be larger than n).. EXAMPLE 2.19: Establish a basis for the space spanned by the three vectors in Example 2.18, In this case ¢ = 3 and n = 4, We find by inspection that the two vectors x; and x, are linearly independent. Hence x; and x2 can be taken as base vectors of the two-dimensional space spanned by Xi, X2, and Xs. Also, using the result of Example 2.18, we have x3 = —4x2 — }x1, Assume that we are given a p-dimensional vector space which we denote as E,, for which x1, X2,. . . , Xp are chosen base vectors, p > 1. Then we might like to consider onlySec. 2.3 Vector Spaces 37 all those vectors that can be expressed in terms of x: and xz. But the vectors x, and x2 also form the basis of a vector space that we call E>. If p = 2, we note that E, and E, coincide. We call Ez a subspace of E,, the concise meaning of which is defined next. Definition: A subspace of a vector space is a vector space such that any vector in the subspace is also in the original space. If &;, 2, . ... Xp are the base vectors of the original space, any subset of these vectors forms the basis of a subspace; the dimension of the subspace is equal to the number of base vectors selected. EXAMPLE 2.20: The three vectors x1, X2, and x3 are linearly independent and therefore form the basis of a three-dimensional vector space Ea: i 1 0 2 0 _|-1 m= Th maT of w=) 9 @ 0 0 7 Identify some possible two-dimensional subspaces of Es. ‘Using the base vectors in (a), a two-dimensional subspace is formed by any two of the three ‘vectors; €.g4 X1 and xz represent a basis for a two-dimensional subspace; x, and x3 are the basis for another two-dimensional subspace; and so on. Indeed, any two linearly independent vectors in E; from the basis of a two-dimensional subspace, and it follows that there are an infinite number of two-dimensional subspaces in Es. Having considered the concepts of a vector space, we may now recognize that the columns of any rectangular matrix A also span a vector space. We call this space the column space of A. Similarly, the rows of a matrix span a vector space, which we calll the row space of A. Conversely, we may assemble any q vectors of order n into a matrix A of order n X q. ‘The number of linearly independent vectors used is equal to the dimension of the column space of A. For example, the three vectors in Example 2.20 form the matrix 1 1 0 2 0-1 Aalto (2.45) oo 1 Assume that we are given a matrix A and that we need to calculate the dimension of the column space of A. In other words, we want to evaluate how many columns in A are linearly independent. The number of linearly independent columns in A is neither increased nor decreased by taking any linear combinations of them. Therefore, in order to identify the column space of A, we may try to transform the matrix, by linearly combining its columns, to obtain unit vectors e,. Because unit vectors e, with distinct i are linearly independent, the dimension of the column space of A is equal to the number of unit vectors that can be obtained. While frequently we are not able to actually obtain unit vectors e, (see Exam- ple 2.21), the process followed in the transformation of A will always lead to a form that displays the dimension of the column space.38 Vectors, Matrices, and Tensors Chap. 2 EXAMPLE 2.21: Calculate the dimension of the column space of the matrix A formed by the vectors x1, x2, and x3 considered in Example 2.20. ‘The matrix considered is 1 1 6 2 0-1 A=lr 0 0 oo 1 ‘Writing the second and third columns as the first and second columns, respectively, we obtain To id | 2 w= lo 0 1 o 1 0 ‘Subtracting the first column from the third column, adding twice the second column to the third column, and finally multiplying the second column by (~1), we obtain 1 0 o 1 oO tC) | But we have now reduced the matrix to a form where we can identify that the three columns are linearly independent; ie., the columns are linearly independent because the first three elements in the vectors are the columns of the identity matrix of order 3. However, since we obtained Az from A by interchanging and linearly combining the original columns of A and thus in the solution process have not increased the space spanned by the columns of the matrix, we find that the dimension of the column space of A is 3. A= peos In the above presentation we linearly combined the vectors x1, . . . , Xp, which were the columns of A, in order to identify whether they were linearly independent. Alternatively, to find the dimension of the space spanned by a set of vectors X1, x2, . . . , Xqy We could use the definition of vector linear independence in (2.43) and consider the set of simultaneous homogeneous equations aun: + ams + +++ 04% = 0 (2.46) which is, in matrix form, Aa=0 (2.47) where @ is a vector with elements a, .. . a, and the columns of A are the vectors x1, X2,...,Xq- The solution for the unknowns + @ is not changed by linearly combining or multiplying any of the rows in the matrix A. Therefore, we may try to reduce A by multiplying and combining its rows into a matrix in which the columns consist only of unit vectors. This reduced matrix is called the row-echelon form of A. The number of unit column vectors in the row-echelon form of A is equal to the dimension of the column space of A and, from the preceding discussion, is also equal to the dimension of the row space of A. It follows that the dimension of the column space of A is equal to the dimension of theSec. 2.3 Vector Spaces 39 row space of A. In other words, the number of linearly independent columns in A is equal to the number of linearly independent rows in A. This result is summarized in the definition of the rank of A and the definition of the null space (or kernel) of A. Definition: The rank of a matrix A is equal t0 the dimension of the column space and equal 10 the dimension of the row space of A. Definition: The space of vectors « such that Aa. = 0 is the null space (or kernel) of A. EXAMPLE 2.22: Consider the following three vectors: _ 2 2 1 3 _hal. _| 72). _|i m=} mal gk =|, 4 2 6 7 ce 4 Use these vectors as the columns of a matrix A and reduce the matrix to row-echelon form. We have 132 2A 1-21 3045 4 26 3-14 Subtracting multiples of the first row from the rows below it in order to obtain the unit vector e; in the first column, we obtain ia 2 0 -5 -1 0 -5 -1 Mel 5-1 0 -10 -2 0 -10 -2 Dividing the second row by (—5) and then subtracting multiples of it from the other rows in order to reduce the second column to the unit vector e2, we obtain 10 A= loooCS ecoooe COC Om40 Vectors, Matrices, and Tensors Chap. 2 Hence we can give the following equivalent statements: 1. The solution to Aa = 0 is a = Fas m= —-bas 2, The three vectors x1, x2, and x3 are linearly dependent. They form a two-dimensional vector space. The vectors x; and x; are linearly independent, and they form a basis of the two-dimensional space in which x, X2, and x; lie. 3, The rank of A is 2. 4, The dimension of the column space of A is 2. 5. The dimension of the row space of A. is 2. 6. The null space (kernel) of A has dimension 1 and a basis is the vector [ ‘Note that the rank of A’ is also 2, but that the kernel of A’ has dimension 4. 2.4 DEFINITION OF TENSORS In engineering analysis, the concept of tensors and their matrix representations can be important. We shall limit our discussion to tensors in three-dimensional space and pri- marily be concerned with the representation of tensors in rectangular Cartesian coordinate frames. Let the Cartesian coordinate frame be defined by the unit base vectors e; (see Fig. 2.3). A vector u in this frame is given by (2.48) Figure 2.3. Cartesian coordinate systems for definition of tensorsSec, 2.4 Definition of Tensors a where the u; are the components of the vector. In tensor algebra it is convenient for the purpose of a compact notation to omit the summation sign in (2.48); i.e., instead of (2.48) we simply write w= wie; (2.49) where the summation on the repeated index iis implied (here i = 1, 2,3). Since i could be replaced by any other subscript without changing the result (¢.g., k or j), it is also called a dummy index or a free index. This convention is referred to as the summation convention of indicial notation (or the Einstein convention) and is used with efficiency to express in a compact manner relations involving tensor quantities (see Chapter 6 where we use this Notation extensively). Considering vectors in three-dimensional space, vector algebra is employed effec tively. The scalar (or dot) product of the vectors u and v, denoted by u - v is given by uv =|ul|¥| cos 6 (2.50) where |u| is equal to the length of the vector u, |u| = Vu. The dot product can be evaluated using the components of the vectors, wey = uo, @.51) The vector (or cross) product of the vectors u and v produces a new vector W = u X ¥ ee & es wedet|u um us (2.52) bi tO Figure 2.4 illustrates the vector operations performed in (2.50) and (2.52). We should note that the direction of the vector w is obtained by the right-hand rule; i.e., the right-hand thumb points in the direction of w when the fingers curl from u tov. x uM 6080 ar ivi IwI = lul ivi sing Figure 2.4 Vectors used in products42 Vectors, Matrices, and Tensors Chap. 2 These vector algebra procedures are frequently employed in finite element analysis to evaluate angles between two given directions and to establish the direction perpendicular to a given plane. EXAMPLE 2.23: Assume that the vectors u and v in Fig. 2.4 are tll Calculate the angle between these vectors and establish a vector perpendicular to the plane that is defined by these vectors. Here we have Jul] = 3V2 |v] =2Vv2 Hence cos 6 = $ and @ = 60°, ‘A vector perpendicular to the plane defined by w and v is given by e@ & e det}3 3 0 022 6 hence w=|-6 6 Using |w] = Viviwi, we obtain |w| = 6V3 which is also equal to the value obtained using the formula given in Fig. 2.4. Although not specifically stated, the typical vector considered in (2.48) is a tensor. Let us now formally define what we mean by a tensor. For this purpose, we consider in addition to the unprimed Cartesian coordinate frame a primed Cartesian coordinate frame with base vectors ej which spans the same space as the unprimed frame (see Fig. 2.3). An entity is called a scalar, a vector (ie., a tensor of first order or rank 1), or a tensor (ie., a tensor of higher order or rank) depending on how the components of the entity are defined in the unprimed frame (coordinate system) and how these components transform to the primed frame. Definition: An entity is called a scalar if it has only a single component ¢ in the coordinates x; measured along e; and this component does not change when expressed in the coordinates x} measured along e!: OG, Xa x3) = O(Rie xh x) (2.53) A scalar is also a tensor of order 0. As an example, temperature at a point is a scalar.Sec. 2.4 Definition of Tensors “a Definition: An entity is called a vector or tensor of first order if it has three components § in the unprimed frame and three components i in the primed frame, and if these components are related by the characteristic law (using the summation convention) Ei = Pak (2.54) where Pa = C06 (e/, €,) _@55) The relation (2.54) can also be written in matrix form as g = PE (2.56) where €', P, and € contain the elements of (2.54). The transformation in (2.54) corresponds to a change of basis in the representation of the vector. To arrive at (2.54) we recognize that the same vector is considered in the two different bases; hence we have Siej = Gee 257) Using the fact that the base vectors in each coordinate frame are orthogonal to each other and are of unit length, we can take the dot products [see (2.50)] on both sides of (2.57) with e/ and obtain (2.54). Of course, analogously we could also take the dot product on both sides with e, to obtain the inverse transformation Sn = COS(Cm, @f) EF (2.58) or in matrix form, g=P (2.59) Hence we note that P~' = P’, and this leads us to the following definition. Definition: A matrix Q is an orthogonal matrix if Q7Q = QQ! = L Therefore, for an or- thogonal matrix, we have Q~' = Q’. Hence the matrix P defined in (2.55) and (2.56) is an orthogonal matrix, and because the elements of P produce a rotation, we also refer to P as a rotation matrix. We demonstrate the preceding discussion in the following example. EXAMPLE 2.24: The components of a force expressed in the unprimed coordinate system shown in Fig, E2.24 are Figure E2.24 Representation of a force in different coordinate systems“ Vectors, Matrices, and Tensors Chap. 2 Evaluate the components of the force in the primed coordinate system in Fig. E2.24. Here we have, using (2.56), 1 0 0 P=|0 cos@ sin@ O -sin@ cos # and then R’=PR @ where R’ gives the components of the force in the primed coordinate system. As a check, if we 30° we obtain, using (a), which is correct because the e}-vector is now aligned with the force vector. To define a second-order tensor we build on the definition given in (2.54) for a tensor of rank 1. Definition: An entity is called a second-order tensor if it has nine components ty, i = 1, 2, 3, and j = 1, 2, 3, in the unprimed frame and nine components ty in the primed frame and if these components are related by the characteristic law 1) = Papate (2.60) As in the case of the definition of a first-order tensor, the relation in (2.60) represents achange of basis in the representation of the entity (see Example 2.25) and we can formally derive (2.60) in essentially the same way as we derived (2.54). That is, if we write the same tensor of rank 2 in the two different bases, we obtain Timmer = fOr) (2.61) where clearly in the tensor representation the first base vector goes with the first subscript (the row in the matrix representation) and the second base vector goes with the second subscript (the column in the matrix representation). The open product! or tensor product, €,€; is called a dyad and a linear combination of dyads as used in (2.61) is called a dyadic, (Gee, for example, L. E. Malvern [A]). Taking the dot product from the right in (2.61), first with e/ and then with e/, we obtain trn@mBy = tes(ei * ej) troDmi By = tel@e + @7)(€1* €}) (2.62) or ty = tupupp "The open product or tensor product of two vectors denoted as ab is defined by the requirement that (ab) + v = a(b+ ¥) forall vectors v. Some writers use the notation a ® b instead of ab.Sec. 2.4 Definition of Tensors 45 Here &y is the Kronecker delta (6, = | for i = j, and 8, = 0 fori # #). This transforma- tion may also be written in matrix form as t= Perr (2.63) where the (i, k)th element in P is given by pa. Of course, the inverse transformation also holds: t= PP (2.64) This relation can be derived using (2.61) and [similar to the operation in (2.62)] taking the dot product from the right with e; and then e,, or simply using (2.63) and the fact that P is an orthogonal matrix. In the preceding definitions we assumed that all indices vary from | to 3; special cases are when the indices vary from 1 to n, with n < 3. In engineering analysis we frequently deal only with two-dimensional conditions, in which case n = 2. EXAMPLE 2.25: Stress is a tensor of rank 2. Assume that the stress at a point measured in an unprimed coordinate frame in a plane stress analysis is (not including the third row and column of zeros) i 1 eel 4. Establish the components of the tensor in the primed coordinate system shown in Fig. E2.25. Figure E2.25 Representation of a stress tensor in different coordinate systems ‘Here we use the rotation matrix P as in Example 2.24, and the transformation in (2.63) is Assume that we are interested in the specific case when 6 = 45°. In this case we have Lt lh “l-Lo 2]46 Vectors, Matrices, and Tensors Chap. 2 and we recognize that in this coordinate system the off-diagonal elements of the tensor (shear components) are zero. The primed axes are called the principal coordinate axes, and the diagonal elements 7/; = 0 and riz = 2 are the principal values of the tensor. We will see in Section 2.5 that the principal tensor values are the eigenvalues of the tensor and that the primed axes define the corresponding eigenvectors. The previous discussion can be directly expanded to also define tensors of higher order than 2. In engineering analysis we are, in particular, interested in the constitutive tensors, that relate the components of a stress tensor to the components of a strain tensor (see, for example, Sections 4.2.3 and 6.6) Ty = Coster (2.65) The stress and strain tensors are both of rank 2, and the constitutive tensor with components Cua is of rank 4 because its components transform in the following way: Cin = PinPinPerP.Cnars (2.66) In the above discussion we used the orthogonal base vectors e, and e; of two Cartesian systems. However, we can also express the tensor in components of a basis of nonorthogonal base vectors. It is particularly important in shell analysis to be able to use such base vectors (see Sections 5.4.2 and 6.5.2). In continuum mechanics it is common practice to use what is called a covariant basis with the covariant base vectors g), i = 1, 2, 3, and what is called a contravariant basis with the contravariant base vectors, g,j = 1, 2, 3; see Fig. 2.5 for an example. The covariant and contravariant base vectors are in general not of unit length and satisfy the relationships ere =o (2.67) where / is the (mixed) Kronecker delta (6 = 1 for i= j, and 8; = 0 fori # j). 3 A Figure 2.5 Example of covariant and gieg'= 1; g2-g9'=0 contravariant base vectors, n = 2 (plotted girg?=0; garg=1 in Cartesian reference frame)Sec. 2.4 Definition of Tensors ar Hence the contravariant base vectors are orthogonal to the covariant base vectors. Furthermore, we have & = Bue! (2.68) with B= Bw (2.69) and B= gg, (2.70) with si=eeg’ (2.71) where gi and g! are, respectively, the covariant and contravariant components of the metric tensor. To prove that (2.68) holds, we tentatively let B= Gugt (2.72) with the ag unknown elements, Taking the dot product on both sides with g,, we obtain Be = dag sg, = audt (2.73) ay Of course, (2.70) can be proven in a similar way (see Exercise 2.11). Frequently, in practice, the covariant basis is conveniently selected and then the contravariant basis is given by the above relationships. Assume that we need to use a basis with nonorthogonal base vectors. The elegance of: then using both the covariant and contravariant base vectors is seen if we simply consider the work done by a force R going through a displacement u, given by R * u. If we express both R and u in the covariant basis given by the base vectors g,, we have (Rigi + Reg: + Rgs) + (u'gr + u2ga + u*gs) Rv'g (2.74) 0 On the other hand, if we express only R in the covariant basis, but u in the contravariant basis, given by the base vectors g/, we have Rew = (Rigi + Rg + Rgs) = (wig! + weg? + usp?) = R'uSt = Ru; (2.75) which is a much simpler expression. Fig. 2.6 gives a geometrical representation of this evaluation in a two-dimensional case. We shall use covariant and contravariant bases in the formulation of plate and shell elements, Since we are concerned with the product of stress and strain (e.g.. in the principle of virtual work), we express the stress tensor in contravariant components [as for the force Rin (2.75)], T= FEB, (2.76) and the strain tensor in covariant components [as for the displacement in (2.75)], e= dee (2.77)“aa Vectors, Matrices, and Tensors Chap. 2 ug Ngai = 4 ligall = 1 ne te ig" cone _ iigtt= usuyg'+u2g? cosa Figure 2.6 Geometrical representation of R and u using covariant and contravariant bases Using these dyadics we obtain for the product of stress and strain W = @™"BnB) * (Eve'e) = PE bmndn (2.78) HE This expression for W is as simple as the result in (2.75). Note that here we used the convention—designed such that its use leads to correct results*—that in the evaluation of the dot product the first base vector of the first tensor multiplies the first base vector of the second tensor, and so on. Instead of writing the product in summation form of products of components, we shall. also simply use the notation Wate (2.79) and simply imply the regult in (2.78), in whichever coordinate system it may be obtained. The notation in (2.79) is, in essence, a simple extension of the notation of a dot product between two vectors. Of cburse, when considering u v, a unique result is implied, but this result can be obtained in different ways, as given in (2.74) and (2.75). Similarly, when 2Namely, consider (ab) - (ed). Let A = ab, B = ed; then A+ B = AyBy = ajbjed, = (aic)(byd) = (a-cXb+d).Sec. 2.4 Definition of Tensors a writing (2.79), the unique result of W is implied, and this result may also be obtained in different ways, but the use of # and é, can be effective (see Example 2.26). Hence we note that the covariant and contravariant bases are used in the same way as Cartesian bases but provide much more generality in the representation and use of tensors, Consider thie following examples. EXAMPLE 2.26: Assume that the stress and strain tensor components at a point in a contin- uum corresponding to a Cartesian basis are 7y and ¢, and that the strain energy, per unit volume, is given by U = } zyey. Assume also that a basis of covariant base vectors g,,i = 1,2, 3, is given. Show explicitly that the value of U is then also given by } 72m. Here we use FB Be = Tyee @ and EmB"B" = Eye) ® But from (a) and (b) we obtain tu = 7(Gn* Gg.) — sumonm and n and €u = Enn(Q"* @)(g"-e) sum on mand n Now since @eve-e)=1 sumonj we also have Ua heen EXAMPLE 2.27: The Cartesian components 7, of the stress tensor 7ye,¢; are 7, = 100, 712 = 60, tm = 200, and the components ¢, of the strain tensor eee, are €1, = 0.001, €i2 = 0.002, €2 = 0.003. Assume that the stress and strain tensors are to be expressed in terms of covariant strain components and contravariant stress components with + a= [}} Ba ve V2. Calculate these components and, using these components, evaluate the product $ 7y¢y. Here we have, using (2.67), el eM To evaluate 77 we use PRB) = Traln€n ‘so that # = Tal€m* BV Cn * 8) ‘Therefore, the contravariant stress components are # = 180; #2 = P= -140V2, 7 = 400 Similarly, Eee! = enalnty & = enm(€m * B:)(Cr * B.)Vectors, Matrices, and Tensors Chap. 2 and the covariant strain components are 1 3 Topi ea = a = 1000 1000V2 Then we have 47/8, = by (180 + 1600 — 840) = 0.47 This value is of course also equal to 4ry¢y. EXAMPLE 2.28: The Green-Lagrange strain tensor can be defined as € = Gee! with the components & = CBr es — Br BD @ ax. a(x +a) ‘where og = an? q an () and x denotes the vector of Cartesian coordinates of the material point considered, u denotes the vector of displacements into the Cartesian directions, and the r; are convected coordinates (in finite element analysis the r, are the isoparametric coordinates; see Sections 5.3 and 5.4.2). 1. Establish the linear and nonlinear components (in displacements) of the strain tensor. 2.. Assume that the convected coordinates are identical to the Cartesian coordinates. Show that the components in the Cartesian system can be written as (au 2 4, te au . (% + an, * an: Ox, © ‘To establish the linear and nonlinear components, we substitute from (b) into (a), Hence Ae 2 2) : (2 21 Nar,” an)” \ar, ‘The terms linear in displacements are therefore 1/au ax, ax ow dhe = 2(38. 2% 5 OH athe = (2-2 +2 @ and the terms nonlinear in displacements are 3 _ lau, aw ah ~ Ase =) © If the convected coordinates are identical to the Cartesian coordinates, we have r, = x,, i = 1, 2, 3, and 4x;/4x; = 8, Therefore, (4) becomes 4am, 3H) shine als + an ® and (@) becomes = Lf ae due yleoninear = x a a) (e) Adding the linear and nonlinear terms (f) and (g), we obtain (c),Sec. 2.5 The Symmetric Eigenproblem Av = Av 51 The preceding discussion was only a very brief introduction to the definition and use of tensors. Our objective was merely to introduce the basic concepts of tensors so that we can work with them later (see Chapter 6). The most important point about tensors is that the components of a tensor are always represented in a chosen coordinate system and that these components differ when different coordinate systems are employed. It follows from the definition of tensors that if all components of a tensor vanish in one coordinate system, they vanish likewise in any other (admissible) coordinate system. Since the sum and differ- ence of tensors of a given type are tensors of the same type, it also follows that if a tensor equation can be established in one coordinate system, then it must also hold in any other (admissible) coordinate system. This property detaches the fundamental physical relation- ships between tensors under consideration from the specific reference frame chosen and is the most important characteristic of tensors: in the analysis of an engineering problem we are concerned with the physics of the problem, and the fundamental physical relationships between the variables involved must be independent of the specific coordinate system chosen; otherwise, a simple change of the reference system would destroy these relation- ships, and they would have been merely fortuitous. As an example, consider a body sub- jected to a set of forces. If we can show using one coordinate system that the body is in ‘equilibrium, then we have proven the physical fact that the body is in equilibrium, and this force equilibrium will hold in any other (admissible) coordinate system. The preceding discussion also hinted at another important consideration in engineer- ing analysis, namely, that for an effective analysis suitable coordinate systems should be chosen because the effort required to express and work with a'physical relationship in one coordinate system can be a great deal less than when. using another coordinate system. We will see in the discussion of the finite element method (see, for example, Section 4.2) that indeed one important ingredient for the effectiveness of a finite element analysis is the flexibility to choose different coordinate systems for different finite elements (domains) that together idealize the complete structure or continuum, 2.5 THE SYMMETRIC EIGENPROBLEM Av = Av In the previous section we discussed how a change of basis can be performed. In finite element analysis we are frequently interested in a change of basis as applied to symmetric matrices that have been obtained from a variational formulation, and we shall assume in the discussion to follow that A is symmetric. For example, the matrix A may represent the stiffness matrix, mass matrix, or heat capacity matrix of an element assemblage. There are various important applications (see Examples 2.34 to 2.36 and Chapter 9) in which for overall solution effectiveness a change of basis is performed using in the transformation matrix the eigenvectors of the eigenproblem Av =Av (2.80) The problem in (2.80) is a standard eigenproblem. If the solution of (2.80) is consid- ered in order to obtain eigenvalues and eigenvectors, the problem Av = Avis referred to as an eigenproblem, whereas if only eigenvalues are to be calculated, Av = Ay is called an eigenvalue problem. The objective in this section is to discuss the various properties that pertain to the solutions of (2.80).52 Vectors, Matrices, and Tensors Chap. 2 Let n be the order of the matrix A. The first important point is that there exist n nontrivial solutions to (2.80). Here the word “nontrivial” means that v must not be a null vector for which (2.80) is always satisfied. The ith nontrivial solution is given by the eigenvalue A, and the corresponding eigenvector v;, for which we have Ay = Aw (2.81) Therefore, each solution consists of an eigenpair, and we write the n solutions as (Ai, vi), Qa, Va), «+5 (Any Va), Where AS sesh (2.82) ‘We also call all n eigenvalues and eigenvectors the eigensystem of A. The proof that there must be n eigenvalues and corresponding eigenvectors can conveniently be obtained by writing (2.80) in the form (A - Adv =0 (2.83) But these equations have a solution only if det (A - AD = 0 (2.84) Unfortunately, the necessity for (2.84) to hold can be explained only after the solution of simultaneous equations has been presented. For this reason we postpone until Sec- tion 10.2.2 a discussion of why (2.84) is indeed required. Using (2.84), the eigenvalues of A are thus the roots of the polynomial pA) = det (A — Al) (2.85) This polynomial is called the characteristic polynomial of A. However, since the order of the polynomial is equal to the order of A, we have n eigenvalues, and using (2.83) we obtain n corresponding eigenvectors. It may be noted that the vectors obtained from the solution of (2.83) are defined only within a scalar multiple. EXAMPLE 2.29: Consider the matrix -1 2 a-[3 a ‘Show that the matrix has two eigenvalues. Calculate the eigenvalues and eigenvectors. The characteristic polynomial of A is ay = at [1 pel Using the procedure given in Section 2.2 to calculate the determinant of a matrix (see Exam- ple 2.13), we obtain P(A) = (-1 ~ AQ = A) — 22) M-A-6 = (A+ 2) - 3) ‘The order of the polynomial is 2, and hence there are two eigenvalues. In fact, we have M=-% b= 3Sec. 2.5 The Symmetric Eigenproblem Ay = Av 53 ‘The corresponding eigenvectors are obtained by applying (2.83) at the eigenvalues. Thus we have for Ai, ie ae 2 -collel 7 (ol o with the solution (within a scalar multiple) “L Cr? 8 . with the solution (within a scalar multiple) ~f A change of basis on the matrix A is performed by using v= PF (2.86) For Aa, we have where P is an orthogonal matrix and ¥ represents the solution vector in the new basis. Substituting into (2.80), we obtain _ AV= AT (2.87) where A=PTAP (2.88) and since A is a symmetric matrix, A is a symmetric matrix also. This transformation is called a similarity transformation, and because P is an orthogonal matrix, the transforma- tion is called an orthogonal similarity transformation. If P were not an orthogonal matrix, the result of the transformation would be Ay = ABE (2.89) where A=PTAP; B= P'P (2.90) The eigenproblem in (2.89) is called a generalized eigenproblem. However, since a generalized eigenproblem is more difficult to solve than a standard problem, the transfor- mation to a generalized problem should be avoided. This is achieved by using an orthogonal matrix P, which yields B = I. . In considering a change of basis, it should be noted that the problem A¥ = ABY in (2.89) has the same eigenvalues as the problem Av = Av, whereas the eigenvectors are related as given in (2.86). To show that the eigenvalues are identical, we consider the characteristic polynomials. For the problem in (2.89), we have BA) = det (PTAP — AP'P) 291) which can be written as (A) = det PT det (A — AI) det P (2.92) and therefore, BOA) = det PT det P p(A) (2.93)4 Vectors, Matrices, and Tensors Chap. 2 where p(A) is given in (2.85). Hence the characteristic polynomials of the problems Av = Avand A¥ = AB¥ are the same within a multiplier. This means that the eigenvalues of the two problems are identical. ‘So far we have shown that there are n eigenvalues and corresponding eigenvectors, but we have not yet discussed the properties of the eigenvalues and vectors. A first observation is that the eigenvalues are real. Consider the ith eigenpair (A,, vi), for which we have Av, = Aw (2.94) Assume that v; and A, are complex, which includes the case of real eigenvalues, and let the elements of ¥, and A, be the complex conjugates of the elements of v, and A;. Then premul- tiplying (2.94) by ¥7, we obtain WA, = AWW (2.95) On the other hand, we also obtain from (2.94), VIA = Wh (2.96) and postmultiplying by vi, we have WAv, = Aviv (2.97) But the left-hand sides of (2.95) and (2.97) are the same, and thus we have Qe — Aj¥Iv; = 0 (2.98) Since v, is nontrivial, it follows that A, = 1, and hence the eigenvalue must be real. However, it then also follows from (2.83) that the eigenvectors can be made real because the coefficient matrix A — AI is real. Another important point is that the eigenvectors that correspond to distinct eigen- values are unique (within scalar multipliers) and orthogonal, whereas the eigenvectors corresponding to multiple eigenvalues are not unique, but we can always choose an orthog- onal set. Assume first that the eigenvalues are distinct. In this case we have for two eigenpairs, Av = Av, (2.99) and Ay = Ay, (2.100) Premultiplying (2.99) by vf and (2.100) by v7, we obtain vJAv = Av]w (2.101) vIAy, = Aviv (2.102) Taking the transpose in (2.102), we have vJAw = Aviv (2.103) and thus from (2.103) and (2.101) we obtain A — Avi (2.104) Since we assumed that A, # Ay, it follows that v/v, = 0, i.e., that v; and v; are orthogonal.Sec. 2.5 The Symmetric Eigenproblem Av = Av 55 Furthermore, we can scale the elements of the vector v; to obtain viv) = 8y (2.105) where 6, = the Kronecker delta; i.e., 8) = 1 when i= j, and 8, = 0 when i # j. If (2.105) is satisfied, we say that the eigenvectors are orthonormal. It should be noted that the solution of (2.83) yields a vector in which only the relative magnitudes of the elements are defined. If all elements are scaled by the same amount, the new vector would still satisfy (2.83). In effect, the solution of (2.83) yields the direction of the eigenvector, and we use the orthonormality condition in (2.105) to fix the magnitudes of the elements in the vector. Therefore, when we refer to eigenvectors from now on it is implied that the vectors are orthonormal. EXAMPLE 2.30: Check that the vectors calculated in Example 2.29 are orthogonal and then orthonormalize them. ‘The orthogonality is checked by forming v{v2, which gives vive = (2)() + (-1)01) = 0 Hence the vectors are orthogonal. To orthonormalize the vectors, we need to make the lengths of the vectors equal to 1. Then we have wipe edt edb] «dll ‘We now turn to the case in which multiple eigenvalues are also present. The proof of eigenvector orthonormality given in (2.99) to (2.105) is not possible because for a multiple eigenvalue, A, is equal to A, in (2.104). Assume that Ar = Avey = +++ = Avem-i3 Les, Aris an m-times multiple root. Then we can show that it is still always possible to choose m orthonormal eigenvectors that correspond to A,, Av+1, + Avsm=1- This follows because for a symmetric matrix of order n, we can always establish a complete set of n orthonormal eigenvectors. Corresponding to each distinct eigenvalue we have an eigenspace with dimen- sion equal to the multiplicity of the eigenvalue, All eigenspaces are unique and are orthog- onal to the eigenspaces that correspond to other distinct eigenvalues, The eigenvectors associated with an eigenvalue provide a basis for the eigenspace, and since the basis is not unique if m > 1, the eigenvectors corresponding to a multiple eigenvalue are not unique. The formal proofs of these statements are an application of the principles discussed earlier and are given in the following examples. EXAMPLE 2.31: Show that for a symmetric matrix A of order n, there are always n orthonor- mal eigenvectors, Assume that we have calculated an eigenvalue A, and corresponding eigenvector v;. Let us construct an orthonormal matrix Q whose first column is v,, Q=% OG Ve=r ‘This matrix can always be constructed because the vectors in Q provide an orthonormal! basis for the n-dimensional space in which A is defined. However, we can now calculate ese=[* 9] @Vectors, Matrices, and Tensors Chap. 2 where Ar = O7A6 and Ay isa full matrix of order (n ~ 1). Ifn = 2, we note that Q”AQ is diagonal. In that case, if we premultiply (a) by Q and let a and hence the vector in Q is the other eigenvector and a is the other eigenvalue regardless of whether A, is a multiple eigenvalue or not. ‘The complete proof is now obtained by induction, Assume that the statement is true for a matrix of order (n — 1); then we will show that it is also true for a matrix of order n. But since ‘we demonstrated that the statement is true for n = 2, it follows that it is true for any n. ‘The assumption that there are (n — 1) orthonormal eigenvectors for a matrix of order (n — 1) gives QIAQ: = A (b) where Q, is a matrix of the eigenvectors of A. and A is a digonal matrix listing the eigenvalues of Ay. However, if we now define 10 7 [2 al we have S7Q’AQS = iB al © Let P=QS; PP=I ‘Then premultiplying (c) by P, we obtain _ py 0 7 {0 4] ‘Therefore, under the assumption in (b), the statement is also true for a matrix of order n, which completes the proof. EXAMPLE 2.32: Show that the eigenvectors corresponding to a multiple eigenvalue of multi- plicity m define an m-dimensional space in which each vector is also an eigenvector. This space is called the eigenspace corresponding to the eigenvalue considered. Let A, be the eigenvalue of multiplicity m; ie., we have Ai = Aer = Reem We showed in Example 2.31 that there are m orthonormal eigenvectors Vi, Visis «++ 5 Wiem-t corresponding to A,. These vectors provide the basis of an m-dimensional space. Consider any vector w in this space, such as WS GY + Oe Vist H+ + itm Viem—d where the a, aisi,..., are constants. The vector w is also an eigenvector because we have AW = AW + GietAVinn $+ +t Oem tAViem=1 which gives AW = AW + ciarAWie +o + Giem tA Hm = AW Therefore, any vector w in the space spanned by the m cigenvectors Vj, Wists «. « » Viem-1 is also an eigenvector. It should be noted that the vector w will be orthogonal to the eigenvectors thatSec. 