0 Bewertungen0% fanden dieses Dokument nützlich (0 Abstimmungen)

86 Ansichten361 Seiten© Attribution Non-Commercial (BY-NC)

PDF, TXT oder online auf Scribd lesen

Attribution Non-Commercial (BY-NC)

Als PDF, TXT **herunterladen** oder online auf Scribd lesen

0 Bewertungen0% fanden dieses Dokument nützlich (0 Abstimmungen)

86 Ansichten361 SeitenAttribution Non-Commercial (BY-NC)

Als PDF, TXT **herunterladen** oder online auf Scribd lesen

Sie sind auf Seite 1von 361

Wilfried Kausel Institute of Musical Acoustics, University of Music (Institut f r Wiener Klangstil, u Universit t f. Musik u. darstellende Kunst) a Singerstr. 26/a, A-1010, Vienna, Austria phone: (43) 1-71155-4311, email: kausel@mdw.ac.at November 10, 2003

Habilitationsschrift zur Erlangung der Venia Docendi fr das Fach Musikalische Akustik, eingereicht im Oktober 2003 an der Universitt fr Musik und darstellende Kunst in Wien.

Concepts, Algorithms & Applications Schriftenreihe des Instituts fr Wiener Klangstil (Musikalische Akustik) an der Universitt fr Musik und darstellende Kunst Wien. Band 7 Copyright 2003 Wilfried Kausel Institut fr Wiener Klangstil (Musikalische Akustik), Wien, (2003). ISBN: 3-900914-05-2

Kurzfassung

Dieses als Habilitationsschrift f r das Fachgebiet der Musikalischen Akuu stik am Institut f r Musikalische Akustik (auch Inst. f. Wiener Klangstil) u der Universit t f r Musik und darstellende Kunst in Wien verfasste Buch a u soll einen Leitfaden f r wissenschaftlich t tige Musiker mit naturwissenu a schaftlichen Vorkenntnissen darstellen, die aktuelle Computersimulationsund Analysemethoden, die sich auf Musikinstrumente und Schallausbreitungsprobleme anwenden lassen, kennen lernen wollen oder beabsichtigen, solche Verfahren selbst einzusetzen. Die empfohlenen Vorkenntnisse sind neben der ublichen Mittelschul mathematik und -physik die elementaren Rechenregeln f r Differentialu gleichungen und partielle Differentialgleichungssysteme sowie Grundwissen uber Vektoren und Matrizen sowie deren zeitliche und r umliche Ab a leitungen. Die allgemeinen physikalischen Grundgesetze werden im ersten Kapitel vorgestellt und besprochen ebenso wie die Vereinfachungen, die unter bestimmten Voraussetzungen zul ssig sind, und die die rechnerische Behanda lung in manchen F llen erst erm glichen. a o Im zweiten Kapitel werden die meisten der gegenw rtig bekannten Mea thoden zur L sung von allgemeinen durch Differentialgleichungssysteme o beschreibbaren physikalischen Problemstellungen vorgestellt. Die gebr uchlichste Methode der Finiten Differenzen wird dabei detaila liert genug behandelt, um eigene Anwendungen beginnen zu k nnen. Die o hier vorgestellten Methoden sind sehr allgemein einsetzbar und keinesfalls auf die musikalische Akustik beschr nkt. a Die im dritten Kapitel vorgestellten Methoden zielen sowohl hinsicht-

lich ihrer Auswahl als auch wegen der gew hlten Darstellung schon sehr a spezisch auf Problemstellungen, die in der Musikalischen Akustik h ua g auftreten. Die Verfahren werden dabei sehr konkret behandelt, mit dem Ziel dem Leser eine konkrete eigene Anwendung zu erm glichen. o Im vierten Kapitel werden schlielich Beispiele gegeben, wie die beschriebenen Konzepte im Rahmen der Musikalischen Akustik eingesetzt werden k nnen, und welche Ergebnisse mit ihnen erzielbar sind. Die Beio spiele entspringen der praktischen Arbeit am Institut f r Wiener Klangu stil und besch ftigen sich mit praxisbezogenen Fragestellungen von Ina strumentenbauern, Musikern und nat rlich auch wissbegierigen Wissenu schaftlern. In diesem Kapitel wird auch ein Uberblick uber besonders interessante in letzter Zeit ver ffentlichte Berechnungsmethoden f r die o u Musikalische Akustik gegeben. Nat rlich ist die Auswahl willk rlich und u u unvollst ndig und erst die weitere Entwicklung kann die wirkliche Bedeua tung der vorgestellten Methoden aufzeigen. Der Anhang h lt schlielich eine Einf hrung in die Signalverarbeitung a u von Abtastsystemen bereit, die zwar von allgemeiner Bedeutung sowohl f r die Diskretisierung von Differentialgleichungen als auch ganz besonu ders f r die akustische Messtechnik und Klanganalyse sind, aber f r viele u u wohl eher eine Wiederholung darstellen werden. Die Tabellen f r die Berechnung von Differenzenquotienten sollen bei u der Implementierung einer Finite-Differenzen Methode helfen. Sie enthalten alle Koefzienten der ublichen Interpolations-Polynome bis zum sechsten Grad einseitig und bis zum zw lften Grad symmetrisch sowie o deren Ableitungen erster bis vierter Ordnung. Stichworte: Musikalische Akustik, Numerische Methoden, Computersimulation, Finite-Differenzen, Lattice-Boltzmann, Optimierung, Physikalische Modelle, Wellengleichung, Signalverarbeitung, Tonerzeugung,

Abstract

This book has been written to obtain the habilitation in the eld of musical acoustics at the Institute of Musical Acoustics (Inst. f. Wiener Klangstil) at the University of Music and Performing Arts in Vienna. It is intended to serve as a guide for scientically interested musicians with mathematical and physical background who want to know or use actual computersimulation and -analysis methods for musical instruments and sound propagation. The recommended prior knowledge exceeds the typical knowledge of lower level undergraduate students majoring in applied mathematics, engineering, physics or anything related only in a few respects. It is assumed that the reader has some elementary knowledge in handling differential equations and partial differential equation systems, and he should be familiar with the basic calculation rules for vectors and matrices, and their time and spatial derivatives. In the rst chapter the fundamental physical laws are introduced and discussed as well as some simplications of them which are valid under certain circumstances and which in some cases open the way for analytical solutions. In the second chapter most of the currently known concepts for solving general physical simulation problems which can be described by systems of partial differential equations are introduced. The most popular method of nite differences is covered in sufcient detail in order to enable the reader to start his own applications. The reviewed methods are very generally applicable and not at all restricted to the eld of musical acoustics. The concepts presented in the third chapter, by their selection and by

their representation, already specically aim at problems and questions which are common in the eld of musical acoustics. The methods are described very concisely with the intention to enable the reader to apply them in applications of his own. In the fourth chapter examples are given showing how the described methods can be applied in the realm of musical acoustics and what kind of results can be expected there. The examples have been worked out as a result of the practical occupation with questions at the Institut f r Wiener u Klangstil, which have been brought up by instrument makers, musicians and, of course, curious scientists. In this chapter it is also attempted to give an overview of especially interesting publications issued in the recent past covering computational methods for musical acoustics. Of course, the selection is in a way arbitrary and incomplete and the future developments will reveal the true signicance of the presented methods. In the Appendix an introduction to signal processing of sampled data systems can be found, which does have general signicance for discretizations of differential equations as well as for the acoustical measurement technique and sound analysis, but which might be redundant for many readers. The tables for the calculation of difference quotients are to help implementing a nite difference method. They contain all coefcients of the common interpolation polynomials up to the sixth degree one-sided and up to the twelfth degree symmetrical as well as for their derivatives of rst up to fourth order. Topics: musical acoustics, numerical methods, computer simulation, nite differences, Lattice-Boltzmann, wave equation, physical modelling, signal processing, interpolation polynomials, sound generation, optimization

Contents

Abstract List of Figures Introduction 1. Aero-Acoustical Theory 1.1. Navier-Stokes Equations . . . . . . . 1.1.1. Continuity Equation . . . . . 1.1.2. Momentum Conservation . . . 1.1.3. Energy Conservation . . . . . 1.2. Boltzmann Equation . . . . . . . . . 1.3. Common Simplications . . . . . . . 1.3.1. Wave Equation . . . . . . . . 1.3.2. Euler Equations . . . . . . . . 1.3.3. Incompressible Viscous Fluid 1.3.4. Stochastic Simplications . .

4 11 14 19 19 22 23 26 28 32 32 35 36 39 43 44 44 46 47 48

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

2. Numerical Concepts 2.1. Computational Methodology . . . . . . . . . . . . . . . . 2.1.1. Finite Difference Method . . . . . . . . . . . . . . 2.1.2. Finite Volume Method . . . . . . . . . . . . . . . 2.1.3. Space-Time Conservation-Element and SolutionElement Method . . . . . . . . . . . . . . . . . . 2.1.4. Finite Element Method . . . . . . . . . . . . . . .

2.1.5. Wave-Digital Filter Method . . . . . . . . . . 2.1.6. Boundary Element Method . . . . . . . . . . . 2.1.7. Lattice-Boltzmann Method . . . . . . . . . . . 2.2. Introduction to Finite Difference Methods . . . . . . . 2.2.1. Differences and Difference Quotient . . . . . . 2.2.2. Explicit Method . . . . . . . . . . . . . . . . 2.2.3. Implicit Method . . . . . . . . . . . . . . . . 2.2.4. Interpolation . . . . . . . . . . . . . . . . . . 2.2.5. Handling Boundary Conditions . . . . . . . . 2.2.6. Accuracy and Consistency . . . . . . . . . . . 2.2.7. Stability and Convergence . . . . . . . . . . . 2.2.8. Linear Algebra . . . . . . . . . . . . . . . . . 2.2.9. Time Stepping Methods . . . . . . . . . . . . 2.2.10. Non-Linear Systems . . . . . . . . . . . . . . 2.2.11. Boundary Value Problems . . . . . . . . . . . 2.2.12. Current Research . . . . . . . . . . . . . . . . 2.3. Introduction to Lattice Boltzmann Methods . . . . . . 2.3.1. Discretization of Phase Space . . . . . . . . . 2.3.2. Relaxation . . . . . . . . . . . . . . . . . . . 2.3.3. Lattice Boltzmann Equation . . . . . . . . . . 2.3.4. Theoretical Validation . . . . . . . . . . . . . 2.3.5. Strength and Weaknesses of Lattice-Boltzmann 3. Specic Methods 3.1. Wave Propagation in Wind Instruments . 3.1.1. Acoustical Input Impedance . . . 3.2. Time-Domain Modelling . . . . . . . . . 3.2.1. Fast Fourier Transform . . . . . . 3.2.2. Impulse Invariant Transformation 3.2.3. Bilinear Transformation . . . . . 3.2.4. Transmission Line Mesh . . . . . 3.2.5. Scattering Approaches . . . . . . 3.2.6. Wave Digital Filters . . . . . . .

. . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . .

49 51 52 53 55 57 61 66 80 84 88 97 108 119 121 121 123 123 124 126 128 130

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

133 . 135 . 136 . 148 . 148 . 151 . 152 . 155 . 159 . 160

3.2.7. Hybrid Approaches . . . . . . . . . . . . . 3.3. The Inverse Problem . . . . . . . . . . . . . . . . 3.3.1. Layer Peeling Algorithm . . . . . . . . . . 3.3.2. Taking Losses into Account . . . . . . . . 3.3.3. Bore Reconstruction from Input Impedance

. . . . .

. . . . .

. . . . .

. . . . .

4. Applications 181 4.1. Sound Generation in Brass Wind Instruments . . . . . . . 184 4.1.1. Lip Model Classication . . . . . . . . . . . . . . 187 4.1.2. Transverse Lip Reed Modelling . . . . . . . . . . 189 4.1.3. Two Dimensional Lip Modelling . . . . . . . . . . 213 4.1.4. Wave Steepening in Brass Wind Instruments . . . 218 4.2. Bore Reconstruction of Tubular Ducts . . . . . . . . . . . 224 4.2.1. Bore Reconstruction by Optimization . . . . . . . 224 4.2.2. Measuring Input Impedance . . . . . . . . . . . . 225 4.2.3. Optimization Algorithm . . . . . . . . . . . . . . 229 4.2.4. Bore Reconstruction Results . . . . . . . . . . . . 237 4.2.5. Sensitivity and Accuracy . . . . . . . . . . . . . . 241 4.2.6. Conclusion and Outlook . . . . . . . . . . . . . . 245 4.3. Aero-Acoustical Simulations . . . . . . . . . . . . . . . . 247 4.3.1. Wood Wind Instruments . . . . . . . . . . . . . . 247 4.3.2. Jet-Acoustics Interaction Edge-Tone . . . . . . . 247 4.3.3. Glottis . . . . . . . . . . . . . . . . . . . . . . . . 249 4.3.4. Side Branch Resonators . . . . . . . . . . . . . . 249 4.3.5. Flow Driven Helmholtz Resonator . . . . . . . . . 251 4.3.6. Flute Simulation with Lattice Boltzmann Method . 252 4.3.7. Starting transients and steady-state oscillations . . 254 Bibliography 261

A. Introduction to Digital Signals 283 A.1. Signal Sampling and Reconstruction . . . . . . . . . . . . 283 A.1.1. Sampling Analogue Signals . . . . . . . . . . . . 284 A.1.2. Aliasing . . . . . . . . . . . . . . . . . . . . . . . 286

Nyquists Law . . . . . . . . . . . . . . . . . . . Signal Reconstruction . . . . . . . . . . . . . . . Mathematical Representation of Sampling . . . . . z-Transformation of a sampled function x(t) . . . . Circuit Analysis in the z-Domain: Up-Sampling by Repetition . . . . . . . . . . . . . . . . . . . . A.2. Down-Sampling and Decimation . . . . . . . . . . . . . . A.2.1. Re-Sampling a Sampled Signal . . . . . . . . . . A.2.2. Decimation and Non-Integer Ratios . . . . . . . . A.3. Up-Sampling and Interpolation . . . . . . . . . . . . . . . A.4. Digital Filtering . . . . . . . . . . . . . . . . . . . . . . . A.4.1. Finite Impulse Response Filters . . . . . . . . . . A.4.2. Comb Filters . . . . . . . . . . . . . . . . . . . . A.4.3. Innite Impulse Response Filters . . . . . . . . . A.4.4. Wave Digital Filters . . . . . . . . . . . . . . . . A.4.5. Multi-Rate Filter Design . . . . . . . . . . . . . . B. Interpolation Tables B.1. Forward Coefcients . B.2. Backward Coefcients B.3. Central Coefcients . . B.4. Relative Accuracy . . . Index

288 290 290 295 296 299 299 302 303 304 305 307 310 312 314 315 315 322 329 338 343

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

10

List of Figures

1.1. Deformation of differential cube by velocity divergence . . 20 Diffusion Simulation with Explicit FDM . . . . . . . . . . Wave Simulation with Explicit FDM . . . . . . . . . . . . Wave Simulation with Implicit FDM . . . . . . . . . . . . Spurious Reections . . . . . . . . . . . . . . . . . . . . Harmonic Interpolation Test Signal . . . . . . . . . . . . . Interpolation of Test Function: Comparison Order 2 and 4 Interpolation of Test Function, Order 6 . . . . . . . . . . . Sixth Order Central Interpolation of First Derivative of Noisy Signal . . . . . . . . . . . . . . . . . . . . . . . . . 2.9. Distorted Wave Number, Central Schemes, 1st Deriv . . . 2.10. Wave Simulation with Explicit Upwind FD Scheme . . . . 2.11. Convergence Speed of Iterative Solvers . . . . . . . . . . 2.12. Stability of Predictor-Corrector (Complex Plane) . . . . . 2.13. Velocity Vectors of D3Q19 Lattice . . . . . . . . . . . . . 2.1. 2.2. 2.3. 2.4. 2.5. 2.6. 2.7. 2.8. 58 61 64 65 70 71 72 76 89 94 106 115 124

3.1. Trumpet section input impedance as calculated by Kemp [Kem02] . . . . . . . . . . . . . . . . . . . . . . . . . . 146 f 3.2. Frequency Warping log( fs ) by Bilinear Transformation . . 153 3.3. Acoustical TLM model . . . . . . . . . . . . . . . . . . . 156 3.4. Screenshots of the VirtualWaveTank program [Wil] . . . . 158 3.5. Some Basic WDF-Elements . . . . . . . . . . . . . . . . 164 3.6. Usual Graphical Representation of WDF-Adaptors . . . . 169 3.7. Signal Flowchart of WDF-Adaptors . . . . . . . . . . . . 169

11

3.8. Digital Wave Guide Element . . . . . . . . . . . . . . . . 170 3.9. Wave Scattering at Discontinuities . . . . . . . . . . . . . 174 Classication of sound production mechanisms . . . . . . 188 Trumpet Sound Production, Equivalent Circuit . . . . . . . 191 Trumpet impedance with varying lip admittance . . . . . . 192 Lip Eigenfrequencies and Trumpet Resonances . . . . . . 195 Mode Selection (sounding frequencies and amplitudes) . . 196 Mode transition by blowing pressure) . . . . . . . . . . . 197 Sounding Pitch . . . . . . . . . . . . . . . . . . . . . . . 197 Stability Analysis Transversal Model . . . . . . . . . . . . 199 Displacement over Force, Mass-Spring System . . . . . . 201 Displacement over Force, Spring-Mass System . . . . . . 203 Displacement over Force, Spring-Mass-Spring System . . 204 Displacement over Force, Spring-Mass-Spring-Mass System204 Displacement over Force, Distributed Mass-Stiffness System, Free End . . . . . . . . . . . . . . . . . . . . . . . . 206 4.14. Displacement over Force, Distributed Mass-Stiffness System, Fixed End . . . . . . . . . . . . . . . . . . . . . . . 207 4.15. Articial Lip, Lumped Circuit Model . . . . . . . . . . . 208 4.16. Articial Lip, Frequecy Response . . . . . . . . . . . . . 208 4.17. Articial Lip, Loop Gain . . . . . . . . . . . . . . . . . . 209 4.18. Trumpet Sound Generation, Simplied Equivalent Circuit . 212 4.19. Self Sustained Oscillations, Minimum Circuit . . . . . . . 212 4.20. Linear Stability Analysis, Minimum Circuit . . . . . . . . 213 4.21. 2D Lip Model, Transmission Line . . . . . . . . . . . . . 214 4.22. 2D Lip Model, Flow Path Coupling . . . . . . . . . . . . 215 4.23. 2D Lip Model, Denitions . . . . . . . . . . . . . . . . . 216 4.24. Time Signals of Pedal Tone . . . . . . . . . . . . . . . . . 217 4.25. Time Signals of rst 4 Resonances . . . . . . . . . . . . . 218 4.26. Lip Movement of Pedal Tone . . . . . . . . . . . . . . . . 219 4.27. Pressure Prole, 4th Resonance . . . . . . . . . . . . . . . 220 4.28. Pressure Prole, 8th Resonance . . . . . . . . . . . . . . . 221 4.1. 4.2. 4.3. 4.4. 4.5. 4.6. 4.7. 4.8. 4.9. 4.10. 4.11. 4.12. 4.13.

12

Displacement Prole, 8th Resonance . . . . . . . . . . . . 222 Wave Steepening of Trombone Sound . . . . . . . . . . . 223 Impedance Measurement Setup . . . . . . . . . . . . . . . 225 Calibration Magnitude . . . . . . . . . . . . . . . . . . . 227 Calibration Argument . . . . . . . . . . . . . . . . . . . . 227 BIAS System . . . . . . . . . . . . . . . . . . . . . . . . 228 Bore Reconstruction Benchmark, Steady State GA . . . . 231 Bore Reconstruction Benchmark, Rosenbrock . . . . . . . 232 Bore reconstruction of 75 cm long object . . . . . . . . . . 238 Results of impedance matching; lower peaks are measured 238 BbTrumpet Reconstruction with Layer Peeling . . . . . . 239 Leadpipe Reconstruction . . . . . . . . . . . . . . . . . . 240 Leadpipe Impedance Matching . . . . . . . . . . . . . . . 240 Stepped Tube . . . . . . . . . . . . . . . . . . . . . . . . 244 Sensitive with 1 cm grid . . . . . . . . . . . . . . . . . . 244 Sensitive with 4 cm grid . . . . . . . . . . . . . . . . . . 245 Vorticity Contours During one Cycle of Glottis Motion . . 250 Cross section of the simulated pipe. Dimensions in lattice units. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252 4.47. Simulated Stopped Organ Pipe . . . . . . . . . . . . . . . 253 4.48. Formation of the jet during start-up. The value of the velocity ranges from blue (u = 0) to red (max. velocity). . . 255 4.49. Density taken at the end of the stopped pipe resonator . . . 256 4.50. Transient behavior of the density at startup, taken at the end of the stopped pipe resonator . . . . . . . . . . . . . . 256 4.51. Frequency spectrum of the starting transients . . . . . . . 257 4.52. Frequency spectrum of the steady state oscillation . . . . . 257 4.53. Jet Velocity in the Mouth, One Oscillation Period . . . . . 258 4.29. 4.30. 4.31. 4.32. 4.33. 4.34. 4.35. 4.36. 4.37. 4.38. 4.39. 4.40. 4.41. 4.42. 4.43. 4.44. 4.45. 4.46. A.1. A.2. A.3. A.4. Applying Observation Window Sampled Analogue Signal . . . Sampled Analogue Signal . . . Sampling at the Signal Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284 285 286 287

13

A.5. Sampling almost at the Signal Rate . . . . . . . . . . . . . A.6. Sampling at the Nyquist Rate (sampling rate is twice the signal frequency) . . . . . . . . . . . . . . . . . . . . . . A.7. Continuous Signal Spectrum and its Mirror Images created by Sampling . . . . . . . . . . . . . . . . . . . . . . A.8. Signal Spectrum distorted by aliased Out-of-Band Components . . . . . . . . . . . . . . . . . . . . . . . . . . . A.9. Limit of Partial Sum ejlT , T = 1 . . . . . . . . . . . A.10.Linear Convolution with Pulse Train . . . . . . . . . . . . A.11.Four Times Up-Sampling by Repetition of Samples . . . . A.12.Transmission Function of Digital Sample and Hold . . . . A.13.Effect of Sampling on Base-Band of Signal Spectrum . . . A.14.Re-Sampling of a Sampled Signal at a Two Times Lower Rate (Decimation) . . . . . . . . . . . . . . . . . . . . . . A.15.Decimation of a Sampled Signal . . . . . . . . . . . . . . A.16.Up-Sampling by a Factor of Two . . . . . . . . . . . . . . A.17.Digital Filter Topologies . . . . . . . . . . . . . . . . . . A.18.FIR or Non-Recursive Filter . . . . . . . . . . . . . . . . A.19.Spectrum of 2nd Order Comb Filter with D=128 . . . . . . A.20.Possible Implementation of 2nd Order Comb Filter . . . . A.21.IIR or Recursive Filter . . . . . . . . . . . . . . . . . . . A.22.Wave Digital Biquadratic Section . . . . . . . . . . . . . . B.1. B.2. B.3. B.4. B.5. B.6. B.7. B.8. Forward/Backward, 1st Deriv, Log(Error) over PPWL . Centered, 1st Deriv, Log(Error) over PPWL . . . . . . Forward/Backward, 2nd Deriv, Log(Error) over PPWL Centered, 2nd Deriv, Log(Error) over PPWL . . . . . . Forward/Backward, 3rd Deriv, Log(Error) over PPWL Centered, 3rd Deriv, Log(Error) over PPWL . . . . . . Forward/Backward, 4th Deriv, Log(Error) over PPWL . Centered, 4th Deriv, Log(Error) over PPWL . . . . . . . . . . . . . . . . . . . . . .

287 288 289 290 292 293 298 298 300 301 302 303 304 306 308 308 310 313 338 338 339 339 340 340 341 341

14

Introduction

This is a book on musical acoustics and it was written for musical acousticians by a musical acoustician, so the content is biased in a way, which is typical for most of the many books on computational physics. Under the title Computational Physics, hundreds of different concepts, methods and algorithms have been published, so that it became virtually impossible to cover all of them in a single book. The selection which had to be made naturally reects the personal taste of the author and his belief in how useful certain concepts would be for applications in musical acoustics. As will be outlined later, musical acoustics has indeed some requirements which are distinct from those of other physical or even acoustical areas. Before starting a book on simulation and analysis of musical instruments, a little bit of historical review is usually appropriate. This allows paying homage to the great scientist Hermann Helmholtz, who laid important foundations for the whole eld of musical acoustics by publishing his famous book On the Sensations of Tone... in 1863 [Hel63]. Another early pioneer in that eld was Lord Rayleigh, who published two volumes of The theory of sound [Ray77, Ray94] in the nineteenth century, full of basic theory of sound radiation. Modelling wave propagation in tubes with varying cross-section has been investigated by Bernoulli, Euler and Lagrange. Between 1760 and 1770 they derived a wave equation for such ducts, but it needed a further publication by its re-discoverer Webster [Web19] in 1919, in order to bring that important result back into the mind of scientists. Other earlier and important work towards the contemporary understan-

15

ding of wind instruments was done by Richardson [Ric29], Morse [Mor36], Webster [Web47], Benade [Ben60, BG68, Ben76] as well as Strong [SC67], Backus [Bac69], Fletcher [Fle79], Keefe [Kee81, Kee82], Elliott [EB82]. This chronological list of authors covers the time up to 1982 and is not exhaustive and just a small selection of numerous publications is cited here. The last twenty years of research on sound mechanisms of brass instruments has been covered by Gilbert in [Gil02]. An even more recent review on brasses with a representative bibliography was published by Campbell in [Cam03]. As an introduction to the whole eld of musical acoustics, the book The Musicians Guide to Acoustics, written by Campbell and Greated [CG87] can be recommended. It explains the qualitative aspects of the operational principles of musical instruments very well. It is dedicated to musicians with moderate mathematical background, who want to understand the operational principles of musical instruments, or to teachers teaching musical acoustics to a broader audience. Quantitative aspects and mathematical derivations can be found in the excellent book The Physics of Musical Instruments written by Fletcher and Rossing [FR90]. In the preface to the second edition the authors write: When we wrote the rst edition of this book, we directed our presentation to the reader with a compelling interest in musical instruments who has a reasonable grasp of physics and who is not frightened by a little mathematics. We are delighted to nd how many such people there are. It is the hope of the author that at least some of such people are also eager to use their personal computers to set up simulations in order to see and observe whats going on inside and around their musical instruments, which is usually invisible. Fletcher and Rossing express their opinion that numerical calculations are less instructive than analytical models. This is indeed true. But rstly, even analytical methods often require numerical evaluations and graphical representations in order to reveal the whole truth that is inherent to

16

a mathematical model. Secondly, experiments are often too expensive to be affordable, especially when sound waves are to be visualized inside or around musical instruments. Numerical simulations that are set-up correctly and that are veried against experimental data can provide enlightening hints about unknown effects and dependencies that were not fully understood before. There are variables that are not observable or at least too difcult to observe in experiments. But it is only moderately difcult to get qualitatively and even quantitatively correct representations of the physical reality using a numerical simulation. The numerical analysis methods that are introduced and explained in this book are applicable to a wide range of musical instruments. Brass instruments will get a certain focus, but recent developments concerning simulation of utes and ue pipes will also be covered in detail. Application examples given here deal with brasses and utes, of course, but could as well deal with woodwind instruments, singing voice, piano, string instruments or other types of instruments. The content ranges from the most theoretical fundamentals, the governing differential equations of uids and wave propagation, to the computational methods that are suitable to nd, or rather stepwise approximate their solution. Practical examples given in the application chapter are to show, how numerical analysis methods can be related to real world problems, taken from the vast eld of musical acoustics. While a few equations on the fundamental theoretical level are sufcient to describe practically all of the effects ever encountered during an entire scientic career, on the computational level one has to dedicate much of time and effort to know, understand and to apply the many concepts, ideas and algorithms that colleagues have developed while striving to obtain practically useful results from these fundamental equations. Therefore, the computational part of this book is much less complete. Instead of giving a short lexical description of all the concepts which are currently being persued, it was decided to concentrate on a few concepts only, which are worked out thoroughly enough to allow someone entering

17

into this eld to apply those methods to the problems he wants to solve. If someone starts doing so, then this book has already fullled one of its main purposes. When he then succeeds and wishes to go further, he may want to consult the publications that are provided in the bibliography of this book to get the answers to the advanced questions which might come up during in-depth studies. The application part contains several reviews of particulary interesting research articles according to the taste and knowledge of the author, of course as well as some more elaborated reports about research done at the Institut f r Wiener Klangstil in Vienna, published by the author and u his colleagues during the last years. Original work is presented in the elds of distributed lip reed modelling, brass instrument bore reconstruction and ue pipe simulation. A time-domain simulation of a real playing situation, including a trumpeters vocal tract, his buzzing lips exhibiting Rayleigh surface waves and a Bb-trumpet with standard mouthpiece is presented. The system simulation framework has been proposed by the author in [Kau02], recent results in [Kau03b]. The method for solving the inverse problem is demonstrated as another application of wave guide modelling of tubular ducts like brass instruments. Computer optimization is used to match a measured input impedance curve with a theoretically computed one. The geometry is varied during the optimization process until matching is achieved. Results of such reconstructions of articial objects, trumpet leadpipes, and even whole instruments are presented. The method was originally proposed by the author in [Kau01]. Another branch of original work which is presented in the application part regards the Lattice-Boltzmann simulation method. A ue pipe simulation including wave and uid propagation in a viscous compressible medium is introduced [K h03]. u

18

1. Aero-Acoustical Theory

1.1. Navier-Stokes Equations

The conservation laws of mass, momentum and energy for a compressible viscous uid ow, like air, are called the Navier-Stokes equations [GHPM01]. They describe the uid macroscopically as a continuum with scalar (pressure p, density ) and vectorial (velocityu) time varying properties assigned to each spatial point r = (x, y, z)T . Before typical differential formulations of these conservation laws are given, an important relation known as Reynolds transport theorem should be understood. It is used to derive the differential conservation laws from more obvious integral formulations, which will be given later, because of their less frequent use in numerical solution algorithms. Reynolds transport theorem is valid for an arbitrary scalar function F = F (r, t) describing any macroscopic property of a owing uid.

d dt

V

F dV

=

V

F + F ( u) t F dV + t

V A

dV = (1.1)

F u n dA,

with V = V (t) denoting a time varying material volume dV = dxdydz enclosed by the time varying surface A = A(t), with n being the surface orthogonal unity vector and u = dr/dt describing the convective translation of the material volume.

19

1. Aero-Acoustical Theory

It can be formally understood as an application of the product differentiation law (uv) = u v + uv , in the actual case with f (t) and V (t) df V d (f (t)V (t)) = V + f dt dt dt (1.2)

It must be noted here that the notion of a nite volume is chosen which contains a certain amount of mass for a nite period of time. Its limits are exible and it is translated by ow and deformed by speed gradients. There are other authors who start their derivations with the notion of innitesimal small xed volumes containing variable amounts of mass uctuating even within nite periods of time [BSL60]. A third possible view is the assumption of innitesimal small point particles as described in [Hin75]. xxx

(x+dx, y+dy, z+dz)

A deeper level of understanding can be obtained from the meaning of the divergence operator. The divergence, sometimes also called source density of a vector eld u = dr/dt expresses how the volume of an observed uid element increases or decreases by volume ow entering or leaving that volume. If the eld u is spatially constant then observed volume elements are translated but not deformed. If there are spatial deriva-

20

tives of the eld vector u which are non-zero then the volume element itself is deformed because opposite border areas may ow with different speeds. If the volume is expressed in cartesian coordinates as a small cube according to Figure 1.1 with a ow component ux (x + dx, y, z) = ux (x, y, z) + dux , (1.3)

(zero ow in all other directions) then divergence u according to its denition /x ux duz dux duy + + (1.4) divu = u = /y uy = dx dy dz uz /z would yield dux /dx and the volume integral (the second term on the right side of Equation 1.1) div u dx dy dz = dux dy dz corresponds to the additional volume created by the strain applied to the differential cube in x-direction caused by the speed gradient of ux . Now the following interpretation of Equation 1.1 can be found. The change rate of some property associated with the uid contained in a nite volume element can be obtained by integrating all property change rates within that volume if there is no deformation of that volume caused by some divergence of the volume ow. If this is the case, some additional property change rate corresponding to the volume difference expressed as div u dx dy dz has to be taken into account.

V V

According to the Gaussian divergence theorem, the volume integral over the divergence of a velocity eld is equivalent to the closed surface integral over the eld vector itself. This equivalence can also be understood by means of Figure 1.1 keeping the denition of the div operator divu = duy dux duz dx + dy + dz in mind. The equivalence is expressed by the second formulation of Reynolds transport theorem in Equation 1.1. A more general formulation of the transport equation including stress

21

1. Aero-Acoustical Theory

and external sources can be written with f = f (r, t) and f = F as d dt

V

f dV = =

A

t

V

f dV +

A

f u n dA = Q(f ) dV.

V

(f ) f n dA +

(1.5)

Assuming steady ow parameters within an innitesimal small volume we can apply the Gaussian integral theorem and we obtain the differential form f df = + f u = ((f ) f ) + Q(f ) . (1.6) dt t The rst term represents the total change rate within a nite volume. The second term gives the change rate of the property itself, the third one is related to the convective deformation of the volume, the fourth one represents stress or diffusion with transport coefcient (f ) and Q(f ) represents distributed external sources of the magnitude f . Appropriate substitutions for f , (f ) and Q(f ) will be chosen in the following sections in order to obtain the mass, momentum and energy conservation equations.

The continuity equation is the rst Navier-Stokes equation. It is the mathematical expression corresponding to the observed fact that matter does not vanish nor appear from nothing. If there is a total momentum out of a volume element it will cause a mass density decrease: + (u) = 0. t (1.7)

The Nabla operator is dened as = ( x , y , z )T . It is also possible to rewrite the second term as div u. It is only zero in the case when the surface orthogonal components of all velocity vectors entering the innitesimal small volume element are compensating each other.

22

In other words, there is no density change when mass ow entering a volume element is perfectly balanced by mass ow leaving that region. It is assumed here that there are no external mass sources or sinks and that 2 there are no conversions between mass and energy according to E = mc . Navier-Stokes equations can also be written in integral form which is more general, because it does not exclude discontinuities of macroscopic ow properties. The integral laws are given here because their discretization is essential for a class of numerical solvers known as Finite Volume Methods (FVM). The integral formulation of the law of mass conservation can be written by substituting f 1, (f ) 0 and Q(f ) 0 in Equation 1.5 [Bre02]: d dt

V

dV =

t

V

dV +

A

(u n) dA = 0

(1.8)

The second Navier-Stokes equation is the momentum equation which is referred to as Euler-Equation in the friction-less case. It expresses the observation that momentum changes are only accomplished by forces. A solid body without any interference by an external force will not change the speed of its movement nor its direction. To change either speed or direction of a car, for example, a force must be applied, and a corresponding acceleration will be felt by the passenger. Actually the momentum equation balances force densities: u + (u uT ) + p = G, t with the dyadic vector product dened as ux uuT = uy (ux uz u2 ux uy ux uz x u2 uy uz . (1.10) uz ) = ux uy y ux uz uy uz u2 z (1.9)

uy

23

1. Aero-Acoustical Theory

Its spatial derivative is

u2 ux uy ux uz /x x u2 uy uz = (u uT ) = /y ux uy y ux uz uy uz u2 /z z ux ( u) (1.11) = uy ( u) = u( u) uz ( u) = . (1.12)

and the inner product of a vector and a matrix is dened as /x 1,1 1,2 1,3 /y 2,1 2,2 2,3 = 3,1 3,2 3,3 /z 1 1,1 /x + 1,2 /y + 1,3 /z = 2,1 /x + 2,2 /y + 2,3 /z = 2 3,1 /x + 3,2 /y + 3,3 /z 3

The second Navier-Stokes equation says that a change of the momentum in a spacial point (the rst term) is caused by convective forces (the second term), by viscous forces described by the viscous stress tensor (the third term), by the pressure gradient (the forth term) and by other forces. G represents external forces like the gravitational force density g. Sometimes a uid stress tensor P is dened which combines viscous and pressure forces according to P = pI which reduces Equation 1.9 to u + (P + u uT ) = G. t (1.14) (1.13)

In the purely aero-acoustical case the viscous stress tensor can often be neglected. If convection is slow compared to the speed of sound then

24

the second order convection term can be omitted. If viscosity cannot be neglected then has to be determined. In the second Navier-Stokes equation it is usually assumed that there is a linear relationship between and the rate-of-strain tensor uT + ( uT )T , with ux /x uy /x uz /x (1.15) uT = ux /y uy /y uz /y ux /z uy /z uz /z and ux /x ux /y ux /z ( uT )T = uy /x uy /y uy /z . uz /x uz /y uz /z

(1.16)

While solids oppose deformation, uids oppose the deformation rate. If the force density is proportional to the deformation rate then the uid is described as Newtonian. Even in this case the expression for becomes very complicated. Assuming that the uid is in local thermo-dynamic equilibrium, an assumption which is referred to as Stokes hypothesis, a considerable simplication can be obtained. In that case can be described as = (u + (u)T ) 2 ( u)I, D (1.17)

with I being the unit tensor and being the dynamic viscosity to be determined experimentally. D is 2 or 3 depending on the spatial dimensions which are considered. Unfortunately depends on T and p. If Stokes hypothesis does not hold, e.g. in the ultra-sound region, then an additional dissipation related to volume changes u can be observed. It can be described by a volume viscosity which is not directly related to [RH02]. This volume viscosity can also become signicant at normal sound frequencies when propagation over long distances or in dusty environment are considered [Tho72].

25

1. Aero-Acoustical Theory

The integral formulation of the law of momentum conservation can be written by substituting f ui , ui being the ith coordinate of u, (f ) , and p u + gi + (1.18) Q(f ) ri ri with ri being the ith coordinate of r and gi being the ith coordinate of g in Equation 1.5 [Bre02]: t

V

ui dV +

A

ui u n dA

A

ui n dA = (1.19)

=

V

p u ( + gi + ) dV ri ri

with ui being the ith coordinate of u and n being the surface orthogonal unity vector for the closed surface integral over A. Using the stress tensor P Equation 1.19 can be formulated as t

V

u dV +

A

(P + u uT ) n dA =

V

G dV.

(1.20)

The rst main proposition of thermodynamics is the law of conservation of energy. This law expressed in the context of uid dynamics is sometimes also referred to as third Navier-Stokes equation. Basically it states that a system which exchanges some kinds of energy with its surrounding world will store any surplus of received energy as an increase of its internal energy and has to pay from this account any deciency resulting from this energy exchange. Sometimes the denition of a system and its outer bordersvaries depending on the authors ideas and intentions, so some confusion about what is internal and what is external may arise. This fact is reected in the

26

form of the resulting energy equation and in what terms have been put at the left side and at the right side of the equation. In the energy equation described by Equation 1.22 the total energy density E of the system associated with the innitesimal small uid element around r has two components. It consists of the internal energy density e (compare Equation 1.32) of the gas itself and the kinetic energy density of some convective motion u. The total energy per unit mass is u2 . (1.21) 2 Its change rate is balanced by the other terms on the right side. The rst term there is related to the divergence of the velocity vector eld and describes the change rate of total enthalpy density caused by isobar convective deformation of the volume element. The total enthalpy per unit p mass H = E + contains the total energy and the mechanical work term p dV both related to unit mass. Its change rate is negative if the velocity vectors leaving the volume element are predominating and thus establish an overall ow out of the volume element. As the divergence is positive in that case, a negative sign is required. In other words, a change of internal plus kinetic energy density of a uid element occupying an innitesimal small volume element can be caused by convective uid transport in or out of the volume element or introduced from outside by thermal (conduction), chemical (release) or mechanical (conversion) sources. E =e+ E = (Hu) + Q (1.22) t Q is the sum of the total work done by other forces. Equation 1.23 gives four examples: viscous stress described by the tensor , external forces like gravity g, energy input by heat conduction, with heat conduction coefcient depending on temperature T , and the heat released by external sources, e.g. due to chemical reactions (heat release rate q), per unit volume and per unit time. Q = ( u) + g u + (T ) + q. (1.23)

27

1. Aero-Acoustical Theory

The Equation of State relating p and can be derived from the third Navier-Stokes equation. In perfect gases where e = cv T it is: p = (cp cv )T = c2 , (1.24)

with T being the temperature and cp and cv being the specic heats at constant pressure and volume. The integral formulation of the energy conservation law can be written using the stress tensor P: d dt

V

E dV =

V

E dV + t

A

Eu n dA = q n dA,

A

=

V

G u dV

A

P u n dA

(1.25)

with q being the heat ux due to heat conduction which can be approximated by q = T . Another form of the integral formulation of the energy conservation law can be obtained using Equation 1.5 with the substitutions f h, (f ) /cp and Q(f ) Dp u with h = e + p/ being the specic enthalpy and D the diffusion constant: t

V

h dV +

A

hu n dA

A

h n dA = cp (1.26)

=

V

D p u dV

Up to now a uid has been described as a macroscopic continuum with time and space varying properties. There are scalar magnitudes like mass density (r, t), pressure p(r, t) and the vector eld u(r, t) describing the

28

local and instantaneous convection speed of the uid, each spatial point being assigned its coordinate vector r = (x, y, z)T . These macroscopic magnitudes can be measured at any time instant and at any spatial point but they are related by complicated non-linear dependencies, the Navier-Stokes equations 1.7, 1.9 and 1.22, which are even difcult and expensive to deal with numerically. Considering the fact that macroscopic quantities are the result of the presence and activity of myriads of uid molecules does not help too much. The mathematical task of predicting positions and velocities of three solid bodies, when their initial positions and velocities are exactly known, has been called three body problem. Although there is only one linking force, gravity, this problem turned out to be a real challenge for mathematicians. The reader can certainly imagine that a 1023 body problem is beyond the scope of any approach, one can think of. But between the microscopic view at the molecular level and the macroscopic view of measurable physical uid properties there is a mesoscopic view, where the laws of statistics can be applied to groups of molecules. The mesoscopic spatial scale is big enough to contain statistically relevant numbers of molecules in each considered volume element, but still small enough that macroscopic quantities can be assumed constant within these volumes. In other words, mesoscopic calculation volumes are points in the macroscopic continuum, but they still contain big numbers of molecules to be described statistically. Fluid molecules are characterized by their position and their velocity. The macroscopic mass density does not take any velocity into account. To determine the mass density according to the mesoscopic view all particles contained in an innitesimal small volume element have to be counted, the resulting number has to be multiplied by the particle mass and divided by the volume of the element itself. The total mass MV contained in a volume V can be obtained by integrating the mass density function of space and time over that volume.

29

1. Aero-Acoustical Theory

MV (t) =

V

(r, t) dV

(1.27)

In order to take the different particle velocities into account, an additional classication has to be introduced. To do so, a particle distribution function f (r, v, t) can be dened, where particle density is not only related to points r = (x, y, z)T in space but also to their velocity vectors v = (vx , vy , vz )T . Because uid states can dynamically change, the distribution function may also vary in time. To obtain total mass MV , again the distribution function must be integrated over the 3-dimensional space as well as over the 3-dimensional velocity space. MV (t) =

V,v

(1.28)

Integrating only over the 3-dimensional velocity space will retrieve the mass density (r, t). (r, t) = f (r, v, t)dvx dvy dvz (1.29)

In other words, each point of the 6-dimensional phase-space, consisting of all possible 3-dimensional velocity vectors at all possible positions in 3dimensional space, is characterized by a particle-density which may vary in time. Averaging all molecular velocities inside an innitesimal small volume element yields the macroscopic uid velocity u(r, t). This is done by integrating all possible particle velocities weighted with the actual particle distribution and dividing the result by the total density at this macroscopic point: u(r, t) = vf (r, v, t)dvx dvy dvz . f (r, v, t)dvx dvy dvz (1.30)

30

From this the mass ow u (r, t)u(r, t) = vf (r, v, t)dvx dvy dvz (1.31)

and e, the internal energy per unit mass can be calculated: e(r, t) = 1 2(r, t) (v u)2 f (r, v, t)dvx dvy dvz (1.32)

(r, t)E(r, t) =

The equation of movement for the distribution function known as Boltzmann-Equation is: + v r f (r, v, t) = t (1.34)

It expresses the fact that the density of particles at a certain point in space having a certain velocity will change in time only if there is a density gradient of such particles in space. The term represents the effect of particle collisions which interfere in the constant ow of particles and introduce velocity transitions. The discretization of the Boltzmann-Equation in time, space and velocity leads to a new class of numerical methods to simulate systems, where uid propagation and acoustical wave propagation are closely related. The method is called the Lattice-Boltzmann Method (LBM). A system developing an innite period of time without any external interference will approach an equilibrium state of constant global temperature, which is described by the Maxwell-Boltzmann!distribution f (v) =

(vu)2 e 2RT D/2 (2RT )

(1.35)

31

1. Aero-Acoustical Theory

with R being the ideal gas constant, D the spatial dimension and ,u and T being the macroscopic (constant) magnitudes mass density, convection speed and temperature.

Solving the Navier-Stokes equations is generally difcult. Fortunately there are many kinds of problems where reasonable simplications to the governing differential equations can be made in order to get simpler forms which can be solved analytically or at least numerically with much less of an effort.

If sound propagation in free air or in musical instruments is to be calculated, then some important simplications can be made in order to describe most of the occurring phenomena adequately well. First, the viscosity for air is small and can normally be neglected. The viscous stress tensor is zero and the operator multiplied with the uid stress tensor reduces to grad p. If viscous effects have to be taken into account, they are often attributed to the boundary conditions and not to the transmission medium itself. This is especially the case in the long and narrow wave guides of many typical wind instruments, where thermo-viscous losses can be concentrated in the boundary conditions. In reality thermo-viscous losses do originate mainly from a thin air layer adjacent to the instruments walls where the uids tangential velocity has to decay to zero, which causes sufcient shear stress in order to dissipate a measurable amount of sound energy. Second, pressure variations caused by propagation of sound waves are small compared to the atmospheric quiescent pressure. Even the most disturbing and already painful sound of a starting ghter aircraft having a sound pressure level of about 140 dB corresponding to an RMS pressure

32

amplitude of pRM S = 200 Pa does not cause any sound pressure amplitudes of more than 0.2 % of a typical atmospheric pressure of 105 Pa. Density variations are still smaller by a factor of which is about 1.4 under typical atmospheric conditions. This allows linearization of pressure and density around their atmospheric mean values. Third, convective uid transport can be neglected in most of the typical acoustical problems. The Mach number, dened as the ratio between air speed and the speed of sound, is usually so small that convection phenomena need not be taken into account. The oscillating sound velocity even at maximum sound levels is small compared to the propagation speed of sound. At 140 dB sound pressure level at 1 kHz the Mach number does not exceed about 1.5 103 during one oscillation period. Under these conditions the maximum acoustical uid particle displacement is about 8 105 m. This is another justication for the replacement of acoustical properties by linear approximations. Finally, external forces are either not present, like chemical heat release or negligible, like the inuences of heat conduction and gravity. External mechanical input of momentum or energy is usually taken into account by the specication of the boundary conditions. Therefore the common way to adapt the Navier-Stokes equations to the typical conditions faced in the eld of linear acoustics is to zero all mean ow related terms, to neglect heat conduction, viscosity and gravityand to postulate only small uctuations of p and around their atmospheric values. Second order terms of all perturbation variables are neglected. What remains of the conservation laws for small perturbations , p and u of a stagnant (u0 = 0, u = u ) uniform uid characterized by p = p + p0 and = + 0 can now be given as + 0 u t u + p 0 t = 0 = 0. (1.36)

Building the time derivative of the mass conservation law and the diver-

33

1. Aero-Acoustical Theory

gence of the momentum equation 2 + 0 ( u ) = 0 t2 t 0 ( u ) + p = 0, t

(1.37)

we can subtract the second equation from the rst and with the proportionality between and p assuming adiabatic state transitions we obtain the wave equation 1 2p = 2 p p , (1.38) c2 t2 with c = (cp cv )T can be obtained. The ticks will be omitted when we are exclusively dealing with linearized quantities. The wave equation can be analytically solved for a long and narrow tube when one-dimensional plane wave propagation is assumed. It can also be solved for the three dimensional axial symmetrical case, when cylinder coordinates are used and a modal decomposition is performed. Modal decomposition can also be applied to get a solution for a duct with rectangular cross-section [Kem02]. The wave equation does not cover the cases where signicant air ow, high sound pressures or other non-linearities are present. Especially in the eld of musical acoustics there are some essential questions which cannot be answered based on such simplications. Sound propagation in brass wind instruments can easily be driven into a nonlinear region for which the wave equation is no longer valid. Wave steepening and shock wave effects are due to the high sound pressure levels generated inside the instrument. To model ow driven instruments like ue pipes or utes the Navier-Stokes equations in their general form have to be solved. Modelling wave propagation does not cover all the nonlinear effects occurring in the primary oscillators of wind instruments either. Waves are propagating on the brass players lips, the singers vocal folds and in the air columns of all wind instruments including the singing voice, but the

34

coupling between the vibrating reeds of clarinets, oboes or tubas to the resonator, the air column of the instrument is highly nonlinear.

Neglecting viscosity of the uid can often be justied in gases. If wall contact is not a predominating condition then viscous losses are very small in air. Only when sound propagation over very long distances is studied, viscous damping can become measurable yet still not signicant. In free air damping is mainly caused by the so called volume viscosity rather than the kinematic viscosity . Kinematic viscosity suddenly starts to become an important factor when the laminar ow region is left or when strong speed gradients orthogonal to the direction of the ow are present. This is the case in guided ows at higher Mach numbers or when sound propagation takes place. In air also buoyancy (gravitational force) and heat conduction terms do not play much role under typical conditions. Gases are light and relaxation into equilibrium pressure conditions does not take much time. Thermal conductivity is bad and pressure disturbances often small. The resulting Euler-Equations + (u) = 0 t u + (u uT ) + p = 0 t E + ((E + p)u) = 0 t (1.39) (1.40)

are useful in various areas. Aerodynamical simulations including sound propagation in windy environmental conditions or ventilation ducts can be made. Analytical solutions are much more difcult to get than for the wave equation, but numerical solutions can be obtained using various algorithms. In musical acoustics the Euler equations are less important because for simulations of sound propagation the wave equation is far more conve-

35

1. Aero-Acoustical Theory

nient. Yet for the remaining effects the Euler equations are still not appropriate. Sound generation in musical instruments depends on non-linear coupling between two oscillating systems. One being the resonator, e.g. the air column in a wind instrument, the other one being a stimulator, e.g. the buzzing lip in a trumpet mouthpiece or the double reed in a bassoon. In utes it is even difcult to separate the stimulator, the air jet, from the resonator because both is uid. Sound generation in ute-like instruments depends on bidirectional interactions between the acoustic eld in the resonator and the air jet oscillating around the labium. On top of that a signicant part of the radiated sound energy stems from vortex shedding, not only in the region of the labium but maybe also from some tone holes.

To eliminate compressibility, mass density is assumed to be constant. This assumption simplies the rst Navier-Stokes equation, the mass conservation law from Equation 1.7 to u = 0, (1.41)

the requirement that velocity must not have any divergence. Therefore the third term of Equation 1.17 has to be zero and the product becomes = ( ( uT ) + ( uT )T ). (1.42)

Using the equations 1.15, 1.16 and 1.12 the intermediate results ( )ux (1.43) ( uT ) = ( )uy = 2 u ( )uz and /x( u) ( uT )T = /y( u) = ( u) = 0 /z( u)

(1.44)

36

can be obtained. From Equation 1.10 and Equation 1.12 it can be seen that (u uT ) contains u which is known to be zero. We could also eliminate the p term, because at constant density there cannot be any spatial pressure gradient. It would relax immediately because the speed of sound in an incompressible medium is innite. With = / the momentum conservation from Equation 1.9 therefore becomes u = 2 u. (1.45) t This is a classical parabolic diffusion equation for the uid velocity eld it is f.e. used by Chattot to give an example of how to determine the viscous uid velocity eld between two moving plates [Cha02] but there is no way to determine other uid characteristics like the pressure. If the pressure is to be calculated, then nite pressure variations must be allowed and therefore a small velocity divergence must be tolerated. An equation where u has been eliminated less rigorously is Equation 1.50. It can be used for further renements of the velocity eld, once the pressure has been calculated by one of the methods given below. Projection Method The divergence of the second Navier-Stokes equation, the momentum conservation law, can be written as u = ( (u uT )) + ( ) 2 p 0. t (1.46)

With the knowledge, that the explicit rst term contains u, which is known to be almost zero, a compatibility law for u and p for the case of nearly zero velocity divergence can be formulated. The equation is known as Poisson-Equation: ( (u uT )) ( ) 2 p = 0 (1.47)

37

1. Aero-Acoustical Theory

By means of this equation a pressure can be estimated which would satisfy the zero divergence condition for a known velocity eld. With this pressure a more accurate velocity distribution can be obtained by means of the momentum Equation 1.50 derived below. This iterative method is known as projection method. Vortex Density Another method to obtain a relation for the pressure variations of an incompressible uid velocity eld is to split the vector eld into an irrotational part and a solenoidal part: u = + , (1.48)

where is a scalar velocity potential and a vectorial velocity potential for which the relation = 0 holds. This way, it can be strictly differentiated between the potential ow, which is guaranteed to be vortex free and the solenoidal part, which cannot develop any sources or sinks. This follows from the mathematical relations () = 0 and ( ) = 0. Another important implication of this split is that only is linked to acoustic aspects because the divergence of the vortex part related to is zero, so all the compressibility issues, like sound propagation, are related to . On the other hand, vorticity = u is only related to the vector stream function , because () is always zero. = ( ) = 2 . (1.49)

Now we can rewrite the momentum equation with = / and p = p/ as u + (u uT ) = + 2 u. p (1.50) t When we replace u by the vorticity which means we take the curl knowing that () = 0 we obtain an equation system ready to be p

38

discretized: + (u ) = 2 t 2 = u = . Pressure To obtain an equation for the pressure we can again use the momentum equation and take the divergence instead of the curl. We obtain u + ( (u uT )) = p + (2 u). t (1.52) (1.51)

In the rst term we can exchange the divergence and the partial derivative and we get a u, which we know to be zero. In the last term we can pull out the scalar Laplacian operator 2 = 2 /x2 + 2 /y 2 + 2 /z 2 which is applied to all three vector components of u and again we get a u. The remaining equation is the missing relation for the pressure p: ( (u uT )) + 2 p = 0. (1.53)

In the mesoscopic view of the Boltzmann-Equation all macroscopic uid characteristics are related to higher order moments (equations 1.29, 1.30, 1.31 and 1.33) of the particle and velocity distribution function f ( v, t). r, This distribution function actually describes the probability density (per unit mass) for nding a particle at a certain point in 3-dimensional space carrying a certain momentum in the 3-dimensional vector space. This distribution function could be simulated on the microscopic level by actually keeping track of all the atoms and molecules and their positions, velocities and collision accidents (if Heisenberg would not interfere in that). Of course, this over-stresses current computer powers by far and

39

1. Aero-Acoustical Theory

we cannot even think of such a simulation, but we do not need to simulate all the atoms and molecules. We just have to ensure that the effect of what we simulate has the same statistic consequences as what we consider the microscopic reality. A simple example can illustrate this idea. If nature throws 1023 different sized and imperfectly shaped dice even exposed to unknown environmental inuences, an engineer can throw one perfect articial die maybe ve hundred times in his lab and he will get a pretty close approximation of the probability density for getting a certain result. Microscopic Level This approach has been successfully taken and the method is usually referred to as lattice gas cellular automata (LGCA). Drastic simplications have been made to the microscopic scenery. Just one kind of particle is populating not continuous space but rather a kind of crystalloidal lattice. They do not move continuously in time, they jump from lattice node to adjacent lattice node after discrete amounts of time. Collisions are simulated with no physical background, just controlled by a random number generator, carefully tuned in order to produce a statistical outcome which approximates the physical reality. And of course the number of simulated particles is orders of magnitudes away from the true physical particle count. The idea of lattice gas cellular automata is quite old. In the early seventies Hardy and Pomeau published already such simulations using a square lattice [HP72]. Unfortunately they could not reproduce realistic macroscopic behavior due to inadequate symmetry of the square lattice. About fourteen years later Pomeau together with Frisch and Hasslacher found the explanation and an appropriate solution [FHP86]. But the statistical output of the microscopic level is still not what we are actually interested in. Who cares whether Boltzmanns distribution function is simulated correctly or not? It is the higher order moments, like uid density, uid velocity, pressure and energy we are really interested in. These macroscopic properties are obtained as certain volume and

40

sub-volume integrals in the 6-dimensional phase space of the distribution function. Mesoscopic Level If these volume integrals correctly represent the physical quantities or if they are at least a good approximation, then the simulation method can be considered to be successful. In order to ensure, that errors are not dramatically accumulating over a longer simulation period, the macroscopic conservation laws must be somehow enforced. That means even drastic simplications on the level of the distribution function are tolerable as long as macroscopic quantities still approximate their correct values. This was already used by the LGCA algorithms, because the simulated distribution function was not continuous but contained integer particle counts at discrete lattice nodes with corresponding velocities taken from a very small set of discrete velocities. It was not even possible for two particles to share the same discrete velocity at the same lattice node. From this, the idea of the Lattice-Boltzmann simulation method came up. Why simulate the microscopic level when there are other ways to determine the distribution function and its variation in time and space? If the distribution function is the fundamental element of the simulation it needs no longer be represented by discrete integer values. The Boltzmann-Equation is the transport relation for the distribution function and it would be very simple if the collision term were not there. The collision term is an extremely complex multi-dimensional function describing the statistical velocity transitions from any one 3-dimensional velocity into any other 3-dimensional velocity caused by random particle collisions. So all simplication attempts will try to tackle the collision term. The Maxwell-Boltzmann-Distribution species an equilibrium prole which will be reached after a very short time locally and much slower, depending on convection and heat conduction on a global level. The assumption leading to the rst simplication is that the statistical effect of all par-

41

1. Aero-Acoustical Theory

ticle collisions is to locally establish a Maxwell-Boltzmann-Distribution quickly. In a rst approximation, it is considered to be an exponential process, with a relaxation constant independent of all the other variables. This simplication of the collision model was proposed in 1954 by Bhatnagar, Gross and Krook [BGK54] and obviously independently by Welander [Wel54]. According to the initials of the inventors it got the name BGKModel. Introducing a constant relaxation time the collision term becomes 1 (f (r, v, t)) = (f (r, v, t) ((r, t), u(r, t), T (r, t))) (1.54)

with being the local Maxwell distribution according to Equation 1.35 for macroscopic mass density , convection velocityu and temperature T .

42

2. Numerical Concepts

As was shown in the previous chapter, nature is described by complicated non-linear partial differential equations for which analytical solutions either do not exist or are still not known today. In spite of this fact scientists have tried in the past and are still trying today to overcome these problems to get increasingly accurate predictions about all the different aspects of uid behavior. While in the past physicists have concentrated on getting simplied versions of the equations describing only some of the various aspects of uid behavior, in order to become able to develop analytical solutions, the situation has changed now. The invention of the computer and the astonishing development of its capabilities has created a new eld of science called computational physics, which deals with numerical analysis performed mainly on computers. The plural can really be taken literally, because often arrays of computers are working in parallel on a single uid-dynamical problem. Getting numerical solutions for a uid-mechanical problem described in three dimensions using the most general formulations of the Navier-Stokes equations is still expensive. For this purpose the calculation range has to be described by an adequately ne grid or mesh of points where state variables like ow speed, pressure, temperature, density. . . are calculated and recorded. The values of the discrete state variables are updated according to an adequate time grid by some sort of iteration algorithms or as in the linear case, at least by solving the linear equation system corresponding to the spatial grid over and over again. Simulating ne spatial grids will create huge matrices possibly with a wide spread of matrix coefcients (stiff systems) which are even difcult to solve numerically. On top of that, there are orders of magnitude between

43

2. Numerical Concepts

the time scales required to describe sound, sound propagation and steady air ow e.g. in an organ pipe. Therefore it is attempted to separate these effects. Whenever possible the simplied versions of the original equations will be taken as starting point even for applying the general computational concepts described in this chapter. Unfortunately especially in musical acoustics, there are many problems where air ow and sound propagation are so closely interrelated that it is not possible to get reasonable models from any of the simple formulations of the uid or wave equations and the most general formulation of the Navier-Stokes equations or alternatively the Boltzmann equation has to be employed. The most general methods which can be applied to simulate any general system described by partial differential equations have therefore received special attention in this chapter. Methods which are applicable only to simplied and more special formulations of the fundamental equations, like the wave equation, will be covered in the next chapter.

The methods explained in this chapter are generally applicable to solve systems of differential equations. However, not all concepts are equally suited for all types of differential equations. Some methods even have inherent limitations of the type of differential equation system which can be solved.

The Finite Difference Method (FDM) is the classical numerical method for solving systems of partial differential equations and will be covered in this chapter rst and in greatest detail, because the understanding of its principles as well as its problems will help to comprehend the underlying concepts and related problems of all the other methods, too.

44

Finite difference methods directly approximate the governing partial differential equations on a commonly but not necessarily equidistant spatial-temporal grid. The time and spatial derivatives within the equations are approximated by algebraic difference quotients involving function values at some discrete points within the computational domain of interest. Knowing the function values at all grid points at a certain point in time, the function values one time step later can be calculated either by solving a linearized equation system (implicit method) for the new function values or by performing a direct transformation of the old values (explicit method). By means of a Taylor-series expansion of the achieved discrete equation system the order of the error term of this approximation can be determined. By means of Fourier error analysis it can be investigated if the width of the spatial mesh or time step has been chosen adequately. It is therefore possible to select the simulation resolution to any desired level of accuracy. This and the fact that nite difference methods are easy to construct, analyze, and work with, have made them very popular. An often-reported drawback of nite difference methods is that they are usually designed for, and are most accurate on, uniform meshes. However, in recent publications the differencing schemes have been generalized in order to allow non-uniform spacing and also non-constant coefcients [Tho96]. But there is another limitation for the nite difference method which is not related to any specic implementation. It is related to the differential formulation of the governing equations themselves. In order to derive those from the more general integral formulations, it was assumed that state variables can be differentiated everywhere in the region of interest. If there are shock waves present in the computational domain then derivatives will tend to become innite and differences will become undeterminable.

45

2. Numerical Concepts

This is the domain of the Finite Volume Methods (FVM). They are based on the integral formulation of the governing conservation laws which are discretized directly in the physical space. This enforces conservation of the macroscopic quantities like mass, momentum and energy even at the discrete level which makes this method appropriate for ows with discontinuous characteristics. Another advantage of the nite volume methods is that they are not at all bound to structured computation grids with possibly equidistant grid spacing, but rather applicable to completely unstructured computational meshes consisting of arbitrary polyhedra in three-dimensional and arbitrary polygons in two-dimensional space. Flexibility, not only in terms of computational grid, permits adapting the discretization scheme to the underlying physical nature. That means knowledge about physics can be built into the scheme in order to increase the accuracy of the approximation. While nite difference schemes directly approximate the differential equations, nite volume schemes do the same with the integral formulations. The integral form of a conservation law consists of three terms. The rst one describes the overall change rate of a physical magnitude, e.g. mass or energy, inside a small volume element enclosed by a closed surface and xed in space. The second term describes the overall ux of this quantity through the boundary surface into or from the neighboring volumes. The third term nally is the source term, which represents the conserved quantity which appears inside the volume out of nowhere or which vanishes from inside the volume into nothingness. In the continuity equation the source term will usually be zero, however there are some problems which can be simplied, if ctitious materialization or dematerialization is assumed. In the momentum and energy equations the source term does have much more signicance, due to the various ways to create or destroy momentum or to convert energy from and to external processes. However, the basic principle of the nite volume method is easy to de-

46

scribe. Accounting what goes out and what comes in during a computational time step, the total content of the nite volume after that time step should be known and can be updated. Of course approximations again have to be introduced. First, only the overall content of the volume, or rather the average density in it, is maintained, so to determine the uxes at the boundaries an inner density distribution must be estimated. This step is usually referred to as reconstruction and has again much to do with interpolation. Second, the approximated ux at the boundary to an adjacent volume is not necessarily the same as that calculated in the other volume element. So this ux discontinuity must be resolved by some technique. Third, the resulting uxes must be numerically integrated over time and surface resulting in the additional or missing quantity which updates the volumes state variable. In recent years, the essentially non-oscillatory method (ENO) has been developed in order to increase the resolution of nite volume methods. This is achieved at the cost of large computational stencils which are difcult to resolve near to boundaries. Atkins and Lockard have been studying this method in the context of aero-acoustics [Atk95, LBA95].

The Space-Time Conservation-Element / Solution-Element (CE/SE) method originally developed by Chang and his colleagues [Cha95] is also based on nite volumes, but has signicantly distinctive conceptual elements, which makes it a separate concept. No reconstruction step is required and time and space are tightly coupled together in a common grid. It has the very interesting feature of being dissipation-free and it involves only two time levels. This is achieved by introducing the solution gradients as additional dependent variables. His differencing scheme is formally second-order accurate, but its non-dissipative nature makes it especially interesting for aero-acoustical simulation. Although the concept is applied primarily to the integral formulation of the conservation laws it can also be applied to the differential form, which

47

2. Numerical Concepts

relates it also to nite difference methods.

The Finite Element Method (FEM) is a technique which is very well established in various elds of engineering. Several commercial software packages are widely used and the theory of this method has been thoroughly elaborated. Originally the nite element method has its origin in the eld of structural mechanics, but meanwhile migration to several other elds has taken place. The basic idea was to minimize an energy function by means of a discretization based on the solution of a variational problem. For this purpose the governing differential equations are reformulated. The unknowns are described at discrete points of the modelled space as linear combinations of interpolation or basis functions in a way to make their coefcients the variational parameters. The interpolation functions are sampled at the nodes of an irregular mesh consisting of elements with arbitrary shape covering the whole region. Complex boundary conditions can be handled easily. The strong theoretical framework of the nite element method allows to analyzing and control of the error of the approximated solution. The basic idea of nite element methods can be applied to uid mechanics, if the velocity eld can be represented as the gradient of a scalar velocity potential. This limitation rules out vorticity, which reduces its general signicance as a uid dynamical simulation concept. By generalization of the original concept some other variants have been developed, which are generally applicable to uid dynamics, but which are no longer easily distinguishable from the other methods. Instead of on a variational problem, the method has been based on the integral formulation of the conservation laws. This makes it a kind of nite volume method. The numerical solution is based upon interpolation by the superposition of basis functions the coefcients of which are determined from some weak form of the governing equations. The approximated continuous function is then integrated over the nite volumes using e.g. the method of weighted residues.

48

Depending on the chosen basis functions, which are often called interpolation functions the method is classied into global and local methods. Spectral and pseudo-spectral methods are a global subset of nite element methods in which the basis functions are high-order functions. They have the highest level of accuracy but require special pseudo-spectral meshes if a general problem has arbitrary boundary conditions. The pseudospectral mesh near boundaries often requires very high resolution in space and time. Time-dependent problems tend to become unstable with higher order time-stepping methods. Pseudo-spectral techniques for aeroacoustic wave propagation problems have been investigated by Kopriva [Kop92, KK95]. As a local variant of the nite element method class that gained signicance in computational uid dynamics, the so-called Galerkin method should be mentioned. With its local and linear basis functions it is basically a nite volume scheme. Other interpolation functions like the Pad e formula (Equation 2.42), which will be treated in Section 2.2, can also be built in some hermitian nite elements. The nite element method has advantages concerning its exibility in terms of mesh construction and its rigorous mathematical framework for proving error bounds, stability, and convergence. Disadvantages are that knowledge about the physical nature of the analytic solution is not easy to build into a given scheme (for example upwinding) and time and space discretizations are not easily coupled together, as would be advantageous for long range wave propagation problems. Work on this has been published by Lowrie [Low95].

The theoretical framework of Wave-Digital Filters (WDF) was developed by Fettweis [Fet86]. Basically it deals with a certain class of digital lters for which most desirable characteristics could be proved mathematically. They are directly related to passive electrical networks sharing their stability and conservation laws. The signal processing counterparts derived in accordance with the wave

49

2. Numerical Concepts

digital lter theory are unconditionally stable and extremely robust against all kinds of imperfections that are introduced by the numerical implementation, like approximation errors caused by the differencing scheme, truncation errors caused by numerical quantization, overow distortions and more. This is achieved because passivity, which means energy conservation, is not only enforced globally but also incrementally in time and space within each computational step. For this purpose wave variables are introduced, combining the two physical quantities which contribute to the power term. Forward and backward travelling waves are dened as having the capability of transferring energy between adjacent points of a unied space-time grid while some kind of relaxation process takes place as described by the partial differential equations. In the case of uid dynamics the two physical quantities are pressure p and volume velocity u. Normally some differencing schemes are applied to both of them independently making the unknown approximation errors associated with the product pu to violate the energy conservation law. The power wave quantities are dened as a = (p + Ru)/(2 R) for incident waves and b = (p Ru)/(2 R) for reected waves, with R being some real-valued port impedance associated with a network junction. Wave scattering is performed according to the scattering matrix approach as described in [Bel68]. While neither pressure p nor volume velocity u can normally be identied as being the cause or the effect, wave quantities a and b have got such a relation. The power transmitted across a port in forward direction is pu = a2 b2 and it can be ensured by the scattering algorithm that an amount of energy which leaves some network port will enter another port without any degradation, even if the chosen difference scheme is not perfect and even in the case of port resistances which vary between scattering events. A multi-dimensional generalization of this principle which was already proposed by Fettweis himself [FN90] leads to possible applications as a tool for solving partial differential equations governing physical systems.

50

These continuous-domain systems are usually passive and therefore stable. By means of the wave digital lter theory corresponding discrete-domain systems described by difference equations of wave quantities can be identied. For these a scattering algorithm can be found that conserves that general passivity and thus all related features like robustness and stability, as well as the massive parallelism and exclusively local nature inherent to all physical systems with nite propagation speed. This concept has been thoroughly reviewed and extended by Bilbao in [Bil01]. A reader who wants to enter into this eld will nd a comprehensive and didactically well outlined elaboration. In this publication Bilbao formulates a unied theory of scattering based approaches including wave digital ltering and digital wave guide networks. Although both methods are very important for the integration of partial differential equations, time and space do not allow to treat them in adequate detail here. The cited work can be honestly recommended to the reader who wants to ll that gap.

The Boundary Element Method (BEM) is related to the nite element method based on the Galerkin approximation. It is derived from a similar integral formulation of the weighted residual of the governing law, which is then transformed using the theorems of Green and Gauss in order to obtain a sum of closed-surface integrals. If the weighting function is replaced by a fundamental solution of the original differential equation then the only remaining volume integral can be integrated analytically. This is possible because in that case the integrand contains a spatial Dirac Delta function concentrated around one point inside or at the edge of the solution domain. The remaining equation contains only boundary integrals and is therefore referred to as the boundary integral equation. It relates the value of the eld variable at any point inside or at the edge of the solution domain to integral expressions involving the eld variable and its gradient in the direction of the unit normal vector of the boundary surface.

51

2. Numerical Concepts

This equation is discretized by splitting the boundary integrals into sums of small and preferably at boundary elements called panels with spatially constant properties. The eld has to be determined at the boundary before points inside the solution domain can be calculated. One of the difculties connected to this method is that this boundary integral contains a singularity but has nevertheless to be calculated very accurately in order not to spoil the nal simulation results. Once the boundary integrals are known, it is straight forward to derive the eld at any other point. The boundary element method has advantages if the simulation domain is too big to be handled by volume methods. Especially if the domain is innite or semi-innite, the BEM is favorable. The differential equation is not approximated inside the boundary; just the boundary conditions are approximated, so just a boundary mesh is required. Anyhow, there exist some limitations which severely restrict the applicability. Firstly, it requires a fundamental solution which is not even known for all practically important linear problems. This means that all non-linear and virtually all nonhomogeneous linear equations are excluded. The boundary element method can be applied e.g. to the Laplace equation, the Wave equation, the Helmholtz equation and the Poisson equation. Secondly the implementation is difcult because sophisticated integration algorithms have to be used. And thirdly the numerical integration of the boundary integrals requires sampling at some retarded time depending on propagation speed and distance between panel and observer point. Therefore long time series have to be stored for all panels and the integrated samples have to be interpolated.

To simulate air-ow, jet oscillations and vortices in a ue-pipe, for example, no simplications of the Navier-Stokes equations are allowed. If a three-dimensional simulation is required e.g. motivated by the aim to recover the three-dimensional nature of turbulences or because of missing symmetries in the boundary conditions, numerical methods as described

52

above might exceed the computing power available to a musical acoustician today. An interesting and promising alternative approach, which is very efcient in terms of memory usage and computing power and which on top of that is much easier to implement than other methods is the LatticeBoltzmann method (LBM), which has recently attracted the attention of acousticians. It is based on a discretization of the particle velocity distribution and its equation of movement, known as Boltzmann equation. The LatticeBoltzmann method approximates the macroscopic behavior by modelling mesoscopically the particle velocity distributions in volume elements that are small enough to still represent the continuum of macroscopic quantities, but already big enough in order to justify the use of statistical particle analysis. The Lattice Boltzmann Method is well suited for simulating uid ow and has proven to be a valid and efcient concept for a variety of complex ow problems. In the eld of aero-acoustics, some fundamental studies have been undertaken to show its validity concerning linear [BGC98] and non-linear wave propagation [BBGG00] and acoustical streaming [HY01]. Only a few studies deal with applied problems [K h03, Wil02, Sko95]. u Therefore the use of the Lattice Boltzmann Method as a tool for computational aero-acoustics can be seen as in its early stage.

When Newton and Leibnitz crossed the border separating the worlds of discrete and continuous descriptions of nature by introducing the differential quotient, they kicked off an avalanche. Mathematicians and physicians conquered new summits exploring new elds and everybody was excited about the completely new set of tools to exactly describe nature. It seems that from this moment on the scientic community put all of its efforts to apply these new tools to all questions related to nature. The new descriptions of nature which have been obtained are the many systems

53

2. Numerical Concepts

of differential and partial differential equations, for example those which have been presented in the previous chapter. For some of these systems analytical solutions could be found, others have resisted all the continuing and persisting attempts of scientists even until today. It is not so long ago that, with the help of modern computers of course, some scientists, instead of putting all their efforts in nding analytical solutions, have focused their attention to the ways similar problems had been solved before calculus methods were discovered. These methods have been and still are based on tabulated functions, interpolations, extrapolations and iterations. Because tabulated function values have usually been precalculated in a xed equidistant grid of discrete argument values this process is referred to as discretization. The discretization grid nowadays does no longer need to be equidistant. It doesnt even need to be regular. Adaptive time stepping as well as spatial multi-grids, hierarchical grids, problem oriented irregular mesh and other discretization concepts like adaptive resolution are implemented in modern simulation programs. Apart from the discretization in time and space, it is essential to estimate the slope of a function, now designated as differential quotient, without knowing the function in between the discrete points. This process is sometimes referred to as differencing and many different interpolation algorithms are competing in terms of prediction accuracy, computational complexity, locality an indication for how many adjacent values are involved and error estimation order. By discretization and differencing a system of partial differential equations is transformed into a linear equation system for a set of new function values provided that a set of previous function values is already known. This allows to distinguish two types of problems: initial value problems, where a set of appropriate starting values is already known before the simulation is started, and boundary value problems, where required initial values are at least partially unknown while other values or derivatives at the simulation boundary are known but initially not needed.

54

Solving systems of partial differential equations numerically by discretization of time and space and by replacing the differential quotients at any discrete point rk in space and at any discrete time tn f (rk , tn ) t f (rk , tn ) x f (rk , tn ) y f (rk , tn ) z (2.1)

by values calculated from certain algebraic approximation functions involving differences of any of f (ri , tj ) is commonly referred to as Finite Difference Method (FDM). Using the notation that fk represents the value of the function f (x) for the discrete argument value xk which we dene as xk x0 +kx with x0 as some initial value and x as the constant spacing of available argument values x, we can dene an forward difference operator fk fk+1 fk . (2.2)

Higher order differences corresponding to the higher order derivatives 2 f /x2 and 3 f /x3 can be dened according to 2 fk fk+1 fk = fk+2 2fk+1 + fk

3 fk 2 fk+1 2 fk = fk+3 3fk+2 + 3fk+1 fk .(2.3) It can be observed that the coefcients are the binomial coefcients, which allows to write the general form n fk with n i =

n

(1)i

i=0

n fk+ni , i

(2.4)

n(n 1)(n 2) . . . (n i + 1) . i!

(2.5)

55

2. Numerical Concepts

(2.6)

i=0 n n

(1)i

n fki . i

(2.8)

In order to calculate an approximation for the nth derivative of a discrete function f with function value fk at its kth breakpoint we have used n adjacent points either from the right side (forward version) or from the left side (backward version). These versions are required when function values are only known for one side of the actual calculation point. This is normally the case for time derivatives - future function values will typically not be known, so the actual derivative has to be computed using past values only. In space this constraint is often non-existent because all points in space are calculated simultaneously from the linear equation system obtained after discretization and differencing. So function values for points left and right of considered calculation positions at a certain time step might enter the linear equation matrix. Therefore a central difference operator is often useful. Analysis shows that the central difference is more accurate than the forward or backward version. Its error estimation term is one order higher than that of the other two. The denition of the central difference operator requires the denition of a central average fk + fk+1 (2.9) fk+ 1 2 2

56

which is used when a missing table value is required. The operator is dened as (2.10) fk fk+ 1 fk 1

2 2

2 2 2 2

2 2 2

(2.11)

2

3 fk 2 fk+ 1 2 fk 1 = fk+ 3 3fk+ 1 + 3fk 1 fk 3 fk 3 fk+ 1 3 fk 1 = fk+2 4fk+1 + 6fk 4fk1 fk2

2 2

generally fk

n

(1)i

i=0

n f n . i k+ 2 i

(2.12)

Replacing the differential quotients by difference quotients as dened above is indeed a possibility and can be done. Examples will be given to understand the principle and to simultaneously illustrate the problems connected to this simple approach.