2.5 The Symmetric Eigenproblem Av = Av 57 correspond to eigenvalues not equal to A,. Hence there is one eigenspace that corresponds to each, distinct or multiple, eigenvalue. The dimension of the eigenspace is equal to the multiplicity of the eigenvalue. Now that the main properties of the eigenvalues and eigenvectors of A have been presented, we can write the n solutions to Av = Av in various forms. First, we have AV = VA (2.106) where V is a matrix storing the eigenvectors, V = [vi, . . . , val, and A is a diagonal matrix with the corresponding eigenvalues on its diagonal, A = diag (A,). Using the orthonormal- ity property of the eigenvectors (i.e., V'V = I), we obtain from (2.106), VAV=A (2.107) Furthermore, we obtain the spectral decomposition of A, A= AV? (2.108) where it may be convenient to write the spectral decomposition of A as A = Daw? (2.109) It should be noted that each of these equations represents the solution to the eigen- problem Av = Av. Consider the following example. EXAMPLE 2.33: Establish the relations given in (2.106) to (2.109) for the matrix A used in Example 2.29, The eigenvalues and eigenvectors of A have been calculated in Examples 2.29 and 2.30. Using the information given in these examples, we have for (2.106), for (2.107), for (2.108), and for (2.109),58 Vectors, Matrices, and Tensors Chap. 2 The relations in (2.107) and (2.108) can be employed effectively in various important applications. The objective in the following examples is to present some solution procedures in which they are used. EXAMPLE 2.34: Calculate the kth power of a given matrix A; i.e., evaluate At. Demonstrate the result using A in Example 2.29. ‘One way of evaluating At is to simply calculate A* = AA, A‘ = A?A?, etc, However, if k is large, it may be more effective to employ the spectral decomposition of A. Assume that we have calculated the eigenvalues and eigenvectors of A; ie., we have A= VAY To calculate A?, we use A? = VAV'VAY" but because V°V= I, we have AP = VA2V" Proceeding in the same manner, we thus obtain AB = VAN? AAs an example, let A be the matrix considered in Example 2.29. Then we have ae tf? YOR Oo 1 V5 1 2jL 0 @)F oe 22 + ] It is interesting to note that if the largest absolute value of all the eigenvalues of A is smaller than 1, we have A¥ > 0 as k— ©, Thus, defining the spectral radius of A, pA) = max |u| ‘we have lim A* = 0, provided that p(A) <1. EXAMPLE 2.35: Consider the system of differential equations X+ Ax =f) @ and obtain the solution using the spectral decomposition of A. Demonstrate the result using the matrix A in Example 2.29 and wl} ef] where ° are the initial conditions. Substituting A = VAV" and premultiplying by V7, we obtain VE + A(x) = VK®) Thus if we define y = V"x, we need to solve the equations y+ Ay = V1)Sec. 2.6 The Symmetric Eigenproblem Av = Av 59 But this is a set of n decoupled differential equations. Consider the rth equation, which is typical: Jr + Aye = VEG) The solution is yt) = %y, eM + Me J eo vitG) dr Jo where °y, is the value of y, at time ¢ = 0. The complete solution to the system of equations in (a) is Wr ) ft As an example, we consider the system of differential equations f-f JE)-f) In this case we have to solve the two decoupled differential equations Sit (Dy = De at 3p = with initial conditions 172 -1)fi fi yews= sat ali) -salil We obtain ye ew be BOs To conclude the presentation, we may note that by introducing auxiliary variables, higher- order differential equations can be reduced to a system of first-order differential equations. However, the coefficient matrix A is in that case nonsymmettic, EXAMPLE 2.36: Using the spectral decomposition of an n X n symmertric matrix A, evalu- ate the inverse of the matrix. Demonstrate the result using the matrix A in Example 2.29. ‘Assume that we have evaluated the eigenvalues A, and corresponding eigenvectors v;, ++”, of the matrix A; iLe., we have solved the eigenproblem Av = Av @60 Vectors, Matrices, and Tensors Chap. 2 Premultiplying both sides of (a) by A~'A~', we obtain the eigenproblem AW Any But this relation shows that the eigenvalues of A“! are 1/A, and the eigenvectors are vi, i= 1,...,m. Thus using (2.109) for A!, we have At = VATIW? 7 wie (Quat ‘These equations show that we cannot find the inverse of A if the matrix has a zero eigenvalue. As an example, we evaluate the inverse of the matrix A considered in Example 2.29. In this case we have [alls i 3] -1 2jLo alli 2 The key point of the tranformation (2.107) is that in (2.107) we perform a change of basis {see (2.86) and (2.88)}. Since the vectors in V correspond to a new basis, they span the n-dimensional space in which A and A are defined, and any vector w can be expressed as a linear combination of the eigenvectors vs i.e., we have we 2 ay, 2.110) An important observation is that A shows directly whether the matrices A and A are singular. Using the definition given in Section 2.2, we find that A and hence A are singular if and only if an eigenvalue is equal to zero, because in that case A~' cannot be calculated. In this context it is useful to define some additional terminology. If all eigenvalues are positive, we say that the matrix is positive definite. If all eigenvalues are greater than or equal to zero, the matrix is positive semidefinite; with negative, zero, or positive eigenval- ues, the matrix is indefinite. 2.6 THE RAYLEIGH QUOTIENT AND THE MINIMAX CHARACTERIZATION OF EIGENVALUES In the previous section we defined the eigenproblem Av = Av and discussed the basic properties that pertain to the solutions of the problem. The objective in this section is to complement the information given with some very powerful principles. A number of important principles are derived using the Rayleigh quotient p(v), which is defined as : av) = ~*~ Qin The first observation is that Ar < pv) S Ae (2.112) and it follows that using the definitions given in Section 2.5, we have for any vector v, if A is positive definite p(v) > 0, if A is positive semidefinite p(v) = 0, and for A indefinite p(v)Sec. 2.6 The Rayleigh Quotient and the Minimax Characterization of Eigenvalues 61 can be negative, zero, or positive. For the proof of (2.112) we use ay, (2.113) it where v are the eigenvectors of A. Substituting for v into (2.111) and using that Av = Avi, viv, = 8), we obtain 3 weet Ane ply) = Mah + Maad ++ + Na aeoeeaee (2.114) Hence, if A: # 0, _ ob + Ca/Anad + +++ + An/Adar pv) = Me ait tae i: Fat (2.115) 7 = a, QilAdad + Gefiedal total and if A, # 0, p(w) = aes tat @.116) But since A: S Ar = + - - = A,, the relations in (2. 114) © 2116) show that (2.112) holds. Furthermore, it is scen that if v = v,, we have p(v) = Considering the practical use of the Rayleigh nae the following property is of particular value. Assume that v is an approximation to the eigenvector vj; i.e., say with € small, we have whe (2.117) Then the Rayleigh quotient of v will give an approximation to A, of order €?; i.e., plv) = A + of€?) (2.118) The notation o(¢”) means “of order e” and indicates that if 5 = o(e?), then |5| < be’, where b is a constant. To prove this property of the Rayleigh quotient, we substitute for v from (2.113) into the Rayleigh quotient expression to obtain (WT + exDA(vi + ex) ply; + €x) Gicganyeren) (2.119) VIAW, + 2eviAx + °x7Ax 7 M0 = iy a Den + ea aa However, since x is an error in v,, we can write x= Day (2.121) and Av, = A;v,, we have v/Ax = 0 and x"v; = 0, and hence But then using v/v; = ted aft, jet l+e Daf it wa pv; + ex) = (2.122)62 Vectors, Matrices, and Tensors Chap. 2 However, using the binomial theorem to expand the denominator in (2.122), we have pv, + ex) = (x + ek aa)ft ~€ (é “) + “(2 a) +e ‘| (2.123) we sei ot or pv + ex) = A + et & ty - 4) + higher-order terms (2.124) i it Yet jet The relation in (2.118) thus follows. We demonstrate the preceding results in a brief example. EXAMPLE 2.37: Evaluate the Rayleigh quotients p(v) for the matrix A used in Example 2.29. Using ¥: and vz in Example 2.29, consider the following cases: 1Lov= vy, + 2y; 2veu; 3B v= vi + 0.02y. 2 1 3 v=] +] : [1] and thus p(y) = +> tN. In case 1, we have Recalling that Ai = —2 and Ar = 3, we have, as expected, AS plv) Sa (4 ey atv) = a =] and so, as expected, p(v) = Ai. Finally, in case 3, we use v=(] + [ote] = [-ose] [2.01 -o9a|) Blleccel Do -osel) cal In case 2, we have and hence and hence pty) = 0.98 = —1.99950005 Here we note that p(v) > A: and that p(v) approximates A, more closely than v approx: imates v).Sec. 2.6 The Rayleigh Quotient and the Minimax Characterization of Eigenvalues 63 Having introduced the Rayleigh quotient, we can now proceed to a very important principle, the minimax characterization of eigenvalues. We know from Rayleigh's principle that pW) =A (2.125) where v is any vector. In other words, if we consider the problem of varying v, we will always have p(v) = Ai, and the minimum will be reached when v = v,, in which case p(v:) = A:. Suppose that we now impose a restriction on v, namely that v be orthogonal to a specific vector w, and that we consider the problem of minimizing p(v) subject to this restriction. After calculating the minimum of p(v) with the condition v'w = 0, we could start varying w and for each new w evaluate a new minimum of p(v). We would then find that the maximum value of all the minimum values evaluated is Az. This result can be generalized to the following principle, called the minimax characterization of eigenvalues, wn (2.126) with v satisfying v"w; = Ofori = 1,...,r ~ 1, r 2 2. In (2.126) we choose vectors w;, -»— 1, and then evaluate the minimum of p(v) with v subject to the condition vw. =0,i=1,...,7 — 1. After calculating this minimum we vary the vectors w, and always evaluate a new minimum. The maximum value that the minima reach is A,. The proof of (2.126) is as follows. Let v= Daw, (2.127) and evaluate the right-hand side of (2.126), which we call R, po 2 2 R= max {min TAA + OPA + Ob ardres +++ + ay (2.128) atts tab tain + ‘The coefficients a; must satisfy the conditions 0 i wyrrl (2.129) Rewriting (2.128), we obtain a(A, — Ai) #5 + F(A, — Arai) = m - Haha, — Arti) + +++ + 2A, — An) Remax jmin) pba taint tal But we can now see that for the condition a,.1 = a,12 = +++ = a, = 0, we have RSA, (2.131) and the condition in (2.129) can still be satisfied by a judicious choice for a. On the other hand, suppose that we now choose w, = v;forj = 1,...,r — 1. This would require that a, = Oforj = 1,...,r — 1, and consequently we would have R = A,, which completes the proof. A most important property that can be established using the minimax characterization of eigenvalues is the eigenvalue separation property. Suppose that in addition to the64 Vectors, Matrices, and Tensors Chap. 2 problem Av = Av, we consider the problems Aly) = Aimiyind (2.132) -where A™ is obtained by omitting the last m rows and columns of A. Hence A™ is a square-symmetric matrix of order (n — m). Using also the notation A® = A, A® = A, v® = y, the eigenvalue separation property states that the eigenvalues of the problem Alms DylrrD = Alm+D y+) separate the eigenvalues of the problem in (2.132); i.e., we have DS AIH SAM SAMY So SAM SAM, SAM, for m = 0, For the proof of (2.133) we consider the problems Av = Av and Ay) = Ay. TE we can show that the eigenvalue separation property holds for these two problems, it will hold also for m = 1, 2,...,n — 2. Specifically, we therefore want to prove that SAM Sa rad. nd (2.134) =o (2.133) Using the minimax characterization, we have a { . vay) ror © max | min Vy (2.135) vwi=0; i= 1,..., yall wearbitrary Similarly, we have vv vw. = 0; i=l, w, arbitrary for i = 1, Wr = Oy AS? = max {rin x Arh (2.136) where w, is constrained to be equal to e, to ensure that the last element in v is zero because e, is the last column of the n X n identity matrix I. However, since the constraint for A,. can be more severe and includes that for A{?, we have APS Ave (2.137) To determine A, we use A= max {rin vw, = 0; (2.138) all w, arbitrary Comparing the characterizations of A!” and A,, ie., (2.136) with (2.138), we observe that to calculate A‘ we have the same constraints as in the calculation of A. plus one more (namely, v"e, = 0), and hence asa (2.139) But (2.137) and (2.139) together establish the required result given in (2.134).Sec. 2.6 The Rayleigh Quotient and the Minimax Characterization of Eigenvalues 65 The eigenvalue separation property now yields the following result. If we write the eigenvalue problems in (2.132) including the problem Av = Av in the form Pmam) = det (A™ — AMD; m=0,...,2-1 (2.140) where p® = p, we see that the roots of the polynomial p(A"*") separate the roots of the polynomial p(A™). However, a sequence of polynomials p(x), i = 1,.. ..q, form a Sturm sequence if the roots of the polynomial p,,,(x) separate the roots of the polynomial p(x). Hence the eigenvalue separation property states that the characteristic polynomials of the problems AV = A”, n = 0, 1,...,n — 1, form a Sturm sequence. It should be noted that in the presentation we considered all symmetric matrices; i.e., the minimax characterization of eigenvalues and the Sturm sequence property are applicable to positive definite and indefinite matrices. We shall use the Sturm sequence property extensively in later chapters (see Sections 8.2.5, 10.2.2, 11.4.3, and 11.6.4). Consider the following example. EXAMPLE 2.38: Consider the eigenvalue problem Ay = Av, where 5-4 -7 ha 4 2 4 fats Evaluate the eigenvalues of A and of the matrices A’, m = 1, 2. Show that the separa property given in (2.133) holds and sketch the characteristic polynomials p(A), ae a ‘and PAA), We have pA) = det (A — AT) = (5 — AL(2 - ANS — A) - 16] + 4[-4( — a) - 28] - 7[16 + 7(2 - A] Hence P(A) = (6 — A) — A)(12 — A) and the eigenvalues are M=-6 HG = Also, PAM) = det (AP — AMD), = (5 ~ A)(2 — A) — 16 oe PRAM) = Ae? — 200 — 6 Hence XP = }- 4V73 = -0.7720 AP = 3+ IVB= 7.772 Finally, pR(A®) = det (A® — AP) =5-,8 Hence Apa 5 The separation property holds because Ar SAW? Sa. SA Sd566 Vectors, Matrices, and Tensors Chap. 2 pala) Has pitigarmy al!) = 0.7720 Bar 40 Figure £2.38 Characteristic polynomials and ANP < AP < Ag? ‘The characteristic polynomials are sketched in Fig. E2.38. 