Explicit differencing schemes result in a linear or non-linear expression for the calculation of a new generation of updated function values only depending on known function values belonging to one or more previous generations. By means of this explicit formula new function values can be computed directly and very efciently. Thermal Diffusion Example Let us demonstrate a very simple thermal diffusion problem described by the following one-dimensional parabolic partial differential equation: 2u u =D 2 t x (2.13)

57

2. Numerical Concepts

25 20 15 10 5 5 25 20 15 10 5 5 25 20 15 10 5 5 10 15 20 25 30 35 40 10 15 20 25 30 35 40 25 20 15 10 5 5 10 15 20 25 30 35 40 10 15 20 25 30 35 40 25 20 15 10 5 5 10 15 20 25 30 35 40 25 20 15 10 5 5 10 15 20 25 30 35 4 25 20 15 10 5 5 10 15 20 25 30 35 40 25 20 15 10 5 5 10 15 20 25 30 35 4 25 20 15 10 5 5 10 15 20 25 30 35 4

A straight forward approximation could be to apply the second order central difference operator at time n in space x and a forward difference operator in time t with un = u(xi , tn ) denoting the ith sample in x and i the nth sample in time: 2 ui un =D . t x2 Substituting the difference operators we obtain D(un 2un + un ) un+1 un i+1 i i1 i i = . t x2 (2.15) (2.14)

With Dt/(x)2 we obtain the explicit simulation formula un+1 = (1 2)un + (un + un ) i i1 i+1 i (2.16)

58

which can directly be used to calculate a new prole ui for the time n + 1 from a known one at time n. Figure 2.1 shows simulated proles for = 0.45. The simulation length was 100 length steps in x-direction, 40 of them being shown in the plots. The stimulation at the left end is a sinusoidal signal with DC-offset according to 20 + 5 sin(2n/10). Therefore the period of the stimulation is 10 time steps. The right end has been forced to zero. After about 100 time steps a steady state condition is reached. The plots are showing almost one period of the excitation signal as 9 consecutive time steps after having reached a steady state. The graphics are arranged in three horizontal rows of three items each. Wave Propagation Example Let us now consider another simple differential equation: u u = c . t x (2.17)

It is the linear convection equation and can be used to study wave propagation. It describes a homogeneous uid moving with constant speed from left to right if c is positive, or from right to left if c is negative. If the left boundary (for positive convection speed c) is stimulated harmonically, then an unattenuated non-dispersive wave should propagate in the convection direction with no reection on the right outow boundary. It is interesting that the linear convection equation can be transformed into the wave Equation 1.38. The transformations according to u u +c t x u u +c t x = 0 | c /x = 0 | /t, (2.18)

59

2. Numerical Concepts

yield two equations which only have to be added c 2 u/tx c2 2 u/x2 = 0 2 u/t2 + c 2 u/xt = 0 +

in order to get the hyperbolic wave equation 2 u/t2 = c2 2 u/x2 in its well known form (compare equations 1.38 and 3.1 for the three- and one-dimensional case). The linear convection equation looks simpler than the previous diffusion equation, because it contains only rst derivatives, but nevertheless it will show the limits of the simple explicit approach, which was demonstrated before. Applying again the forward difference operator to replace the time derivative and a central difference operator to replace the spatial derivative we obtain ui un +c = 0. (2.19) t x Substituting the difference operators we obtain

n n un+1 un c(ui+ 1 ui 1 ) i i 2 2 + = 0. t x

(2.20)

With k = ct/(2x) and with the central average un 1 = (un + i+1 i+ un )/2 we obtain the explicit simulation formula i1 un+1 = un k(un un ) i i+1 i1 i

2

(2.21)

which can directly be used to calculate any value ui of a prole for the new time step n + 1 from known values ui , ui+1 , ui1 calculated the time step before. While this method worked pretty well for the diffusion case, it is a disaster for the wave propagation example. Figure 2.2 shows that the simulation is unstable and amplitudes grow exponentially after a very short period of time. In Section 2.2.7 we will address the reasons.

60

0.2 0.15 0.1 0.05 5 10 15 20 25 30 35 40 -0.05 -0.05 -0.1 -0.15 5 10 15 20 25 30 35 40 3 2 1 -1 -2 -3 5 10 15 20 25 30 35 40

0.1 0.05

75 50 25 -25 -50 5 10 15 20 25 30 35 40

2 10 5 10 15 20 25 30 35 40 -2 10 -4 10

5 10 5 10 15 20 25 30 35 40

-1 10

Although the explicit method is simple, extremely convenient and computationally efcient, there are major drawbacks concerning stability. The advantages that no equation system has to be solved and that simulation can be performed by evaluating a simple formula for each spatial point, time step by time step, are counterbalanced by the fact that for many types of problems stability is a real issue. In the next section we will use an implicit approach to simulate wave propagation in our linear convection example.

Implicit methods do not have the restriction that equations for new function values must not contain other unknown function values. New values are calculated from a system of linear equations where all values of the new time step n + 1 are variables and all values of the previous time step

-500000 6 -1 10 6 -1.5 10

-5 10

1.5 10 6 1 10 500000

4 10

1 10

9 8

8 9

5 10 15 20 25 30 35 40

61

2. Numerical Concepts

n are constants. With r spatial points the equation system for at time step n + 1 looks like n+1 u1 a1,1 a1,2 a1,3 . . . a1,r a2,1 a2,2 a2,3 . . . a2,r un+1 2 a3,1 a3,2 a3,3 . . . a3,r un+1 3 . . . . . .. . . . . . . . . . . . ar,1 ar,2 ar,3 . . . ar,r un+1 r

Because the number of spatial points does not change from time step to time step, the coefcient matrix must be square. The coefcients ai,j do not change either, only the vector un is replaced after each time step by a more recent one. As can be seen in the following example, the matrix ai,j is not a general matrix but contains non-zero coefcients only in the proximity of the main diagonal. All these facts are very advantageous for efcient numerical treatment and many clever algorithms have been invented just for this special case. Stability of the simulation can be analyzed by means of the Von Neumann method as will be demonstrated in Section 2.2.7. So much for the good news. The bad news is that the size of the coefcient matrix is k2 if k is the number of spatial points. So for a three dimensional aero-acoustical problem with a resolution of 100 points per dimension we have to calculate 100 100 100 = 106 points which requires a coefcient matrix of (106 )2 = 1012 coefcients corresponding to 8 Terabytes of core memory! Well, as we said, those coefcient matrices are sparse, that means the biggest part of it contains nothing more than zeros, which do not have to be stored nor multiplied, but are 100 points an adequate resolution for 3-dimensional acoustical problems? Wave Propagation Example (Implicit) Let us reconsider the linear convection equation 2.17 and the differencing according to Equation 2.19. Applying again the forward difference operator to replace the time derivative and a central difference operator for the

62

spatial derivative, but this time we take the spatial differences at the time step n + 1 instead n, so we obtain

n+1 n+1 un+1 un c(ui+ 1 ui 1 ) i i 2 2 + = 0. t x

(2.23)

Applying the central average to the non-existing values un+1 and un+1 i 1 i+ 1 we obtain the calculation scheme with k = ct/(2x) as kun+1 + un+1 + kun+1 = un . i i1 i i+1

2 2

(2.24)

n+1 This expression cannot be transformed into an explicit formula for u i because it contains a simultaneous dependence of the new neighbor values un+1 and un+1 . But we can write a simultaneous linear equation system i1 i+1 by tabulating this implicit relation for all spatial points:

kun+1 i1 0 0 0

+un+1 i kun+1 i +0 +0

+0 +0 +0 +kun+1 i+4

= = = =

Here is the matrix notation for this linear equation system: n+1 n u1 u1 1 k 0 0 0 0 ... 0 0 un+1 un k 1 k 0 0 0 ... 0 0 2 2 0 k 1 k 0 0 ... 0 0 un+1 un 3 3 0 0 k 1 k 0 . . . 0 0 un+1 un 4 4 0 0 0 k 1 k . . . 0 0 un+1 = un 5 5 . . . un+1 un . . . . .. . . . 6 . . . . . 6 . . . . . . . . . . . . . . 0 . 0 0 0 0 0 ... 1 k . 0 0 0 0 0 0 ... k 1 un+1 r

un r (2.26) It can now be observed that only about 3 n regular positions are occupied by one of three different values, therefore the matrix does not have to

63

2. Numerical Concepts

be stored at all. Anyhow the system has to be solved and methods to do that will be discussed in a later section. Boundary conditions, like the vibrating wall at the left side can be din rectly enforced in the vector u. In the present example the rst element u 1 has been reset to (2.27) un = 0.1 sin(n2P i/), 1 with = 30 before calculating the next time step by solving the linear equation system.

0.1 0.075 0.05 0.025 -0.025 -0.05 -0.075 -0.1 20 40 60 80 100 120 140 0.1 0.075 0.05 0.025 -0.025 -0.05 -0.075 -0.1 20 40 60 80 100 120 140 0.1 0.075 0.05 0.025 -0.025 -0.05 -0.075 -0.1 20 40 60 80 100 120 140

0.1 0.075 0.05 0.025 -0.025 -0.05 -0.075 -0.1 20 40 60 80 100 120 140

0.1 0.075 0.05 0.025 -0.025 -0.05 -0.075 -0.1 20 40 60 80 100 120 140

0.1 0.075 0.05 0.025 -0.025 -0.05 -0.075 -0.1 20 40 60 80 100 120 140

0.1 0.075 0.05 0.025 -0.025 -0.05 -0.075 -0.1 20 40 60 80 100 120 140

0.1 0.075 0.05 0.025 -0.025 -0.05 -0.075 -0.1 20 40 60 80 100 120 140

0.1 0.075 0.05 0.025 -0.025 -0.05 -0.075 -0.1 20 40 60 80 100 120 140

Using the equation solver based on LU-Decomposition, which is implemented in Mathematica [Wol88], the simulation was run. Results are shown in Figure 2.3. More details together with an in-depth discussion about the consequences of various differencing and interpolation schemes on the accuracy of far-eld sound propagation simulations can be found in [WR97].

64

The simulation is apparently stable and the resulting proles look familiar at a rst glance, but having a closer look reveals some strange and unexpected issues. First, there is some strong attenuation in the resulting wave form, which would require signicant damping to be modelled by the simulated differential equation. But the convection equation we started from is linear and it does not contain any loss term at all and remember, we are simulating a one dimensional world where energy does not spread out. So there must be some kinds of numerical losses that become dramatically signicant in even this simple kind of simulation.

0.1

0.075

0.05

0.025

25

50

75

100

125

150

175

200

-0.025

-0.05

-0.075

-0.1

Second, some dispersion seems to take place wave propagation speed does not seem to be constant. Indeed, a deeper investigation shows that wave speed is related to frequency, distorting multiple-frequency wave forms. Finally, artifacts like spurious high spatial frequency reections can be observed as shown in Figure 2.4.

65

2. Numerical Concepts

It has to be noted that all these effects start getting signicant after the sound waves have propagated some distance in space. The same numerical anomalies taking place in a simulation of aero-dynamic loading of a body might not cause problems there because results are required in close proximity of the excitation and not in the far-eld. This is one of some signicant differences between aero-dynamics and aero-acoustics explaining, why aero-acoustical simulations are much more difcult to perform and why acousticians are still struggling with problems which do not seem to play any role in other elds or which scientists of other elds claim to having solved already. Another issue special to musical acoustics is the close and bidirectional interaction between ow and sound as will be addressed in later chapters. For now it seems that we have to deal with improvements to the basic differencing schemes and to understand their inuence on accuracy and stability.

2.2.4. Interpolation

Replacing differential quotients by difference quotients is fairly straightforward and proceeding towards innitely ne grid, it will perfectly approach the exact value for the slope of the tangent. This was the idea of Newton and Leibnitz who nally found an analytical expression for the limit of the difference quotient for x 0. This expression was then called the differential quotient. For us, increasing the grid resolution is a very expensive step because doubling the spatial grid will increase the complexity of a three dimensional problem by a factor of 8. If the time grid has to be doubled, too, we have increased the simulation cost in terms of memory and CPU requirements by about one and a half orders of magnitude. Polynomial Approximation Therefore it is much more efcient to approximate the function, which is known by a few equidistant points only, by a polynomial expression

66

which we can differentiate as many times as we need to. The polynomial coefcients are calculated from the requirement that the breakpoints of the unknown function must exactly lie on the approximating curve. The three coefcients of a second order parabola are therefore determined by three known points. In general a nth order polynomial approximation can be found for every set of n + 1 points of the original function. As an example it will be demonstrated how an unknown function, represented by three points, can be approximated by a second order polynomial expression describing a parabolic curve. The exact derivatives of the interpolation function will then be compared with the original function. The starting equation system for the three coefcients a, b and c of the polynomial approximation a + b(x xk ) + c(x xk )2 based on forward samples fk ,fk+1 ,fk+2 at positions xk , xk+1 = xk + x and xk+2 = xk + 2x can be indicated as a + b(x xk ) + c(x xk )2 = fk+2 | x xk + 2x a + b(x xk ) + c(x xk )2 = fk+1 | a + b(x xk ) + c(x xk )2 = fk | x xk + x x xk . (2.28)

The solution of equation system 2.28 for a, b and c can be found as fk 2fk+1 + fk+2 . 2(x)2 (2.29) Knowing the coefcients, the approximation function and its derivatives can be written as function of x = (x xk ) according to a fk b c F () = fk + x x (3fk + 4fk+1 fk+2 )+ 2(x) x2 (fk 2fk+1 + fk+2 ) (2.30) + 2(x)2 3fk + 4fk+1 fk+2 2x

67

2. Numerical Concepts

F () = fk + x x (3fk + 4fk+1 fk+2 )+ 2(x) x2 (fk 2fk+1 + fk+2 ) (2.31) + 2(x)2

x F () =

1 (3fk + 4fk+1 fk+2 )+ 2(x) x (fk 2fk+1 + fk+2 ) (2.32) + (x)2 1 (fk 2fk+1 + fk+2 ). (x)2 (2.33)

x F () =

A general expression for polynomial approximation functions of any order m using forward differences and normalized x = x/(x) can be given as

x Fm () = fk +

x x 2 fk + fk + = 1 2

m

= fk +

i=1

x i fk + R((x)m+1 ), (2.34) i

with R((x)m+1 ) being the residue, which is a function of (x)m+1 . Using backward differences a similar expression is obtained:

x Fm () = fk + x fk +

x2 + x 2 fk + = 2! m x + i 1 i fk + R((x)m+1 ). (2.35) = fk + i

i=1

This method of deriving the coefcients of a polynomial approximation is referred to as the Newton or Newton-Gregory interpolation scheme in

68

literature. Using central differences it is often named Sterling formula [Ves93]. It looks a bit different because the interpolation order must be even and it contains fk terms with odd exponents which have to be replaced by the difference of central averages according to Equation 2.9.

x2 2 2! fk x3 x 3 3! fk k x4 2 x 4 4! fk

x F2n () = fk + x fk + = fk +

m i=1

x+i1 2i1

2i1 f

x 2i 2i fk

+ R((x)2n+1 ). (2.36)

Instead of giving the lengthy expressions for the rst up to the fourth x x x derivative of the interpolation functions Fm (), Fm () and F2n () the polynomial coefcients have been tabulated in the Appendix. The tables entries look like the example given below which represents the second order approximation for the rst derivative using forward values. This polynomial is repeated here in order to demonstrate how to interpret the tabulated information. It was derived in Equation 2.30: 3fk + 4fk+1 fk+2 x(fk 2fk+1 + fk+2 ) + . 2.37) ( 2(x) (x)2

x F () =

fk+2 1 1

fk fk+1 3 4 1 2

69

2. Numerical Concepts

The matrix contains one row of coefcients for each term, the corresponding variable names being prepended as row zero. The numerator and denominator values of the multiplicator for each term are indicated above each coefcient matrix.

1.2

0.8

0.6

0.4

0.2

50

100

150

200

Let us now evaluate the quality of some of these approximation functions by means of simple examples. Let us start with a seemingly smooth and predictable function, e.g. f (x) = sin(t/20)2 + 0.3 sin(t/3). In Figure 2.5 this curve is plotted together with points marking the interpolation breakpoints. In the Figure-Tables 2.6 and 2.7 interpolation results can be compared. Curves have been included for 2nd order, 4th order and 6th order, each plot showing forward (red), backward (green) and central (blue) differencing. The target curve has been plotted dashed and in grey. The approximation center k has been marked by the cross-hairs.

70

1.2

2

1

1.5

0.8

1

0.6

0.5

0.4

10

20

30

40

0.2

-0.5

-1

10

20

30

40

(a) F (x),Order=2

0.4

(b) F (x),Order=4

0.2

0.8

0.6

10

20

30

40

-0.2

0.4

-0.4

0.2

-0.6

10

20

30

40

(c) F (x),Order=2

0.3

1

(d) F (x),Order=4

0.2

0.5

0.1

10

20

30

40

10

20

30

40

-0.1

-0.5

-0.2

-1

-0.3

(e) F (x),Order=2

(f) F (x),Order=4

71

2. Numerical Concepts

0.4

1

1.5

0.2

0.5

1

10

0.5

20

30

40

10 20 30 40

-0.2

-0.5

10 20 30 40

-0.4

-1

-0.5

-0.6

(a) F (x)

(b) F (x)

(c) F (x)

Alternative Solutions As interpolation and differencing are crucial topics in order to successfully apply a nite difference method to any real-world problem, many books and articles have been written on that. Many theoretical aspects of computational uid dynamics are covered in depth in the book Fundamentals of Computational Fluid Dynamics by Lomax and Pulliam [LP01]. More implementation oriented aspects can be found in publications of Eckhoff [Eck03]. Often the Lagrange interpolation formula is mentioned. For a second order interpolation polynomial it can be written as: (xk+1 x)(xk+2 x) + (xk+1 xk )(xk+2 xk ) (xk x)(xk+2 x) + + fk+1 (xk xk+1 )(xk+2 xk+1 ) (xk x)(xk+1 x) . + fk+2 (xk xk+2 )(xk+1 xk+2 )

F i(x) = fk

(2.38)

If this formula is expanded and evaluated for equidistant xk the Newton coefcients for second order forward differences according to Equation 2.30 are obtained. So there is obviously nothing special with the Lagrange interpolation method, it is just another way to describe the one and only second order polynomial dened by three points.

72

But it can be stated that the requirement for equidistant breakpoints which was introduced in Equation 2.28 does not exist for the Lagrange formulation. If we had not introduced this simplication we could have ended up with the Lagrange formulation. The Lagrange formulation can be still further generalized. An interpolating function F im (x) of any order m can be constructed from m + 1 (m) polynomials Pk in x of degree m and m + 1 known values of the interpolated function fk :

m

F im (x) =

k=0

Pk

(m)

(x)fk .

(2.39)

Hermitian interpolation generalizes this concept even further. Known function values as well as known derivatives are involved: F im (x) = Pk

(m)

(x)fk +

Qk (x)fk +

(m)

Rk (x)fk . . . (2.40)

(m)

As a simple application of this concept a combination of some basic differences could be considered. As an example, the second order backward approximation for the rst derivative at x 0 according to Section B.2 on page 322 is fk = (fk2 4fk1 + 3fk )/(2x), the corresponding forward approximation at x 0 according to Section B.1 on page 315 is fk = (3fk + 4fk+1 fk+2 )/(2x) and the central approximation according to Section B.3 on page 329 is fk = (fk1 + fk+1 )/(2x). Building the sum fk1 + 4fk + fk+1 and introducing the shorter abbreviation h x we obtain fk1 4 fk fk+1 fk1 fk =( 3 4 = ( 4 (1) =( 1 4 6 fk1 fk+1 1 41 3 ) / (2 h) ) / (2 h) ) / (2 h)

73

2. Numerical Concepts

By this manipulation we got an implicit expression, known as spline approximation of the rst derivative, being member of the class of so called Pad formulae, e fk1 + 4fk + fk+1 = 3 fk+1 3 fk1 , h (2.42)

which we can no longer solve explicitly. Admittedly, this is a disadvantage but on the other side the equation is now accurate to the 4th order of h x, as we will show now. If the Taylor expansion

(kh)n

n=1

1 n f (x0 ) n!

(2.43)

(kh)n

n=1

1 n+1 f (x0 ) n!

(2.44)

is applied to the terms of Equation 2.42 up to fth order we can write hfk h2 fk fk 1 = 1 (1) 1 (1)2 2 1 1 21 = 1 (+1) 1 (+1) 2 0 0 0 1 1 1 h3 f 3 k h4 f 4 k h5 f 5 k 1 1 (1)3 1 (1)4 24 (1)5 120 6 3 1 (+1)4 1 5 1 (+1) 6 24 (+1) 120 (1)3 1 6 0 31 (+1) 6

1 (1)4 24 0 4 1 (+1) 24

fk1 fk+1

fk1 = fk = fk+1 =

(2.45) We multiply the Equation 2.42 with h x and taking the correct coefcients into account now, we can rewrite the table in order to compare

74

left side and right side term by term: hfk1 4xfk hfk+1 = = = = fk hfk h2 fk h3 f 3 k h4 f 4 k h5 f 5 k 1 1 1 0 1 1 2 6 24 0 4 0 0 0 0 1 1 1 0 1 1 2 6 24 2 0 6 0 1 0 24 3 3 6

3 2 3 2 3 6 3 6 3 24 3 24 3 120 3 120 6 120

(2.46)

3 fk+1 = 3 3 fk1 = 3 = 0

As we can see both sides do exactly match up to the fourth order term, that is why we are talking about fourth order accuracy. The approximation error is fth order because it is proportional to f(5) . Its coefcient is h4 , because we have to take the initial multiplication of the equation by h x into account. This is a much better accuracy than what we have achieved before using any other three point scheme. It should be mentioned that the well-known spline approximation for the second derivative also falls into this category of implicit Pad formulae. It e is dened by fk1 + 4fk + fk+1 = or alternatively by 3(fk1 fk+1 ) (2.48) x and is considered to offer a supreme global continuity with minimum curvature and very good stability characteristics. However its accuracy is poor, the error term is of second order only. Modifying two coefcients makes it a fourth order operator fk1 + 4fk + fk+1 = fk1 + 10fk + fk+1 = 12(fk1 2 fk + fk+1 ) , (x)2 (2.49) 6(fk1 2 fk + fk+1 ) (x)2 (2.47)

75

2. Numerical Concepts

as can be checked using the method according to equations 2.45 and 2.46.

0.2

0.1

10

15

20

-0.1

-0.2

Figure 2.8.: Sixth Order Central Interpolation of First Derivative of Noisy Signal

Cubic spline polynomials have become so popular because they create very smooth-looking interpolation curves. Higher order polynomials tend to wiggle unnecessarily and can develop strong curvatures depending on the interpolated data set. The high sensitivity caused by nonlinear amplication of small data uctuations is demonstrated in Figure 2.8. This effect can be a severe stability issue of underlying FD simulations. The idea of cubic splines was to restrict the polynomial order to three, but enforce perfect continuity of position, slope and curvature at the intersection points between its validity range and that of its neighbors. The four polynomial coefcients are derived from the two section endpoints and the matching requirements for f and f with the left and right neighbor spline.

76

From these conditions six equations can be formulated containing the polynomial coefcients ak , bk , ck and dk and the rst and second derivatives of the interpolated function at the end points of the kth section which is the interval from xk to xk+1 : ak x3 + bk x2 + ck xk + dk = fk k k ak x3 + bk x2 + ck xk+1 + dk = fk+1 k+1 k+1 3ak x2 + 2bk xk + ck = fk k 3ak x2 + 2bk xk+1 + ck = fk+1 k+1 6ak xk + 2bk = fk 6ak xk+1 + 2bk = fk+1 (2.50)

Similar equations can be formulated for the previous interval from xk1 to xk : ak1 x3 + bk1 x2 + ck1 xk1 + dk1 = fk1 k1 k1 ak1 x3 + bk1 x2 + ck1 xk + dk1 = fk k k 3ak1 x2 + 2bk1 xk1 + ck1 = fk1 k1 3ak1 x2 + 2bk1 xk + ck1 = fk k 6ak1 xk1 + 2bk1 = fk1 6ak1 xk + 2bk1 = fk (2.51)

By eliminating fk1 , fk , fk+2 all coefcients ak , bk , ck , dk and ak1 , bk1 , ck1 , dk1 can be derived and Equation 2.47 can be obtained. Eliminating fk1 , fk , fk+2 the coefcients ak , bk , ck , dk and ak1 , bk1 , ck1 , dk1 can be derived again and Equation 2.48 can be obtained. The implicit Pad equations can be formulated for the whole range of e interpolation breakpoints and arranged in matrix form. If this is done with the spline Equation 2.48 the following system can be obtained:

77

2. Numerical Concepts

4fk 0 0 0 +fk+1 +0 +fk+1 +fk+1 +0 +fk +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 +0 = = = =

fk1 +4fk

+4fk+2 +fk+3

= = = = =

3 x ( 3 x ( 3 x ( 3 x ( 3 x (

+0 +0 +0 +0

+fk +0 fk +0 +0 +0 +0

+0

+0 +0

+0 +0 +0

) ) ) )

fk1 +0

+fk+1 +0 fk+1 +0

+fk+2 +0 fk+2 +0

+fk+3 +0

+fk+4 )

Zero boundary conditions for f and f of the rst and last element have been assumed in order to end up with a square matrix where the number of equations matches the number of unknowns. The matrix notation for this linear equation system is sometimes written as B(1, 4, 1)f = or as f = 3 B(1, 4, 1)1 B(1, 0, 1)f x (2.54) 3 B(1, 0, 1)f x (2.53)

with B(a, b, c) meaning the banded matrix with the constants a, b and c contained in the three bands clustered around the main diagonal. Such a matrix is also called tridiagonal matrix and we will encounter this type very often. Fortunately there are very efcient algorithms to deal with banded matrices, e.g. to invert them, which means to solve the corresponding system of linear equations.

78

The full notation of Equation 2.53 would be: 4 1 0 0 0 . . . 0 0 0 0 0 0 0 1 0 4 1 . .. . . . 0 0 0 0 0 0 0 0 0 0 0 1 4 1 0 0 . . . 0 1 4 1 0 . . . 0 0 1 4 1 . . . ... ... ... ... ... . . . ... ... 0 0 0 0 0 . . . f1 0 0 f2 0 f3 0 f4 0 f5 . f6 . . . . 4 1 . fr 1 4 0 0 0 0 0 . . . = (2.55)

0 1 0

0 f1 0 f2 0 f3 0 f4 0 f5 . f6 . . . . 1 . fr

Algorithms for inverting banded matrices will be reviewed in the algebra section below. Here it should only be noted that solving the implicit equation system for a spline represented by a constant tridiagonal matrix is pretty straight forward and has very little to do with inverting a general n n matrix. Finally, it has to be stated that more alternative interpolation methods do exist. Especially when some properties of the interpolated function are known in advance, then this knowledge can help in the interpolation process tremendously. Consider the case of a strictly periodical system. Transients have decayed and a steady state of oscillation has been reached. As in most natural systems there is damping and therefore some low-pass characteristic in the feedback loop. That means, we have perfect conditions for a harmonic

79

2. Numerical Concepts

interpolation, which means that a Fourier decomposition of the interpolated function can be attempted and the Fourier coefcients can be derived either directly or by a numerical t. Once the complex amplitudes of the fundamental and some higher partials have been determined, derivatives of any order can easily be calculated. The same concept can be applied spatially when standing waves between ideally reecting boundaries are to be investigated. Interpolation can be based on a modal decomposition of the discrete signals. The functions describing the different modes can be evaluated and differentiated analytically wherever this is needed. The weighting coefcients can be determined by a numerical t. For time-dependent hyperbolic problems like advection and aero-acoustics where the physical nature of the analytic solution is known at the differential equation level, the discretization can be biased in a manner which best represents the known physics. Upwind differencing, which may also be used in nite volume methods, is a primary example of this. In other cases it is possible to determine some analytic functions, so called trial functions, from which the unknown function can be composed. These kinds of variational problems are formulated and solved in the so called Finite Element Method (FEM). Although FEMs are not very popular in computational uid dynamics, there are some promising attempts. Finite element methods for uids are usually based on the integral formulation of the conservation laws, and therefore very much related to Finite Volume Methods.

Differencing always involves neighbors of the actually interpolated function value or rather its derivative: That is not a problem as long as we have not approached the boundary of the computational mesh. In time we are always at the boundary of the known region as we normally do not have any knowledge about the future. One way to solve this difculty is to dynamically switch to another differencing scheme whenever required function values are not available.

80

We could for example choose a fourth order central scheme requiring fk2 ,fk1 ,fk ,fk+1 ,fk+2 in space to process the interior part of a mesh. Approaching the left boundary we could switch to an unsymmetrical third order scheme which does not require fk2 . The left-most value could then be derived from a suitable forward differencing scheme if it is not already enforced by the external boundary conditions bc. The same concept can be applied to the right boundary. Fortunately, it is tolerable to decrease the interpolation order near the boundary by one without degrading the global accuracy which is equal to that of the interior scheme. In matrix notation the approximation of the rst spatial derivative in one dimension by the fourth order central differencing scheme, taking the boundary conditions into account can be written as: f = A f + bc (2.56) with 4u0

f2 f 3 f 4 f = f5 . . . f r3 fr2 fr1

f1

f1

u0 0 0 1 0 bc = 12x . . . 0 ur 4ur

(2.57)

81

2. Numerical Concepts

and 1 A= 12x 6 12 2 0 0 ... 0 8 0 8 1 0 . . . 0 1 8 0 8 1 . . . 0 0 1 8 0 8 ... 0 0 0 1 8 0 . . . 0 . . . .. . . . . . . . 0 0 ... 0 1 8 0 0 0 ... 0 0 2 12 . 8 6 0 0 0 0 0 . . .

(2.58)

The coefcients of the fourth order central scheme can be looked up in the Appendix. For the boundary scheme used in the rst and last row of the square matrix A Equation 2.59 can be referred. Boundary Schemes Various boundary schemes can be obtained by combining two or more of the basic schemes as listed in the Appendix. Taylor analysis as described in Section 2.2.6 on page 85 can be performed to determine the resulting approximation order. A common third order numerical boundary scheme which complements a fourth order central interior scheme is given by fk = 2fk1 3fk + 6fk+1 fk+2 . 6x (2.59)

This unsymmetrical differencing scheme can be obtained by combining the second order central scheme with the second order forward scheme: fk1 2fk = 4fk = 6fk = 2 2 fk fk+1 fk+2 3 4 1 2 1 3 6 1

(2.60)

82

Extrapolation In order to extrapolate missing function values at the boundaries of the equidistant integration mesh to any required order, the following extrapolation formula can be used: (1 E 1 )m fk+1 = 0 Efk = fk+1 (2.61)

The accuracy of this extrapolation scheme is one order less than the exponent m. E is a formal shift operator which adds or subtracts one from the subscript. As an example, a forward extrapolation scheme for m = 2 meaning extrapolation order one, is derived: (1 2E 1 + E 2 )fk+1 = 0 fk+1 2fk + fk1 = 0 fk+1 = 2fk fk1 For 2 < m 6 corresponding results can be obtained: fk+1 = 3fk 3fk1 + fk2 fk+1 = 4fk 6fk1 + 4fk2 fk3 fk+1 = 5fk 10fk1 + 10fk2 5fk3 + fk4 fk+1 = 6fk 15fk1 + 20fk2 15fk3 + 6fk4 fk5 If the extrapolation formula is resolved for downward differences the following schemes can be obtained: fk+1 + 2fk = fk1 fk+2 3fk+1 + 3fk = fk1 fk+3 + 4fk+2 6fk+1 + 4fk = fk1 fk+4 5fk+3 + 10fk+2 10fk+1 + 5fk = fk1 fk+5 + 6fk+4 15fk+3 + 20fk+2 15fk+1 + 6fk = fk1 (2.64) (2.63) (2.62) (1 E 1 )2 fk+1 = 0

83

2. Numerical Concepts

The extrapolated value can be inserted in any other scheme. If all values of one side of a symmetrical scheme are replaced by extrapolated values a corresponding forward or backward scheme is often obtained. The above unsymmetrical third order scheme can also be obtained by inserting the 3rd order extrapolation for fk2 = 4fk1 6fk + 4fk+1 fk+2 into the central fourth order scheme 12xfk = fk2 8fk1 + 8fk+1 fk+2 .

The improvement achieved by higher order differencing can be clearly observed. Anyhow, it is difcult to generally predict the achievable matching. We must not forget that an approximation function does not have any knowledge about the curve it has to approximate, other than the function values at some equidistant points. It does not have any information about the shape of the curve in between these breakpoints nor about derivatives there. It is obvious that there is some connection to Nyquists sampling theorem. If the function contains frequency components with half wave lengths in the same order of magnitude as grid distance or even shorter, then such an approximation attempt would terribly fail. This can happen, if the simulated system contains little damping and high loop amplications caused e.g. by strong non-linearities or discontinuities or if wave propagation is simulated at too low resolution of the spatial grid. The effects of undersampling on periodic signals are called aliasing and are illustrated in the Appendix. Signal processing strategies like the comb lter which is also described there could be used to smoothen any time or spatial signal before differencing is applied. This can assure that the chosen grid resolution is adequate and it can help to make the interpolation stencils smaller which means that fewer points are involved in getting one interpolated value.

84

Determining Error Order An important tool to estimate the accuracy of a differencing scheme has already been shown. It is the Fourier series decomposition of all involved function points except the one which is to be determined or the derivative of which is to be approximated. This method shows the degree of accuracy. This means that it shows the maximum polynomial degree which can be approximated exactly. If a cubic polynomial can by exactly represented by a certain differencing scheme, but a fourth order one can only be approximated with a certain remaining error, then this scheme is said to be third order accurate. Its error term will contain fourth and higher order derivatives and the exponents n of the grid spacing terms (x)n will be three and higher. It can generally be written as = O((x)3 , f (4) ) + O((x)4 , f (5) ) + . . . . As an example let us consider the popular leapfrog scheme which is the central difference with second order accuracy. It got its name from the fact that it is continuously jumping between the odd and the even samples of the tabulated function (which can indeed become a problem if even and odd solutions decouple and diverge!). Let us apply the Taylor expansion given by equations 2.43 and 2.44 to the differencing scheme fk + = We obtain fk+1 fk1 = fk + fk (x) + fk (x) 2! 2 = fk + fk (x) + fk (x) 2!

2

fk+1 fk1 . 2x

3

(2.65)

(x)2 6

fk + with =

= 0 + O((x)2 , f

fk ).

fk

+ ... (2.66)

85

2. Numerical Concepts

The second order parabolic curve y(x) = ax2 + b and y (x) = 2ax can exactly be represented by the scheme: (2x)y (k) y(k + x) y(k x) (2x)(2ak) (a(k + x)2 + b) (a(k x)2 + b) 4akx a 2kx a (2kx) = 4akx (2.67) A nite difference scheme is said to be consistent, if the truncation error (x) 0 as x 0. In the case of partial differences the truncation error (t, x1 , x2 , x3 , . . . ) 0 as all discretization steps go to zero independent from each other. An example for a scheme which is not unconditionally consistent can be found in the book [Cha02]. It is the application of the so called DuFort-Frankel scheme to the thermal diffusion equation 2.13. Basically it consists of the central leapfrog scheme for the time derivative and of the simplest central scheme for the second derivative, thus second order accurate in time and third order accurate in space. The difference of the DuFort-Frankel scheme is that time and space differences are mixed in a way to average the central value used by the spatial differencing over time. D(un (un+1 + un1 ) + un ) un+1 un1 i+1 i1 i i i i = . 2t x2

(2.68)

Its error term contains the factor (t)2 /(x)2 which means that it will not go to zero independently of how the other spacings go to zero. If x goes to zero faster than t then the error will increase rather than decrease. If t = Kx then the factor will be K2 regardless how small the deltas are made. The DuFort-Frankel scheme is consistent with the diffusion equation only for t = xk for k > 1. The error order is important information but it does not help to determine a suitable grid resolution for an actual problem. For more specic analysis another method must be used.

86

Fourier Error Analysis Any periodic function can be decomposed into its Fourier components. The differential quotient of this periodic function is the sum of the differential quotients of its components. The nth Fourier component of any function f (x) has the form An ejkn x which can easily be differentiated yielding An jkn ejkn x . The idea of the Fourier error analysis is to determine the approximation error as a function of the wave number k of any Fourier component of any interpolated periodic function. If this is done with the leapfrog scheme from above we can derive: ejk(x+x) ejk(xx ) 2x jkx (ejkx ejk(x) ) e 2x 2j sin(kx) 2jkx sin(kx) 1 kx

r

= = = =

(2.69)

If the logarithm of the result, which is the relative error r of the scheme, is plotted over the number of sampling points per wave length PP WL = 2P i/kx the Figure B.2 in Appendix B.4 can be obtained. In this Figure the accuracy of the second order leapfrog scheme can be compared with the accuracy of higher order schemes. In order to obtain a maximum relative error better than 103 = 0.1 % the leapfrog scheme requires about 80 points per wave length. Enlarging the stencil by not more than one point on both sides (fk2 , fk1 , fk , fk+1 , fk+2 ) the fourth order scheme can be obtained which requires not more than 15 points per wave length to achieve the same accuracy. Doubling the order once more reduces the required resolution to about 5 points per wave length.

87

2. Numerical Concepts

If we are interested in what the approximation error means in terms of wave propagation speed or wave dispersion, then we can look at the matching between the wave number k and a distorted wave number k for which the relation (2.70) jkejkx + = jk ejkx holds. With this assumption the equation sequence 2.69 would look like jk ejkx = ejk(x+x) ejk(xx) 2x ejkx (ejkx ejk(x) ) jk ejkx = 2x 2j sin(kx) jk = 2x k x = sin(kx)

(2.71)

Figure 2.9 shows this result for the rst derivatives approximated by central differences of the orders 2,4,6,8,10,12. The grey curve has been e calculated for the implicit 4th order Pad formula of Equation 2.42 according to jk ejk(xx) + 4jk ejkx + jk ejk(x+x) = 3ejk(x+x 3ejk(xx x 3ejkx (ejkx ejk(x) ) jk ejkx (ejk(x) + 4 + ejkx ) = x 3 2j sin(kx) jk x = 4 + 2 cos(kx) 3 sin(kx) (2.72) k x = 2 + cos(kx)

There is a theorem, which has not been proved for the general non-linear case but is assumed to be true for properly posed initial value problems

88

2.5

1.5

0.5

0.5

1.5

2.5

3 2 1

89

2. Numerical Concepts

with consistent nite difference schemes, stating that stability is a necessary and sufcient condition for convergence. Convergence means here that the difference between the true exact solution and the discrete approximation goes to zero as time and spatial increments go to zero. Stability means that the approximation error does not boundlessly increase from step to step, which means its magnitude must not be amplied by a factor greater than one. As we have already seen in our simple examples, stability is not necessarily the case. There are many types of simulation problems which are unstable, meaning the approximation error boundlessly increases from step to step, as soon as the grid distance exceeds a certain threshold. There are other problems which are unstable regardless of how small a grid distance is chosen. It is therefore essential to have some methodology in order to predict if a certain differencing scheme is stable in conjunction with a certain type of differential equation. If it is not possible to prove stability analytically then the approximation error must be continuously monitored during time stepping. If it rises above a certain threshold some measures must be taken, like decreasing the stepping distance in time or space, which will help if the scheme is consistent with the problem. This strategy is followed especially in complex non-linear problems where analytic stability analysis is difcult or impossible. The approximation error can be estimated from the sensitivity of the evaluation to a variation in step size. If the result does not change much, then not much further improvement can be expected from another increase in resolution. Dynamic stepping is often connected with the Runge-Kutta algorithm, which has the advantage that an error estimate can be easily obtained and that it does not need any information about past time steps, so it starts by itself and the time step can be adjusted as often as necessary for no extra cost. The Runge-Kutta time stepping algorithm will be covered in Section

90

2.2.9 on page 108. It is for instance implemented in the famous electrical circuit simulator SPICE, originally developed at Berkeley University. Error Amplication Explicit schemes are often referred to as equations in update form, because the values can be updated without much effort time step by time step without having to solve any implicit equation system. In such explicit schemes with one independent variable stability can be analyzed directly from its denition. The amplication of the approximation error must not be greater than one when the scheme is updated by computing a new generation of values. The types of update schemes which can be analyzed using this method have the form un+1 = un + fn t + O[(t)2 ] T (un ). When the approximation error is included in the scheme we obtain un+1 + en+1 = un + en + f (un + en )t = T (un + en ). (2.74) (2.73)

The transformation T is assumed to be well behaved and the error e to be small. So we can develop T at un to separate the error: T (un + en ) T (un ) + en T (un ). (2.75)

Knowing that T (un ) was dened to be un+1 we can obtain the error amplication factor G = T (un ) by reformulating Equation 2.74 with substitution of Equation 2.75: en+1 = en T (un ) en G. (2.76)

If we are dealing with vectors u, e and f , we will need matrices T and G. In that case the stability condition has to be stated for the amplication matrix G. It says that all eigenvalues of G must be inside of the unit circle.

91

2. Numerical Concepts

As an example let us consider the harmonic oscillator d2 x 2 = 0 x, dt2 (2.77)

which can be transformed into a rst order differential equation system 2 by substituting dx/dt = v and dv/dt = 0 x. We can write the matrix formulation du = L u, dt with u x v and L = 0 1 2 0 0 (2.78)

The rst order forward difference quotient (un+1 un )/t, approximating the rst derivative, applied to a rst order ordinary differential equation creates an explicit scheme, which is also known as the Euler-Cauchy scheme. It can be written as un+1 = un + fn t + O[(t)2 ] (2.79)

This scheme introduced into the matrix formulation of the harmonic oscillator yields, as can be easily proved, the vector relation un+1 = [I + tL] un T (un ). The amplication matrix G results to G dT (u) du = I + tL. (2.81) (2.80)

un

The eigenvalues of G are g1,2 = 1 j0 t with the magnitudes |g1,2 | = 1 + (0 t)2 (2.82)

Regardless how small we choose t, the eigenvalues will never be inside the unit circle. The Euler-Cauchy formula will never be stable with the harmonic oscillator.

92

Fourier Decomposition Partial differential equations require a different strategy. It is related to the Fourier error analysis covered in Section 2.2.6 on page 87. It is usually referred to as Von Neumann analysis and can be applied to explicit as well as to implicit linear or linearized partial differential equations. The occurring signals are replaced by their spatial Fourier decomposition and developed at the calculation locus. If a general transformation rule for all Fourier components can be found, a necessary stability condition is that its magnitude must be less or equal to one in all cases. Consider our diffusion example from Equation 2.13. The transformation rule to obtain a new generation of function values from previous generation data samples can be found as Equation 2.16. This explicit difference equation is sometimes referred to as diffusion equation in update form. It is repeated here: un+1 = (1 2)un + (un + un ) i i1 i+1 i (2.83)

If we decompose the discrete function ui at time n+1 and at time n into its spatial Fourier components developed around xi and if we can relate these components at time n + 1 and at time n by means of a set of spectral transformation rules gk then stability depends on the amplication of the gk . If their magnitude is less or equal to one for all k then the scheme is stable and will converge if it is consistent. n Knowing that un = k Uk ejkxi with xi = ix we can decompose i our equation into a relation for the Fourier components:

n n n n gk Uk ejkxi = (1 2)Uk ejkxi + (Uk ejkxi1 + Uk ejkxi+1 )

(2.84)

n In the next step the common factor Uk ejkxi can be pulled out and cancelled. We obtain:

gk = (1 2) + (ejk(x) + ejkx )

(2.85)

93

2. Numerical Concepts

direction in the complex plane. Their imaginary parts j sin(kx) therefore cancel out for all kx. What remains is our transformation rule gk = 1 2 + 2 cos(kx) = 1 2(1 cos(kx)). (2.86)

The scheme is unstable if the magnitude of gk is greater than one which is the case for > 0.5.

0.1 0.075 0.05 0.025 -0.025 -0.05 -0.075 -0.1 20 40 60 80 100 0.1 0.075 0.05 0.025 -0.025 -0.05 -0.075 -0.1 20 40 60 80 100 0.1 0.075 0.05 0.025 -0.025 -0.05 -0.075 -0.1 20 40 60 80 100

Now let us deal with the wave propagation example from Equation 2.17 which did not work with the explicit scheme. It seemed to be stable with the implicit scheme, so let us analyze this one rst. This is also a demonstration for the applicability of the Von Neumann analysis to implicit schemes. The starting point is the implicit difference equation which was already given as Equation 2.24. The constant k was replaced by because k is needed here as the index of the spectral component. un+1 + un+1 + un+1 = un . i i1 i i+1 (2.87)

94

Introducing the spectral transformation gk for the transition of the kth spectral component of the function f from time nt to time (n + 1)t we can write

n n n n gk Uk ejkxi1 + gk Uk ejkxi + gk Uk ejkxi+1 = Uk ejkxi

(2.88) (2.89)

which is always inside the unit circle. Therefore the given implicit scheme for the linear convection equation seems to be unconditionally stable. A stronger statement cannot be made on the basis of Von Neumann analysis because the |gk | 1 condition is a necessary but not sufcient condition. If a reader wants to delve deeper into the vast eld of stability, the book Fundamentals of Computational Fluid Dynamics by Lomax, Pulliam and Zingg [LP01] can be warmly recommended. Let us compare the explicit scheme now. The substitution of the Fourier components into the update form of the original scheme 2.21 un+1 = un k(un un ) i i+1 i1 i yields

n n n n gk Uk ejkxi = Uk ejkxi k(Uk ejkxi+1 Uk ejkxi1 )

(2.91)

(2.92)

and after cancelling the common factors gk = 1 k(ejkx ejk(x) ) which is nally gk = 1 2kj sin(kx). (2.94) It is quite obvious that |gk | > 1 for all kx so the scheme must be unstable. (2.93)

95

2. Numerical Concepts

The results of the very similar explicit scheme given by Equation 2.95 are shown in Figure 2.10. They are simply perfect no attenuation, no dispersion and no spurious artifacts! The scheme is not at all special, it is even simpler than the rst one, because the second order central difference in space has been replaced by a rst order backward difference. Why? Let us analyze stability. The difference equation in update form is un+1 = un 2k(un un ). i i i1 i (2.95)

Its stability analysis according to Von Neumann can be performed according to:

n n n n gk Uk ejkxi = Uk ejkxi 2k(Uk ejkxi Uk ejkxi1 )

gk = 1 2k(1 ejk(x) ) gk = 1 2k + 2k cos(kx) + j2k sin(kx). For simplicity we have chosen k = 0.5 which lets remain only the trigonometric part describing the unit circle in the complex plane. If k > 0.5 then the scheme is unstable, the circle bigger than unit. For k < 0.5 it is stable, too. The circles are smaller than the unit circle but wholly contained inside it touching the unit circle at the rightmost point. But now the scheme looses the property of perfectness. It introduces attenuation, dispersion and spurious effects as we would expect for a rst order scheme. Why is there a perfect case? An analysis of the truncation error by means of the methods already discussed would show that the scheme is rst order accurate if k = 0.5. For k = 0.5 the second order error term becomes zero indicating an accuracy better than one. In fact it can even be shown that all higher order error terms cancel out, making the scheme exact. The explanation for the intriguing fact is that with the actual convection speed it takes exactly one t to shift a value downstream the distance x. By looking at the adjacent upwind sample the update algorithm quasi

96

reads the future and can therefore predict it correctly. The steady ow makes that possible. If c was negative, a forward differencing scheme would be adequate to retain the upwind principle because the ow would change its direction. This example illustrates in a nice way how the knowledge about the underlying physics can signicantly improve the approximation process. More details about this technique and how to apply it to complex and realistic aero-acoustical problems can be found in [Tho96].

Now it is time to think about solving the equations systems which are created by implicit differencing schemes or interpolation methods like the Pad formulations. e Any linear equation system for N unknownsx = (x1 , x2 , x3 , . . . , xN )T must consist of N equations in order to be determined properly. The matrix containing all coefcients must therefore be N N quadratic. The linear equation system can be written in matrix notation Ax = b. (2.96)

In order to solve this system the matrix A must be made triangular with all values below the main diagonal being zero. In other words the equation system a11 a12 a22 a32 . . . a13 a23 a33 . . . ... ... ... .. . a1N x1 b1

a 21 a31 . . .

a1N a1N . . . aN N

x2 x3 = . . . xN

b2 b3 . . . bN

(2.97)

aN 1 aN 2 aN 3 . . .

97

2. Numerical Concepts

must be transformed into an equivalent equation system x1 b1 a11 a12 a13 . . . a1N x b 0 a a1N 2 2 22 a23 . . . 0 a33 . . . a1N x3 = b3 0 . . . . . .. . . . . . . . . . . . . . . 0 0 0 ... aN N xN bN

(2.98)

Gaussian Elimination The last row of the new system can easily be made explicit yielding xN = bN /aN N . This now known value can be substituted into all other equations. This results in an updated vector b and an eliminated last column of A . Now the row N 1 is explicit and the next column can be eliminated. This process is sometimes referred to as back substitution and is the second part of the task. During the rst part the zeros in the lower left triangle have to be created. Swapping two rows of the matrix A does not change the solution vector x if the same rows are simultaneously changed in the vector b. Moreover, it is allowed to replace one line by a linear combination of it with other lines. Swapping some columns of the matrix A is also permissible if the corresponding rows of the solution vector x are swapped accordingly. All these manipulations can be used to create the required zeros in the left triangle of the coefcient matrix A. Usually this is done systematically starting with the rst column according to the following algorithm: 1. Find the biggest entry in column one. 2. Swap this row with the original row one to have the biggest element on top. 3. Subtract a proper multiple of row one from all other rows to make the rst column values zero.

98

4. Go to the second row now and repeat these steps for the second column 5. Proceed accordingly until the lower triangle has been zeroed By sorting up the biggest leftmost coefcient, partial pivoting has already been performed. Pivoting is a strategy to keep the numerical roundoff error as small as possible. Complete pivoting would also require sorting of columns to avoid adding or subtracting numbers with too different orders of magnitude. The described algorithm is called Gaussian elimination and is very exhaustive. It can also be used to nd the inverse of a matrix. According to (2.99) A A1 = I we can do Gaussian eliminations for all columns of A1 using the corresponding columns of matrix I as vectors b. If this is done simultaneously for all vectors b, the triangulation has to be done only once. Just the back substitution has to be repeated for all columns of A1. LU-Decomposition A much more efcient and modern approach is called LU-Decomposition and was published by Banachiewicz, Cholesky and Crout. This algorithm tries to split the matrix A into a product of two triangular matrices L U with L being a lower triangular matrix with zero coefcients above the main diagonal and U being an upper triangular matrix with zero coefcients below the main diagonal. The idea was to split the original problem A x = L (U x) = b into the two problems Ly =b and U x = y. (2.102) (2.101) (2.100)

99

2. Numerical Concepts

Because of the triangular shapes of the matrices L and U it is straight forward to solve for y. Once y is known it is just as simple to determine x, actually this is equivalent to the nal back substitution step of the Gaussian elimination procedure. Finding the L and U matrices is not too difcult either. 2 N2 coefcients are to be determined, N2 N of them an upper and a lower triangle excluding the main diagonal have to be zero. From the remaining N 2 + N coefcients we can choose N because the system is underdetermined. We can set the main diagonal coefcients of the matrix L to one. The algorithm given by Crout is:

for j {

1 = = =

1 ujj i1 k=1

lik ukj

j1 k=1

aij

lik ukj

As a side product the determinant of matrix A can easily be calculated as the product of the main diagonal entries of matrix U. The main advantage of the LU-decomposition is the efciency of recalculating the equation system several times for different vectors b. While the Gaussian elimination would start from scratch with each update of the constant vector b, the LU-decomposition can determine a new vector x by one forward and one back substitution using the unchanged diagonal matrices L and U. When the inverse of the matrix A is to be calculated, a reevaluation

100

for each column of A1 with a corresponding column of the unit matrix I is required. This is already an application where the advantages of the LU-decomposition take effect. It has to be noted that in order to write competitive program code, the introduction to that subject which was given here is not sufcient. Numerical issues will become apparent and pivoting will have to be introduced, too. Pivoting is much more difcult to implement for the LU-decomposition algorithm than it is for the Gaussian elimination. For actual programs the use of professional mathematical libraries or more specialized textbooks is strongly encouraged. Banded Matrices For us a most important subsection of the linear algebra is the solution of equation systems described by diagonal, tridiagonal or, even more generally, banded matrices. Up to now all implicit difference schemes and all implicit interpolation schemes which have been mentioned lead to banded matrices. It makes a huge difference if we are dealing with a banded matrix or if we have to deal with a general square matrix. The latter has N N elements, while a tridiagonal matrix has 3 N 2 elements. Banded matrices can be solved by a recursive approach. In a forward traversal of the main diagonal of the matrix A a vector of N temporary intermediate results is created, which is then used in a backward traversal step to calculate all the unknown function values. The matrix A does not have to exist. Usually just the diagonal entries will have to be stored, which means memory for e.g. 3 N numbers has to be provided for tridiagonal matrices. If the bands consist of repeated, shifted constant groups, f.e. (1, 4, 1) as in the spline case or (k, 1, k) as in our implicit example, then all the memory can be saved to be allocated for the matrix representation. If the vector for the result is used to keep half of the intermediate results during the forward pass, then extra memory for only N intermediate numbers has to be provided. For the case of a tridiagonal matrix an often applied approach is the

101

2. Numerical Concepts

Thomas algorithm. It includes pivoting and is therefore practically applicable. The direction of the two sweeps and the order of the operations has been carefully optimized in order to minimize storage requirements and computation overhead. It assumes that the three diagonals of matrix A are stored in the vectors q1..N , representing the main diagonal, r1..N 1 , representing the upper diagonal and p2..N , representing the lower diagonal and that the elements p1 and rN have been initialized to zero in order to smoothen the algorithm. b1 b2 b3 b4 = b 5 . . . . . . . pN1 qN1 rN1 . . ... 0 pN qN bN (2.103) The vector b1..N can be overwritten by b during the rst pass and may keep the solution x after the second pass. The vector r can be overwritten by the intermediate result . r A 0th element which is initially reset to zero has to be prepended to the vectors and b. The recurrence formulae for the upward sweep are: r q1 r1 0 0 p2 q2 r2 0 0 p3 q3 r3 0 0 p4 q 4 0 0 0 p5 . . . . . . . . . . . . 0 0 0 0 0 0 0 0 ... ... ... r4 q5 .. . 0 0 0 0 r5 .. . 0 0 0 0 0 .. . 0 0 0 0 0 . . . x1 x2 x3 x4 x5 . . . . . . xN ri = ri qi pi ri1 i = bi pi bi1 b qi pi ri1 i = 1 . . . N. (2.104)

The diagonal term is used as a pivot and normalized to one. The condition qi pi ri1 = 0 must be fullled for all i. This is the ensured if

102

the matrix is diagonally dominanated i.e. |qi | |pi | + |ri |. (2.105)

Similar formulae can be derived for 5-diagonal, 7-diagonal or 9-diagonal matrices, which are created when higher order implicit schemes are applied. Iterative Methods Another popular approach for solving sparse matrices, matrices containing many zero elements like banded matrices, is the iterative approach. It can also be applied to general matrices, especially if they are stiff (coefcient spread over many orders of magnitude) or close to being singular. In such cases where pivoting is difcult or ineffective, the accuracy of an available solution can be increased by an iterative approach until the remaining error is below a given threshold. While all the methods described above yield theoretically exact results and are practically limited in accuracy only by the available numerical precision and the associated truncation error propagation, iterative methods do not even pretend to be exact. By denition they provide approximations to the exact solution with a remaining error which tends to become zero with an increasing number of iteration steps. In the general case about N3 operations are necessary to solve a N N matrix exactly. The commonly used iterative algorithms need only N2 mathematical operations per iteration step and are therefore much faster for big matrices than exact methods. With x being the exact solution of Equation 2.96 and x being an estimate with error E (2.106) x x + E, substitution of x into 2.96 yields AE=bAx, (2.107)

103

2. Numerical Concepts

which is already an iteration instruction, because after solving for the error E a new and theoretically even exact estimate forx can be obtained. Once the LU-decomposition of A is known, each iteration step mainly consists of a back substitution and a matrix / vector product, both operations with a complexity N 2 . The approach described by Equation 2.107 theoretically yields the exact solution after only one iteration step, regardless how accurate or wrong the original estimate has been. This is not quite the idea behind iterative approaches it is not necessary to get the exact result after one step. If we can ensure that the iteration result converges to the exact result of the original equation system, we can exchange the matrix A by a matrix B which is similar but easier to handle. Actually we have done this already by mentioning the LU-decomposition in the description above. We can be sure that the result of the LU-decomposition will not be exactly equivalent to the original matrix as soon as we implement the algorithm on a real computer with nite numerical precision. But there are still simpler methods to get an admittedly less accurate approximation for A1, which still provide convergence to the accurate solution x. Convergence to the solution of A x = b can be ensured for the recurrence formula B (xi+1 xi ) = b A xi or explicitly xi+1 = B1 b + B1 (B A) xi (2.109) if and only if all eigenvalues of B1 (B A) are inside the unit circle. The different iterative methods make different choices of the matrix B. The Method of Jacobi picks out only the main diagonal elements of matrix A in order to obtain a pure diagonal matrix D, which can be inverted trivially. The recurrence formula is given by D xi+1 = b + (D A) xi . (2.110) (2.108)

104

The convergence rate is determined by the spectral radius which is the eigenvalue with the biggest magnitude max of the matrix B1 (B A) in this case D1 (D A) according to r |xi+1 xi | |max 1| |xi x| (2.111)

The method of Jacobi works best for diagonally dominated systems, but even then it does not converge too fast. Faster convergence can be achieved by the Gauss-Seidel-Relaxation Method. The difference to the Jacobi method is that together with the main diagonal elements of the matrix A the elements below the main diagonal are also preserved in the matrix B. The iteration step is now a bit more complicated because an implicit system has to be solved. But because of the triangular shape of B just the nal back substitution step of a Gaussian elimination procedure is required. It can be shown that the convergence rate rG is much better than that of Jacobi rJ . The spectral radius max,G is related to that of Jacobi by max,G = 2 max,J which makes the convergence rate to rG |2 max,J 1|. (2.112)

The Gauss-Seidel method can be improved in terms of convergence speed by overcorrecting the intermediate solution into the predicted direction. The improved estimate x is calculated from the Gauss-Seidel i+1 estimate xi+1 according to x = xi (1 + (xi+1 xi )) i+1 (2.113)

with the relaxation parameter . This method is referred to as Successive Over-Relaxation Method (SOR). The optimum value of can be found in terms of the Jacobi radius J according to opt = 2 1+ 1 2 j . (2.114)

105

2. Numerical Concepts

The SOR method exhibits overshooting, in order to converge in a minimum number of iteration steps. Therefore it can happen, that the error temporarily even increases, especially if is signicantly greater than one. To avoid this, several strategies have been published proposing a steady acceleration of the convergence by increasing starting at unity and asymptotically reaching the optimum value (Tschebyscheff Acceleration).

log(Err) # Iterations

200 -2.5 -5 -7.5 400 600 800 1000

Jacobi

-10 -12.5 -15 -17.5

Block Iterative Methods are dividing unknowns in blocks or lines which are solved by an implicit algorithm like the Thomas algorithm. The Alternating Direction Implicit (ADI) method is specially well suited for multidimensional discretizations of the Poisson or Laplace equations where it converges very well. In a two-dimensional problem, the solution vector x actually represents the n rows and n columns of the two-dimensional mesh X, which can be attened either row by row or column by column. A central second order

106

discretization of e.g. the Poisson equation applied to both directions leads to a sparse matrix, which is basically tridiagonal but contains two other diagonals n positions away from the main diagonal. If the horizontal and vertical discretizations are separated, then two tridiagonal problems can be obtained, which can be iterated alternatingly: Ar u + Ac v = b. (2.115)

This is achieved by introducing a row vector u and a column vector v both of length N = n n, u being a concatenation of all n rows and v being a concatenation of all n columns of X. We dene a transformation matrix T by v = T u, which is a sparse matrix consisting of only ones and zeros. It interprets the vector u as concatenation of rows and reorders the elements in order to create the vector v as concatenation of columns. The relation of Equation 2.115 can be iterated in semi-steps in alternating direction. The recurrence formulas for the two semi-steps can be written using T as un+1/2 = (1 )un + (A1 (b T (Ac (T un )))) r (2.116)

1 un+1 = (1 )un+1/2 + (T (Ac T (b Ar un+1/2 ))) (2.117) 1 The inverse matrices Ar and A1 can easily be calculated because they c are tridiagonal and the Thomas algorithm can be applied. They have to be computed only once before the iteration is started. The ordering operation T does not really take time or memory. In a computer program it will be implemented implicitly, simply by the way how the array elements are addressed. The parameter is difcult to determine analytically. For some typical problems an optimum value is known. For the two dimensional Poissonequation 2u 2u + 2 = (x, y), (2.118) x2 y

which has been used as a benchmark problem in the performance comparison shown in Figure 2.11, it was determined experimentally as 0.971.

107

2. Numerical Concepts

The benchmark was based on second order central differences in both directions according to ui,j+1 2ui,j + ui,j1 ui+1,j 2ui,j + ui1,j + = (x, y) (2.119) (x)2 (y)2 with i = 1, 2, . . . n and j = 1, 2, . . . n. N = n n was chosen to be 64. The convergence rate has been calculated for the Jacobi method as 0.03, for the Gauss-Seidel method as 0.06, for the SOR algorithm as 0.4. The implicit versions of the ADI formulas given in equations 2.116 and 2.117 for the benchmark Equation 2.118 are ui,j

n+1/2

= un + i,j +

ui+1,j 2ui,j +ui1,j (x)2

n+1/2 n+1/2 n+1/2

+ (x, y) (2.120)

for step one and un+1 = ui,j i,j + for step two.

n+1/2

+

n+1/2 n+1/2

n+1/2

+ (x, y) (2.121)

Basically there is not a big difference between spatial derivatives and time derivatives. Both could be approximated by various differencing and interpolation schemes which have been discussed so far. However, higher order schemes using bigger stencils, which are working well for spatial differencing, will often be too expensive in the time-domain because they lead to heavily implicit equations needing future values that are not yet known, or past values which have to be stored and which are difcult to initialize during startup.

108

Starting a simulation is often an issue, because as stencils are reaching boundaries they have to be replaced by smaller stencils thus avoiding the need of unavailable data. And the start of a simulation is denitely a boundary for the time marching algorithm. The same applies when the simulation time step is being changed on the y. New boundary conditions are created by this situation and time stepping has to restart. Modern algorithms for simulating highly non-linear systems require a change of the step width very often, so self-starting time stepping algorithms are very important. Some examples for time stepping have already been given in previous sections. A summary of useful linear multi-step methods for rst order ordinary differential equations is given below. The indicated names correspond to those which are commonly used in literature. Multi-Step Methods Partial differential equations become ordinary differential equations after the spatial derivatives have been resolved by the application of some differencing scheme. Ordinary differential equations of second order can be transformed into rst order systems, so the methods listed below will usually be sufcient to deal with most types of physical problems. The expressions given below can be used for actual problems by spatially discretizing the differential equations that are to be integrated. This way time-continuous ordinary differential equations can be obtained that can be made explicit. Now formulas for the time derivatives exist and can be used to substitute the derivatives un , un1 . . . by corresponding expressions usually involving function values ui , ui . . . at some spatial n n1 positions i and possibly other functions and variables. Applying the leapfrog scheme to the linear convection example from Equation 2.17, the explicit and spatially discretized formulation un = c(ui+1 ui1 )/(2x) n n (2.122)

could be used for substituting un or correspondingly un1 and un2 in any one of the time stepping schemes below.

109

2. Numerical Concepts

Explicit Methods Explicit methods are often called update formulas because the array holding a generation of function values can be updated by direct numerical evaluation of the explicit expression which depends on previously calculated function values only. Explicit Euler un+1 = un + tun Leapfrog un+1 = un1 + 2tun Adams-Bashforth, Order 2 un+1 = un + Adams-Bashforth, Order 3 un+1 = un + Implicit Methods Implicit differencing schemes can be used by means of the equation solving algorithms outlined in the linear algebra section above. Implicit Euler un+1 = un + tun+1 Trapezoidal (Adams-Moulton, Order 2) un+1 = un + Backward Difference, Order 2 1 un+1 = (4un un1 + 2tun+1 ) 3 (2.129) t (u + un+1 ) 2 n (2.128) (2.127) t (23un 16un1 + 5un2 ) 12 (2.126) t (3un un1 ) 2 (2.125) (2.124) (2.123)

110

Adams-Moulton, Order 3 un+1 = un + Generalization All examples mentioned above can be represented by a common formula for two-step linear time quantization schemes. It was given together with the table below in [LP01]: (1+)un+1 = (1+2)un un1 +t(un+1 +(1 )un un1 ) (2.131) Depending on the actual values of the coefcients , and various methods can be derived. Some of them are unconditionally stable (Astable). They are marked with a +: 0 1 1/2 1 3/4 1/3 1/2 5/9 0 0 0 1/3 5/12 1/6 0 0 0 1/2 0 1/2 1/2 1/6 1/2 0 5/6 1/6 0 1/2 0 0 0 0 1/4 1/3 1/2 2/9 0 1/2 1/3 0 1/12 1/6 Method Explicit Euler Implicit Euler Trapezoidal (Adams-Moulton) Backward 2nd order Adams type Lees type Two-step trapezoidal A-contractive Leapfrog Adams-Bashforth-2 Most accurate explicit Third-order implicit Adams-Moulton-3 Milne Order 1 1 2 2 2 2 2 2 2 2 3 3 3 4 St. + + + + + + + t (5un+1 + 8un un1 ) 12 (2.130)

111

2. Numerical Concepts

proximation error. It contains a second order error term proportional to (t)2 (2 + 2 2 1), a third order error term proportional to (t)3 (3 3 + 1) and a fourth order error term proportional to (t)4 (4 + 4 2 1). (2.134) (2.133) (2.132)

Second order accurate schemes must therefore meet the requirement 1 =+ . 2 For third order accurate schemes the additional relation = 2 5 6 (2.136) (2.135)

must hold. Enforcing all of these terms to become zero yields the only fourth order accurate scheme of Milne. Predictor-Corrector Methods As already discussed in Section 2.2.9, depending on the spatial differencing scheme which is applied to a partial differential equation, the time derivative un at time nt is usually represented by a weighted sum of some spatially adjacent values of ui which are sampled at the same time. n It can be calculated only if all required values ui are already known. n The Euler formula un+1 = un + un t is a simple but typical explicit two-step time-marching scheme based on one-sided rst order differencing. It implicitly assumes that the derivative un does not change between nt and (n + 1)t. In the context of a spatially discretized differential equation, like the one described by Equation 2.122, this assumption is not justied.

112

Higher order time stepping methods will have to employ a more accurate approximation algorithm. Memory efciency does not allow to store several generations of past values for all the many lattice nodes of a multi-dimensional simulation, so simple higher order backward differencing schemes cannot be applied. The solution to this problem is the predictor-corrector concept. The predictor, an explicit scheme, is used to create estimates for future function values un+1 using a backward interpolation scheme for extrapolation into the time range beyond the last known value un . In a second step the corrector can evaluate originally implicit expressions involving u n+1 and un+1 based on the estimates delivered by the predictor. This two step scheme has a better overall accuracy and error order than the predictor and is still explicit. The improved Euler-Cauchy method is used to demonstrate this concept. According to the rst order Euler-Cauchy formula an estimate for the new function value un+1 is calculated un+1 = un + un t. (2.137)

Now the expression for u derived from the spatially discretized differen tial equation can be evaluated at time tn+1 using the predicted value un+1 and an estimate for un+1 can be obtained. By means of the corrector for mula t (u + un+1 ) (2.138) un+1 = un + 2 n an improved result can be obtained. The effect of the corrector can be interpreted as replacing the one-sided approximation by a more balanced one, thus introducing the trapezoidal rule into the time stepping algorithm. In other words, the difference quotient (un+1 un )/t is now related to the average of a known old value and a roughly estimated new value of the time derivative (un + un+1 )/2 rather than an old one alone. Expanding this concept towards higher orders requires the general interpolation formula 2.35, which can be used to extrapolate the function u beyond the last known value at tn . This extrapolation can be averaged

113

2. Numerical Concepts

between tn and tn+1 by formally integrating the polynomial interpolation formula over that interval and dividing the result by the time step. The resulting equation is a general explicit predictor of higher order called Adams-Bashforth predictor. The two- and three-step versions have already been given as Equation 2.125 and 2.126. As an estimate for un+1 is available, the expression for u can now be evaluated in the corrector step at time tn+1 yielding a more accurate approximation for un+1 . Corrector formulas can be obtained from downward interpolation schemes which are applied at tn+1 . This time the polynomial expression is not used for extrapolation purpose. The interpolation range is now inside the range dened by known and estimated points and correspondingly more accurate. Commonly used corrector schemes are the Adams-Moulton schemes given in Equations 2.128 and 2.130. The un+1 terms have to be evaluated using the un+1 values delivered by a suitable predictor scheme. Useful predictor corrector combinations are shown in the following table. Predictor Euler-Cauchy Adams-Bashforth Adams-Bashforth Order 1 2 3 Corrector Adams-Moulton Adams-Moulton Adams-Moulton Order 2 3 4 Result 2 3 4

The fourth-order Adams-Moulton corrector is added below. As a three step method it was not included in the two-step tables above. un+1 = un + t (9un+1 + 19un 5un1 + un2 ) 24 (2.139)

The stability of predictor-corrector schemes is close to the stability of the implicit corrector. It cannot be calculated separately from the actual differential equation which is to be solved. Once this differential equation has been spatially discretized and reformulated into an explicit ordinary differential equation, as it was shown for the example in Equation 2.122, the Van Neumann analysis can be applied to the time and space discretization.

114

k The ui have to be replaced by k Un exp(jkxi ) with xi = ix and the n spectral transformations gk have to be determined. Because the stability condition |gk | < 1 must be valid for all k simultaneously the sum is not written and all equations are formulated for a single spectral line k.

0.2

2

0.2

0.15

0.15

0.1

0.1

0.05

0.05

-1

-0.5

0.5

-0.05

-0.05

-1

-0.1

-0.1

-0.15

-0.15

-2

-0.2

-0.2

(a) x = 1c = 2

The starting point is the spatially discretized differential equation and its spectral counterpart

i un =

c i+1 (u ui1 ) n 2x n

(2.140)

115

2. Numerical Concepts

c k U 2j sin(kx). (2.141) 2x n The Euler-Cauchy predictor for the actual equation and its spectral counterpart are ct i+1 (u ui1 ). (2.142) ui n+1 = ui n n 2x n ct k 2j sin(kx)). (2.143) uk = Un (1 n+1 2x The second order corrector formula is

k un = i ui n+1 = un +

t i (u + un+1 ), i 2 n

(2.144)

i i which requires the derivatives un and un+1 . The former is known from our original equation, the latter has to be calculated. We differentiate the predictor formula, substitute Equation 2.140 at i + 1 and at i 1 and we obtain for time and spectral domain

un+1 = i

(2.145) (2.146)

k un+1 = Un k

Now we can substitute and solve. With un+1 = gk un we obtain gk = 1 c2 (t)2 c2 (t)2 ct sin(kx). (2.147) + cos(2kx) j 2 2 4(x) 4(x) x

This result is not straightforward to interpret. With t = x and c = 2 we obtain a gk according to Figure 2.12(a). If we choose t = x/2 and c = 0.5 we get a nearly stable result according to Figure 2.12(b). The worst case magnitude is 1.0005 but on the other side there is no k for which |gk | < 1. The best we can do is to choose x = . This makes gk exactly to 1 for all k. The continuous curve for k = 0..10 is shown in Figure 2.12(c).

116

Runge-Kutta The Runge-Kutta method is a self-starting one-step method which increments time by two semi-steps. The second order accurate version uses a predictor generated semi-step un+1/2 in order to calculate the corrector for the function value un+1 . With h = t and f (tn , un ) = un being the right side of the explicit rst order ordinary differential equation the procedure is usually given as k1 = hf (tn , un ) h k1 k2 = hf (tn + , un + ) 2 2 un+1 = un + k2 (2.148) (2.149) (2.150)

The popular fourth-order Runge-Kutta uses two successive predictor/corrector evaluations. It is commonly given as k1 = hf (tn , un ) h k1 k2 = hf (tn + , un + ) 2 2 h k2 k3 = hf (tn + , un + ) 2 2 k4 = hf (tn + h, un + k3 ) 1 un+1 = un + (k1 + 2k2 + 2k3 + k4 ) 6 (2.151) (2.152) (2.153) (2.154) (2.155)

In our usual notation with predictor and corrector indications an equivalent formulation would be un+ 1

2

un+ 1

2

un+1

h u 2 n h = un + un+ 1 2 2 = un + hn+ 1 u = un +

2

un+1 = un +

117

2. Numerical Concepts

Actually this Runge-Kutta method combines an explicit Euler predictor, an implicit Euler corrector, an explicit Leapfrog predictor and an implicit Milne corrector. A general derivation and a prove of its fourth order accuracy are too ambitious to be given here. We will treat a simple example instead which demonstrates application and error order. If it is a fourth order method it must integrate up to fourth order polynomials exactly. Let us assume that u = t4 with u0 = t4 then our starting differential equation 0 can be written as (2.160) u (t) = 4t3 . Applying the Runge-Kutta method for the transition from u0 to u1 yields k1 = 4ht3 0 k2 k3 k4 h = 4h(t0 + )3 2 h 3 = 4h(t0 + ) 2 = 4h(t0 + h)3 . (2.161) (2.162) (2.163) (2.164)

Substituting, expanding and simplifying the result for the next time step yield ut0 +h = t4 + 4ht3 + 6h2 t2 + 4h3 t0 + h4 = (t0 + h)4 0 0 0 (2.165)

which is valid for any actual value of t0 and therefore general. A disadvantage of the Runge-Kutta method is, that four evaluations of the function f are required per time step. An advantage is that the approximation error can be estimated using the half-step result. Although it is always possible by substitution of variables to transform a second order ordinary differential equation into a system of rst order differential equations, it is sometimes efcient and advantageous to apply a suitable time stepping method directly. A fourth order accurate RungeKutta scheme for ordinary differential equations of second order which can be described by (2.166) u (t) = f (u(t), u (t)),

118

can be given according to Abramowitz [AS65] as k1 = f (un , un ) h k2 = f (un + un + 2 h k3 = f (un + un + 2 k4 un+1 un+1 (2.167) h2 k1 , un + (2.168) (2.169) (2.170) (2.171) (2.172)

A collection of more methods for second order differential equations can be found in the book Computational Physics by Vesely [Ves93]. Additional integration methods can also be found in the book Numerical Initial Value Problems in Ordinary Differential Equations by Gear [Gea71].

Solving non-linear differential equations does not make too much difference as far as explicit methods are concerned. The numerical procedure is the same but convergence can become an issue. In a consistent scheme accuracy can always be improved by decreasing the stepping distance. Non-linearities can change the sensitivity of the approximation error to the grid resolution signicantly. Therefore dynamic step-size monitoring and adjustment should be considered. For this a self-starting time stepping scheme like the Runge-Kutta scheme is required. If implicit schemes cannot be avoided, it is possible to linearize the equations by replacing the non-linear terms by a Taylor series centered around the last known value. This procedure is called local linearization because the operating point which is the center of the linearization is up-

119

2. Numerical Concepts

dated after each time step. If the Taylor expansion order matches the error order of the time marching scheme then no accuracy is lost. Consider the trapezoidal scheme applied to a non-linear rst order differential equation u (t) = F (u, t). If F (u, t) was not linearized the implicit expression could not be solved easily. So we replace F (u, t) by the Fourier series F (u, t)n = Fn + (t tn ) F F + (u un ) + O(h2 ) t u (2.173)

around time tn and apply the trapezoidal rule un+1 = un + We obtain un+1 = un + h 2 Fn + Fn + h F F + (un+1 un ) t n u n + O(h3 ) h (Fn+1 + Fn ) + h O(h2 ). 2 (2.174)

(2.175) which is accurate to second order just as the time stepping algorithm. By local linearization which is updated after each time step no accuracy is lost and a non-linear implicit scheme becomes solvable with the methods provided by linear algebra. In this simple example the implicit expression can even be made explicit by a simple division but this is not possible in the general case, especially when spatial differencing is involved. But there is another possibility for simplication. If Equation 2.175 is rearranged it becomes 1 h F un 2 u n = hFn + h2 F . 2 t n (2.176)

In many uid dynamical applications the non-linear function F is not explicitly depending on t. The partial derivative with respect to t is zero and the right hand side simplies to hFn . In general Equation 2.176 will contain a matrix-vector product A u on the left side and a vector hFn on the right side. A simple iteration procedure can be used to solve for u, hFn and A in turn.