2.7 VECTOR AND MATRIX NORMS ‘We have discussed vectors, matrices, eigenvalues, and eigenvectors of symmetric matrices and have investigated the deeper significance of the elements in these entities. However, one important aspect has not been discussed so far. If we deal with single numbers, we can identify a number as being large or small. Vectors and matrices are functions of many elements, but we also need to measure their “size.” Specifically, if single numbers are used in iterative processes, the convergence of a series of numbers, say x1, x2,. . ., Xt, toa number x is simply measured by Him |x — x] = 0 (2.141) or, in words, convergence is obtained if the residual y, = |x, — x] approaches zeroas k—>™, Furthermore, if we can find constants p = 1 and c > 0 such that [xe — a] tim 121 El Tm 3P (2.142) we say that convergence is of order p. If p = 1, convergence is linear and the rate of convergence is c, in which case c must be smaller than 1. In iterative solution processes using vectors and matrices we also need a measure of convergence. Realizing that the size of a vector or matrix should depend on the magnitudeSec. 2.7. Vector and Matrix Norms. 7 of all elements in the arrays, we arrive at the definition of vector and matrix norms. A norm is a single number that depends on the magnitude of all elements in the vector or matrix. Definition: A norm of a vector v of order n written as || v|| is a single number. The norm is 4a function of the elements of v, and the following conditions are satisfied: 1. || vl] = 0 and || vl| = 0 if and only if v = 0. (2.143) 2 Hlevl] = | clll vl for any scalar c. (2.144) 3. [Iv + wll < [vl] + ll wl for vectors v and w. (2.145) The relation (2.145) is the triangle inequality. The following three vector norms are commonly used and are called the infinity, one, and two vector norms: [vl]. = max | ©] (2.146) vb = 2 Jeu (2147) llvk = (= lou (2.148) a {| vil is also known as the Euclidean vector norm. Geometrically, this norm is equal to the length of the vector v. All three norms are special cases of the vector norm V5, o:[?, where for (2.146), (2.147), and (2.148), p = ~, 1, and 2, respectively. It should be noted that each of the norms in (2.146) to (2.148) satisfies the conditions in (2.143) to (2.145). ‘We can now measure convergence of a sequence of vectors %1, Xz, X,..., Xe toa vector x. That is, for the sequence to converge to x it is sufficient and necessary that lim Ine - x (2.149) for any one of the vector norms. The order of convergence p, and in case p = 1, the rate of convergence c, are calculated in an analogous manner as in (2.142) but using norms; ie., we have jem UXt+1 = XIl lim nT = 2.1, we |x, — xIP (2.150) Looking at the relationship between the vector norms, we note that they are equivalent in the sense that for any two norms || ¢ ls, and || ¢ |b, there exist two positive constants a and a, such that Whey = aall vss (2.151) and Iv lls: = all vs, (2.152) where 51 and sz denote the co-, 1-, or 2-norms. Hence it follows that cull vl = lv lby = call vl, (2.153) where c: and c; are two positive constants that may depend on n, and of course also 1 1 Bll, = Ivlh, = Ave68 Vectors, Matrices, and Tensors Chap. 2 EXAMPLE 2.39: Give the constants c and c2 in (2.153) if, first, the norms s; and s2 are the - and |-norms, and then, second, the co- and 2-norms. Then show that in each case (2.153) is satisfied using the vector 1 -3 2 In the first case we have Ivlk $[Ivh allyl @ with c: = 1, 2 = n, and in the second case we have Ive vb s Valve ) with c: = 1 and cz = Vn. These relations show that the 1- and 2-norms are equivalent to the co- norm. We can easily show that lower and upper bounds on || y||: in (a) and || vj} in (b) cannot be closer because the equal signs are reached for the vectors v7 = [1 1... 1] and v" = e; (and any scalar multiples thereof). If we apply (a) and (b) to the given vector v, we have Ive = 3 Ivh=1+3+2=6 Iv =VI4944 = Vid and the relations in (a) and (b) read 3565()G); 3< VI4s(V3)3) In analogy with the definition of a vector norm, we also define a matrix norm. Definition: A norm of a matrix A of order n X n, written as || All, is a single number. The norm is a function of the elements of A, and the following relations hold: 1, || All = 0 and || Al] = 0 if and only if A = 0. (2.154) 2. [cl] = | ||| Al] for any scalar c, (2.155) 3. |A + Bl mal 6 -1 -1 4], om -1 u D Sots Use these vectors as the columns of a matrix A and determine the rank and kernel of A. 2A. Consider the following matrix A. Determine the constant k such that the rank of A is 2 and then determine the kernel of A. 1-1 0 A 1itk -1 o -1 1 2.5. Consider the following two vectors defined in the three-dimensional Cartesian frame with basis vectors e, 2 1 u=[3}; v=]2 4 3. (a) Evaluate the angle between these vectors. (b) Assume that a new basis is to be used, namely, the primed basis in Example 2.24, Evaluate the components of the two vectors in this basis. (©) Evaluate the angle between the vectors in this new basis. 2.6. A reflection matrix is defined as P awa E where visa vector (of order) normal to the plane of reflection. . (a) Show that P is an orthogonal matrix. (b) Consider the vector Pu where u is also a vector of order n, Show that the action of P on u is that the component of u normal to the plane of reflection has its direction reversed and the component of u in the plane of reflection is not changed, 2.7, The components of the stress tensor in the x1, x2 coordinate system of Fig. E2.25 are at a point 10 Al 6 20. (a) Establish a new basis in which the off-diagonal components are zero, and give the new diagonal components.™ Vectors, Matrices, and Tensors Chap. 2 (b) The effective stress is defined as & Va where the Sy are the components of the deviatoric stress tensor, Sy = 1) ~ Tndy and Tm is the mean stress tm = J. Prove that Fis a scalar. Then also show explicitly for the given value of + that is the same number in the old and new bases. an 4 Xt x2, 2.8. The column q is defined as where (x1, x2) are the coordinates of a point. Prove that q is not a vector. 2.9. The components of the Green-Lagrange strain tensor in the Cartesian coordinate system are defined as (see Section 6.2.2 for details) ©=4xx-) where the components of the deformation gradient X are xy a strain tensor is a second-order tensor. 2.10, The material tensor in (2.66) can be written as [see (6.185)] Cire = ABySu + w(8ir5e + 6u8,) @ where A and ys are the Lamé constants, a eae Genta "2TH This stress-strain relation can also be written in the matrix form used in Table 4.3, but in the table the use of engineering strain components is implied. (‘The tensor normal strain components are equal to the engineering normal strain components, but the tensor shear strain components are one-half the engineering components). (a) Prove that Cy, is a fourth-order tensor. (b) Consider the plane stress case and derive from the expression in (a) the expression in Table 4.3. (©) Consider the plane stress case and write (2.66) in the matrix form C’ = TCT’, where C is given in Table 4.3 and you derive T. (See also Exercise 4,39.) 2.11. Prove that (2.70) holds. 2.12, The covariant base vectors expressed in a Cartesian coordinate system are Slr Slr ‘The force and displacement vectors in this basis are R= 3g: + 4g2; u= —2g: + 3gSec.2.8 Exercises 5 (@) Calculate R + u using the covariant basis only. (b) Calculate R + u using the covariant basis for R and the contravariant basis for u. 2.13. Assume that the covariant basis is given by g: and g2 in Exercise 2.12. Let the stress and strain tensor components in the Cartesian basis be 5 [0 10 10 200J* Evaluate the components # and é,, and show explicitly that the product +-€ is the same using ‘on the one side the Cartesian stress and strain components and on the other side the contravari- ant stress and covariant strain components. 2.14, Let a and b be second-order tensors and let A and B be transformation matrices, Prove that a+ (ADB?) = (AaB) « b. (Hint: This proof is easily achieved by writing the quantities in component forms.) 2.15. Consider the eigenproblem Av = AV with 2-1 a [ i] (a) Solve for the eigenvalues and orthonormalized eigenvectors and write A in the form (2.109). (b) Calculate A‘, A“! and A~?, 2.16. Consider the eigenproblem 0.01 0.05 0.05 0.02, = 0 l}y= ar 2 lonN ‘The smallest eigenvalue and corresponding eigenvector are AL sl-sl-sI- Also, Az = 2, As = 4, Calculate the Rayleigh quotient p(v) with 1 veytOl/1 0. and show that p(¥) is closer to A; than v is to vi. 2.17. Consider the eigenproblem 2-1 0 -1 4 -1\v=av o-1 876 Vectors, Matrices, and Tensors Chap. 2 Evaluate the eigenvalues of the matrices A and A“, m = 1,2, where A™ is obtained by omitting the last m rows and columns in A. Sketch the corresponding characteristic polynomials (see Example 2.38). 2.18. Prove that the 1- and 2-norms of a vector Vv are equivalent. Then show explicitly this equivalency for the vector 2.19. Prove the relation (2.157) for the 1-norm. 2.20. Prove that | Av|j: = || Alh ll vih. 2.21. Prove that for a symmetric matrix A we have || Al} = p(A). (Hint: Use (2.108).) 2.22, Prove that (2.177) and (2.178) hold when we use the dual norm of the L-norm for the R-norm.@@ CHAPTER THREE Some Basic Concepts of Engineering Analysis and an Introduction to the Finite Element Method 3.1 INTRODUCTION The analysis of an engineering system requires the idealization of the system into a form that can be solved, the formulation of the mathematical model, the solution of this model, and the interpretation of the results (see Section 1.2). The main objective of this chapter is to discuss some classical techniques used for the formulation and solution of mathematical models of engineering systems (see also S. H. Crandall [A]). This discussion will provide a valuable basis for the presentation of finite element procedures in the next chapters. Two categories of mathematical models are considered: lumped-parameter models and continuum-mechanics-based models. We also refer to these as “discrete-system” and “continuous-system” mathematical models. In a lumped-parameter mathematical model, the actual system response is directly described by the solution of a finite number of state variables. In this chapter we discuss some general procedures that are employed to obtain the governing equations of lumped- parameter models. We consider steady-state, propagation, and eigenvalue problems and also briefly discuss the nature of the solutions of these problems. For a continuum-mechanics-based mathematical model, the formulation of the gov- ering equations is achieved as for a lumped-parameter model, but instead of a set of algebraic equations for the unknown state variables, differential equations govern the response. The exact solution of the differential equations satisfying all boundary conditions is possible only for relatively simple mathematical models, and numerical procedures must in general be employed. These procedures, in essence, reduce the continuous-system math- ematical model to a discrete idealization that can be solved in the same manner as a lumped-parameter model. In this chapter we summarize some important classical proce- dures that are employed to reduce continuous-system mathematical models to lumped- parameter numerical models and briefly show how these classical procedures provide the basis for modern finite element methods. n73 Some Basic Concepts of Engineering Analysis © Chap. 3 In practice, the analyst must decide whether an engineering system should be repre- sented by a lumped-parameter or a continuous-system mathematical model and must choose all specifics of the model. Furthermore, if a certain mathematical model is chosen, the analyst must decide how to solve numerically for the response. This is where much of the value of finite element procedures can be found; that is, finite element techniques used in conjunction with the digital computer have enabled the numerical solution of continuous- system mathematical models in a systematic manner and in effect have made possible the practical extension and application of the classical procedures presented in this chapter to very complex engineering systems, 3.2 SOLUTION OF DISCRETE-SYSTEM MATHEMATICAL MODELS In this section we deal with discrete or lumped-parameter mathematical models. The essence of a lumped-parameter mathematical model is that the state of the system can be described directly with adequate precision by the magnitudes of a finite (and usually small) number of state variables. The solution requires the following steps: 1. System idealization: the actual system is idealized as an assemblage of elements 2, Element equilibrium: the equilibrium requirements of each element are established in terms of state variables 3. Element assemblage: the element interconnection requirements are invoked to estab- lish a set of simultaneous equations for the unknown state variables 4. Calculation of response: the simultaneous equations are solved for the state variables, and using the element equilibrium requirements, the response of each element is calculated. These steps of solution are followed in the analyses of the different types of problems that we consider: steady-state problems, propagation problems, and eigenvalue problems. The objective in this section is to provide an introduction showing how problems in these particular areas are analyzed and to briefly discuss the nature of the solutions. It should be realized that not all types of analysis problems in engineering are considered; however, a large majority of problems do fall naturally into these problem areas. In the examples in this section we consider structural, electrical, fluid flow, and heat transfer problems, and we emphasize that in each of these analyses the same basic steps of solution are followed. 