120

Boundary value problems cannot be solved directly because initial values which are required to start a time marching process are not given. Other boundary conditions are given instead which cannot be processed initially because they are actually simulation output rather than input. In other words the simulation is expected to produce some specic output conditions after having run for some while but the start up condition of the simulation is not determined properly. A usual way to cope with this kind of problems is the so-called shooting method. Missing initial conditions are guessed somehow, the simulation is run until the boundary is reached and the results are stored. For a second shot the estimated initial conditions are modied by a small amount. After the second simulation the dependence of the resulting boundary values on the initial values can now be estimated in rst order and a NewtonRaphson strategy can be followed. The new shooting direction is computed as if the system was linear. This way the distance between the simulated boundary conditions will converge towards targeted ones. After each shot the direction will be adjusted until the required error measure is met. The shooting method can be formulated as recurrence formula ic

n n+1

= ic

n1

A1 en1

with Aij

(2.177)

n

bc being the boundary conditions reached by the nth iteration, ic being n the initial conditions of the nth iteration and en = bc bc being the remaining error of the boundary conditions achieved by the nth iteration.

Aero-acoustics, as already indicated, is very demanding in terms of accuracy and stability. Especially the fact that most numerical methods introduce numerical dissipation rendering these methods useless for simulation of sound propagation over longer distances.

121

2. Numerical Concepts

Two popular routes for devising highly accurate and dissipation-free nite difference methods are currently observable. One approach increases the order of spatial differencing schemes proposing large computational stencils which are, as already stated, difcult to handle when boundaries are approached. With spatially extended stencils higher-order accuracy can be achieved or matching between actual numerical behavior and that of the uid can be optimized. In the other approach more compact differencing is optimized. An interesting approach is the upwind leapfrog scheme which has been elaborated by Thomas [Tho96]. Another high-order scheme suitable for aero-acoustics has been proposed by Goodrich [Goo93]. He uses spatially extended meshes based on central differencing techniques. Tam [TW92] also proposes an extended stencil aiming at correctly representing the dispersion relation. Optimized Runge-Kutta methods exhibiting low dissipation and dispersion were proposed by Zingg [ZLH93] and Manthey et.al. [HHM95]. All these techniques offer excellent resolution for high frequency data but most of them require very small time-steps. Examples of compact differencing schemes are Leles [Lel91] Pad e approximation based techniques, and some recent publications by Davis [Dav95], Kim and Lee [KL96]. However, a typical drawback of compact methods is that they usually require expensive matrix inversions. A good overview about various nite differences techniques which are currently employed for solving uid dynamical differential equations can be found in Chattots introductory book Computational Aerodynamics and Fluid Dynamics [Cha02].

122

The Lattice Boltzmann method (LBM) is based on a specic discretization of the Boltzmann equation. It approximates the Navier-Stokes equations and incorporates uid dynamicsand acoustics intrinsically.

In order to discretize the six-dimensional phase space of the Boltzmann equation, the innite velocity space is simplied drastically. Instead of the innite number of possible directions and velocity magnitudes at each point in space only a handful directions and still fewer magnitudes are considered. Any set of lattice velocities ci has to satisfy specic symmetry and isotropy requirements in order to recover the desired macroscopic behavior as described by the Navier-Stokes equations. The simplest velocity sets have seven or nine elementary vectors in two dimensions and 15 or 19 in three dimensions. Several discrete sets of velocities have already been proved mathematically to be adequate for a two- or three-dimensional approximation of the Navier-Stokes equations [Wol94]. Two useful sets of directions and velocities which are adequate for three-dimensional (D3Qx) uid simulations are shown in Figure 2.13. The Figure was taken from a publication by K hnelt [K h02] which includes further details about the Latticeu u Boltzmann method and its application to ute simulations. The velocity directions and magnitudes as indicated in the Figures are directly related to a certain spatial lattice, because the velocity magnitudes are favorably chosen in such a way that a translation from one lattice node to another takes place during exactly one time step. The lattice node distance di , the time step t and the corresponding velocity ci are therefore connected according to ci = di /(t). This permits three different velocity magnitudes to be present in the schemes shown in Figure 2.13. The six inner velocity magnitudes c(1..6) in both schemes are a/(t) with a being the lattice constant. The eight outer velocity magnitudes

123

2. Numerical Concepts

Figure 2.13.: Velocity Vectors of D3Q19 Lattice c(7..14) of the D3Q15 scheme result in a 3/(t). The twelve outer ve locity magnitudes c(7..18) of the D3Q19 scheme result in a 2/(t). The remaining velocity c0 of both schemes is zero. For the ute simulation in Chapter 4.3 the D3Q19 scheme has been used, so we will refer to this whenever an actual implementation detail is mentioned. Each of the lattice nodes keeps the accounts of the particle density distribution for 19 discrete velocity vectors now, which are updated whenever particle densities have moved from one lattice node to another according to the discretized Boltzmann equation, the Lattice Boltzmann equation, which will be derived below.

2.3.2. Relaxation

Just like in the continuous case, as already described by Equation 1.54, a relaxation into a local equilibrium distribution can be implemented, which approximates the collision term . While the authors of the BGK-Model have chosen a local Maxwell-Boltzmann equilibrium distribution as a lo-

124

cal relaxation target, this is not absolutely required for a discretized version of the Boltzmann equation. There is some freedom in choosing the local relaxation target as long as some important requirements are met. Firstly the sum of all 19 contributions fi to the distribution function f which are associated with the discrete velocities ci at any lattice node must be identical to the local density (r, t) at any time.

18

i=0

(2.178)

The second requirement deals with the conservation of the momentum. The sum of the 18 nonzero momenta has to be the macroscopically observable uid momentum at any lattice node and any time:

18

i=1

(2.179)

The third requirement concerns the second order momentum which must represent the macroscopic stress tensor of the uid. Using Einsteins summing convention for the repeated indices and , with being the Kronecker Delta representing a diagonal matrix, this requirement can be written as:

18

i=1

(2.180)

Finally isotropy is required for certain fourth order moments. Anyhow it is safe and more straight forward to approximate the physical equilibrium distribution as described by the Maxwell-Boltzmann distribution by a polynomial expression [HL97]. The Taylor expansion up to order two yields: f (0) (v) = (v u)2 = exp 2RT (2RT )D/2

125

2. Numerical Concepts

= v2 exp 2RT (2RT )D/2 1+ v u (v u)2 u2 + RT 2(RT )2 2RT + O(u3 )

(2.181) With c2 = RT the discretized version of the equilibrium distribution s can be derived according to fi

(0)

= wi 1 +

+ O(u3 )

(2.182)

The weighting factors wi can be determined by equating the macroscopic equilibrium quantities derived as the moments of the distribution function with those obtained from the continuous version [WG00]. Following Qian, who has calculated weighting coefcients for various schemes [QdL92, QO93], the D3Q19 scheme has zero odd order moments. From the even order moments which are given by 0th order moment: i wi

= w0 + 6w1 + 12w2 =

c2 s

2nd order moment: 2 = i wi ci,x 4th order moment: 4 i wi ci,x 2 2 i wi ci,x ci,y

2w1 + 8w2

(2.183)

= =

= 3c4 s = c4 s

The distribution fi of the particle density is moving on a regular grid in discrete directions with discrete speeds ci . The Boltzmann equation can

126

be written for a set of discrete velocities: fi + ci fi = , t i = 1, 2, . . . , N. (2.185)

We approximate the time derivative by a simple rst order forward difference fi (r, t + t) fi (r, t) fi (r, t) (2.186) t t and the spatial derivative by a rst order backward scheme ci fi ci fi (r, t) fi (r ci t, t) a (2.187)

with a being the lattice constant. For the collision operator we choose the same rst order backward difference fi (r, t) fi (r ci t, t) fi (r, t + t) fi (r, t) +ci = i (rci t, t), t a (2.188) to get an explicit formula formula for the new distributions. With time step and lattice spacing according to ci = a/(t) the equation simplies to fi (r, t + t) fi (r ci t, t) = i (r ci t, t). t t (2.189)

Proceeding one step in space we arrive at r + a getting the evolution equation of the Lattice-Boltzmann method: fi (r + a, t + t) = fi (r, t) + t i (r, t). (2.190)

To include the collision operator we implement the single relaxation time (BGK) model 1 fi (r, t) fieq ((r, t), u(r, t)) , (2.191) with the relaxation time and the equilibrium distribution fieq , which only depends on the conserved macroscopic quantities. The low Mach number fi (r + a, t + t) = fi (r, t) t

127

2. Numerical Concepts

expansion of the Maxwell-Boltzmann distribution, second order in u as derived above is commonly used as equilibrium distribution in order to approximate the Navier-Stokes equations for small Mach numbers, small Knudsen numbers and states not too far from equilibrium. This equation can be processed sequentially. The new distributions at the next future time point can be calculated explicitly taking propagation of streaming uid from one lattice cell to its downstream neighbor into account. Once the new distributions are known the collision operators can be evaluated locally. Its results will be needed during the next streaming step. It is straight forward to implement Equation 2.190 for all nodes of the simulation domain. It is well suited for parallel computing and can be run on specialized hardware. The conserved macroscopic quantities, density and momentum density j = u are the zeroth and rst moments of the distributions fi : (r, t) =

i

fi (r, t) ci fi (r, t)

i

(2.192) (2.193)

(r, t)j(r, t) =

Note that in the isothermal, or more correctly, non-thermal model used here, the temperature is not a conserved quantity. As equation of state that of an ideal gas is chosen: p = c2 . s

For wave vectors k = 0 the macroscopic behavior of the LBM can be deduced using a multi-scale (Chapman-Enskog) expansion, which leads to the compressible Navier-Stokes equations (neglecting terms of higher order in u): (2.194) t + u = 0 t u + (c2 +u u ) = c2 s s 1 2 ( u + u ) (2.195)

128

cs = 1/ 3 is the sound speed of the model, = 1 c2 and s 2 2 = d c2 1 are the kinematic and the bulk viscosity in d spacial s 2 dimensions. These dimensionless equations can be connected to the real world by x2 x real = , (2.196) creal = cs , s t t with the lattice spacing x and the time step t. For wave numbers k > 0 a linear stability analysis [LL00] shows that up to second order of k the acoustic wave propagation is Galilean invariant and the shear and bulk viscosities are independent fromu. Additionally, the stability analysis shows that transverse modes are generally more stable than longitudinal modes and that sound waves propagating in the direction of the mean ow velocity can be unstable. As the unstable modes have a small spatial scale, spacial or temporal ltering can be used as a practical means to limit the development of instabilities. Lattice-Boltzmann models with collision terms with multiple relaxation times [dGK+ 02] are more stable than BGK models. Their free parameters are adjusted to minimize the effect of the lattice symmetry on the macroscopic behavior or the model. Especially the separate treatment of volume viscosity and shear viscosity can stabilize an aero-acoustical LatticeBoltzmann simulation signicantly when realistic properties of the propagation medium are assumed. In these models, the distributions fi are transformed from the vector space, based upon the discrete velocity set into a vector space based upon the moments of fi the hydrodynamic quantities and their uxes. The relaxation parameters of the moments are directly related to the various transport coefcients. Implementing the transformation matrices in a computer program has to be done carefully to avoid severe numerical cancellation. Simulating low viscosity ows with LBM meets a principal problem of stability. Due to the truncated expansion of the equilibrium distribution, the H-theorem of the continuous Boltzmann equation, assuring unconditionally stability, gets lost. Nevertheless, there results a region of stability

129

2. Numerical Concepts

which turns out to be rather large for isothermal LBM.

A major strength of the Lattice Boltzmann method is the very easy implementation of realistic, irregular boundary conditions. The simplest model of a rigid wall is just bouncing-back, reecting all incident populations by exactly 180 degrees, fout = fin , cout = cin , leading to zero velocity at the wall. This simple, computationally efcient and often-used procedure is of second order for boundaries in the direction of the main axes, but of rst order for inclined or curved boundaries, which lowers the accuracy in the whole simulation domain. Many other boundary conditions of higher order have been proposed, but the last word on universal boundary conditions for LBM has not yet been spoken. In short, the LBM is an approximation of the Navier-Stokes equations. It is fully local in space and time, second-order accurate in space and time, incrementally and locally conservative, conditioned to low Mach numbers, unconditionally linearly, yet only conditionally nonlinearly, stable, efcient on serial and parallel computers, and can easily handle irregular boundary conditions (within the limits of the regular grid). A recent study [KADA+ 03] shows that a new kind of LB model, which secures compliance with the H-theorem and thus is numerically much more stable than the other LB models, exhibits a naturally built-in subgrid model so that it can be used for high resolution turbulence modelling. The LBM is a developing eld of research, and new, promising models that can be used for the simulation of uid-acoustic interaction are to be expected in the near future. Comparing the lattice Boltzmann method with other computational methods that directly operate on the Navier-Stokes equations, several differences are revealed: While the Navier-Stokes equations are second-order partial differential equations, the Lattice-Boltzmann method is based on a set of

130

rst-order partial differential equations. Navier-Stokes solvers need to deal with the nonlinear convective term uu, whereas there is simple advection handling in the LBM. In the incompressible Navier-Stokes equations, the pressure eld serves as a dynamic constraint that instantaneously enforces the zero-divergence condition of the ow. Changes in the pressure eld distribute immediately because sound waves propagate at a virtually innite speed in an incompressible ow. Therefore the Poisson equation for the pressure ( p = (u )u) has to be solved after every time step, which is a very costly operation involving global data communications. In the LBM the pressure is obtained through an equation of state, and data communication is purely local during the collision step. In the LBM the Courant-Number (CF L = t/x) is equal to one, based on the lattice units x = t = 1. Consequently the time dependent LBM is inefcient for solving steady-state problems because its speed of convergence is dictated by acoustic propagation, which is rather slow (csound < x/t). Boundary conditions involving complex geometries require careful treatment in Navier-Stokes as well as Lattice-Boltzmann solvers. In LBM, boundary conditions can be set locally in a very easy way, thus allowing irregular realistic boundaries. This is a major strength of the Lattice-Boltzmann Method. The spatial discretization in the LBM is dictated by the discretization of the particle velocity space. This coupling between discretized velocity space and conguration space leads to a regular square or cubic grid. This is a limitation of the LBM, which can be partially circumvented by the use of grid renement methods and conditions for curved boundaries.

131

2. Numerical Concepts

132

3. Specic Methods

Flow acoustics incorporates all phenomena of uid mechanics and acoustics, specically sound generation by ows, sonic ow generation, owacoustics interactions and sound propagation in ows. A description of all these effects is included in the most general formulation of the NavierStokes equations. If the uid is air, as is normally the case in musical acoustics, viscosity is sometimes neglected and the much simpler Euler equations can be used, which still describe all dynamic effects of an inviscid ow. A further simplication arises from neglecting the compressibility of the medium. Unfortunately it is just the sound propagation which depends on the compressibility of the medium. If density is constant independent of pressure then the speed of sound must be innite. It seems therefore that this approach is generally not viable in acoustics. But this is not quite true, because it is possible to separate the simulation of an unsteady ow from the analysis of sound wave propagation. An acoustical analogy relating the unsteady ow to a certain distribution of equivalent acoustical sources due to Lighthill [Lig78] serves as the onedirectional link between the ow simulation and the sound sources of the acoustic simulation. It is therefore assumed that ow does generate sound, but sound does not inuence the ow. Unfortunately it is the latter effect which causes a ute or organ pipe to create self sustained sound. Computational aero-acoustics (CAA) deals with the numerical simulation of the generation and radiation of sound by unsteady ows. Predicting the noise generated by airplane fans, helicopters and cars are typical problems which can be solved using the acoustical analogy. In problems where the closed feed back loop of ow and sound is essential, as is the case in utes and ue pipes, either the compressible unsteady

133

3. Specic Methods

Navier-Stokes equations have to be solved, or the Boltzmann equation has to be simulated. In both approaches ow and acoustic waves will be obtained as direct results. In order to reduce the complexity of a problem signicantly, the problem is often simplied dimensionally. Using underlying symmetries a three-dimensional problem can often be reduced to a two- or even onedimensional problem. This helps a lot, but one should be aware that structures like vortices are three-dimensional in nature which will not be properly reected in a simulation at reduced spatial dimensions. Again, it is the eld of musical acoustics where this common simplication of dimensionality sometimes distorts the simulation results significantly. Vortex shedding is an important sound source e.g. in utes or other wind instruments, and must not be neglected. In order to accurately analyze sound generation and radiation of e.g. a ute, the computational algorithm should be setup in three spatial dimensions with a grid resolution ne enough to resolve all the scales of motion from the largest scales imposed by the boundary conditions to the smallest scales associated with viscous dissipation. This is a challenging task, which in many cases surpasses the todays computational possibilities. On top of that, available software for compressible ow is often inefcient or inaccurate at low Mach number conditions. Therefore it is the Lattice-Boltzmann method which is currently attracting much attention when such problems are to be solved. In the following chapter results of such a simulation for a recorder-like ute will be presented. However, the vast eld of musical acoustics also offers much simpler problems, which can be solved with more conventional methods and much less computational efforts. Computing wave propagation in uniform, often axis-symmetrical ducts - a rather technical description of all the members of the wind instrument family - might be part of the daily exercises of a musical acoustician. Wave propagation can be studied by solving the wave equation analytically by means of harmonic decomposition. The steady state frequency-

134

domain solutions can be transformed back into the time-domain using Fourier transformation but they do not allow the study of non-linearities and transient phenomena. The solutions of the wave equation can be adapted to several special cases of boundary conditions. In the eld of musical acoustics wave propagation in tubular objects, strings and membranes are most important. Modal decomposition makes use of existing symmetries and gives analytic results in the frequency-domain even in tubular ducts with axialsymmetric but arbitrarily varying cross-sections. The wave equation is also well suited for treatment in the time-domain by a nite difference method. Wave guide modelling and a related technique called Wave Digital Filter modelling can be applied efciently to synthesize sound of various musical instruments. Even non-linear elements like a reed or a piano-hammer can be included in the feedback loop involving wave guides and wave digital lters. In addition to the physical modelling as described above, there is also a kind of behavioral modelling, which does not make use of physical laws, but which rather tries to mimic a nal effect, usually the radiated sound, by means of mathematical methods. Specic numerical methods are also known for solving the so called Inverse Problem, the task to reconstruct the inner dimensions of a duct from an acoustical measurement made at the mouth of the duct. Practically this is applied to wind instruments or their parts in order to determine their bore by means of an acoustical pulse response or input impedance measurement.

Analysis of wave propagation can be made in the time-domain or in the frequency-domain. Both methods should be equivalent and exchangeable as long as the assumptions hold, which were made when the wave equation (Equation 1.38) was derived. The pulse response r(t) is the inverse Fourier-transform of a transmission function H(j).

135

3. Specic Methods

Practically we are dealing with discrete fourier transforms (DFT) of limited numerical precision and up to a certain bandwidth. We are bound to a certain sampling frequency and cannot wait an innitely long time period to make sure transient effects have decayed. Finally, the linearity of our systems is limited and we can easily drive them into a state where non-linear effects start to become noticeable. To self-sustain oscillations in wind instruments a highly non-linear element (lip-valve, reed, air-jet) is part of a feedback loop which contains the resonating air column. A realistic sound synthesis will have to model the non-linear element in the oscillator. It has turned out that a small signal analysis of a linearized oscillator-resonator system is problematic if the normal operating region is highly nonlinear. Advantages and disadvantages of time and frequency-domain methods are discussed by several authors e.g. [Sch81]. A good example how nonlinear effects like wave steepening can be included in a frequency-domain model is given by Thompson and Strong in [TS01].

In wind instruments properties like intonation, response, efciency, even sound timbre are strongly related to the acoustical impedance Z(j), often measured at the entry of the resonating air column. It is a complex function of frequency, dened as the quotient of sound pressure p(j) and volume velocity u(j) at this point. Air column resonances corresponding to the playable tones of the instrument show up as local maxima (or minima in the case of utes) of the input impedance magnitude. Their position along the frequency axes, quality factor and absolute value together with the group delay at these frequencies correlate with musical features like intonation, response, efciency and sound timbre. Bore reconstruction starting from acoustical input impedance measurements is another application where accurate modelling of input impedance is essential. A practical method based on optimization, which has been developed by the author [Kau03a] will be discussed in detail in Section 4.2.

136

How to solve the wave equation in one and three dimensions for cylindrical and rectangular ducts has been superbly demonstrated by Kemp in [Kem02]. He also collected and extended related expressions for the multimode radiation impedance and for mode conversions at discontinuities. Single Mode, Cylindrical Pipe Wave propagation inside cylindrical pipes can be described by a onedimensional theory. Waves set up inside approximately cylindrical instruments have plane wavefronts of nearly identical pressure perpendicular to the wall. They propagate just like in open space but are partially reected and partially transmitted by any change of cross-sectional area within the pipe. They are described by the one-dimensional (z-axis) wave equation 2p 1 2p = 2. c2 t2 z The complex solution for the pressure p(z, t) is p(z, t) = (Aejkz + Bejkz )ejt (3.2) (3.1)

with wave number k, wave length while = 2 and = 2f . A and k B are the complex amplitudes of the forward resp. backward travelling pressure waves. To calculate the particle velocity vz we use the one dimensional momentum Equation 1.9 but we neglect convection, friction and all external forces p vz = . (3.3) t z Substituting p from Equation 3.2 into Equation 3.3 we obtain a similar solution S (3.4) U (z, t) = (Aejkz Bejkz )ejt , c with u being the volume velocity dened as u = vz S and S being the cross-sectional area of the tube. That means we have sinusoidal pressure and velocity proles along the propagation axes z.

137

3. Specic Methods

To derive the acoustical properties p0 and u0 at any point z of the duct assuming they are known at another point z1 , a distance d away, requires substitution of z = z1 d into the equations 3.2 and 3.4. With Zc = c S we obtain (3.5) p0 = cos(kd)p1 + j sin(kd)Zc u1 and u0 = j sin(kd)Zc 1 p1 + cos(kd)u1 . (3.6) We can dene the acoustical impedance and again derive its projection along the duct Z0 = p0 cos(kd)Z1 + j sin(kd)Zc . = 1 u0 j sin(kd)Zc Z1 + cos(kd) (3.7)

It should be mentioned here that a discontinuity in the tube, that means a sudden change in cross-sectional area from S = SL , left of it, to S = SR , right of it, cannot cause any change in p, u and therefore in Z. This follows from the continuity equations for mass and momentum. What changes is the characteristic impedance Zc which has the effect that waves get partially reected and transmitted at the discontinuity. By means of Equation 3.2 and 3.4 we can write two equations matching the left and right sided pressures and volume velocities. For this purpose we substitute S = SL , A = AL , B = BL on the left sides and and S = SR , A = AR , B = BR on the right sides. Simplied we obtain AL + BL = AR + BR and (AL BL )SL = (AR BR )SR . The reection coefcient AL /BL for a forward travelling wave (D = 0) can now be calculated to SL /SR 1 AL . (3.8) = BL SL /SR + 1 The same way a transmission coefcient AR /AL can be determined 2SL /SR AR . = AL SL /SR + 1 (3.9)

138

Single Mode, Varying Cross-Section The input impedance of arbitrary tubular ducts, terminated by a known impedance at the far end, can be calculated by projecting the termination impedance section by section back to the input. The results obtained above would already be sufcient to calculate input impedance curves of wind instruments, arbitrarily well approximated by a sequence of cylindrical slices. However, the termination impedance has to be known and viscous losses have been neglected up to now. Both of these issues are going to be addressed now. Another practical problem with the above solution is that in order to model longer conical parts accurately, those conical sections have to be split into many cylindrical slices, which means, projection has to be calculated very often. The calculation model that is reviewed below, resolves this issue by giving projection formulae for cones. Originally it was published by MapesRiordan [MR93] but its notation is adapted here. It is based on work published by Keefe [Kee81] [Kee90] [Kee84], Causs [Cea84] et al. and Bee nade [Ben88]. The model describes transmission matrices Ai (j) representing conical or cylindrical wave guide elements and it takes thermo-viscous losses into account. Its derivation would go beyond the scope of this work, so just the results are presented. It is assumed that any known duct geometry can be decomposed into n simple elements yielding n complex frequency dependent wave guide transmission matrices. ai,11 (j) ai,12 (j) ai,21 (j) ai,22 (j)

xi+1 xi 1 Zc D C xi+1 CL + (xi )2

Ai (j) = =

= (3.10)

xi xi+1 Zc D xi xi+1 D C + xi

xi+1 1 xi (xi )2

139

3. Specic Methods

where C = cosh(L), D = sinh(L),

1 1 = k(1.045rv + j(1 + 1.045rv )), , k = c

rv =

Sm ,

Zc,i = c/Si ,

with being the equilibrium gas density, the radian frequency, the shear viscosity coefcient, c the speed of sound, Sm the planar crosssectional area at the center and Si the spherical area at the input end of the conical element, xi the radius of the input spherical sector, xi+1 the radius of the output spherical sector and L the distance between the two spheres. This matrix describes the relationship between sound pressure p and volume velocity u in front of and after a conical slice by pi (j) ui (j) The product A(j) =

L i=1

= Ai (j)

(3.11)

characteristic of the complete instrument. Radiation Impedance The ratio between sound pressure and volume velocity at the open end of the aring bell is enforced by the termination or radiation impedance Z T

140

which is the characteristic impedance of the open mouth of the instrument. It is here modelled by a rst order approximation for a piston radiator without bafe as proposed by Levine [LS48] pT (j) = ZT = Zc,T uT (j) ( 0.61RT RT 2 ) + j( ) 2c c (3.12)

with RT = ST /, Zc,T = c/ST and ST being the cross-sectional area of the open end. If the radiation impedance of the open end of a brass wind instrument is transformed by the chain matrix A from the open end back to the mouthpiece, then the acoustical input impedance is obtained. Because of the one-dimensional assumptions higher oscillation modes are not taken into account up to now. The validity of the plane wave model is limited by the rst cut-off frequency co , which is usually most restrictive at the open mouth of the instrument. It can be calculated from RT co /c 1.84. It should be mentioned here that Causs proposes an approximation e for the radiation impedance of a piston radiator without bafe valid up to RT /c < 3.5 in [Cea84]. Rayleigh gave an expression for a piston radiator terminated in an innite bafe [Ray94]. Multi-Modal Propagation Modal decomposition is a way to solve the 3-dimensional wave Equation 1.38 for cylindrical or rectangular ducts without plane wave assumption. In cylindrical coordinates (r, , z) the Laplacian operator becomes = + 1 2 2 2 1 2 + 2 =( 2 + ) + ( 2 ). 2 2 z r r r r z (3.13)

If we express pressure p and axial velocity vz using innite series we get the solutions p(r, , z, t) =

i

(3.14)

141

3. Specic Methods

and vz (r, , z, t) = 1 S(z) Ui (z)i (r, ) exp(jt),

i

(3.15)

with pressure prole Pi and velocity prole Ui of the ith mode along the z-axis and the prole i orthogonal to it. We know, that modal proles along the z-axis Pi (z) = exp(jki z), Ui (z) = exp(jki z) (3.16)

are sinusoidal with modal wave numbers ki . The proles i across the longitudinal axis are the classical eigenfunctions obeying the transverse Laplacian ( ) eigenproblem i = i 2 i (3.17)

with boundary condition i /r = 0 for r = R, R being the radius of the cylindrical duct. The eigenvalues in the axis-symmetrical case are given by i (z) = i /R with i being the successive zeros of the Bessel function of order one. Substituting p from Equation 3.14 into the wave Equation 1.38, resolving the differential operator from Equation 3.13, dividing by p and substituting Pi and i from equations 3.16 and 3.17 we can get rid of all partial derivatives and obtain the relation ki 2 = k2 i 2 , (3.18)

with the free space wave number k = /c and i = 2/ki . Propagation of the different modes along a cylindrical piece of duct does not differ from the plane wave case, described by equations 3.2 and 3.4, yet each mode having got its own wave number ki and its own characteristic impedance Zc,i = kc/ki S. If the wave number ki becomes imaginary (for k < i ), exponential damping is observed. The modal projection equations from plane 1 to plane 0, a distance d away, are (3.19) P0,i = cos(ki d)P1,i + j sin(ki d)Zc,i U1,i

142

and U0,i = j sin(ki d)Zc,i 1 P1,i + cos(ki d)U1,i . (3.20)

If these equations are written using a matrix notation and are substituted into the following equation dening a multi-modal impedance matrix Z P = ZU we obtain the projection equation for impedance matrices Z0 = (Z1 + jDZc )(jDZc 1 Z1 + I)1 (3.22) (3.21)

with D(n, m) = tan(kn d) for n = m and D(n, m) = 0 for n = m, and with Zc (n, m) = kc/(kn S) for n = m and Zc(n, m) = 0 for n = m. Assuming axis-symmetric pressure distributions only, the solution of the eigenproblem in Equation 3.17 is i = J0 (i r/R) J0 (i ) (3.23)

with J0 being the Bessel function of the rst kind of order zero. Lossy multi-modal wave propagation was studied by Bruneau et al. in [BBHK87]. Starting with a lossy boundary condition he obtained complex modal wave numbers ki = with = (1 (i )2 /(k2 R2 )) v + t (3.25) with v = (1 + j)2.03 105 f and t = (1 + j)0.95 105 f (simplied, refer to [BBHK87, Kem02]).

i

k2 (

i 2 2k ) + ( )(Im( i ) jRe( i )) R R

(3.24)

143

3. Specic Methods

Multi Mode Conversion While different modes in a uniform duct propagate independently of each other, mode conversion takes place where the ducts cross-sectional area changes. The ith mode after the discontinuity will be composed from a weighted sum of all modes before the discontinuity. The pressure and volume velocity mode amplitude vectors P and U at one side of the discontinuity (cross-section S0 ) are related to the corresponding vectors at the other side (cross-section S1 ) by P0 = FP1 , and P1 = VP0 , U0 = VT U1 , S0 > S1 . (3.27) The matrices F and V for cylindrical and rectangular ducts have been derived by Kemp in [Kem02]. The results for the cylindrical case are repeated here as Fn,m () = ( 2 m 2 2m J1 (m ) , n 2 )J0 (m ) (3.28) U1 = FT U0 , S0 < S1 (3.26)

with = R1 /R2 and F (0, 0) = 1 and Vn,m () = Fn,m (1/). From the equations 3.26 and 3.27 the projection of the impedance matrix across a discontinuity can be calculated as Z0 = FZ1 FT , and S0 < S1 , (3.29)

Z0 = V1 Z1 (VT )1 ,

S0 > S1 .

(3.30)

Multi Mode Radiation In [Zor73] Zorumski published a numerically applicable multi-modal solution for the radiation impedance of a cylindrical pipe terminated in an innite ange.

144

Multi-modal wave propagation in axis-symmetrical or rectangular ducts has superbly been reviewed and extended by Kemp in [Kem02]. Important studies of multi-modal wave propagation have been published by Pagneux [PAK96] and Amir [APK97]. The multi-modal result for the radiation impedance of a circular opening in an innite bafe is repeated here: Zn,m = + jc S where

0 c S

2

(3.31)

2 J1 ( kR) Di ( ) = (i /(kR))2 2

(3.32)

In a recent study published by H lie and Rodet in [HR03] the multie modal radiation impedance has been calculated for a pulsating portion of a sphere without any solid wall or object other than the rest of the sphere which remains motionless. The resulting expression for the impedance Z as a function of frequency f , angle off the axis of symmetry and mode number n is Zn (f, ) = jZc n (2f r0 /c0 )n (0 )Pn (cos ) where (3.33)

Pn1 (0 ) Pn+1 (0 ) (3.34) 2 with r0 being the radius of the sphere, Zc = 0 c0 being the characteristic specic impedance, 0 and c0 the mass density and speed of sound, theta0 being the maximum opening angle at the edge of the pulsating portion of the sphere, Pn being the Legendre polynomials and n (z) = hn (z) with hn (z) hn representing the outgoing spherical Hankel functions. Another version of Equation 3.33 averaged over all angles theta and therefore useful to terminate one-dimensional conical wave guides with n (0 ) =

145

3. Specic Methods

10

10

10

10

200

400

600

800

1400

1600

1800

2000

146

spherical wave fronts was given as Z(f ) = 2jZc 1 cos 0

+ n=0

(3.35)

The cited publication also contains three approximations without expensive functions, which are much better suited to time-domain simulations. They should be considered as termination impedance of wave guide networks as described in Section 3.2. Taking multi-modal propagation into account can improve modelling of musical instruments signicantly. Even below the cut-off frequency, modal conversion makes quite a notable difference as shown in Figure 3.1. Comparisons with FEM simulations made by Amir, Pagneux and Kergomard [PAK96] [APK97] show good agreement.

147

3. Specic Methods

Time-domain methods are usually based on a time-domain formulation. Time-domain methods are inherently more suited to the prediction of perceptually relevant transient phenomena [WC03], while frequency-domain methods are often more accurate to deal with the steady state. Timedomain modelling is usually applied in musical instrument sound synthesis. With the desktop computer power available today even real-time sound synthesis is possible, at least if the algorithms are optimized for this purpose. Sound synthesis by means of physical modelling has attracted the attention of researchers for a long period of time now. Pioneering work has been published by DePoli who wrote a tutorial on digital sound synthesis techniques in 1983 (reprint in [DP91]). In 1991 he edited a collection of articles on that subject [DPP91], other papers are [DPP89, DPR99], with Borin [BDPS90, BDPS92, BDP96], with Magalotti [MBDP95], and with Avanzini et.al. [ABB+ 01]. Physical models of various instruments with the aim of sound synthesis have been developed by J. O. Smith e.g. [Smi92, SI95, SVD95, Smi96, SI97, Smi98, SI03], Barjau [BKC99, BG01, BG02, BGC02, BG03],Bensa [BBKMSI03], Scavone [SSI97], and others.

Time-domain descriptions can be obtained from frequency-domain transmission functions by applying an inverse Fourier transformation. The continuous time representation and the continuous frequency representation are dual. In practice we are mostly dealing with discrete time and discrete frequency representations. We have sequences of signal samples recorded by a sound card or obtained by a data acquisition system. We can use the Fast Fourier Transform (FFT) or more generally the Discrete Fourier Transform (DFT) to move between a discrete time representation and a discrete spectrum. Filtering a discrete time signal requires convolution of the sampled data stream with a suitable lter kernel, often

148

a window function like the Gauss distribution. The same operation can be performed by multiplying the spectral representation of the signal by the spectral representation of the lter kernel. Therefore convolution and multiplication are dual operations in discrete time- and frequency-domain. If an acoustical system can be spatially split into simple sound processing elements represented by known transmission functions for the acoustical state variables p and u or Z = p/u, then the overall characteristic can be obtained by multiplying all partial transmission functions along the acoustical propagation path. If a time-domain representation is required, an FFT back-transformation can be made. Once the time-domain characteristic, the systems pulse response, has been determined, its effect on acoustical input signals can be simulated. The sampled data stream of a stimulus signal has to be convolved with the systems pulse response in order to generate an output data stream representing the ltering effect of the system on the stimulus. By measuring for example the pulse response of a concert hall at a certain point in the auditorium, stimulated at some point on the stage, the halls transmission characteristic between these two points can be determined. By convolving any audio data stream with that measured pulse response, the effect of the hall can be simulated and the audio information, which would be received by a listener at that point in the auditorium can exactly be synthesized. Convolution can be computationally more expensive than multiplication especially if the involved signals are already available in the spectral domain. Therefore, calculating a multiplication in the frequency-domain instead of a convolution in the time-domain, will often be tried. Synthesizing the sounds of musical instruments using physical models is usually done in the time-domain, especially if the lter kernels to convolve with can be kept short. Instruments are often composed from small wave guide elements as described in the previous section. The transmission functions of these slices can be multiplied in the frequency-domain as long as physical dimensions do not change during the simulation. The overall instrument is then transformed back into the time-domain and con-

149

3. Specic Methods

volved with a stimulus stream. The main advantage of this concept is that there is no need to know any transmission function of pulse response analytically. It is sufcient to have measured magnitudes and phases of a complex spectrum representing the transmission function of an instrument sampled in the frequency-domain or to have a sampled pulse response, directly recorded in the time-domain. There is one issue associated with measuring a pulse response in the time-domain, which is also crucial to the measurements required in the next section. It is practically not possible to stimulate a system with an idealized Dirac-Delta-Function. Real pulses which we can generate will never be innitely short nor innitely high. Not to speak about the inability of any real system to linearly respond to such a monster. Therefore we will have to use some realistic pulse-like signal with nite duration and nite amplitude. We can view the actual stimulus signal as an ideal Dirac pulse, pre-ltered by a kind of reality lter. This reality lter function can be measured because we can record the signal when it enters our system as we can record it when it leaves. The recorded overall pulse response is the the response of a combination of two systems to an ideal Dirac pulse. The recorded input signal is the pulse response of our reality lter to an ideal Dirac pulse. In the frequency-domain we had to divide the complex spectrum obtained from the output signal of our analyzed system by the complex spectrum obtained from its input signal. In the time-domain we have to deconvolve the overall pulse response and the recorded stimulus pulse. Aware of this possibility, we can drive things further. The actual shape of a stimulus pulse does not matter at all, because the response to an ideal Dirac pulse can exactly be recovered by deconvolution with the known actual input signal. Therefore we can choose another kind of stimulus signal which can easier carry more energy, like a short burst of white noise, a chirp or an MLS signal. These signals have a spectral distribution which is similar to that of a Dirac pulse, but because of their longer duration, we will get a better signal to noise ratio with our measurement. Care must be taken when the signal entering the analyzed system is

150

recorded. It must be ensured that reections coming back from the system under test do not mix with the stimulus. The stimulus signal must terminate before the leading edge of the excitation signal returns from the rst discontinuity where it is reected. For this reason, coupling to the test object must be nonreective and some reection-free source tube should be used between the microphone for the stimulus pulse and the actual device under test. A reader who wants to go deeper into this kind of measurements called pulse reectometry could refer to Sharp [Sha96] or Kemp [Kem02] with latest updates in [FSKL03].

If an analytical description of an acoustical system is available in the continuous frequency-domain, then a recursive discrete time formula can be obtained by means of the bilinear transformation. This transformation is expressed within the theoretical framework of the z-transformation. An introduction to the theory of sampled data systems and the associated ztransformation is given in the Appendix A of this book. For the moment we only have to know that the sampled time variable z exp(jT ), with T being the sampling period, can be interpreted as time shift operator. Dividing a sampled signal given in the z-domain by z is equivalent to delaying it by one sample. Dividing by z2 means delaying the signal by two samples. The recursion formula for applying a second order backward difference fk = 1 (3fk 4fk1 + fk2 ) 2T (3.36)

This z-domain formulation has a corresponding expression in the frequency-domain, which can be obtained using the denition of z e(jT ) .

151

3. Specic Methods

Substituting for z we obtain f (jT ) = 1 2T 3f (jT ) 4 f (jT ) f (jT ) + exp(jT ) exp(2jT ) . (3.38)

This is an expression, which is periodic in T = 2f /fs , fs being the sampling frequency. It makes sense to analyze gain and phase for frequencies f , which are below the Nyquist frequency fs /2. According to Nyquist law, there must not be any spectral components above half of the sampling frequency. Otherwise, sampling would lead to severe signal distortion caused by aliasing of the under-sampled spectral components rendering the transformation invalid. The transition from the z-domain to the s = j domain as demonstrated above is called the impulse invariant transformation because the pulse response obtained by a Fourier transformation of the resulting spectrum is equivalent to the pulse response of the continuous time system except for the fact that the discrete version is periodic. The impulse invariant transformation is especially well suited for analyzing recurrence formulas and differencing schemes for sampled data sequences according to spectral gain and phase. Using the transformation in the opposite direction as a tool to nd a sampled data representation of a given continuous frequency description is much less straightforward and it does not work with all systems.

A convenient way to transform a system description from the continuousfrequency domain into the sampled-time domain and back is by means of the bilinear transformation. It is dened as the substitution s= 2 T 1 z 1 1 + z 1 , (3.39)

with the complex frequency variable s = j = j2f , or in the opposite direction 1 + (T /2)s . (3.40) z= 1 (T /2)s

152

The idea of introducing complex frequency variables in the continuousand discrete-time domain allows us to bring the power of complex variable theory to bear on problems of discrete-time signals and systems. According to this theory, the bilinear transformation is essentially a mapping between the imaginary j axis of the complex s-plane onto the unit circle of the complex z-plane. Once a problem is described in the z-domain, all the theory is available to do stability, convergence and consistency analysis of that sampled-data system. When we use the bilinear transformation for discretizing a continuousdomain system specication derived from the governing differential equations, we should know, even if we are not aware of all that theory behind it, that the frequency scaling is nonlinear.

1

0.5

(1 z 1 ) (1 + z 1 )

-1.5 -0.5

-1

-0.5

s 2

-1

-1.5

-2

For low frequencies the mapping is pretty accurate. Approaching the Nyquist frequency the frequency error becomes bigger and bigger. It is

153

3. Specic Methods

said that a warping of the frequency axis has taken place, and if this error is not tolerable, it has to be taken into account, when the continuous-time specication is created. The relation between the original of the continuous description and of the discrete-time formulation obtained by the bilinear transformation is 2 tan(/2) (3.41) T Figure 3.2 shows the relative frequency error plotted logarithmically up to the sampling frequency fs . The sampled version exhibits the expected pole at the Nyquist frequency. Now let us have a look at how this method can be practically applied in order to get a recurrent discrete-time formulation for a simple ordinary differential equation. Let us take the equation of movement = mx (t) = kx(t) (3.42)

of a mass m attached to a spring k. Its complex-frequency-domain description can be obtained by replacing functions of time, here x(t), by related functions of complex frequency s, here X(s). Differentiation with respect to time becomes a multiplication with s, integration over time would transform to a division by s. More transformation rules can be looked up in any collection of mathematical formulas. We obtain (3.43) mX(s)s2 = kX(s). Now we can apply the bilinear transformation by replacing the functions of s, here X(s), by corresponding time series, here xn , and by substituting s according to Equation 3.39. With xn z 1 xn1 and xn z 2 xn2 we get after some simplications 4m xn 2xn1 + xn2 = kxn . T 2 xn + 2xn1 + xn2 Now we can solve for xn and we obtain xn = 4m + kT 2 2xn1 xn2 , 4m kT 2 (3.45) (3.44)

154

which is a nice recurrence formula for integration of the original differential equation. We can get the same result by applying the trapezoidal rule at times n and n 1 to v(t) and x(t). We start from the original differential equation dening the velocity v(t) = x (t) as second unknown function v (t) = k x(t). m (3.46)

With the trapezoidal rule from Equation 2.128 we can write vn = vn1 + T (kxn /m + kxn1 /m)/2 vn1 = vn2 + T (kxn1 /m + kxn2 /m)/2 xn = xn1 + T (vn + vn1 )/2 xn1 = xn2 + T (vn1 + vn2 )/2. From these four equations we can eliminate vn , vn1 and vn2 . The remaining equation can be solved for xn , which yields exactly the expression from Equation 3.45. It can be shown that the application of the bilinear transformation to a polynomial expression in s = j is equivalent to the integration of the corresponding differential equations using a nite difference scheme based on the trapezoidal rule. The characteristic frequency warping is of course also a property of the trapezoidal rule.

Wave propagation can be described by Huygens principle. It states that a wavefront can be decomposed into a number of point sources, which give rise to spherical wavelets. The envelope of these wavelets forms a new wavefront which again gives rise to a new generation of spherical wavelets. To nd a discrete equivalent to this principle without loss of generality, lets consider a two-dimensional mesh of orthogonal acoustic tubes of equal length l.

155

3. Specic Methods

An incident pressure pulse P that travels in a tube is scattered at the next junction where part of it will be reected. The remaining part will be split and propagated by the three continuation tubes as shown in Figure 3.3(a). The reected part of the original pressure pulse has a relative amplitude of 1/2, the relative amplitude of the transmitted waves is 1/2. Therefore mass and momentum are conserved by a single scattering process (Figure 3.3(b)).

The Huygens principle requires superposition of wavelets travelling in all four directions N , S, E and W . Incident wavelets from any direction have to be reected and propagated independently as described above. The total wave quantity in a junction is therefore the sum of four directional contributions. The scattering rule for any junction i of a TLM is a recurrence formula relating outgoing quantities at time n + 1 to incoming quantities at time n. The four directions have to be specied separately, so a matrix formulation is obtained. It relates the vector P of the four outgoing directional components P N, , P S, , P E, and P W, to the vector P+ consisting of

156

P P P P

the four incoming wave components. In vector notation it can be written as P = S P+ , n n+1 its components are given by N, Pn+1 1 1 1 1 S, 1 1 1 1 Pn+1 E, = 1/2 1 1 1 1 Pn+1 W, 1 1 1 1 Pn+1 (3.48) (3.47)

In this model the speed of sound is given by c = 1/ 2 l/t. The TLM-method is valid as long as transmitted wavelengths are large compared to the grid spacing l. If this condition is met, the discrete nature of the model will not become evident. The dispersion relation is exact to the second order in = l/L (L representing the macroscopic physical length scale of the problem) and anisotropy does not appear until the third order in . Until now lossless propagation has been considered, but it is not difcult to adapt the transmission line matrix in order to include local viscous losses. Boundary nodes can be preconditioned as reecting or absorbing nodes, approximating the effects of rigid walls or open boundary conditions. This exibility in terms of local properties and boundary conditions makes it easy to mimic the complex and possibly time varying arrangement of boundary conditions and local parameter changes which is inherent to most real world problems which is a crucial advantage of this method. The simple and efcient computational algorithm and the economic use of memory makes it well-suited for three dimensional sound propagation simulations. Virtual Wave Tank [Wil], is a demonstration program for the TLMmethod. It was used for the following examples of wave propagation in horn-like geometries.

157

3. Specic Methods

The TLM-method can be viewed from several sides. It can be obtained as a result of the discretization of the Huygens principle [KTFF98] as outlined above. It can also be considered as a linear lattice Boltzmann method [L t98] or it can be seen as the limit of a lumped element apu proximation, for instance in the context of spatially distributed electrical network models. The latter view was the starting point for Pelorson, who was simulating the vocal tract in three dimensions which is a problem with rather complicated boundary conditions [PBSEM02], and for the author, who used this kind of approximation to model the cross-section of the buzzing lip [Kau03b] as described in Section 4.1 on page 184. Another even more important view was elaborated by Bilbao in [Bil01]. He puts this approach which is often treated as isolated and non-theoretically into the context of the nite difference theory. He also uses the framework of the scattering theory to equally embed transmission line mesh, digital wave guide networks and wave digital lter techniques. Reconsider Equation 3.48 describing the scattering process. According to Huygens the macroscopic wave is the superposition of all incident wavelets. The wave quantity W in a node with discrete coordinates i, j at

158

time n is therefore determined by

i,j N,+,i,j S,+,i,j E,+,i,j W,+,i,j + Pn + Pn + Pn Wn = Pn

(3.49)

,+,i,j are of course identical to correspondThe incoming wavelets Pn ing outgoing wavelets associated with neighboring nodes. These again originate from incoming wavelets according to Equation 3.48, which is orthogonal, so there is no need to invert S. Equation 3.49 can now be formulated for the nodes (i 1, j), (i + 1, j), (i, j 1), (i, j + 1) at time n 1. Applying once more Equation 3.48 we can eliminate the incoming wavelets at these four nodes. If we proceed recursively one more step in time and space then we are able to eliminate all wavelet quantities from the equation system and we can obtain i,j i,j Wn + Wn2 =

(3.50)

which is a nite difference approximation of the two-dimensional wave equation 2W 2W 1 2W = + , (3.51) c2 t2 x2 y 2 third order accurate in space and time. It can be obtained by using the (1, 2, 1) second order backward scheme for the time derivative and the (1, 2, 1) second order central scheme for the spatial derivatives. If the i,j spatial differencing is formulated at time n 1 then the fn1 terms cancel and the scheme gains one order in accuracy. The TLM is therefore equivalent to a nite difference method applied to the two-dimensional wave equation, which justies its usage as a tool to simulate wave propagation. The discrete Huygens principle can be considered as a numerical integrator of the wave equation.

The TLM has been originally discovered intuitively, and being able to prove that it is indeed a numerical approximation of the wave equation

159

3. Specic Methods

does not help in discovering similar explicit scattering formulations for the integration of other partial differential equations. However, Fettweis [Fet86] and others, recently especially to mention Bilbao [Bil01], have developed and contributed to a theoretical framework that allows to apply scattering methods methodically and systematically for solving common types of ordinary and partial differential equations. As a member of Julius O. Smiths group at Stanford University who is dealing with scattering methods for some while now [Smi92], Bilbao studied the wave digital lter method that was not so well-known in the US. He unied its theory with the digital wave guide theory and could show that, for a given simulation problem, a general digital wave guide network and a multidimensional wave digital ltercan be developed in a very similar way. The two approaches even share the good stability and robustness properties already mentioned in relation to wave digital lters, and they can often be used alternatively. Robustness and numerical stability are derived from the fact that the scattering process is locally conservative thus ensuring not only global passivity but also differential passivity. As already mentioned, the sum of wave quantities entering a scattering junction is accurately balanced by the wave quantities leaving it one moment later. If energy is associated with the wave variables, then conservation of energy is naturally taken care of locally and at any moment. Even if some propagation coefcients, material parameters or spatial dimensions are modied between the scattering events, no energy can vanish or be created by the scattering process itself.

Let us now have a look at how these properties are achieved and how wave digital lter (WDF) methods can be derived from the differential equations describing a physical system. The application of such methods in the eld of musical acoustics was proposed by Sarti and Borin of the group of DePoli [SDP99] [BDPR00]. In spite of the apparent advantages of this method, it took almost twenty years in order to propagate from circuit and systems engineering to other

160

elds of science. It might have to do with the fact that many acousticians are not so familiar with inductors, capacitors, transformers, or Kirchhofs voltage loops and current nodes that they lost interest before having got the point. To avoid this her, it will be attempted to develop the ideas of the wave digital lter theory without stressing the analogy of electrical networks too much or rather to avoid the usage of such terms at all. We start our derivations with the equations of movement of a single mass and a single spring. The acceleration of a mass is proportional to the force acting on it and the force of the spring is proportional to the displacement from the equilibrium position, which is the time integral of the velocity. The two equations are

t

f (t) = mv (t)

f (t) = k

=0

v( )d.

(3.52)

In the complex frequency-domain these two equations can be written as f (s) = ms v(s) f (s) = k v(s) . s (3.53)

Applying the bilinear transform according to Equation 3.39 yields for the mass 2m v(z)(1 z 1 ) (3.54) (1 + z 1 )f (z) = T and for the spring 2 (1 z 1 )f (z) = k(1 + z 1 )v(z). T (3.55)

This corresponds to the trapezoidal integration scheme of the original differential equations fn + fn1 = and 2m (vn vn1 ) T (3.56)

(3.57)

161

3. Specic Methods

If the driving force fn represents a known boundary condition then we have explicit equations for vn and vice versa. Up to now there is not much new, it was just a repetition of the bilinear transformation. The resulting difference scheme could already be used and it is, as we know, second order accurate. In the rst example, we are approximating the time derivative of v by the difference, in the second example, the same is true for f . The power that is transferred by this action is f v, so the approximation error associated with the differencing of one of these variables modies that instantaneous power and the energy balance will become incorrect. Theoretically, the mechanical energy f x should be conserved as kinetic energy mv2/2 of the mass. In the other case, the mechanical energy should be stored in the spring as potential energy. The whole process is reversible, all the energy should be recoverable by retarding the mass or releasing the spring. If the instantaneous power has some error, then energy is no longer conserved. Another problem arises when the mass is attached to the spring. The velocity is common to both elements but the external force has to drive mass and spring now. the latter fact is expressed mathematically by equating the external force fx = fm + fk to the sum of the two forces acting on the mass and on the spring. The system of equations is no longer explicit. We have to solve an implicit system now if this is possible at all. Often this is difcult, expensive in the matrix case or impossible in the non-linear case. In the wave digital lter terminology the circuit is said to contain delay-free loops and is non-realizable. Therefore, the wave digital lter method introduces a variable substitution, before the differencing scheme is evaluated. The proposed substitution to avoid the term voltage we will call it force-waves is a(t) f (t) + v(t)R b(t) f (t) v(t)R. (3.58)

From this denition the inverse substitution can be derived as f (t) = a(t) + b(t) 2 v(t) = a(t) b(t) . 2R (3.59)

162

An alternative substitution by so-called power-waves, which should be used if coefcients are not constant or if non-linear equations are involved is dened as a(t) f (t) + v(t)R 2 R f (t) v(t)R . b(t) 2 R v(t) = f rac(t) a b(t) R. (3.60)

(3.61)

The conversion that can be applied locally where ever needed in the resulting equation system is straightforward b(t) = 2b (t) R. (3.62) a(t) = 2a (t) R Having introduced the concept of wave variables now, we can perform the substitution in our two difference equations 3.56 and 3.57 according to Equation 3.59, which was derived from the force-wave denitions: 2m an bn an1 bn1 an + bn an1 + bn1 + = ( ) 2 2 T 2R 2R (3.63)

an bn an1 bn1 2 an + bn an1 + bn1 ( )= + . (3.64) kT 2 2 2R 2R The parameter R, the so-called port-resistance, is a free parameter of the substitution. It can be chosen according to the requirements of the problem. In our case it is advantageous to choose it as 2m/T in the mass equation and as kT/2 in the spring equation. Now we obtain bn = an1 and bn = an1 . (3.66) The results are so-called strictly causal; that means, there exists a simple cause-and-effect relation between an input wave a and an output wave (3.65)

163

3. Specic Methods

b. The equations are explicit and realizable because there is a unit-delay between input and output. Figure 3.5 shows some basic WDF-elements, their usual circuit representation and the associated signal owcharts, derived according to the two examples that have been given. The friction element can be obtained by substituting from Equation 3.59 into f (t) = v(t). Substitution yields R(an + bn ) = (an bn ). Choosing R = we obtain the condition bn = 0. The velocity node is an element that enforces zero velocity. A force node is an element with zero force input. From Equation 3.59 it is immediately obvious that the wave-variable relations are bn = an for the former and bn = an for the latter. A nonzero external force input results in bn = 2fx an and an externally enforced velocity in bn = an 2Rvx as can be easily veried.

a

z

R=2m/T

-1

a

z

-1

R=kT/2

a b

R=

Mass

Spring

Friction

(a)

a b

R=any 2 fx f-source

(b)

a b

R=any 2Rvx v-source

(c)

a b

R=any

v-node

(d)

(f)

As we will work out now, the introduction of wave variables has solved the energy leakage problem, too. The instantaneous power f (t) v(t) 2 2 b results in a (t)b (t) in the case of force-waves and in a2 (t) 2 (t) in the 4R case of power-waves. In the case of the mass we can replace bn by an1 thus b2 by a2 . n n1

164

1 For the instantaneous power we obtain fn vn = 4R (a2 a2 ) or in the n n1 2 a2 case of power-waves fn vn = an n1 . The same is true for the spring. This instantaneous power times the interval T can be interpreted as the energy or work which is delivered by the force to the mass or the spring. It is stored as kinetic or potential energy and is equivalent to the difference a2 a2 . This difference is now exactly equivalent to the square of n n1 the value that has to be stored in the register representing the unit delay. This memory element therefore holds a physical quantity representing the kinetic energy of the mass or the potential energy of the spring. In the case of power-waves there is not even a port resistance contained in the proportionality term, so it does not matter if that parameter changes between the sampling events. In the case of force-waves the port resistance is a kind of scale factor for the energy stored in the element. The only problem that still has to be solved is a procedure how to handle combinations of such elements. In the force-velocity case things are clear. There is one kind of coupling where the velocity of two or more elements is equal and where the external force is split into partial forces acting on the parts. The other kind of coupling is where the external force is applied to all parts identically while partial velocities sum up yielding the external velocity at the point driven by the force. An example for the rst case is the system consisting of a mass and an ideal spring with an external force acting on the mass. Mass and spring are enforced to move at the same velocity but the external force is split into a partial force accelerating the mass and a remaining part squeezing the spring. If a velocity-proportional friction force is included, then this system can be described by the ordinary differential equation t

f (t) = mv (t) + k

=0

v( )d + v(t),

(3.67)

or, after a Laplace transformation, in the s-domain as f (s) = msv(s) + k v(s) + v(s) s (3.68)

165

3. Specic Methods

An example for the second case is a system consisting of a mass and an ideal spring again, but with an external force acting on the spring. The mass is horizontally gliding without any friction this time. The spring propagates the external force to the mass - actio est reactio - and its length must always reect that external force. The length of the spring will change with the displacement difference of mass and driving point. The time derivatives of the displacements are the related velocities. Therefore the velocity difference between mass and driving point will be equivalent to the springs velocity or strictly speaking the spring length change rate. With vm = vx + vk , vx being the velocity of the driving point at the spring, this system can be described by a system of ordinary differential equations

t

=0

vk ( )d

(3.69)

or, after a Laplace transformation, in the s-domain as f (s) = msvm (s) = k vk (s) . s (3.70)

Let us recapitulate the situation. In both examples we have some distinct force-velocity dependencies, mass m, spring k and friction fm (s) = msvm (s) fk (s) = k vk (s) s f (s) = v (s). (3.71)

In the rst example these three distinct elements are coupled by the condition that the three velocities vm , vk and v are equal to the velocity of the external driving point vx while the forces including the external driving force fx are balancing each other. In the second example the distinct elements are coupled by the condition that the element forces fm and fk are equal to the external driving force fx while the velocities are balancing each other. Care must be taken in any case to determine the right signs for the quantities balancing each other. Internal elements absorb and store energy, their

166

force-velocity product must be negative. At an external port that is driven from outside, the force-velocity product must be positive as energy is supplied to the combination. These coupling conditions can be taken into account by formulating scattering expressionsfor the set of wave variables a and b or alternatively for a and The versions for the common-force and common-velocity b. cases can be easily derived with Equations 3.59 or 3.61 from the conditions f1 (t) = f2 (t) = f3 (t) = . . . = fN (t)

N

vi (t) = 0

(3.72)

N

fi (t) = 0.

(3.73)

The coupling element is usually referred to as an adaptor. It provides signal pairs an and bn for the N elements connected to it. Here the index n designates the port number. The time relationship of all port signals is instantaneous. Time delays must be present in the elements that are coupled to the adaptor. The adaptor expression for a common-force connection of N elements is for the nth port bn = an + 2

N

N j=1 Gj j=1

Gj aj

n = 1...N

(3.74)

with Gj = 1/Rj . A similar expression for the nth port of a commonvelocity connection of N elements is obtained as bn = an 2Rn

N

N j=1 Rj j=1

aj

n = 1 . . . N.

(3.75)

167

3. Specic Methods

where b = (b1 , . . . , bN )T and a = (a1 , . . . , aN )T and Sf = I + 1 GT f I Rv 1T common force common velocity (3.77) (3.78) Sv = 2

N j=1 Gj

Rv =

2

N j=1 Rj

For power-waves the scattering conditions are b=Sa with Sf = I + I Gf Rv GT f RT v common force common velocity (3.82) (3.83) (3.81)

Sv =

It can be shown that in all cases the scattering process conserves the power. In other words, the total input power of all aj is equal to the total output power of all bj . In the case of power waves this relation is direct, i.e. without any proportionality factor related to the port resistances R . j The usual circuit representation of a WDF-adaptor is illustrated in Figure 3.6. The corresponding signal ow graphs are shown in Figure 3.7. With these tools, the signal owcharts of our two simple examples can be drawn and implemented directly. Evaluation is straightforward. In the scattering step, the new input values for the basic elements are calculated. Then the elements themselves can be evaluated for the next time step tn+1 . With the new values another scattering step can be performed and so on. It has to be noted that real world problems are not always as straight forward as the examples presented here. If a common-force adaptor is connected to a common-velocity adaptor, then zero-delay loops are created.

168

a1

b1

a1

b1

b2 a2

b2 a2

b3

a3

b3

a3

a1

b1

a1

b1

G1

2 2

b2

2

b2

1 Gj

R2 2

R1

1 Rj

a2

G2

a2

R3 2

G3

b3

a3

b3

a3

169

3. Specic Methods

As it was said already, adaptors have an instantaneous time relation; the outputs bi depend directly and without delay on the inputs ai . Fortunately it is usually possible to get rid of delay-free loops between adaptors. If the port resistance Rm at port m of a common-velocity adaptor is chosen such that it is the sum of all other port resistances N j=1,j=m Rj , then the corresponding output bm does no longer depend on its input am . It is merely the sum of all other inputs aj . Such a port is said to be reectionfree. In the circuit representation this fact is usually marked by a short cross-line indicating a dead end at the corresponding input. In a common-force adaptor a reection-free port m is obtained if the port admittance Gm is chosen such that it is the sum of all other port admittances N j=1,j=m Gj . Again the corresponding output bm does not depend on am .

a b

-N

b a

T-line, R=Zw z

-N

Flow chart diagrams, similar to those which are shown here, have been published by Karjalainen in [Kar03]. In this paper the close relationship between digital wave guides and wave digital lters has been worked out. The implementation of a unied and modular system is presented, which allows free interchange between WDF elements and digital wave guide elements. The basic digital wave guide element is shown in Figure 3.8. It represents a bidirectional delay line with a signal delay of N times the sampling period. It is fully compatible with all the WDF-elements and can be connected to any adaptor port just like other elements. The transmission line network described in Section 3.2.4 could be implemented using a regular

170

grid of adaptors and T-line elements. If delays have to be more accurate than integer multiples of the primary sampling period, fractional delay lters have to be implemented. Fractional delays are usually implemented as all-pass networks.

Combining the efciency of frequency-domain wave guide modelling and the accuracy of a nite difference method in the time-domain was proposed by Noreland [Nor02], to calculate the input impedance of brass instruments. The rapidly aring bell of the instrument and a part of its environment was simulated in the time-domain. A coordinate transformation was applied in order to obtain a coordinate grid better reecting the bells geometry. A pulse response of the bells input impedance was computed and transformed into the frequency-domain. It was used as terminating impedance for the one dimensional wave guide model of the rest of the instrument. Very accurate results have been obtained because the plane wave assumption was only made for slowly aring and narrow parts of the instrument. Even the radiation impedance was not approximated by the usual innite bafe model but contained the effect of the real environmental condition. A hybrid approach such as the one described above is especially useful when instruments are to be optimized. Usually the bell region is not allowed to be modied by the optimizer, so it does not need to be recalculated at each optimization step.

171

3. Specic Methods

Accurate bore information of tubular objects is essential for the analysis of their acoustical characteristics or their uid propagation properties. Especially if the duct is folded or bent it is often difcult to get its dimensions by mechanical measurements. If the cross-section is not perfectly circular an additional difculty arises. In this case acoustically or uid dynamically equivalent circular cross-sections are often required for subsequent mathematical analysis. Bore reconstruction starting from acoustical measurements is a desirable approach applicable to many technical elds. Of special interest is such an approach in the eld of musical acoustics where actual dimensions of wind instruments are required whenever such an instrument is to be documented, analyzed, improved or repaired. Their tubular length of up to 6 m and more, together with their complicated 3-dimensional folding and bending, as well as their signicant deviations from perfectly circular cross-sections almost rules out attempts using calliper and gauges. As described in Section 3.1 it is possible to calculate the input impedance function of a brass wind instrument from its geometrical dimensions and some assumptions about the termination impedance which describes the transition of the guided wave into the open space of the environment. Depending on the boundary condition at the near end there is some method to calculate the pulse response at the reecting entry plane (the closed lips) or the reection function at the same place when the instrument is coupled to a sound source in a non-reecting way. The inverse problem is to calculate the geometrical dimensions from data which have been obtained by acoustical measurements usually made at the leadpipe or mouthpiece entry of a brass wind instrument either in the time-domain or in the frequency-domain. Solving the inverse problem in the context of musical instruments has been tackled by N. Amir [ARS94] and D. Sharp [Sha96] starting from the input reection function measured in the time-domain [BS81] with a nonreecting sound source connected to the near end. A pulse is transmitted

172

into the instrument and all received reections arriving at the entry level are recorded. A recursive algorithm known as layer peeling algorithm is then applied to the reection function to get a table with equidistant bore diameters. The length resolution is directly related to the time resolution (sampling frequency) of the reection function and the speed of sound.

As described in Section 3.1.1 plane waves travelling in a one dimensional wave guide are scattered at discontinuities of the cross-sectional area. A part of the wave is reected and only the remaining part is transmitted. The reection and transmission coefcients r and t are only related to the ratio of the cross-sectional areas left and right of the discontinuity as given by equations 3.8 and 3.9. Between the discontinuities waves are propagating in both directions with the speed of sound. In the lossy case this is also connected with frequency dependent attenuation in amplitude and some phase modulation as expressed by a complex valued wave number. The discretization of the loss-less case is shown in Figure 3.9. The scattering coefcients r and t are given for waves travelling from left to right. In the opposite direction the reection coefcient r1 = r and the transmission coefcient t1 = 1 r 1 = 1 + r b which can be veried easily by swapping the indices of S. The time axis has been drawn on the left side with T0 at the top. It is assumed that a discrete Dirac pulse is entering the rst cylindrical section at time T0 . At time T1 = T0 + t the pulse has travelled a distance x = cs t with cs being the speed of sound. Now scattering takes place. Depending on the change in cross-sectional area a certain fraction of the incident wave is reected and only the remaining part is propagated. Now a recursive process will start. The transmitted part will be scattered again after another time interval t when it reaches the end of the second cylindrical section where the spatial discretization has concentrated all the changes in cross-sectional area of this region.

173

3. Specic Methods

r1 = (S1S2)/(S1+S2) t1=1 r1 r2 = (S2S3)/(S2+S3) b1=1 + r1 t2=1 r2 r3 = (S3S4)/(S3+S4) b2=1 + r2 t3=1 r3 r4 = (S4S5)/(S4+S5) b3=1 + r3 t4=1 r4 r5 = (S5S6)/(S5+S6) b4=1 + r4 t1 t5=1 r5 b5=1 + r5 t1 r 2 t1 t2 t1 r2 r1 t1 t2 r3 b2 t1 r22 r1 t1 t2 r3 b2 b1 t1 r22 r1 b1 t1 t2 r3 b2 r1 +t1 r22 r12 t1 t2 r 3 t1 t2 r3 r2 t1 r2 r1 t2 t1 t2 t3 r4 b3 t1 t2 r32 r2 t1 r2 r1 t2 r3 t1 t2 t3 t1 t2 t3 r 4 t1 t2 t3 r4 r3 t1 t2 r3 r2 t3 t1 r2 r1 t2 t3 t1 t2 t3 t4 t1 t2 t3 t4 r 5 t1 t2 t3 t4 t5

T0 T2 T4 T6

1 r1 t1 r2 b1

X0

X1

X2

X3

X4

X5

The reected pulse will return to the entry plane at T2 where it can be observed as second sample of the pulse response function. The length resolution in axial direction which can be achieved is therefore directly related to the sampling frequency of the pulse response according to f = s 1/(T2 T0 ) = cs /(2x) or x = cs /(2fs ). Reection at the entry plane back into the duct must be avoided, so no change in cross-sectional area is allowed at X0 . This must be taken care of as actual measurements are being made. The measured object must either be coupled non-reectively to a source tube which should ideally be of innite length, or the reections caused by the coupling to the sound source must be removed analytically from the pulse response obtained by such a measurement. Inspecting the signal received at the left end X0 of the measured object at times T2 , T4 , T6 ,. . . it becomes immediately obvious that the unknown reection coefcients r1 , r2 , r3 . . . and their related counterparts ti and bi

174

can easily be derived from the pulse response. The response at T2 is equal to r1 . From this t1 = 1 r1 and b1 = 1 + r1 can be calculated as well as the unknown cross-section S2 = S1 (1 r1 )/(1 + r1 ) assuming S1 is known. Substituting the known quantities into the expression for the response at T4 , the next reection coefcient can be determined. By repeating this procedure all the originally unknown Si could be reconstructed. This procedure might work for a small number of reecting discontinuities, resulting in a pulse response containing only a few signicant events. Originally this method was developed by geologists for the study of the earths crust with its geologic layers of different density, especially for the exploration of oil. In order to make the method applicable to distributed reecting areas, in our case non-cylindrical segments like cones, it must be modied so that higher resolutions in time and space can be handled. For this purpose, the pulse response has to be split into a pulse sequence travelling from left to right and one travelling backwards. Both signals are sampled at the same sampling frequency and must have the same number of samples. Initially we are starting with the Dirac pulse so the forward travelling sequence has a one and many zeros. The backward travelling sequence can be initialized with the recorded pulse response. The forward travelling Dirac pulse, which has been created as a stimulus signal, will most probably be also recorded by the microphone. It marks the beginning of the sequence for synchronization purpose, but as it is not part of the backward travelling wave it has to be deleted from the recorded sequence. The forward sequence starts at T0 , the rst sample being the only nonzero sample of the stimulus signal. The backward sequence should start at T2 , the rst sample being the very rst reection of the stimulus pulse at the rst axial grid point X1 . The ratio of the rst samples of the outgoing and incoming sequence is the rst unknown reection factor r1 . S1 must be known, so it can be used to determine S2 . By means of the scattering equation for a scalar signal u

175

3. Specic Methods

and its directional components at the discontinuities 0 1 and 1 2 u1,2 u2,1 = 1 1 r1 u0,1 1 r1 r1 1 u1,0 (3.84)

we can transform the forward travelling sequence known at X0 into a corresponding sequence which we could have measured at the cross-section X1 , a time step later. The same way we can proceed with the backward travelling sequence we have measured at X0 . After the transformation we have the signal which we could have measured at the cross-section X , a 1 time step earlier. We delete the rst sample of the backward travelling sequence as this was already processed to obtain r1 . We also delete the last sample of the forward travelling sequence because the remaining unknown length of the duct is now shorter by one slice. Actually we have obtained a new virtual starting position one slice inside of the unknown duct, with forward and backward travelling wave sequences equal to those that we would have measured at the entry of the remaining still unknown part of the duct, at least if we would have been able to properly separate the forward and backward parts of a recorded signal. We have virtually peeled off the rst layer of the problem and we can repeat this step with all samples of our wave sequences, each step delivering one unknown cross-sectional dimension. The scattering Equation 3.84 which we have used is easy to understand. It is the solution of the two equations describing the superposition of a transmitted and reected part of the waves travelling in both directions: u1,2 = (1 r1 )u0,1 r1 u2,1 u1,0 = r1 u0,1 + (1 + r1 )u2,1 (3.85)

Including losses in the layer peeling algorithm can be accomplished as proposed by Amir et.al. [ARS94] based on a frequency-domain formula-

176

tion due to Keefe [Kee84]. The idea was to replace the simple time shift between the scattering procedures which represents the propagation delay of waves in cylindrical wave guides by a more sophisticated model. In the frequency-domain, the effect of losses can be described in terms of a complex valued wave number as f.e. given in Equation 3.24. The rotating phase model given by Keefe can be written according to [Kem02] as k = + j (3.86) with = /vp and being the frequency dependent attenuation, vp the phase speed and = 2f . The expressions given in [Kee84] are: = c 1+ A C 3 rv rv = c A B C + 2+ 3 rv rv rv (3.87)

with rv = R ( /) being the normalized boundary layer thickness depending on R, the radius, , the density and , the viscosity of air. A,B and C can also be related to thermodynamical constants according to: t t2 1+t B =1+t (3.88) A= 2 2 2 and 1 C= 2 with t c = = 7 t t t2 t3 t2 +t 2 + + 8 2 8 2 2 2 ( 1)/ cp / 0.8410 (1 0.0002T ) 1.4017 (1 0.00002T ) 1.846 105 (1 + 0.0025T )Pa s 1.1769 (1 0.00335T )kg/m3 347.23 (1 + 0.00166T )m/s () = ejkx . . (3.89)

(3.90)

177

3. Specic Methods

This lossy propagation must be modelled by a digital lter which can be applied to the forward travelling waves. The backward travelling waves have to be transformed by the inverse lter function, so both topologies are needed. In general it is not ensured that an inverse lter exists for arbitrary lter functions. Practically, typical attenuation and rotating phase lters as well as all-pass lters do have an inverse. The simplest implementation is to tabulate the real and imaginary part of the transmission coefcient over frequency with a frequency increment matching that of the FFT of the forward and backward wave sequences. The ltering function in the time-domain is actually a convolution with an FIR lter core and can therefore be performed as multiplication in the frequency-domain. The inverse lter being a deconvolution can be performed as division in the frequency-domain. The product resp. the quotient has to be back-transformed into the time-domain before the next scattering step can be performed.

There is a unique and reversible relationship between the complex-valued input impedance and the pulse response of an acoustical system, so the idea came up to use the input impedance function instead of the reection function as the starting point for bore reconstruction. Frequency-domain measurements are often easier to accomplish and there exist ready measurement systems for input impedance which are being used by brass wind instrument manufacturing companies for quality control and engineering [Wid99, Wid95, WW94, Wid93, PWW92, WPO89]. Basically it is possible to calculate the required reection function once the complex input impedance spectrum has been measured. A transformation of the measured spectrum into the time-domain will yield the instruments pulse response, but the reecting termination at the near end will add periodic components, which are not present when measuring with a matched, non reecting sound source. With such data the layer peeling algorithm, as it is known today, does not work. So another transformation has to be applied rst. It is a bilin-

178

ear transformation IIR(e

j

)=

1 1+

Z0 Zin ej Z0 Zin ej

(3.92)

which performs a sort of unrolling in the spectral domain, removing additional periodic reections, caused by the near end termination. The inverse Fourier transform of the spectrum IIR(ej ), being the discretized frequency, is the reection function, which can serve as input to the layer peeling algorithm. Practical experiments using this procedure have shown that this approach is not straight forward. The initial transformation is very sensitive to the value of Z0 , the characteristic impedance of the entry cross-section. It also requires accurate impedance measurements in the low frequency range, which is difcult to achieve. One will not gain anything from a high resolution in the frequency-domain, because this would only increase the reconstruction length, which is already determined by the physical length of the instrument. A good resolution of the reconstruction grid requires a high sampling rate of the reection function, but high bandwidth measurements are difcult. First, because at higher frequencies more energy is radiated and less energy is reected to the input, and second, at higher frequencies plane wave assumptions are no longer valid and higher order modes must be taken into account. The higher the frequencies are, the smaller the discontinuities can be to signicantly disturb the plane wave assumption. On top of that, the recursive layer peeling algorithm exhibits numerical weaknesses and it accumulates any numerical errors as well as errors in its input data. Many of these difculties can be avoided if an iterative approach is chosen. Bore reconstruction using optimization techniques does not require an explicit solution of the inverse problem and can therefore handle much more sophisticated models of wave propagation in the time-domain as well as in the frequency-domain. Even hybrid models are possible and recommendable. The approach is described in more detail in Section 4.2 where

179

3. Specic Methods

practical reconstruction examples are presented and results can be found.

180

4. Applications

This chapter presents results and practical aspects of the application of computational methods to typical simulation problems in the eld of musical acoustics. Musical acoustics does not just deal with sound. It deals with intentionally created expressive sounds, in other words with music, its performance, production, propagation, perception and impression. It is, of course, centered around musical instruments, made by instrument makers following a certain tradition or by machines developed and programmed to meet certain quality standards and production tests. The instrument is means of musicians to communicate with their audience on an acoustical, perceptional, emotional and intellectual level. Musical acoustics typically includes areas of physiological acoustics, psycho-acoustics, music performance and perception, perhaps some concert hall acoustics and, of course, the acoustics of musical instruments that represent the focus of this book. Computational methods which have been presented here are usually related to the analysis and optimization of sound production in musical instruments, the synthesis and analysis of musical sounds and systematic approaches for designing, improving, diagnosing and repairing musical instruments. For simulations of sound production and propagation in musical instruments there are typically two discernible aims which are motivating scientists to write computer programs. The rst aim is to understand the mechanisms of sound production well enough in order to help builders and players of musical instruments to better control sound characteristics or to improve playing techniques. For this purpose accuracy of sound modelling is not primarily relevant, yet may prove validity and relevance of the modelled physical effects.

181

4. Applications

For the sake of understanding the basic interactions between player and instrument, initial models tend to be as simple as possible. With growing understanding, more higher-order effects are usually included, yielding a more and more realistic synthetic sound and explaining more of the possibilities of interaction between artist and instrument. A second aim is the production of virtual instrument sounds, possibly in real time and as realistic as possible, which would offer all the natural expressivity to composers or performers. The end justies the means - the implemented algorithms do not absolutely need to be physical but have to be computationally efcient. Many successful results have been worked out by J.O. Smith III and his colleagues in Stanford e.g. [SI03] the audio examples of which can be found on the internet. Only a few of many possible representative applications have been included in this chapter. Some of them are reviewed briey, three others which are concerning actual research projects of the author and his colleagues are elaborated in greater detail. The rst section deals with sound production mechanisms in brass wind instruments. After a review of some actual publications in this eld it is demonstrated, how a simulation of the whole system consisting of a players lung, his vocal tract, the mouth cavity, the vibrating lips, the mouth piece and the instrument itself can be carried out. It was noticed by observations that lips do not oscillate like swinging or sliding doors. Stroboscopic studies revealed complex multi-dimensional wave patterns travelling on the surface of the lips which have been recently studied by Yoshikawa and Muto in great detail [YM03]. Therefore a two dimensional distributed lip model [Kau03b] which is able to exhibit surface waves travelling between the teeth and the mouth piece rim is proposed. Surface waves interact with forces originating from the pressures in the mouth, in the mouthpiece and in the lip orice which is taken care of by the model. Phase speed depends on the tension of the lip and is the independent simulation parameter. The actual simulation was carried out in the time-domain using the electrical circuit simulator SPICE. The equivalent circuit is presented and dis-

182

cussed. Resulting wave forms are shown and pressure signals and lip displacements at several points along the lip surface cross-section between teeth and mouthpiece rim are plotted over time. Mode transitions can be observed in the simulation when a glissando over the rst four natural tones is played. Surface waves on the lip can be observed in an animation sequence of the upper and lower lip cross-section. The second section of this chapter deals with so-called bore-reconstruction. Brass wind instruments are basically tubes with varying nearly circular cross-sections. The fact that most brasses are bent and folded in a way to get their total length of up to six meters into a reasonably small and practically manageable volume does not matter too much. The curve of the acoustically effective inner or open cross-sectional area over the distance from the mouth piece plane determines most of the characteristics of the instrument. Bore reconstruction is an attempt to reconstruct this varying cross-section, the bore, by means of an acoustical measurement. One way would be to do an impulse response measurement in the mouth piece plane and to reconstruct the geometry analytically from the recorded echo sequence using a recursive algorithm called layer peeling algorithm which has been described already. The approach described in this section is a different one. It starts with a measured input impedance curve of the instrument, and performs an impedance matching optimization of the measured curve with a theoretical curve being recalculated during each optimization step according to a frequency-domain wave guide model as described in Section 3.1. If appropriate matching can be achieved, the optimum bore sequence must be more or less identical with the unknown geometry which was to be reconstructed. The problematic nature of lumped models for the description of the sound generation mechanism in ue instruments has been worked out by Fabre and Hirschberg in their comprehensive paper: Physical modelling of ue instruments: a review of lumped models [FH00]. Therefore, the third section deals with computational uid dynamics

183

4. Applications

applied to wind instruments. A review of some recent publications in this eld is presented. Of especial interest are simulations of air-jets interacting with acoustical elds, vortex generated sound, edge-tones and selfsustained oscillations in utes and ue-pipes. Finally, the Lattice-Boltzmann method is introduced as a means to simulate the sound generation mechanism of ute instruments. Under rather realistic conditions and with reasonable assumptions for airspeed and ute dimensions a three dimensional simulation shows stable and expectable results, including vortex shedding, observable jet-acoustic eld interactions, mode transitions and a nearly realistic sound spectrum. Intermediate results have already been published by K hnelt [K h03]. u u Actual work at the Institut f. Wiener Klangstil is being done to further improve resolution, stability and realism of the model in order to permit bigger simulation spaces, to approach the true viscosity of air and to introduce more sophisticated boundary conditions like oscillating walls and non-reective outow boundaries.