3.2.1 Steady-State Problems The main characteristic of a steady-state problem is that the response of the system does not change with time. Thus, the state variables describing the response of the system under consideration can be obtained from the solution of a set of equations that do not involve time as a variable, In the following examples we illustrate the procedure of analysis in the solution of some problems. Five sample problems are presented: 1, Blastic spring system 2, Heat transfer systemSec. 3.2 Solution of Discrete-System Mathematical Models 73 3. Hydraulic network 4, De network 5. Nonlinear elastic spring system. The analysis of each problem illustrates the application of the general steps of analysis, summarized in Section 3.2. The first four problems involve the analysis of linear systems, whereas the nonlinear elastic spring system responds nonlinearly to the applied loads. All the problems are well defined, and a unique solution exists for each system response. EXAMPLE 3.1: Figure E3.1 shows a system of three rigid carts on a horizontal plane that are interconnected by a system of linear elastic springs. Calculate the displacements of the carts and the forces in the springs for the loading shown. Us, Ry Us, Pa a ee (a) Physical layout YW Z ky A ky Uy FI Uy uy Y, Us Fe ke Fe AY ke Fo 1 - [ -1 7 abe if Hl [o]-fee | uy & Fo ks Ae ks “Che «CE Es] (b) Element equilibrium relations Figure E3.1 System of rigid carts interconnected by linear springsSome Basic Concepts of Engineering Analysis Chap. 3 We perform the analysis by following steps 1 to 4 in Section 3.2. As state variables that characterize the response of the system, we choose the displacements U;, Uz, and U. These displacements are measured from the initial positions of the carts, in which the springs are unstretched. The individual spring elements and their equilibrium requirements are shown in Fig. E3.1(b). To generate the governing equations for the state variables we invoke the element intercon- nection requirements, which correspond to the static equilibrium of the three carts: FD + FP + FP + FP =R FP +P +P eR, @ FO +P eR ‘We can now substitute for the element end forces F\”; 1,2, 3;7 = 1,..-, 5; using the element equilibrium requirements given in Fig. E3.1(b). Here we recognize that corresponding to the displacement components U;, U2, and Us we can write for element 1, kh 0 Ou FI” 0 0 O}}U.J=] 0 0 0 OFLUs 0 or KU = FO for element 2, er | a -h ok Ol]U)=| FP 0 0 Oflu 0 or K®U = F®, and so on. Hence the element interconnection requirements in (a) reduce to KU=R wo where u'=[U1 Ur Us) [" thtkt+kh) -(e +k) ake | K= —(k +k) (lat ky ths) ks wks wks (ka + ks) and RT=(Ri RR) Here it is noted that the coefficient matrix K can be obtained using 5 k=2K? © m where the K are the element stiffness matrices. The summation process for obtaining in () the total structure stiffness matrix by direct summation of the element stiffness matrices is referred to as the direct stiffness method. The analysis of the system is completed by solving (b) for the state variables Ui, U2, and Us and then calculating the element forces from the element equilibrium relationships in Fig. E3.1. EXAMPLE 3.2: A wall is constructed of two homogeneous slabs in contact as shown in Fig. E3.2. In steady-state conditions the temperatures in the wall are characterized by the external surface temperatures 6; and 6; and the interface temperature 6. Establish the equi-Sec. 3.2 Solution of Discrete-System Mathematical Models 81 % & 8 ‘Surface costicient coefficient 3k 2k % a Conductance Conductance 2k 3k Figure E3.2 Slab subjected to temperature boundary conditions librium equations of the problem in terms of these temperatures when the ambient temperatures and 6, are known. The conductance per unit area for the individual slabs and the surface coefficients are given in Fig. E3.2. The heat conduction law is q/A = k A@, where q is the total heat fiow, A is the area, Ad is the temperature drop in the direction of heat fiow, and k is the conductance or surface coefficient. The state variables in this analysis are @;, 62, and 6s. Using the heat condue- tion law, the element equilibrium equations are for the left surface, per unit area: qi = 3k(% — 6) for the left slab: G2 = 2K(6: — 6) for the right slab: qs = 3k(@, — 6) for the right surface: qa = 2(6s — 64) To obtain the governing equations for the state variables, we invoke the heat fiow equilibrium requirement 4) = gz = 93 = qa. Thus, 3k(@ — 6:) = 2k(@, — 6) 2k(8, — 6) = 3k( — 05) 3k(6, — @:) = 2k(@s — O) ‘Writing these equations in matrix form we obtain Sk —2k 0 Tf 3k 6 2k Sk —3k ||} =] 0 @ 0 -3k Sk j] 2k64, These equilibrium equations can be also derived in a systematic manner using a direct stiffness procedure. Using this technique, we proceed as in Example 3.1 with the typical element equilibrium relations 1 -1]f6) _ far 1; ile) - (2)82 Some Basic Concepts of Engineering Analysis Chap. 3 where q., 4 ate the heat flows into the element and 6,, 6, are the element-end temperatures. For the system in Fig. E3.2 we have two conduction elements (each slab being one element), hence we obtain ¢ 2k -%e O7176 ‘3k(& — 61) -% sk —3k |) | = 0 © 0 -3k ak} es 2k(@. — 63) Since 0; and @, are unknown, the equilibrium relations in (b) are rearranged for solution to obtain the relations in (a). Itis interesting to note the analogy between the displacement and force analysis of the spring system in Example 3.1 and the temperature and heat transfer analysis in Example 3.2. The coefficient matrices are very similar in both analyses, and they can both be obtained in a very systematic manner. To emphasize the analogy we give in Fig. 3.1 a spring model that is governed by the coefficient matrix of the heat transfer problem. 3k 2k 3k 2k Uy Py Uz Fa Us, Pa Figure 3.1 Assemblage of springs governed by same coefficient matrix as the heat transfer problem in Fig. E3.2 ‘We next consider the analyses of a simple flow problem and a simple electrical system, both of which are again analyzed in much the same manner as the spring and heat transfer problems. EXAMPLE 3.3: Establish the equations that govern the steady-state pressure and flow distribu- tions in the hydraulic network shown in Fig. E3.3. Assume the fluid to be incompressible and the pressure drop in a branch to be proportional to the flow q through that branch, Ap = Rg, where Ris the branch resistance coefficient. In this analysis the elements are the individual branches of the pipe network. As unknown state variables that characterize the flow and pressure distributions in the system we select the c R=3b o Figure E33 Pipe networkSec. 3.2 Solution of Discrete-System Mathematical Models 83 Pressures at A, C, and D, which we denote as Pa, Pc, and Pp, and we assume that the pressure at Bis zero. Thus, we have for the elements ais, Po Po B00 Pa Pe Po Pc po . gh = BRS glow = Fs we = ‘The element interconnectivity requirements require continuity of flow, hence Q=ntna (b) alc= 9st qe gales = a5 + Ge ‘Substituting from (a) into (b) and writing the resulting equations in matrix form, we obtain 3-2 Olf~ 100" -6 31 -25|)pc]=| 0 -1 1 i fhpo. 0 oe Cpa 30bQ” or -6 3 -25/lp-|=] 0 © o ~25 31 \Lpo 0 The analysis of the pipe network is completed by solving from () for the pressures pa, Pc, and Po, and then the element equilibrium relations in (a) can be employed to obtain the flow distribution. The equilibrium relations in (©) can also be derived—as in the preceding spring and heat transfer examples—using a direct stiffness procedure. Using this technique, we proceed as in Example 3.1 with the typical element equilibrium relations apy o-a | _f% R-1 illel~ Lo, where qi, q) ate the fluid flows into the element and p,, p) are the element-end pressures. EXAMPLE 3.4: Consider the de network shown in Fig. E3.4, The network with the resistances shown is subjected to the constant-voltage inputs E and 2E at A and B, respectively. We are to determine the steady-state current distribution in the network. In this analysis we use as unknown state variables the currents /,, /2, and Js. The system elements are the resistors, and the element equilibrium requirements are obtained by applying ‘Ohm’s law to the resistors. For a resistor R, carrying current J, we have Ohm’s law AE = RI where AF is the voltage drop across the resistor. The clement interconnection law to be satisfied is Kirchhoff’s voltage law for each closed loop in the network, 2E = 2Rh + 2R(h — b) E = 4R(h — bs) 0 = ORI; + 4R(s — hh) + 2R( — h)84 Some Basic Concepts of Engineering Analysis © Chap. 3 Fa @) ‘ll © sles 2R 4k |. ® @ 2E E— Figure E34 De network ‘Writing these equations in matrix form, we obtain 4R 0 -2RTTh 2E 0 4k -4R||h|=| E (@) —2R -4R 12R jh 0 ‘The analysis is completed by solving these equations for Ii, /2, and Js, Note that the equilibrium equations in (a) could also have been established using a direct stiffness procedure, as in Examples 3.1 to 3.3. ‘We should note once again that the steps of analysis in the preceding structural, heat transfer, fluid flow, and electrical problems are very similar, the basic analogy being possibly best expressed in the use of the direct stiffness procedure for each problem. This indicates that the same basic numerical procedures will be applicable in the analysis of almost any physical problem (see Chapters 4 and 7). Each of these examples deals with a linear system; ie., the coefficient matrix is constant and thus, if the right-hand-side forcing functions are multiplied by a constant a, the system response is also « times as large. We consider in this chapter primarily linear systems, but the same steps for solution summarized previously are also applicable in nonlinear analysis, as demonstrated in the following example (see also Chapters 6 and 7). EXAMPLE 3.5: Consider the spring-cart system in Fig. E3.1 and assume that spring (2) now has the nonlinear behavior shown in Fig. E3.5. Discuss how the equilibrium equations given in Example 3.1 have to be modified for this analysis. As long as U; Ay, the equilibrium equations in Example 3.1 are applicable with ki = k. However, if the loads are such that U; > Ay, i.e., Fi” > F,, we need to use a different value for k,, and this value depends on the force F{ acting in the element. Denoting the stiffness value by ,, a8 shown in Fig. E3.5, the response of the system is described for any load by the equilibrium equations KU=R @ where the coefficient matrix is established exactly as in Example 3.1 but using k, instead of ki, (i th+ht+k) (ke +k) ak K, — (ke + ks) (atk ths) ks ky wks (ke + ks), )Sec. 3.2 Solution of Discrete-System Mathematical Models 8s Figure E35 Spring © of the cart-spring system of Fig. E3.1 with nonlinear elastic charac- teristics Although the response of the system can be calculated using this approach, in which K, is