In order to understand the mechanism of sound production in brass wind instruments, valuable theoretical and experimental work has been published by several authors. They proposed to model lips as outward or inward striking doors [Hel63], opened by positive or negative pressure differences between mouth and mouthpiece, or as sliding doors, driven by the Bernoulli pressure between the lips [STY87]. All three simple models exhibit self sustained oscillations when combined with a pressure source (lung) and an input impedance (instrument). An elaborate study of the operation of the lip reed in brass wind instruments according to some simple models has been published by Adachi in [AS95]. In this article the dependence between the various pressures acting on the lip and the ow through the lip which is modulated by the open cross-section of the lip orice has already been reviewed very clearly. One

184

of the investigated lip models is the so called transversal model, where the lip valve is modelled by a spring-mass system activated only by the Bernoulli pressure between the lips. The pressure drop between mouth cavity and lip orice has been derived from momentum and energy conservation laws assuming laminar ow. The pressure drop between the lip orice and the mouthpiece layer, where a plane wave has already been established, has been derived from the momentum conservation law only. It was assumed that a jet is formed in this region and energy is dissipated by the jet. Anyhow, the fact that a human player can easily play above as well as below the air resonances of an instrument cannot be reproduced by any of the simple models [Cam99, CG87]. As was shown by Campbells group, who are intensively experimenting with an articial mouth, that the lip oscillator cannot be reduced to one dominant mechanical resonance [RCG03, CC00, BRC03]. At least two resonances with different phase characteristics are required to explain the fact that stable self sustained oscillations can take place below and above the air columns resonance. Therefore the simple model has to be extended by at least one more degree of freedom. A combination of two simple models (stretchable outward striking door) has been proposed by Adachi [AS96]. While the stretching action of the lip is still controlled by the Bernoulli pressure in the lip orice, a swinging motion of the lip is activated by the pressure difference between mouth and mouthpiece. This combined model exhibits the observed bi-directional pitch controllability at least for the two lowest played notes. Results of a system simulation involving the pulse response of a real trumpet have also been presented. Trumpet sound was generated and claimed to be quite realistic. Some observations observed by trumpet players have been correctly reproduced. Simulated wave forms are in good agreement with measured quantities. Nevertheless, the fact that players can easily lip notes up and down from the center of their corresponding air column resonance is still an open modelling issue. Especially the idea of a regime of oscillation [CG87] meaning that

185

4. Applications

higher harmonics do contribute in establishing a self sustained oscillation, is very important. It deals with the fact that players can play tones even when there is no natural resonance of the air column present. A realistic model should explain the mechanisms behind this phenomenon. All these open issues have been motivation enough to think about a system simulation environment which might help to nd some answers to these questions. The idea was to combine partial models of rather different types and level into a single simulation environment which can do timedomain transient analysis as well as frequency-domain small signal and stability analysis. In order to be able to include complex and more realistic therefore highly non-linear models, a simulation environment based on electrical equivalent circuits has been chosen. A one dimensional distributed wave guide model has been used in order to get familiar with the simulation environment and with the electro-acoustical analogies [Kau02]. The electro-acoustical analogy has been used in order to be able to make use of SPICE, a public domain circuit simulator, originally developed at the University of Berkeley, which is well-known among electrical engineers. It solves the big non-linear systems of partial differential equations associated with complex electrical networks consisting of hundreds and even thousands of resistors, capacitors, inductors, transformers, transmission lines and various other linear and non-linear controlled sources of voltages or currents. The driving force behind the more than thirty years of development of circuit simulators like SPICE was and is the industry developing and producing integrated circuits of higher and higher complexity. State of the art circuit simulators are therefore based on most advanced computational concepts implemented with all the current knowledge of numerical mathematics and are containing the whole tool-set of contemporary computational physics. The sophisticated component models which are implemented in the more recent derivates of SPICE can be used to model almost any nonlinear partial differential equation which might be part of an actual sim-

186

ulation problem. They can be described by customizing some model parameters like is done with lossy transmission lines, switches and some dependent voltage/current sources. Other components can be characterized by entering polynomial coefcients or even arbitrary expressions of any number of other circuit variables. Still other components accept tables, either related to time-domain, frequency-domain or to the z-domain of sampled systems. Small signal analysis in the frequency-domain, operating point analysis, transient analysis in the time-domain and even sensitivity analysis can most efciently be made with the same equivalent circuit. Considering the fact that engineers are nowadays simulating integrated circuits with thousands of non-linear elements (e.g. transistors) with such circuit simulators, it will become clear that almost any model of vibrating lips can be implemented with little effort. This section can also serve as an example how an actual acoustical and mechanical model described by a set of differential equations can be transformed into an electrical equivalent circuit, ready to be analyzed in frequency or time-domain.

Although it seems obvious at the rst glance that lip reeds of brass wind instruments are operating like outward striking doors, opening with the blowing pressure, it was found that some experimental evidence is in contradiction to that simple assumption [CG87]. According to that model, tones would always sound sharper than the corresponding air resonance and it would be impossible to lip them down. To excite air resonances, lip resonances are always required to be slightly lower. Treating the lip as inward striking door tending to close with increased mouth pressure would yield the opposite result. A lip tension resulting in a lip resonance well above the air column resonance would be required and the sounding pitch would always be lower than the air resonance. Experimental evidence indicates that both mechanisms are in action simultaneously and good players seem to be able to control which effect is

187

4. Applications

dominating. Anyhow, it is not straight forward to understand how the lip orice can be closed by a higher blowing pressure. In 1982 Elliott and Bowsher [EB82] raised the question about the role of the time-varying Bernoulli pressure, which has the ability to close the valve by applying a force perpendicular to the direction of air ow. This idea was elaborated by Saneyoshi et al. [STY87] introducing a third concept, which is now referred to as the transverse model. The three concepts are illustrated in Figure 4.1, which was taken from the mentioned publications of Adachi [AS95, AS96], where a much deeper investigation of these models can be found. In [AS95], a direct comparison between the sliding and outward-striking door models has been made. In [AS96], Adachi produced simulated trumpet sound using a model combining both effects.

p0

(a)

p0

(b)

p0

(c)

A linear theory of self-oscillation was published, reviewed and extended in [Fle79, EB82, AS96]. A unied discussion of oscillation conditions for all three classes of oscillators was presented by Fletcher in [Fle93]. The inward striking door model of Figure 4.1 (a) is usually applied to woodwind instruments like clarinet and oboe. The door closes just because of a positive pressure difference between upstream and downstream region. The reed resonance must always be higher than the played frequency in order to meet the phase condition required for self sustained oscillation. The outward striking door model of Figure 4.1 (b) was originally applied to lip reed instruments. The door opens with a positive pressure difference between upstream and downstream region. The lip reed resonance

188

must always be lower in order to meet the - phase condition.

The transverse model [AS95] of Figure 4.1 (c) is related to the inward striking door model. Closing the valve is accomplished by the Bernoulli force due to the underpressure caused by quickly owing uids. Stationary, the Bernoulli pressure plip is proportional to the square of the air ow U through the area of the lip orice Slip according to p0 plip = U 2 1 d U ( , ) + 2 Slip Slip t (4.1)

where is the average air density, d the length of the ow path through the lip orice and p0 the blowing pressure. On the other side the Bernoulli pressure plip is connected to the mouthpiece pressure p by plip p = U 2 ( 1 1 2 ), Scup Slip Scup (4.2)

with Scup as the cross-sectional area of the mouthpiece entryway. Adding Equation 4.1 and Equation 4.2 yields an equation for the airow U . After the decay of some transients (neglecting the inertia) it will become proportional to the square root of the pressure difference p p 0 between mouth and mouthpiece: U = Scup Slip

2 (Scup

(4.3)

If the sliding door tends to close with increasing air ow U , it closes with increasing pressure difference p0 p. This explains why the transverse model is considered to show similar behavior like the inward striking reed [EB82]. In [AS95] f.e. it exhibits self sustained oscillations only below the corresponding air resonances even if the lip resonance frequency is well above

189

4. Applications

the sounding frequency. The center of the air resonance is in fact an upper bound for lip tension induced pitch variations. This effect will become understandable by studying the results of a stability analysis. Actually it is not in conict with the experience of many trumpet players and instrument makers, especially of those who are building natural horns. While it is comparably easy to lip a note down and play it at compared to the pitch offered by the air-column resonance of the instrument, it is much harder to play a note sharp. There are trumpet players reporting that it is not possible to lip a note up without changing the shape of the mouth cavity or the coupling between mouth and mouthpiece. Without entering into arguments here it should be stated that it is obviously much easier to play at than to play sharp. If the door tends to open with increasing pressure difference p0 p then the characteristics of an outward striking door can be achieved. Self sustained oscillations will occur above the corresponding air resonances with lip admittance peaks well below the sounding frequencies. The airresonance frequency is here a lower bound for pitch variations induced by lip tension modications. In a rst approximation the mouth pressure is often assumed to be nearly constant and the Bernoulli pressure in the lip orice usually closely follows the pressure in the mouthpiece. The force driving a one-dimensionally modelled lip valve can therefore be associated with the pressure between the lips. If the opening force is proportional to the lip pressure p we will get inward striking or sliding door characteristics. If the opening force is linked to the negative lip pressure p we will get outward striking behavior. This simplication ignores the fact that in a swinging door model the direction of the force is initially parallel to the ow and not perpendicular as in a transversal model. Therefore the relation between the effective opening area and the lip mass velocity is not described correctly especially for small amplitudes of the lip displacement. But as this is not the only simplication of reality, we can not expect quantitatively correct results anyway. For now we have to be satised if we can get a principal understanding of

190

the underlying physical principles. One-Dimensional TLine-Model The system to be simulated consists of a players lung, his air column, his mouth cavity, the vibrating lips and an open Bb trumpet with mouthpiece. The electrical equivalent circuit is shown in Figure 4.2.

It is divided in three functional sub-circuits: The path of airow from lung to trumpet mouthpiece, the oscillating lips, and their effect on the airow. Little attention has been paid on the actual trumpet sound. It would be very easy to connect an additional element to the mouthpiece pressure pmp, which represents the complex transmission function of the trumpet - including even auditory acoustics, if one likes.

191

4. Applications

This element would be driven by a table of frequency, magnitude and phase triplets, just as the element H1, representing the trumpets input impedance. The element H1 converts an input current (the acoustical volume ow u) into an output voltage (the sound pressure level pmp in the mouthpiece) according to a given frequency-domain table, which has been calculated from an actual trumpet bore using the lossy transmission line model published by Mapes-Riordan [MR93] which is summarized in Section 3.1.

The magnitude spectrum is shown in Figure 4.3. The mapping between voltage [V ] and pressure [P a] and between current [A] and volume ow [unitf racm3 s] requires acoustic impedance [W ] to be specied following the mks convention. The air ow originates in the lung, driven by the pressure source V 1. For certain simulations the blowing pressure was varied between 1 and

192

6 kPa. As a vocal tract model the one proposed by Scavone [Sca03] could be implemented. It exhibits one major resonance which can be tuned in order to demonstrate the presence or absence of an inuence on the sounding pitch. A realistic throat resistance should be in the order of 4 M depending on throat area, the mouth cavity capacitance can be derived from C = V /(c2 ) with V being the volume, the density of air and c the speed of sound. This was not done for the simulations in this section, the effect of the mouth cavity was taken into account in later simulations. The elements E2, L1 and E3 represent the modulation of the ow by the time varying lip orice. Two distinct modulation regions have been differentiated: The upstream contraction region, i.e., the mouth cavity and the lip orice and the downstream expansion region, i.e., a thin region in the mouthpiece cup, where a jet is formed and energy is dissipated by turbulences until laminar ow is established again. The conservation laws for energy and momentum in the laminar contraction region and for momentum only in the turbulent expansion region give the required relations between the pressures pm (mouth pressure), plip (the Bernoulli pressure applying the driving force onto the lip) and pmp (the mouthpiece pressure, which is linked to the ow u by the instruments input impedance), the ow u and the area of the lip orice slip (r = air density, D = length of lip orice, SCUP = cross-sectional area at mouthpiece entry) as shown in equations 4.1 and 4.2. Electrically speaking, the voltage drop between pm and plip is caused by a nonlinear voltage source E2 and an inductor L1 with time dependent inductance L = D/slip and the voltage drop between plip and pmp is again caused by a nonlinear voltage source E3. In the simulation schematic slip denotes only the AC component of the time varying area, therefore the average lip opening X0 B (X0 = equilibrium lip opening distance, B = breadth of lip orice) has to be added. Minimum and maximum lip opening areas are limited to avoid area numbers less or equal to zero. The actual lip model is implemented by

193

4. Applications

the transmission line TLIP . It is excited by the voltage source E1, which now represents a force rather than a pressure (Volt = Newton) and which duplicates the Bernoulli pressure plip multiplied by the lip area D B. For the one dimensional simulations below the outward striking mode was selected. In previous publications of the author [Kau02, Kau03b] this fact was not quite clear, so the way these results are presented here should be taken for reference. Please apologize also a mistake in the linear stability section contained in [Kau03b] which has been corrected and extended in Section 4.1.2. Like any distributed dissipative mass / stiffness system, a transmission line conducts, damps and reects input pulses and modulates the input current by its complex impedance, which depends on termination, length and material properties. It exhibits several resonance modes and it can even be arranged in a network to represent 2-dimensional, even 3-dimensional, distributed oscillatory systems as was described in Section 3.2.4. Acoustically the currents have to be interpreted as surface or node velocities. The impedance or rather input admittance of the lip has to be estimated. Hints can be obtained from measurements of the opening area over frequency made on articial lips [BRC03, RCG03]. However these frequency-domain measurements have been made with chirp excitations recording the opening area of the lips. Due to the non-linear characteristic of the opening area itself which involves saturation on closing as well as on opening it is difcult to extract a simple mechanical admittance. In this simulation reasonable assumptions have been made to get a resonance quality according to [YM03] and the absolute scaling has been chosen in a way to get lip displacements in the expected order of magnitude. As a tuning parameter for the rst resonance mode the factor K has been introduced which scales the original length of the line. Mechanical admittance is the reciprocal value of impedance, it is therefore the minima not the maxima where self oscillations can occur. Lip velocity vlip has to be integrated (X1, gain=1, initial value=0) to get lip displacement xlip. In order not to integrate any DC offsets a simple high-pass lter has been

194

put in front. The resulting displacement is limited and multiplied with the breadth of the rectangular lip orice (X2, low=X0B, high=2X0B, factor=B) to get the AC lip orice opening area slip. Stepped Transient Analysis, 1D-Model

Eigenfrequencies vs. parameter K

2nd Imp. Min

[Hz]

1.60K

Trumpet Resonances

1.20K

0.80K

1000m K 6

0.40K

0.00K 500m

Looking at the results shown in Figure 4.5 it can be observed that mode selection takes place. All natural tones with the exception of the pedal tone (which is even difcult to play in reality) are played properly, when the rst anti-resonance of the lip oscillator is swept between about 100 Hz and 700 Hz. Comparing the lip resonance curves shown in Figure 4.4 with the sounding frequencies, it can be seen that the rst resonance frequency of the lip is always lower than the sounding trumpet resonance. This explains the

195

4. Applications

missing pedal tone, because a frequency well below 90 Hz has not been reached during simulation. It is interesting that the actually sounding frequency is nevertheless even higher than the trumpets resonance.

[Pa]

6.0K

[Hz]

1.00K

Sound frequency / amplitude vs. parameter K Sound frequency vs. parameter K (related to eigenfrequency)

7th resonance

4.8K

0.80K

6th resonance 5th resonance

3.6K

0.60K

4th resonance

2.4K

2nd resonance

1.2K

0.20K

Pedal tone

0.0K

0.00K

4

Press(pmp) Press(plip)

This fact can be recognized more easily by looking at Figure 4.7. It shows the sounding pitch related to the second and third air column resonance which has been plotted against the lip resonance frequency. There is no doubt that the sounding pitch is higher than the air column resonance frequency and that an even still lower lip resonance frequency is required to induce that stable oscillation. As already mentioned this behavior is usually attributed to the swinging door model. Comparing with Adachis results [AS95] it can be veried that the transversal model can act like a swinging-door model, when the mechanical lip admittance is driven by the negative lip pressure. In order to understand this fact better a linear stability analysis has been made. A further simulation detail is shown in Figure 4.6. By increasing the

196

-2.0K

-3.5K

50.000m pmp

52.800m

55.600m

58.400m

61.200m

64.000

400

350

300 Frequency [Hz]

rd

250

200

150

100 100

150

200

300

350

400

197

4. Applications

blowing pressure it is possible to switch to the next resonance mode without changing the tension of the lips, a fact that trumpet players probably know and most probably dislike. Stability Analysis, 1D-Model In order to exhibit self sustained oscillations in a closed feedback loop, it is necessary that the open loop gain of the linearized system exceeds unity at an oscillation frequency, where the phase shift is zero (2). The oscillation is stable if the open loop phase is positive below and negative above the oscillation frequency because only in this case the oscillation frequency will lock. If the static open loop gain is < 1, that means the circuits DC-gain is > 1 but inverting, then the circuit will unconditionally oscillate as soon as it is closed. The starting condition is a frequency below the resonance and a phase above zero. A positive phase shortens the cycle and raises the frequency until the phase is zero and the resonance frequency is reached. If the frequency is above resonance then a negative phase will lengthen the cycle and slow down the oscillation until zero phase is achieved again. In all other cases, a small temporary stimulus signal has to be applied to excite a self sustained oscillation. If a broadband excitation signal like a short pulse is applied, usually the frequency with the strongest oscillation condition will be triggered. Let us start the investigation of a lip reed oscillator according to the transversal model at the air pressure plip in the lip orice. This pressure applies a force to the lip perpendicular to its surface. This force is proportional to the area of the lip surface which is exposed to this pressure. The lip itself acts as a distributed massstiffness system with a complex input admittance function. A maximum of the input admittance magnitude corresponds to a mechanical resonance of the lip. In a one dimensional approximation it is assumed that this input admittance is identical at all points of the lip surface. The lip admittance converts the time varying force flip applied by the Bernoulli pressure into a surface velocity vlip , which can be integrated in

198

order to obtain the lip displacement xlip . This modulates the cross-section of the lip orice Slip . The cross-section of the lip orice modulates the acoustic ow U , which enters the mouthpiece, where it meets the input impedance of the instrument. The input impedance of the instrument converts the ow U into the mouthpiece pressure p. The pressures p0 and p together with the ow U and the area of the lip orice Slip must fulll Equation 4.3 (this neglects the inuence of the inertia of the air in the lip orice). Now a new Bernoulli pressure of the open loop can be calculated according to Equations 4.1. If there are frequencies meeting the oscillation condition, self sustained oscillations will occur.

The results of a linear stability analysis are shown in Figure 4.8. In the frequency range between 100 Hz and 400 Hz it displays phases (top)

199

4. Applications

and gains (bottom) of some intermediate signals along the open loop as described above, when the lip resonance is varied from below to above the air column resonance of about 230 Hz. The curves labelled ph(pmp/u) and db(pmp/u) represent phase and dB gain of the input impedance function of the instrument. The curves labelled vlip, slip, u and plip correspond to the signals vlip , the lip surface velocity, Slip , the cross-sectional area of the lip orice, U , the ow and plip , the resulting Bernoulli pressure. There is an oscillation condition at points of plip , where the gain is greater than 1, when the open loop phase crosses the bold 360 line. We can see that the phase of vlip below the lip resonance frequency is 90 to become 270 above it, crossing the x-axes not very steeply from above to below. The slope is moderate because signicant damping v has been assumed. We can dene a lip mobility plip which has a rislip ing velocity (acceleration phase) during negative resp. positive pressures depending on outward or inward striking mode and a falling velocity (retardation phase) during the other pressure polarity for frequencies below a resonance frequency. Above this resonance frequency, the opposite is true. The integration of vlip into xlip resp. Slip subtracts 90 from the loop phase, while U is almost in phase with Slip . The conversion of U to plip , mainly by the input impedance, has zero phase at the resonance, 90 below and 90 above. Much less damping is involved causing a steeper zero crossing. In total we have 90 open loop phase below both resonances and at about 450 above both of them. The 360 oscillation threshold can be reached safely when air column and lip cooperate. Neither of them can create an oscillation condition alone. Each of the partners can contribute 180 which leaves a stability margin of another 90 . The lip can trigger an oscillation only above the air resonance when the air has already added 90 phase shift. This is exactly what the time-domain simulation was showing. The small signal stability analysis is consistent with the simulation.

200

Comparison of Basic Lip Models It is still worthy to study the effect of different lip models in a more general way to learn what a lip model can contribute to the overall oscillation conditions. We have seen that the air-column rotates the open loop phase from +90 to 90 centered around its resonance frequency. To have an oscillation condition there, the rest of the feedback loop must contribute either zero or 360 . Studying the single mass model of other authors it can be found that it contributes 0 below and 180 above the lip resonance. This is due to a mechanical admittance corresponding to a phase rotation of 90 below above one single resonance that has to undergo one integration and 90 step in order to get the displacement instead of the velocity. As we will see later, this integration step is the reason for the unsymmetrical behavior around the air column resonance.

1 0.1 -0.5 0.5 -1 -1.5 0.1 -0.5 0.2 0.3 0.4 0.5 -2 -2.5 -3 0.2 0.3 0.4 0.5

The mobility curve of a single mass, single spring system, when force and damping act on the mass, is shown in Figure 4.9. It is based on the force balancing equation (4.4) mxs2 = mksx/Q + Fext kx with mass m = 1, spring k = 1, quality factor Q = 10, s = j2P if

201

4. Applications

and force Fext being the independent, displacement x being the dependent variable. With this model the open loop would have 90 below both resonances and 270 above both of them. The way to the oscillation condition is only 90 , a quarter circle. The lip can trigger the 0 oscillation condition below the air-resonance alone just as the air-column can, but there is no way to get oscillation conditions above the air resonance. If the single mass is driven by the negative lip pressure which is the case in a typical swinging door model but this can also be directly achieved in our transverse model, then the loop phase contributions are 180 below both and 360 above the mechanical resonance. This makes the total open loop phase 90 below both resonances and 450 above. The way to an oscillation condition is now 270 which is three quarters of a full circle long. Therefore it needs both resonances in order to achieve oscillation. There is no way to get oscillation below any of the to resonance frequencies. The different characteristics of the various approaches are caused by the phase diagram of the mechanical model of the lip orice. While a single mass-spring system opening the gap by positive lip pressures cannot sound above the air resonance, one opening with negative lip pressures cannot sound below. A lip model which can do both had to mirror the behavior of the air column. If it rotated the open loop phase from +90 to 90 centered around its own resonance frequency then it would be in full control of the sounding pitch. Below both resonances the loop would exhibit +180 . By adjusting the lip resonance slightly above both resonances 180 below or slightly above the air resonance, a pitch lower or higher that of the air-column could be achieved. Theoretically such a model can be constructed. But is such a behavior physical? Let us investigate other mass spring arrangements. First let us consider a single mass, single spring system where the force acts on the spring instead of the mass. As real lips are distributed massstiffness systems this view is equally reasonable. The notion that the force

202

acts on a solid lip which is xed into position by a spring is equally simplied as the notion that the force cannot act on the mass because there is elastic tissue in between. Figure 4.10 shows why the former notion is usually preferred. The latter concept creates a band reject lter which would not allow oscillations at its resonance frequency because of the low gain there.

2 1.5 -0.5 1 0.5 0.1 -0.5 -1 -1.5 -2 -3 0.2 0.3 0.4 0.5 -2 -2.5 -1 -1.5 0.1 0.2 0.3 0.4 0.5

The mobility curve of a single mass, single spring system, when the force acts on the spring, is based on the force balancing equation Fext = (x xm )k (4.5) mxm s2 = mkxm s/Q + k(x xm ) with mass m = 1, spring k = 1, quality factor Q = 10, s = j2P if and force Fext being the independent, displacement x being the dependent variable. xm reects the displacement of the mass which is not equal to the displacement of stimulation point. A combination of both characteristics is obtained when the mass is suspended between the force driven spring kf and a xed second spring km . The describing equations are mxm s2 = rxm s + kf (x xm ) xm km Fext = (x xm )kf (4.6)

with friction coefcient r = 0.1 of mass m = 1 and spring constants km = 0.5 and kf = 0.5 being used in the mobility curves shown in Figure 4.11.

203

4. Applications

0.1 1 -0.5 0.75 0.5 0.25 0.1 -0.25 -2.5 0.2 0.3 0.4 0.5 -1 -1.5 -2

0.2

0.3

0.4

0.5

1 -0.5 0.5 -1 0.1 -0.5 -1 -3 0.2 0.3 0.4 0.5 -1.5 -2 -2.5

0.1

0.2

0.3

0.4

0.5

204

But how about multi-mass systems? Can a more sophisticated arrangement of masses and springs behave the way we would like it to behave? A mass-spring-mass-spring model is described by m1 x1 s2 = r1 x1 s + k2 (x x1 ) x1 k1 m2 xs2 = r2 xs + Fext k2 (x x1 ). (4.7)

The corresponding mobility curves have been plotted in Figure 4.12 with inner mass m1 = 5, driven mass m2 = 0.5, springs k1 = 1 and k2 = 1 and friction coefcients of both masses r1 = 0.1 and r2 = 0.1. The distributed mass-stiffness system which was used for the timedomain simulations of this section is described by the the telegraph equations F () = (R + jL)v() v () = (G + jC)F (). (4.8)

Here they are formulated with force F () and velocity v(). is the position inside the distributed system with = 0 at the stimulation location and = L at the far end which is terminated by some impedance ZL . R and G are loss coefcients per unit length, L is associated with mass and C with stiffness per unit length. From this differential equation system the mechanical input impedance Z0 can be calculated if the boundary condition ZL at the far end L is known. With the abbreviations a = G + j C and b = R + j L we obtain ZL a b cosh(L a b) + b sinh(L a b) (4.9) Z0 = a b cosh(L a b) + ZL a sinh(L a b) For an free and oscillating far end (ZL = 0) the mechanical input admittance Y0 = 1/Z0 becomes Y0,f ree a = coth(L a b). b (4.10)

205

4. Applications

1 0.5 0.5 -0.5 -1 -1.5 -2 -2.5 -3 0.5 1 1.5 2 2.5 3 -1.5 -2 -2.5 1 1.5 2 2.5 3 -0.5 -1

Figure 4.13.: Displacement over Force, Distributed Mass-Stiffness System, Free End

The displacement of the stimulated end as response to a unit force has been plotted over frequency. The resulting curves are shown in Figure refgTLopen. Note that this requires integration of Y0 causing 90 phase rotation, because mechanical input admittance is dened as velocity over force. It can be seen that there is no fundamental difference compared with the single mass model in terms of open loop phase. What could be essential is the fact that there is not only one resonance but several which could interact with higher resonances of the air column. If a solid wall (the mouthpiece rim?) enforces a velocity node at the far end (ZL = ) we obtain Y0,f ixed a = tanh(L a b). b (4.11)

The updated plots are shown in Figure 4.14. The main difference is the shift of the resonance frequencies. Now we have resonances where we had anti-resonances before. But the phase rotation at resonance peaks is just the same as before. All these models can introduce a resonance condition only below the air-column resonance because none of them can add a positive phase.

206

1 0.5 0.5 -0.5 -1 -1.5 -2 -2.5 -3 0.5 1 1.5 2 2.5 3 -1.5 -2 -2.5 1 1.5 2 2.5 3 -0.5 -1

Figure 4.14.: Displacement over Force, Distributed Mass-Stiffness System, Fixed End

Lumped Model of Measured Lip Admittance Let us nally analyze something which is more realistic. Starting from the measured input admittance curve of an articial lip that was published by Richards, Campbell, Gilbert et.al. in [RCGN02], we can try to model the dominant poles and zeros by coupled harmonic oscillators. According to the authors three main resonances are essential in order to self-sustain oscillations below and above the center of an air-column resonance. A resonance with 90 phase rotation below the air-resonance is usually related to the outward striking mode. It should be involved when a sounding pitch higher than the air resonance is to be enforced. Another resonance signicantly higher than the air-column resonance rotates the loop phase by 270 at its admittance peak. It is related to the inward striking mode and said to be responsible for sounding pitches below the center of the air resonance. A third resonance, which is pretty well centered around the air-column resonance, has been modelled with the measured phase shift of 180 in order to well mimic the phase response in the frequency range in question. The lumped model and its frequency response are shown in Figures 4.15 and 4.16. The circuit is composed of three RLC impedance terms representing the three resonances. They are stimulated by the voltage f lip

207

4. Applications

-160. Artificial Lip, Frequency Response

166.08 ,-172.17

-205.

166Hz, -172dB

-250.

50 db(xlip3/flip3)

100

500

frq [Hz]

F

180.

0.

-180.

100 F

500

frq [Hz]

208

representing the force acting on the lip, which is derived from the lip pressure plip and the contact area D B. The three partial currents are added in a 1 shunt resistor where the output signal vlip can be taken from. This signal is replicated by the controlled voltage source vlip in order to create the required additional phase rotation around the center resonance.

Air Column

Open Loop

Outward Striking

Lip Model

Two integration stages nally introduce a constant phase shift of 180 that is demanded by the measured phase curve. At least the second one has a simple physical interpretation because it represents the relationship between the lip velocity and the displacement which is directly proportional to the cross-sectional area which was measured. In the range between 150 Hz and 250 Hz both amplitude and phase response are fairly good approximations of the published measurements.

209

4. Applications

The open loop stability analysis shown in Figure 4.17 is indeed different from that of the simpler models. The inward striking mode does cause oscillation conditions at the high side of air column resonances. But unfortunately not on the same but on the next one. There is no continuous transition and the pitch will jump to the high side of the next higher resonance when an attempt is made to pull a note up too far. But although many trumpet players might swear the correctness of this simulation, for they know too well, how easy it is to miss the note they want to play, it does still not quite reect the experimental evidence. Campbell et al. have observed continuous self sustained oscillation and a smooth transition from outward to inward striking mode while shifting the air-resonance of a trombone across the resonance frequencies of articial lips by pushing in the instruments slide. On the other side, it is evident that even the realistic lip admittance in conjunction with our simple model of its interaction with the lip pressure, cannot create such a behavior. The inward striking resonance is the highest of the three resonances of the articial lip. In order to produce a smooth transition from the low side to the high side of an air-column resonance, the outward striking mechanism has to act rst. It must anticipate the 90 phase rotation of the air-column by its own corresponding phase transition. This is the easy part. Due to their inherent 0 180 phase characteristic all lip-models discussed so far have this ability. The difcult part starts when the lip-model has to postpone the phase transition of the air-column resonance. The most natural way to do that would require a positive DC-phase of the lip admittance with a +90 90 transition at the resonance frequency. If the DC phase starts at 0 as this is the case with all analyzed lip models, a smooth pitch transition is almost impossible. It would require a very strange phase transition changing temporarily its direction at the resonance. Below the resonance a clockwise phase rotation would be required to

210

play the low side of the air-resonance. Above the resonance a counterclockwise phase rotation would be required in order to compensate the aircolumn resonance and push the self-oscillation point smoothly upwards. A smooth transition is even a difcult problem when a more sophisticated arrangement of the pressure-lip interaction is investigated. If the inward and outward striking oscillators are separated in a way to allow them to control their own opening areas, then the problem arises how to unite the different lip pairs again. In two mass models with two or more degrees of freedom a minimum relation is often used. The smaller opening area modulates the air ow, but this results in an area function which does not have a continuous derivative, which is a strict requirement for a linear stability analysis. In cases where the loop gain depends on functions with discontinuous derivatives only large-signal time-domain simulations can be used in order to obtain reliable predictions. Some concerns against typical two-mass models which have been initially used in vocal fold modelling e.g. [Luc93], have recently also been issued by Vilain, Hirschberg et.al. in [VPH+ 03]. They pointed out that positive coupling between the two mass-spring systems would violate the volume conservation condition of the lip while the assumption of negative coupling does not create a necessary oscillation condition. Minimum Model To conclude the section of one-dimensional lip modelling, in a nal experiment the model has been gradually simplied in order to identify components which essentially participate in establishing self-sustained oscillations of a brass players lips. The complexity of the circuit was reduced step by step until the model shown in Figure 4.18 has been obtained. With a pressure sign-condition according to an outward striking model the simulated trumpet player was still blowing at 242 Hz, a frequency slightly higher than the corresponding air resonance at 230 Hz with a lip resonance frequency of 190 Hz. The self sustained oscillations of all involved signals can be clearly

211

4. Applications

212

identied in Figure 4.19. A linear stability analysis is shown in Figure 4.20. In spite of all the simplications, the linear behavior is not much different than before. It reects the known stability properties of a typical outward striking model with a satised oscillation condition only above the air column resonances excited by a lip resonance well above that frequency.

Equivalent Circuit Extending the electrical equivalent circuit shown in Figure 4.2 a two dimensional lip model has been added. The crosssection of the lip surface has been modelled as a transmission line with two reecting terminations

213

4. Applications

as shown in Figure 4.21. One end is the point where the lip touches the teeth, the other one, where the lip is xed by the mouthpiece rim. The voltage waves travelling back and forth the transmission line between the two grounded end points are interpreted as Rayleigh type transversal surface waves, similar to those that can be seen when buzzing lips are observed under the light of a slightly desynchronized stroboscope as presented by Ayers et. al. in [ABE99].

The waves are stimulated by and interacting with forces related to the mouth pressure, the mouthpiece pressure, and the Bernoulli pressure at several equally spaced positions along the ow path through the lip orice. The forces are represented by currents while surface velocities correspond to voltages. The parameters of the transmission line have been chosen according to experimental data published by Yoshikawa and Muto in [YM03]. The phase speed of the surface waves is the independent parameter and is determined by the lip tension. The uid dynamical Equation 4.1 describing the relationship between pressure drop, laminar ow and cross-sectional area has been implemented ve times as shown in Figure 4.22. Equation 4.1 has been split into two parts. The part proportional to the derivative of the ow, which represents the inertia of the air between the lips, has been implemented as inductor with time varying inductance controlled by the lip displacement xlip . The remaining part has been modelled by a nonlinear voltage source, controlled by the quadratic term containing U and xlip . The lip displacement xlip can vary along the lip orice just like the pressure between the lips.

214

The expansion region, where a jet is formed and ow energy is dissipated, is again implemented as dependent voltage source controlled by Equation 4.2. The load impedance is a theoretical input impedance of a Bb-Trumpet specied in the frequency-domain. Results A transient analysis of the circuit described above has again been made with SPICE. In a rst attempt the built in lossy transmission line model was used for the segments between the points, where the pressures interact. Unfortunately this method does not allow varying the TL parameters during one transient analysis run. But it is possible to step them and run the simulation automatically for many different lip tensions. An example of the resulting wave forms is shown in Figure 4.24. In this case the pedal tone has been stimulated in fortissimo (6 kPa blowing pressure). The top graph contains the pressure wave forms starting with pm (mouth pressure) downstream to pmp (mouthpiece pressure). The pressure drop in between can be traced as signals pl1 to pl5, drawn with increasing gray value. With the assumptions used for the throat impedance and mouth capacitance, the mouth pressure looks still fairly constant at this playing condition.

215

4. Applications

The other graphs show lip surface velocities, displacements and lip opening areas at ve points along the lip orice. Anyhow, this approach does not simulate a real playing situation, where a player varies his embouchure in order to play a sequence of tones adjusting his intonation. Therefore, an equivalent circuit for a lossy transmission line has been built, similar to those that are used by telecom engineers to simulate the conditions of long telephone lines. With this approach it was possible to vary the values of lip surface phase speed and lip damping even during one simulation run. By continuously increasing the lip tension a whole harmonic series starting with the pedal tone has been produced. Figure 4.25 shows the time-domain signals of this musical event. Now the mouth and throat parameters have been adjusted to reasonable values, so even the mouth pressure exhibits oscillations in the expected

216

range. On top of the pressure signals the ow is plotted, appropriately scaled in order to make it visible. Below the displacements of the various lip surface points are plotted. The xlq signal is the quiescent lip opening distance, which is controlled by a circuit in order to keep the average lip opening area proportional to the amplitude. Ideally it should set the bias value such that the lips close just for a moment during any period. The negative signal parts of the xl i are clipped, before the area Slip is calculated. The average of the clipped signal parts is driving the circuit controlling xlq. Figure 4.26 shows the moving lip cross-section during one period of the rst resonance found in Figure 4.25. The left side is the entry side where the teeth are xing the lip edge. The right lip edge is xed by the mouthpiece rim. The ow magnitude is represented by the length of the arrow and the pressure distribution along the air path by the gray value

217

4. Applications

of the background. If the shape is animated, nice surface waves can be observed travelling along the lip orice.