referred to as the secant matrix, we will see in Chapter 6 that in general practical analysis we actually use an incremental procedure with a tangent stiffness matrix. These analyses demonstrate the general analysis procedure: the selection of unknown state variables that characterize the response of the system under consideration, the identification of elements that together comprise the complete system, the establishment of the element equilibrium requirements, and finally the assemblage of the elements by invok- ing interelement continuity requirements. A few observations should be made. First, we need to recognize that there is some choice in the selection of the state variables. For example, in the analysis of the carts in Example 3.1, we could have chosen the unknown forces in the springs as state variables. A second observation is that the equations from which the state variables are calculated can be linear or nonlinear equations and the coefficient matrix can be of a general nature. However, it is most desirable to deal with a symmetric positive definite coefficient matrix because in such cases the solution of the equations is numerically very effective (see Section 8.2). In general, the physical characteristics of a problem determine whether the numerical solution can actually be cast in a form that leads to a symmetric positive definite coefficient matrix. However, even if possible, a positive definite coefficient matrix is obtained only if appropriate solution variables are selected, and in a nonlinear analysis an appropriate linearization must be performed in the iterative solution. For this reason, in practice, it is valuable to employ general formulations for whole classes of problems (e.g., structural analysis, heat transfer, and so on—see Sections 4.2, 7.2, and 7.3) that for any analysis lead to a symmetric and positive definite coefficient matrix. In the preceding discussion we employed the direct approach of assembling the system-governing equilibrium equations. An important point is that the governing equi- librium equations for state variables can in many analyses also be obtained using an extremum, or variational formulation. An extremum problem consists of locating the set (or86 ‘Some Basic Concepts of Engineering Analysis Chap. 3 sets) of values (state variables) U,, i = 1, . . . n, for which a given functional TI(U;, . . . U,) is a maximum, is a minimum, or has a saddle point. The condition for obtaining the equations for the state variables is en =0 Ga) 5 _ al aml and since aT = 5p, Ui + +--+ Sy Us 2) we must have a 0 for ” 3) ‘We note that 5U; stands for “variations in the state variables U; that are arbitrary except that they must be zero at and corresponding to the state variable boundary conditions.” The second derivatives of IT with respect to the state variables then decide whether the solution corresponds to a maximum, a minimum, or a saddle point. In the solution of lumped- parameter models we can consider that Il is defined such that the relations in (3.3) generate the governing equilibrium equations.” For example, in linear structural analysis, when displacements are used as state variables, [I is the total potential (or total potential energy) 1m=u-W G4) where %U. is the strain energy of the system and ‘W is the total potential of the loads. The solution for the state variables corresponds in this case to the minimum of IT. EXAMPLE 3.6: Consider a simple spring of stiffness k and applied load P, and discuss the use of (3.1) and (3.4). Let u be the displacement of the spring under the load P. We then have = yk W = Po and T= bau? - Pu Note that for a given P, we could graph II as a function of u. Using (3.1) we have, with as the only variable, an aM = (ku — P) 8u; 0 Se tu P which gives the equilibrium equation ku =P @ ‘Using (a) to evaluate ‘W’, we have at equilibrium W = ku; i.e, W = 2U and Tl = —}ku? = —}Pu, Also, 8°1/au? = k and hence at equilibrium I] js at its minimum. EXAMPLE 3.7: Consider the analysis of the system of rigid carts in Example 3.1. Determine TI and invoke the condition in (3.1) for obtaining the governing equilibrium equations. Using the notation defined in Example 3.1, we have u = $U"KU @ ' More precisely, the variations in the state variables must be zero at and corresponding to the essential boundary conditions, as further discussed in Section 3.3.2. In this way we consider a specific variational formulation, as further discussed in Chapters 4 and 7.Sec. 3.2 Solution of Discrete-System Mathematical Models 87 and W=UR (b) where it should be noted that the total strain energy in (a) could also be written as s w= yo(S Keo = }UKOU + JUKOU +. + FUTKOU = Uy + Uz te + Uy where 4; is the strain energy stored in the ith element. Using (a) and (b), we now obtain Tl = 3U°KU - UR © Applying (3.1) gives KU=R Solving for U and then substituting into (c), we find that TI corresponding to the displacements at system equilibrium is T= Since the same equilibrium equations are generated using the direct solution approach and the variational approach, we may ask what the advantages of employing a variational scheme are. Assume that for the problem under consideration II is defined. The equilibrium equations can then be generated by simply adding the contributions from all elements to I and invoking the stationarity condition in (3.1). In essence, this condition generates auto- matically the element interconnectivity requirements. Thus, the variational technique can be very effective because the system-governing equilibrium equations can be generated “quite mechanically.” The advantages of a variational approach are even more pronounced when We consider the numerical solution of a continuous system (see Section 3.3.2). How- ever, a main disadvantage of a variational approach is that, in general, less physical insight into a problem formulation is obtained than when using the direct approach. Therefore, it may be critical that we interpret physically the system equilibrium equations, once they have been established using a variational approach, in order to identify possible errors in the solution and in order to gain a better understanding of the physical meaning of the equations. 3.2.2 Propagation Problems The main characteristic of a propagation or dynamic problem is that the response of the system under consideration changes with time. For the analysis of a system, in principle, the same procedures as in the analysis of a steady-state problem are employed, but now the state variables and element equilibrium relations depend on time. The objective of the analysis is to calculate the state variables for all time t. Before discussing actual propagation problems, let us consider the case where the time effect on the element equilibrium relations is negligible but the load vector is a function of time. In this case the system response is obtained using the equations governing the steady- state response but substituting the time-dependent load or forcing vector for the load vector employed in the steady-state analysis. Since such an analysis is in essence still a steady-state analysis, but with steady-state conditions considered at any time #, the analysis may be referred to as a pseudo steady-state analysis.88 Some Basic Concepts of Engineering Analysis Chap. 3 In an actual propagation problem, the element equilibrium relations are time- dependent, and this accounts for major differences in the response characteristics when compared to steady-state problems. In the following we present two examples that demon- strate the formulation of the governing equilibrium equations in propagation problems. Methods for calculating the solution of these equations are given in Chapter 9. EXAMPLE 3.8: Consider the system of rigid carts that was analyzed in Example 3.1. Assume that the loads are time-dependent and establish the equations that govern the dynamic response of the system. For the analysis we assume that the springs are massless and that the carts have masses rm, mz, and ms (which amounts to lumping the distributed mass of each spring to its two end points). Then, using the information given in Example 3.1 and invoking d’Alembert’s principle, the element interconnectivity requirements yield the equations FD + FP + FP + FP = RO — mis FP + F) + FP = RQ — mth FY + FY = Rst) — ms Os i=1,2,3 ‘Thus we obtain as the system-governing equilibrium equations MU + KU = R(®) @ where K, U, and R have been defined in Example 3.1 and M is the system mass matrix m 0 0 M=]0 m 0 0 0 ms ‘The equilibrium equations in (a) represent a system of ordinary differential equations of second order in time. For the solution of these equations it is also necessary that the initial conditions on U and U be given; i.c., we need to have °U and °U, where PU =Ujecs °C = Uh -o Earlier we mentioned the case of a pseudo steady-state analysis. Considering the response of the carts, such analysis implies that the loads change very slowly and hence mass effects can be neglected. Therefore, to obtain the pseudo steady-state response, the equilibrium equations (a) in Example 3.8 should be solved with M = 0. EXAMPLE 3.9: Figure E3.9 shows an idealized case of the transient heat flow in an electron tube. A filament is heated to a temperature 6, by an electric current; heat is convected from the filament to the surrounding gas and is radiated to the wall, which also receives heat by convection from the gas. The wall itself convects heat to the surrounding atmosphere, which is at tempera- ture 6. It is required to formulate the system-governing heat flow equilibrium equations, In this analysis we choose as unknown state variables the temperature of the gas, 6;, and the temperature of the wall, 8. The system equilibrium equations are generated by invoking the heat flow equilibrium for the gas and the wall. Using the heat transfer coefficients given inSec. 3.2 Solution of Discrete-System Mathematical Models 89 Walt Heat capacity _ ¢, of gas ‘ Haat capacity of wall =G Atmosphere Filament Convection Atmosphere {ky 0 Figure E3.9 Heat transfer idealization of an electron tube Fig. E3.9, we obtain for the gas 6, C1 = FQ, — 81) — bo, ~ 64) and for the wall 5B = (C63 ~ (03) + lO, ~ 8) — HO ~ 4) ‘These two equations can be written in matrix form as cé6+Ke=Q @ Ke a cl where 0 Bb -ans al K{(O)* — (83)*) + ka8a, ‘We note that because of the radiation boundary condition, the heat flow equilibrium equations are nonlinear in @, Here the radiation boundary condition term has been incorporated in the heat flow load vector Q. The solution of the equations can be carried out as described in Section 9.6. Although, in the previous examples, we considered very specific cases, these examples illustrated in a quite general way how propagation problems of discrete systems are formu-90 Some Basic Concepts of Engineering Analysis Chap. 3 lated for analysis. In essence, the same procedures are employed as in the analysis of steady-state problems, but “time-dependent loads” are generated that are a result of the “resistance to change” of the elements and thus of the complete system. This resistance to change or inertia of the system must be considered in a dynamic analysis. Based on the preceding arguments and observations, it appears that we can conclude that the analysis of a propagation problem is a very simple extension of the analysis of the corresponding steady-state problem. However, we assumed in the earlier discussion that the discrete system is given and thus the degrees of freedom or state variables can be directly identified. In practice, the selection of an appropriate discrete system that contains all the important characteristics of the actual physical system is usually not straightforward, and in general a different discrete model must be chosen for a dynamic response prediction than is chosen for the steady-state analysis. However, the discussion illustrates that once the discrete model has been chosen for a propagation problem, formulation of the governing equilibrium equations can proceed in much the same way as in the analysis of a steady-state response, except that inertia loads are generated that act on the system in addition to the externally applied loads (see Section 4.2.1). This observation leads us to anticipate that the procedures for solving the dynamic equilibrium equations of a system are largely based on the techniques employed for the solution of steady-state equilibrium equations (see Sec- tion 9.2). 3.2.3 Eigenvalue Problems In our eatlier discussion of steady-state and propagation problems we implied the existence of a unique solution for the response of the system. A main characteristic of an eigenvalue problem is that there is no unique solution to the response of the system, and the objective of the analysis is to calculate the various possible solutions. Eigenvalue problems arise in both steady-state and dynamic analyses. Various different eigenvalue problems can be formulated in engineering analysis. In this book we are primarily concerned with the generalized eigenvalue problem of the form Av = ABy G5) where A and B are symmetric matrices, A is a scalar, and v is a vector. If A, and v, satisfy (3.5), they are called an eigenvalue and an eigenvector, respectively. In steady-state analysis an eigenvalue problem of the form (3-5) is formulated when it is necessary to investigate the physical stability of the system under consideration. The question that is asked and leads to the eigenvalue problem is as follows: Assuming that the steady-state solution of the system is known, is there another solution into which the system could bifurcate if it were slightly perturbed from its equilibrium position? The answer to this question depends on the system under consideration and the loads acting on the system. We consider a very simple example to demonstrate the basic idea. EXAMPLE 3.10: Consider the simple cantilever shown in Fig. E3.10. The structure consists of a rotational spring and a rigid lever arm. Predict the response of the structure for the load applications shown in the figure. ‘We consider first only the steady-state response as discussed in Section 3.2.1. Since the bar is rigid, the cantilever is a single degree of freedom system and we employ A, as the state variable,Rotational spring, stiffness k Rigid, weightless bar P, P, P, Multiple solutions k Pan = £ ae wo aw a, {c) Displacement responses 4 w : Pew : w P P o a an {d) Loading conditions (small load W) P, a a ae ye a Aydue to Wat P= 0 Aydueto Wat P=0 A due to Wat P=0 a= (Pe Wit Wh for P< Patt k @) ) o mm an (e) Displacement responses Figure E3.10 Analysis of a simple cantilever model 192 Some Basic Concepts of Engineering Analysis Chap. 3 In loading condition I, the bar is subjected to a longitudinal tensile force P, and the moment in the spring is zero. Since the bar is rigid, we have A, =0 @) Next consider loading condition i. Assuming small displacements we have in this case PL? Aa (b) Finally, for loading condition IIT we have, as in condition I, A, =0 © We now proceed to determine whether the system is stable under these load applications. To investigate the stability we perturb the structure from the equilibrium positions defined in (a), (b), and (¢) and ask whether an additional equilibrium position is possible. Assume that A, is positive but small in loading conditions I and I. If we write the equilibrium equations taking this displacement into account, we observe that in loading condition the small nonzero A, cannot be sustained, and that in loading condition II the effect of including ‘A, in the analysis is negligible. Consider next that A, > 0 in loading condition III. In this case, for an equilibrium configuration to be possible with A. nonzero, the following equilibrium equation must be satisfied: PA, But this equation is satisfied for any A.,, provided P = k/L. Hence the critical load Pas: at which ‘an equilibrium position in addition to the horizontal one becomes possible is L Pa = In summary, we have P< Psu only the horizontal position of the bar is possible; equilibrium is stable P= Pe horizontal and deflected positions of the bar are possible; the horizontal equilibrium position is unstable for P = Pas. To gain an improved understanding of these results we may assume that in addition to the load P shown in Fig. E3.10(b), a small transverse load W is applied as shown in Fig. E3.10(@). If we then perform an analysis of the cantilever model subjected to P and W, the response curves shown schematically in Fig. E3.10(¢) are obtained. Thus, we observe that the effect of the load W decreases and is constant as P increases in loading conditions I and 11, but that in loading condition ITI the transverse displacement A, increases very rapidly as P approaches the critical load, Pai. The analyses given in Example 3.10 illustrate the main objective of an eigenvalue formulation and solution in instability analysis, namely, to predict whether smal distur- bances that are imposed on the given equilibrium configuration tend to increase very substantially. The load level at which this situation arises corresponds to the critical loading of the system. In the second solution carried out in Example 3.10 the small disturbance wasSec. 3.2 Solution of Discrete-System Mathematical Models 93 due to the small load W, which, for example, may simulate the possibility that the horizontal oad on the cantilever is not acting perfectly horizontally. In the eigenvalue analysis, we simply assume a deformed configuration and investigate whether there is a load level that indeed admits such a configuration as a possible equilibrium solution. We shall discuss in Section 6.8.2 that the eigenvalue analysis really consists of a linearization of the nonlinear response of the system, and that it depends largely on the system being considered whether a reliable critical load is calculated. The eigenvalue solution is particularly applicable in the analysis of “beam-column-type situations” of beam, plate, and shell structures. EXAMPLE 3.11: Experience shows that in structural analysis the critical load on column-type structures can be assessed appropriately using an eigenvalue problem formulation, Consider the system defined in Fig. E3.11. Construct the eigenvalue problem from which the critical loading on the system can be calculated. ‘Smooth hinges at A,B,and C KUy + kU P Figure E3.11 Instability analysis of a columnSome Basic Concepts of Engineering Analysis Chap. 3 As in the derivation of steady-state equilibrium equations (see Section 3.2.1), we can employ the direct procedure or a variational approach to establish the problem-governing equa- tions, and we describe both techniques in this problem solution, Inthe direct approach we establish the governing equilibrium equations directly by consid- ering the equilibrium of the structure in its deformed configuration. Referring to Fig. E3.11, the moment equilibrium of bar AB requires that PL sin(a + B) = kU; L cos(a + B) + ka (a Similarly, for bars CBA we need PL[sin(a + f) + sin B] = kU; L[cos(a + B) + cos B] + kU2L cos B () We select U; and Us as state variables that completely describe the structural response. We also assume small displacements, for which Lsin(a + B) = U; — U2; LsinB = Ur Us = 2, L ‘Substituting into (a) and (b) and writing the resulting equations in matrix form, we obtain ke _ ake e+e 27 l]ul _ pt —1/ la ake kL AU 1 ol ly. We can symmetrize the coefficient matrices by multiplying the first row by —2 and adding the result to row 2, which gives the eigenvalue problem Loo(a+f)*L; Leosp+k a P © Tt may be noted that the second equation in (c) can also be obtained by considering the moment equilibrium of bar CB. Considering next the variational approach, we need to determine the total potential TT of the system in the deformed configuration. Here we have Th = kU} + 2kU3 + $ha? — PL[1 — cos(a + B) + 1 — cos A] @ As in the direct approach, we now assume small displacement conditions. Since we want to derive, using (3.1), an eigenvalue problem of form (3.5) in which the coefficient matrices are independent of the state variables, we approximate the trigonometric expressions to second order in the state variables. Thus, we use (a+ pF costa + B) = 1 ~ = © and ©Sec. 3.2 Solution of Discrete-System Mathematical Models 95 Substituting from (¢) and (f) into (d) we obtain - 20; 1 1 1 = hut + kui + Sa Applying the stationarity principle, an au the equations in (c) are obtained. Considering now dynamic analysis, an eigenvalue problem may need to be formulated in the solution of the dynamic equilibrium equations. In essence, the objective is then to find a mathematical transformation on the state variables that is employed effectively in the solution of the dynamic response (see Section 9.3). In the analysis of physical problems, it is then most valuable to identify the eigenvalues and vectors with physical quantities (see Section 9.3), To illustrate how eigenvalue problems are formulated in dynamic analysis, we present the following examples. EXAMPLE 3.12: Consider the dynamic analysis of the system of rigid carts discussed in Example 3.8. Assume free vibration conditions and that U = é sin(or — y) (@) where & is a vector with components independent of time, « is a circular frequency, and is a phase angle. Show that with this assumption an eigenvalue problem of the form given in (3.5) is obtained when searching for a solution of ¢ and «. ‘The equilibrium equations of the system when considering free-vibration conditions are MU + KU=0 () where the matrices M and K and vector U have been defined in Examples 3.1 and 3.8. If U given in (a) isto be a solution of the equations in (b), these equations must be satisfied when substituting for U, -w™Mé sin(o — Y) + Ke sin(wr — Y) = 0 ‘Thus, for (a) to be a solution of (b) we obtain the condition Ko = w*Md © which is an eigenvalue problem of form (3.5). We discuss in Section 9.3 the physical character- istics of a solution, «? and ¢,, to the problem in (c). EXAMPLE 3.13: Consider the electric circuit in Fig. E3.13. Determine the eigenvalue problem from which the resonant frequencies and modes can be calculated when L, = Lz = L and C=GQ=C. Our first objective is to derive the dynamic equilibrium equations of the system, The element equilibrium equation for an inductor is dl on @96 Some Basic Concepts of Engineering Analysis Chap. 3 =C Figure E3.13 Electric circuit where L is the inductance, I is the current through the inductor, and V is the voltage drop across the inductor. For a capacitor of capacitance C the equilibrium equation is av T= Ca (b) As state variables we select the currents /, and / shown in Fig. E3.13, The governing equilibrium equations are obtained by invoking the element interconnectivity requirements contained in Kirchhoff’s voltage law: Ve, + Vig + Ve, = 0 7 © Vi, + Vig + Ve, = 0 Differentiating (a) and (c) with respect to time and substituting into (c) with Li = L2 = Land C. = C= C, we obtain 1 1h] .af2 fn] _ feo i alli] +a dt] : [0] @ ‘We note that these equations are quite analogous to the free-vibration equilibrium equations of a structural system. Indeed, recognizing the analogy 1 > displacement; ke stiffness; L—> mass the eigenproblem for the resonant frequencies is established as in Example 3.12 (and an equiv- alent structural system could be constructed). 3.2.4 On the Nature of Solutions In the preceding sections we discussed the formulation of steady-state, propagation, and eigenvalue problems, and we gave a number of simple examples. In all cases a system of equations for the unknown state variables was formulated but not solved. For the solution of the equations we refer to the techniques presented in Chapters 8 to 11. The objective in this section is to discuss briefly the nature of the solutions that are calculated when steady- state, propagation, or eigenvalue problems are considered. For steady-state and propagation problems, it is convenient to distinguish between linear and nonlinear problems. In simple terms, a linear problem is characterized by the fact that the response of the system varies in proportion to the magnitude of the applied loads. All other problems are nonlinear, as discussed in more detail in Section 6.1. To demonstrate in an introductory way some basic response characteristics that are predicted in linear steady-state, propagation, and eigenvalue analyses we consider the following example.Sec. 3.2 Solution of Discrete-System Mathematical Models 7 EXAMPLE 3.14: Consider the simple structural system consisting of an assemblage of rigid weightless bars, springs, and concentrated masses shown in Fig. E3.14. The elements are con- nected at A, B, and C using frictionless pins, It is required to analyze the discrete system for the loading indicated, when the initial displacements and velocities are zero. The response of the system is described by the two state variables U; and U2 shown in Fig. E3.14(¢). To decide what kind of analysis is appropriate we need to have sufficient informa- tion on the characteristics of the structure and the applied forces F and P. Let us assume that the structural characteristics and the applied forces are such that the displacements of the element assemblage are relatively small, L Smooth hinges at AB,and C {a) Discrete system {b) Loading conditions {c) External forces in deformed configuration Figure E3.14 A two degree of freedom system