To conclude the section about simulating sound generation mechanisms in brass wind instruments, an often neglected fact is to be mentioned here. While there is a common understanding that the operation of the lip reed oscillator cannot be described without taking the non-linear dependencies between air ow, air pressures, and cross-sectional area of the lip orice into account, the instrument itself is usually modelled by any application of the linear theory. There are good reasons for this. First, sound and especially music is considered to be a very small disturbance of the quiescent atmospheric

218

pressure, which validates the Taylor development leading to the linear theory. And second, because taking non-linear wave propagation into account, is usually a very expensive undertaking. Basically it means that Fourier decomposition, and therefore frequencydomain analysis, had to be given up. Only a few computational methods described in chapter one would remain applicable to the general formulations of the governing equations which describe non-linear propagation effects adequately. So most people relinquish that usually small range of operating conditions where linear theory is no longer applicable. On the other side, music can be a quite signicant disturbance of the quiescent atmospheric pressure, at least inside of some brass wind instruments when they are played in fortissimo. It is clear that the linear range of sound propagation is exceeded under such conditions. Wave steepening and the resulting formation of shock waves at high playing pressures has already been suspected to be the cause of the so called brassy sound of

219

4. Applications

horns, especially noticeable in the trombone. And, according to a publication recently presented by Thompson and Strong, taking such non-linear effects into account, does not necessarily mean giving up frequency-domain analysis. The model [TS01] that was proposed by these authors was used to demonstrate wave steepening in a trombone sound production simulation. The interesting fact is that this kind of non-linearity has not been implemented in a time-domain model but in the frequency-domain. Based on a one-dimensional lossy wave guide approach, the instrument was split into about 150 cylindrical slices propagating waves according to the linear wave equation. Thermo-viscous losses have been taken into account by allowing the wave number to become complex valued. A non-linear wave steepening correction was applied to the outgoing

220

pressure-wave spectra whenever a slice has been passed. This is a simplication of the real situation because the reected waves have not been corrected. The simplication was justied by the fact that the reection coefcient is much smaller than one at small discontinuities of the crosssection of a typical instrument. As input to the sequence of slices composing the bore of the trombone, a measured pressure spectrum has been used. The calculated resulting pressure spectrum and its time-domain wave form as shown in Figure 4.30 have been compared with corresponding data, recorded at an observers point somewhere in front of the bell, when the stimulus spectrum was measured simultaneously. As can already be recognized from comparing the time-domain wave forms, inclusion of wave steepening in the model greatly improves the matching between predicted and measured signals.

221

4. Applications

222

223

4. Applications

In order to get accurate bore information of tubular objects like brass wind instruments from acoustical input impedance measurements, an optimization approach has been proposed and investigated by the author [Kau01, Kau03b]. It avoids some of the difculties of the analytical approach (Section 3.3.1) and it is easier to extend for more sophisticated wave guide models like multi-modal propagation (Section 3.1.1) or even hybrid approaches (Section 3.2.7) involving high resolution computational methods (Chapter 2). The waveguide model used by the optimization program presented here has been reviewed in Section 3.1 and the impedance measurement setup is described here. Reconstruction results are presented and discussed. Sensitivity of input impedance magnitude to bore variation indicates that axial accuracy can be improved, without increasing the sampling rate. Advantages of the method are its increased numerical stability, its tolerance to white noise on top of measured data, no requirement to measure phase, and the fact that measured data need not to be contiguous in the frequency-domain. Frequency bands which are difcult to measure can be omitted.

Bore reconstruction by optimization is based on an optimization program originally developed for optimize various characteristics of the input impedance function of brass wind instruments described in musical terms. It is mainly used by instrument makers to improve the intonation of certain notes by allowing smaller or bigger modications in some parts of the bore. The program also permits to directly work on the shape of impedance peaks, their magnitude, quality factor and position. The so called impedance matching function can be used in bore reconstruction. The computer program, following a reliable optimization strategy [KA99], modies a starting geometry continuously in order to improve the matching between its theoretically obtained impedance spectrum and a measured one of an instrument, which is to be reconstructed.

224

Once the matching is achieved, the model geometry should be a reconstruction of the measured instrument. Because of the averaging nature of the optimization cost function CF = n p i=1 wi (|Zmeas (if )| |Ztheo (if )|) (with f = fsmp /nF F T , fsmp = sampling frequency, nF F T = FFT buffer length, wi = weighting factors, p = weighting exponent, Zmeas Ztheo being measured and theoretical impedance values), there is a relative insensitivity to white noise. It is also possible to mask certain impedance values, which are known to be uncertain, by zeroing the corresponding weighting factors. This is especially useful below 30 Hz, where most small speakers do not generate much output. Nevertheless it is essential to avoid any systematic measurement inaccuracies.

micr (psrc) micr (pinp)

capillary speaker

mouthpiece

The basic principle of the measurement system BIAS (Brass Instrument Analysis System), which has been used, is shown in Figure 4.31. The instrument or tubular duct is attached to the measuring head by means of a mouthpiece or an adapter, which is tightly pressed to the plane, where the high impedance capillary ends and the sound pressure pinp is measured using a small microphone. The sound pressure psrc at the other end of the capillary is measured simultaneously. The speaker in the closed chamber

225

4. Applications

is driven by the sound card of a PC, which also samples the response measured by the microphones. Stimulus Signals Any wide-band stimulus signal which contains sufcient sound energy at all frequencies of interest is adequate. Experiments have been made with white noise, MLS, chirp and sweep stimuli, each having its own advantages and drawbacks. Chirp and sweep signals turned out to be most suitable, because they can easily be amplitude modulated in order to balance the sound level over frequency, even when strong resonances in the measuring head are present. A frequency-domain-generated periodic chirp signal with a repetition period of 65536 samples at 24000 Hz (about 2.7 s), an envelope compensating for head resonances and a matching FFT buffer size of 65536 samples was used to make all the measurements for this paper. Because of the periodicity of the stimulus chirp a rectangular window can be applied in the time-domain. Each object was measured ve times and the resulting data were averaged. Calibration Although the capillary has a high acoustic impedance compared to most objects which are usually measured, it cannot be neglected, if results are to be used for bore reconstruction. Calibration of the whole measurement system with reference objects having a fairly well known impedance is important. Small volumes, which can be theoretically analyzed as very short closed tubes, have their rst resonance above the frequency range of interest. Up to 12 kHz they represent a fairly at magnitude and phase characteristic, which is insensitive to temperature over the whole frequency range. The calibration object which was used is shown in Figure 4.34. Its theoretical impedance is shown in Figure 4.32 and 4.33.

226

Impedance Magnitude 1e+009 Calibration Volume 5mm x 15mm Calibration Volume 2mm x 15mm 1e+008 [Ohm] 1e+007 1e+006 0 2000 4000 6000 [Hz] 8000 10000 12000

Impedance Argument -1.55 Calibration Volume 5mm x 15mm Calibration Volume 2mm x 15mm -1.555 [rad] -1.56 -1.565 -1.57 0 2000 4000 6000 [Hz] 8000 10000 12000

227

4. Applications

For this purpose the capillary is treated as an acoustical wave guide with complex frequency dependent characteristic. By solving the chain matrix system, the input impedance Zinp (ej ) ( being the discretized two matrix elements A11 (ej ) and A12 (ej ) by

p

frequency) can be derived from the sound pressure quotient pinp(ej ) and src

(ej ) (ej )

(ej )

Zinp (ej ) =

p

(4.12)

The two frequency dependent complex vectors A11 (ej ) and A12 (ej ) can be derived from two calibration measurements of known objects. p Ri are the measured complex pressure spectrum ratios pinp,i and Zi the src,i known complex impedance spectra of the two calibration volumes V1 and V2 (all are f (ej )). A11 = R1 Z2 R2 Z1 R1 R2 (Z2 Z1 ) A12 = Z1 Z2 (R2 R1 ) R1 R2 (Z2 Z1 ) (4.13)

228

The one dimensional theory of tubes seems to be satisfactorily accurate for the calibration volumes up to frequencies of about 10 kHz. There 2 17 mm, which is already close to the dimensions of the duct. Even below these frequencies mechanical vibrations, tiny leakages, body sound paths or non linear effects in the capillary can be present, which cannot be removed by this type of calibration. So care must be taken in order to design the mechanical components properly. Figure 4.34 shows the BIAS measuring head without the upper part, which is screwed on top of it, to tightly press the objects to the rubber surface. The microphone is centered and three capillary outlets are arranged symmetrically around this center in order to minimize the inuence of stimulating or recording higher order modes.

Computer optimization is sometimes considered rather an art than a science especially when actual simulated genetic or annealing approaches are referred to [Wal01]. The reason is that much experience, insider knowledge and sometimes intuition is usually required to select the right optimization variables and their best variation range, to create a target function which really reects the intentions of its creator and to nd a good combination of settings for all the various tuning parameters of the optimization algorithm. Several different optimization strategies have been tested for brass wind instrument optimization and bore reconstruction. It turned out that advanced optimization strategies did not compete so well, because it is difcult to present the data set to be optimized in a way which is compatible with the optimization concepts. The algorithms which have been tested were primarily out of the group of the so called genetic algorithms. Simulated annealing has not been tested because it was considered that too many target function evaluations would be required in order to get a reasonably successful result. The same turned out to be the weakness of genetic algorithms. In order to create sufciently large populations as required by the simulated evolution many

229

4. Applications

evaluations of the target function are necessary. Genetic Algorithms All of genetic algorithms basically simulate the rules of Darwins evolution [Sch94]. Trumpet individuals are reproducing themselves by mating to form populations, which are exposed to the pressure of selection. Children inherit basic properties from both parents and bad trumpets have a much smaller chance to reproduce themselves as they may die early. Random mutation is implemented as well as independent parallel populations with some individuals migrating from one to another. The rst approach, which has been implemented, is the simple genetic algorithm as described by Goldberg [Gol88]. It uses non-overlapping populations and optional elitism. Elitism means that the best individuals are directly moved from one generation to the next, making them somehow immortal - at least until better individuals take their place. Each generation, the algorithm creates an entirely new population of individuals. The second approach is a steady-state genetic algorithm that uses overlapping populations. It can be specied how much of the population should be replaced in each generation. The third variation is the incremental genetic algorithm, in which each generation consists of only one or two children. The incremental genetic algorithms allow to specify replacement methods dening how the new generation should be integrated into the population. So, for example, a newly generated child could replace its parent, replace a random individual in the population, or replace an individual that resembles it closely. The fourth type that was tested is the Deme genetic algorithm. This algorithm evolves multiple populations in parallel using a steady-state algorithm. In each generation, the algorithm migrates some of the individuals from each population to one of the other populations. The last type of a genetic algorithm that has been implemented is a deterministic crowding scheme based on the steady-state genetic algorithm as proposed by Goldberg. Like the other genetic algorithms its implementation has been taken from the GAlib genetic algorithm package, written

230

by Matthew Wall at the Massachusetts Institute of Technology [Wal01]. The genome type, which was associated with the trumpet individuals, was the so called binary string genome. Each optimization parameter was quantized within its variation range according to a specied minimum parameter resolution and the required number of bits was then mapped to the next empty piece of the binary string. Populations are initialized randomly and mutations are random single bit errors created in the binary string with a given probability. The default crossover method, which was normally used, determines a random bit position in the parents chromosomes and then creates a son with its fathers lower string and its mothers upper string and a daughter composed from the remaining segments.

Figure 4.35.: Bore Reconstruction Benchmark, Steady State GA Some disadvantages of that simple implementation are immediately obvious. One is that an already good starting position is completely lost during initialization of a population. The specied variation range determines the range for the random initialization as well as that for random mutations. Another one is that the probability for a single bit mutation to create an improvement is almost zero. Looking at some results it becomes obvious that there is much too much freedom to create meaningless and crazy geometries. In order to make

231

4. Applications

genetic algorithms to compete well, a completely different data representation would be necessary. Rosenbrock Algorithms The optimization algorithm which has nally been selected for brass instrument optimization and bore reconstruction follows a pretty old concept. It was proposed in the early sixties and has almost been forgotten and superseded by the variety of modern approaches that have already been mentioned.

In a direct comparison with the other concepts based on the data representation which was available, the Rosenbrock algorithm turned out to be the winner. This was not so much a surprise because the author has already evaluated this concept in the early eighties for the purpose of optimizing digital lter coefcients and integrated circuit devices. At that time the concept turned out to be faster and more stable than most of the gradient based methods available at that time. It has been proven to be a very useful strategy for optimization of target functions with many parameters which are very expensive to evaluate.

232

The Rosenbrock algorithm was reviewed by Schwefel in [Sch77]. Originally it was published by Rosenbrock in 1960 [Ros60]. The search strategy is based on a very stable 0th order search algorithm which does not require any derivatives of the target function although it approximates a gradient search. Therefore it combines advantages of 0th order and 1st order strategies. In the rst iteration it is a simple 0th order search in the directions of the base vectors of an n-dimensional coordinate system. In the case of a success, which is an attempt yielding a new minimum value of the target function, the step width is increased, while in the case of a failure it is decreased and the opposite direction will be tried. Once a success has been found and exploited in each base direction the coordinate system is rotated in order to make the rst base vector point into the direction of the gradient. Now all step widths are initialized and the process is repeated using the rotated coordinate system. Initializing the step widths to rather big values enables the strategy to leave local optima behind and to go on with search for more global minima. It has turned out that this simple approach is more stable than many sophisticated algorithms and it requires much less calculations of the target function than higher order strategies, which was already found out by Schwefel in [Sch77]. Because of this inherent stability and because of some well working heuristics in the calculation of the step widths this algorithm is even suitable and has already proven to be valuable for optimization problems involving highly non-linear and non monotonous target functions. Finally, somebody who does not want to become an optimization expert, too, has a real chance to understand it and to set and tune its parameters properly. Heuristic controlling procedures have been implemented in the optimization program BIOS [KA99] which was used to generate the results presented below. The optimization program is continuously monitoring the optimization process and adjusts optimization parameters at run time as soon as the optimization progress slows down or gets stuck. This allows to run success-

233

4. Applications

ful optimizations without any user intervention. Another feature which has turned out to be essential for unattended optimization runs is a bore smoothing step which is automatically performed between successive iterations. Representation of Instrument Geometry The right representation of the instruments geometry is already a crucial point. The instrument representation used in the optimizer assigns data structures called segments to physical parts of the instrument like mouthpiece, slides, bell and so on. These segments contain sequences of elementary conical elements described by coordinate pairs representing diameter d and its position x along the segment axis or optionally diameter increment and relative position. Each coordinate value (x or d) is linked to an instruction if and how much this value is allowed to be modied during the optimization run. Mixing absolute and relative coordinates freely allows to specify cylindrical or conical sleeves with a certain length which are inserted at an absolute position. Position, length and bore of the sleeve can be released for optimization. Releasing the last x value of a segment allows optimization of the segment length. This can be essential when the tuning slide of an instrument is to be modelled. Another optimization parameter which effects the overall tuning is the air temperature. It can be released between specied limits just like other coordinates. Making coordinate values of an instruments geometry optimization parameters is a very simple and exible way to give the optimizer enough freedom to nd any shape in order to come to an optimum. The Rosenbrock optimization algorithm was indeed able to deal with that high degree of freedom and, as will be shown below, gave good results even with 100 or more coordinate parameters. Genetic optimization methods did not perform well with that many parameters. It turned out that neither the standard initialization, crossover and mutation methods nor the standard parameter mapping to binary string genomes are suitable for this kind of coordinate optimization.

234

On the next higher level, segments are now arranged to bigger structures just like instruments are built in the physical reality. An instrument conguration, which corresponds to a certain pattern of valves engaged is represented by a data structure called arrangement. Most instruments allow the modication of their acoustical lengths by means of valves. When the player presses one of these valves a corresponding tube segment is inserted in a certain place increasing the total acoustical length and lowering the resonance frequencies of the instrument. This way chromatic scales can be played even in the lowest register. Optimization of a horn has to take that into account. There are certain parts like mouth piece, leadpipe, tuning slide or bell which are always contributing to the acoustical length of the instrument. Any modication there will equally inuence all played notes regardless of which valve is engaged. Other segments, the so called slides, are only active as long as their corresponding valve is depressed. Treating different valve combinations of an instrument like different instruments is not a solution unless modications in common parts of the instrument are synchronized properly. If different valve combinations were optimized separately it might be impossible to reunite the results back into one physical instrument because they might contain contradicting proposals for modications in one and the same common part. Therefore it is essential to deal with all valve combinations at once. Arrangements therefore contain an ordered list of segment references reecting the sequence of tubular instrument parts aligned along the total acoustical length of the instrument. An instrument with three valves is represented by eight different arrangements of segment instances. Optimization Target Function Each arrangement is associated with an input impedance list which is continuously recalculated during the optimization whenever a change is made to any of the segments involved. It is computed using the transmission line model according to Section 3.1.1 but higher order modes according to 3.1.1 could be included as well.

235

4. Applications

It has to be noted that the phase relationship between sound pressure and sound ow in the mouthpiece, which is the argument of the complex input impedance, is not considered for optimization purpose. This can be done because the phase information of the complex impedance function is not independent of its magnitude. In a minimum phase system it can be reconstructed from the magnitude by means of the Hilbert transform, using the requirement that the impulse response of a real instrument must be causal, that means there must not be any reection preceding the excitation pulse. The question, if a real system is truly a minimum phase system at least to the accuracy required by the application can only be answered by comparing measured phases with theoretical values. Recent measurement results have again strengthened the conviction of the author that the difference between actual phases and phases obtained for minimum phase systems is at least smaller than the measurement accuracy which can be achieved today. It was already noted that the frequency range of the calculation is limited by the model because it includes only the fundamental mode of a cavity or duct. As diameters are increasing this condition is only met for lower frequencies. That means that especially the bell region of a horn will introduce modelling errors at higher frequencies. The upper frequency limit for a typical trumpet is close to 1500 Hz, which is fortunately beyond the range of all played notes. The actual aims of an optimization run are expressed by the optimization target function. Basically it combines the contribution of all imperfections in a way to become zero when all optimization targets are equally reached. The optimization program BIOS provides user assistance for dening appropriate optimization target functions. For what is considered a standard task a suitable target function is created automatically. In the optimization target function, all specications for the optimized instrument are weighted and combined. One simple case is the optimization of the matching of impedance magnitudes which typical for bore re-

236

construction optimizations. By command or by using an editor the user creates a table of frequencies which will be attached to a certain arrangement. Corresponding target magnitude values have to be supplied. The user can load a measured impedance curve, or he can take a previous simulation of a reference instrument. Another typical application is the intonation optimization. Even during bore reconstruction optimizations it makes sense to add intonation targets. For this purpose centering attributes can be assigned to certain frequencies and the number of the resonances which should be centered at these frequencies can be attached. All target function contributions (normalized deviations from the rated value) are raised to a specied power called progression the higher this power, the more relative weight will be put on the biggest deviations, the lower this power, the more evenly the relative weights will be distributed multiplied with the user specied weight factors and added to get the nal result. In the ideal case this result will become very small during the optimization being zero when all targets have been reached perfectly.

Starting from input impedance measurements as described above, bore reconstruction of the measured tubular objects has been attempted. For this purpose an impedance magnitude matching optimization has been set up. The optimization cost function included about 1000 frequency points between 40 Hz and 7.5 kHz (later 9 kHz) with a frequency step of 10 Hz. Stepped Tube Reconstruction Figure 4.37 shows the stepped tube system which was reconstructed rst in order to test the achievable accuracy. The entry bore of 15 mm t perfectly to the cylindrical projection of the measuring head, so minimum errors caused by the coupling have been expected.

237

4. Applications

The impedance matching which has been achieved is shown in Figure 4.38. When these measurements have been made, a smoothing lter was applied to the measured curves in order to eliminate noise. Unfortunately the lter also decreased the quality factor of the resonance peaks, which is the reason that the theoretical peaks are generally higher than the corresponding measured ones. Trumpet Reconstruction A reconstruction obtained from the layer peeling algorithm applied to a reection function calculated from a measured trumpet input impedance following Equation 3.92 (without up-sampling as explained below) is shown in Figure 4.39. For comparison, another reconstruction of the same instrument obtained by David Sharp using the pulse reectometry system introduced in [Sha96] is presented.

238

Trum pet in Bb, Bore Reconstruction From Input Im pedance, BW=8kHz From Pulse Reflectometry 18

16

diameter [mm]

14

12

10

20

40

60 length [cm]

80

100

It can be seen that Sharps reconstruction contains much more detail, which is clear, when his sampling rate of 48 kHz is compared to the 8 kHz bandwidth of the impedance measurements. The differences at the left end originates from the different setup. To connect to the BIAS head an adapter was used, which has been roughly reconstructed together with the instrument. For pulse reectometry the instrument without mouthpiece was smoothly coupled to the source tube. Reconstruction of Trumpet Lead Pipe A recent reconstruction result is shown in Figure 4.40. The reconstructed object is an experimental lead pipe of a BbTrumpet. The bold curve represents the dimensions as indicated by the maker. The average deviation between the curves is 0.2mm. The tolerance of the manufacturing process itself can be estimated as around 0.1mm. The leadpipe was attached to the BIAS system by means of an adapter. This adapter is similar to a mouthpiece, but it provides a much smoother and less reective transition from the 15 mm head projection diameter to

239

4. Applications

240

the entry bore of the lead pipe, which is below 10mm. Without an adapter it would not be possible to attach any tube with an inner diameter less than 15 mm hermetically to the head. After careful calibration of the system using the two calibration volumes, the adapter was mounted and a cylindrical reference tube was attached. The combination of adapter and tube was measured and an acoustical equivalent of the adapter was reconstructed by optimization. This reconstructed adapter was then used for the reconstruction of the lead pipes. The achieved impedance matching is shown in Figure 4.41. The mismatch between the theoretical and the measured curve is almost invisible in the plot. The average of the relative mismatch of all 895 frequency lines is about 60 ppm, or 40 in absolute terms. These numbers also give an indication for the required measurement accuracy.

When the reconstruction results from Figure 4.37 and 4.39 are compared, then it can be observed that the optimized shape exhibits much better axial resolution than the one resulting from the layer peeling algorithm. This is interesting because both reconstructions have been made from frequencydomain spectra with comparable bandwidth. This will now be investigated in more detail. Axial Accuracy and Sampling Rate Layer peeling requires a reection function, which can be measured in the time-domain by pulse reectometry, but which can also be derived from the input impedance spectrum as the inverse fourier transform of IIR according to Equation 3.92. The resolution of the impedance spectrum f = 1/T can be increased by lengthening the sampling interval T . The inverse FFT of a ne spectrum gives a long reection function. The layer peeling algorithm does not make use of a long input vector. When the far end reection of the pulse returns to the near end the algorithm will stop. The duct is then reconstructed all the way long to the

241

4. Applications

far end. Each sample of the pulse response reconstructs an axial length c L = 2fs . Resolution and therefore axial accuracy can be gained by increasing the sampling rate or by decreasing the speed of sound c. This is very clear as far as the reection function is concerned. When reconstructing from impedance, things are a bit different. Accuracy and resolution or not directly coupled. The impedance spectrum consists of discrete frequency lines each of them representing a standing wave with a certain wavelength i and a corresponding frequency if related to the position i of the spectral line. The standing wave itself originates from sound reections between the ideally reecting input plane and another reecting discontinuity somewhere along the axis of the acoustic wave guide. The position of this discontinuity is therefore directly related to the frequency of the standing wave. If it does not coincide with the spectral grid, then any windowing function will fold its sound energy into the near or even farer neighborhood and some information is lost. A good frequency resolution together with a good windowing function therefore preserves the information of the accurate position of reecting discontinuities along the duct. The inverse FFT of a high resolution impedance spectrum transforms high resolution into length rather than time resolution, which we would like to get. The key to get the time resolution back is to not truncate the pulse response but to make use of the multiple and secondary reections contained in the tail of the long pulse response. If a reected pulse returns exactly in between two sampling events, the distance of the reecting discontinuity cannot be determined accurately. But if we wait for another round trip of the pulse, it will now accurately coincide with a sampling moment and the information about its original wave length will be accurately available. And this is exactly what happens when the bilinear transform (Equation 3.92) is applied to the impedance spectrum before the inverse FFT. All the secondary and multiple reections, which would arrive at the microphone long after the primary reection of the instruments far end has arrived,

242

are folded into the primary reection period. Unfortunately they often cannot contribute, because most people do not up-sample the spectrum before the transform. In order to make use of the available extra information, the spectrum must be up-sampled (zero padded) before the transformation is applied. Because of the nature of time to frequency-domain transitions periodic events within the recording interval are transformed into discrete spectral lines and single events will produce distributed spectral effects ( pulse at magnitude). This indicates that bore discontinuities which are not evenly spaced in very close distances, will leave their footprints in the whole impedance spectrum and can be traced accurately even if parts of it are missing. Yet, true resolution cannot be gained, as it is dened as the capability to separate two closely neighboring features. Such a case will only create high frequency spectral lines, which need an adequate sampling rate in order to contribute to the reconstruction. Sensitivity and Bandwidth Dependance However, reconstruction by optimization does make use of all the information contained in a high-resolution impedance spectrum, and it tolerates gaps in the measured spectrum. The sensitivity of magnitude matching with respect to variations of single diameter values has been investigated. For this purpose a simple articial duct with a length of about 90 cm consisting of two cylindrical and one conical segment has been modelled. As with magnitude matching optimizations, the sum of all absolute values of impedance magnitude differences over a certain frequency band has been calculated and plotted over the position of the modied diameter value. The overall frequency band of 12 kHz was split into low, medium and high bands and the investigation has been repeated for three different axial resolutions. It was expected that the effect of a single parameter modication in a ne axial grid would mainly effect high frequency regions of the spectrum while the same done in a course axial grid would be noticeable even in the

243

4. Applications

Bore Values

4,5 4 3,5 3 2,5 2 1,5 1 0,5 0 0 20 40 60 80 100 axial length [cm]

diameter [cm]

0..12kHz 1600 Peak Mag Mismatch 1400 1200 1000 800 600 400 200 0 0 20 40 60 80 100 Modified Point (1cm Grid) 0..4kHz 4..8kHz 8..12kHz

244

0..12kHz 2000 1800 1600 1400 1200 1000 800 600 400 200 0 0 5 10 15 20 25 30 Modified Point (4 cm Grid) 0..4kHz 4..8kHz 8..12kHz

low frequency band. This behavior would justify the widespread opinion that wide band measurements are required if ne axial resolution is to be achieved. The results are shown in Figure 4.43. Diameter steps in a 1 cm axial grid have the most effect in the medium band but can be traced in any other band as well. The sensitivity is signicantly smaller in the vicinity of bore discontinuities. A 4 cm axial grid makes the base band the most sensitive region. According to the spectrum of a triangular pulse higher bands are effected to a much lesser extent.

Using the proposed method useful reconstruction results have been obtained. As with pulse reectometry, it is difcult to obtain accurate reconstruction results when the measured object is long. Flaring bells do not

245

4. Applications

reconstruct well because cut-off frequency is low when the bore diameter is big. Small details along the axis require high-bandwidth measurements, which are not easy to achieve and which require multi-mode modelling to get good impedance matching above the cut-off frequency. So there are two directions for future research. First, the measurement bandwidth should be increased above 8 kHz, which currently seems to be a practical limit for conventional measurement setups. Second, a future reconstruction model should take higher order modes into account. Finally, the expression for the radiation impedance should match the conditions which are present during the measurement. In order to speed up the optimization process, propagation of higher order modes that are usually evanescent inside the instrument, could probably be neglected. Taking account of the mode conversion at bore discontinuities should already be sufcient to improve the accuracy signicantly. Another possible way to improve reconstruction results, which has not been tested yet, is to introduce an acoustically hard and airtight termination somewhere inside the bell. If this is practically possible, several factors will improve. Reection of high frequencies will generate information bearing resonances, which will increase SNR in the higher frequency band. Smaller maximum bore diameter will increase the cut-off frequency of the rst mode and extend the validity range of the plane wave model. And nally the model for the radiation impedance can be omitted. It usually attributes most of the modelling uncertainty. In order to increase the measurement bandwidth, the calibration procedure seems to be the most critical issue. Active calibration objects presenting accurate impedances without strong resonances could be one approach. Multi-mode algorithms, which calibrate not only the plane wave mode but also a certain number of higher modes, could be another branch of study. Alternative measurement setups, like the one Walstijn is currently working on, also seem promising [WC02]. They consist of arrays of microphones and a non reective coupling to the measured object. The latter helps to ensure single mode propagation at the reference plane.

246

4.3.1. Wood Wind Instruments

An extensive computational analysis of the clarinet was recently published by Facchinetti et al. [FBC03]. Coupled uid and solid three-dimensional nite element models for the reed and the air load in the rst 10 cm of the pipe and a lumped elements model for the main part of the pipe were used. In a rst step, eigenmodes were computed for an isolated reed rigidly clamped on the section corresponding to the ligature that had a stress-free boundary elsewhere. As a second step, the dynamics of the reed inuenced by air loading was studied using a coupled uid-solid model. The system then was composed of the reed, the mouthpiece and the barrel. It was found that air coupling changes the normal modes of the isolated reed. Therefore, the modes of the whole system have to be taken as source for the acoustic eld in the mouthpiece. An experimental modal analysis on reeds by means of holographic interferometry showed the validity of the numerical model of the reed coupled to air within 10 20 % of the measured resonance frequencies. Torsion modes of reeds generate a strong but, due to acoustical short-circuits, very localized acoustic eld in the mouthpiece, which only weakly couples to the acoustic eld in the clarinet. In the cylindrical part of the mouthpiece, acoustic waves can be considered already plane within a very good approximation. Therefore, the nite element model of the rst 10 cm of the pipe was connected with a lumped-element model for the rest of the clarinet in order to simulate the modal behavior of the whole instrument.

Adachi [Sei02] used a two-dimensional nite element method to simulate the acoustical deection of a jet in an external sound eld perpendicular to the jet. It was shown that the method used is able to simulate the soundow interaction.

247

4. Applications

In a recent update [Ada03] he compared the deection parameters extracted for some velocity proles at several positions along the jet with measurements made by Nolle. He concluded that CFD simulations are in good agreement with experimental results and can be referred to in order to extend available models. Non-linearities are predicted well by NavierStokes based CDF simulations. In his previous article, Adachi mentioned a recent study by Ito et. al. [IFY02], where edge tones were simulated with a nite element method. Adachi reported that it was found that the simulated sound frequency was in good agreement with that obtained by experiment. Unfortunately no material concerning this approach is yet available in English. Bamberger et al. [BBS01] simulated edge tones using a two-dimensional model of the incompressible, isothermal Navier-Stokes equations using an adaptive nite element method and compared the results with measurements. They examined edge tones in the fundamental uid dynamic mode varying the height d of the nozzle, the edge distance w and the jet velocity. A good proportionality of the frequency with the ow velocity was found, both in experiments and simulations. There was almost no deviation from the general law in terms of the Strouhal number Sd = f d = C (d/w)n , with n = 1 for laminar ow U0 and the maximum velocity at the exit of the nozzle U0 . The constant C was found to agree within 10 15 % between experiment and simulation. One difference between experiment and simulation was noticed concerning the threshold for the distance w where oscillations occur: especially at low velocities and for a wider nozzle the regime of oscillation of the rst hydrodynamic mode was shifted to smaller edge distances in the simulation. The authors hold the missing noise in the simulation responsible for this behavior. In [TSK02], an article unfortunately in Japanese with an abstract in English, the two-dimensional direct numerical simulation of the edge-tones by the nite difference lattice Boltzmann method (FDLBM) is reported. Very small pressure uctuations result from periodically oscillation of jet around the edge were successfully captured, showing that lattice Boltz-

248

mann methods could mature to a powerful tool for low Mach number computational aeroacoustics in the near future.

4.3.3. Glottis

Zhao et al. [ZZFM02] simulated the ow and the acoustic eld in an idealized vocal tract with time-varying geometry. The compressible NavierStokes equations in a two-dimensional, axis-symmetrical and moving coordinate system were solved. A nite difference method was applied using a compact sixth-order differencing scheme in space and a fourth-order Runge-Kutta method in time. The computation was compared with an acoustic analogy to identify the contributions of sound sources within the ow an unsteady volume velocity monopole source associated with the motion of the vocal fold in the throat region, a dipole source due to the oscillating force, which arises from the interaction between vortical structures shed from the glottis and the velocity eld, and a quadrupole source due to net unsteady ow inside the vocal tract. The assumption of a two-dimensional axis-symmetrical geometry prevents the generation of turbulence, which is expected to dramatically alter the ow pattern inside and downstream of the glottis, as well as the subsequent acoustic radiation. Therefore, the computed contribution of the monopole source that is believed to be negligible in speech production, was overestimated, whereas the quadrupole source was underestimated about one order of magnitude compared to large eddy simulations of the sound radiation from subsonic turbulent jets. Figure 4.45 shows a time sequence of vorticity contour plots for one cycle of glottis motion as simulated and published by Zhao [ZZFM02].

In engineering, coaxial side branches attached to a main duct are commonly used as a silencer. They act as quarter-wave resonators taking sound

249

4. Applications

energy from the main duct. In presence of mean ow, high noise levels can be excited in the main duct and large wave amplitudes in the side branches due to coupling of ow and acoustics at certain ow conditions. The excitation of acoustical modes in such a system is of interest for musical acousticians because the mechanism of vortex induced sound generation in a low Mach number ow (M a < 0.1) is similar to that in utes and ue pipes. Radavich et al. [RSN01] computed such a system, solving the twodimensional, unsteady, compressible Navier-Stokes equations with the k turbulence-closure-model using an implicit, non-iterative, operator splitting algorithm. The experimentally measured region of oscillation was successfully reproduced. Although the simulations were not fully comparable with the experiments due to differences in the boundary conditions, the soundpressure level differed about 6 dB at maximum. A better spatial resolution of the computation was said to be able to drop the difference to 3 dB. The comparison of computed and experimental visualizations of the vortical structures in the ow in the vicinity of the side branches suggests that the ow-acoustic coupling is captured well in the computations. A large net acoustic source caused by the interaction between the vortex, the velocity eld and the acoustical velocity was localized over the junction between the main duct and the side branch with a concentrated acoustic

250

sink near the upstream corner. In her thesis, Dequand [Deq01, DHK+ 03] covers a larger area of owacoustic interactions, from the aeroacoustic response of sharp bends in ducts, self-sustained oscillations in a cross-junction, to the excitation of a Helmholtz resonator by a low Mach number ow, which were all studied experimentally and computationally. The numerical method used here is based on the Euler equations for two-dimensional inviscid compressible ows. To simulate the generation of vorticity in the wall boundary layer, the Kutta-condition was imposed. The Euler equations were spatial discretized using a second-order accurate, cell-centered nite volume technique. The time-integration was based on a second-order accurate four-stage Runge-Kutta method. The simulations of self-sustained oscillations in a cross-junction could reproduce the region of oscillation. Comparison with measurements, others than the above mentioned, showed that the oscillation frequency was predicted within 2 %, but the calculated pulsation amplitude was overestimated by 40 %. Since previous calculations by Hofmans [Hof98] of a similar pipe system based on a method including visco-thermal losses, also overestimated the same parameter by about the same amount, the author concludes that the absence of visco-thermal losses cannot explain the difference. There must be other reasons, maybe experimental problems like wall vibrations.

For musical acoustics, the most interesting parts in Dequands thesis are the simulations and experiments of self-sustained oscillation in a Helmholtz-like resonator [Deq01]. The inuence of the shape of the resonators neck on the ow behavior and the generation and absorption of sound by vortex shedding was extensively examined in a series of different congurations. The Euler computations predict the resonance frequency within 5 %. A comparison to experiments shows, that the computations performed best in the conguration having a neck with a 90 angle and both the upstream and

251

4. Applications

the inner edge were chamfered. Here the predicted pulsation amplitude was about 20 % too low. In the congurations with a 60 neck and sharp or chamfered upstream edge the predicted amplitudes are 40 % too low, and in the conguration with a sharp upstream edge and a chamfered inner edge the predicted pressure amplitude was at least a factor 4 too low.

The sound generation in a small stopped ue pipe with recorder-like proportions was simulated using a three dimensional LBM with multi-relaxation-times. Here some results of a recent simulation are presented. The dimensions of the stopped pipe are 60 mm 7 mm 7 mm for the resonator, 15 mm 4.5 mm 1.2 mm for the ue. The ratio of the distance from ue exit to the labium to the height of the ue exit is about 2.8 and the angle of the labium is approx. 14 .

Figure 4.46.: Cross section of the simulated pipe. Dimensions in lattice units. The whole pipe is embedded into a volume of 75 mm 7 mm 19 mm, which was simulated only partially. Due to the limited computational resources the lattice spacing had to be chosen rather large, x = 0.22 mm. The speed of sound was cs = 340 m/s used, resulting in a time step of t = 3.74 107 s. The time for rising the ow in the ue was 5000 t =

252

253

4. Applications

1.87 ms, which is rather short compared to the rising time found in ue instruments. A higher viscosity than air was assumed to enhance the stability of the calculation. Formation Jet at Startup When the pressure rises at the inlet and a ow through the ue develops, at the ue exit a jet is formed due to the effect of viscosity, as shown in Figure 4.48. The ow separates and vortices above and below the ue exit are clearly visible. This shows that with the LBM effects of viscosity emerge naturally, with no needs of special boundary conditions.

Figure 4.49 shows the density near the stopped end of the resonator. During startup the rst mode is excited rst but not supported, so it decays visible in Figure 4.50. The second pipe mode then grows until it reaches a saturation. The frequency of the fundamental mode is 1197 Hz. The frequency of the second mode is 3235 Hz and is about 10 % deeper than three times the fundamental frequency. Figure 4.51 shows the spectrum of the decaying rst mode and the growing second mode, Figure 4.52 shows the spectrum during steady state oscillations. The velocity in the mouth of the ute during a oscillation period is shown in Figure 4.53

254

Figure 4.48.: Formation of the jet during start-up. The value of the velocity ranges from blue (u = 0) to red (max. velocity).

255

4. Applications

1.0025 1.002 1.0015 1.001 Density 1.0005 1 0.9995 0.999 0.9985 0 0.01 0.02 0.03 0.04 Time (s) 0.05 0.06 0.07 0.08

Figure 4.49.: Density taken at the end of the stopped pipe resonator

1.0016 1.0014 1.0012 1.001 Density 1.0008 1.0006 1.0004 1.0002 1 0.9998 0 0.005 0.01 0.015 0.02 Time (s) 0.025 0.03 0.035 0.04

Figure 4.50.: Transient behavior of the density at startup, taken at the end of the stopped pipe resonator

256

0.0035 0.003 0.0025 0.002 0.0015 0.001 0.0005 0 0 2000 4000 6000 8000 10000 12000 14000 16000 Frequency (Hz)

0.025

0.02

0.015

0.01

0.005

0 0 2000 4000 6000 8000 10000 12000 14000 16000 Frequency (Hz)

257

4. Applications

258

Acknowledgments

It did not take long to decide, whom to express my gratitude rst. I sincerely and whole-heartily want to thank my dear family for their patience and understanding. I have to apologize to my wife and my children for not being present, physically and with my mind, for quite a while now. I can understand the complaints of my little boy who became almost half an orphan during the time when this book was conceived and written. The same is true for my daughter, who did not get as much of an assistance she needed, whether in writing her math homework or in practicing the bass guitar. I also have to thank my wife, for not having looked for a boyfriend during the time when I really neglected her. Quite on the contrary, she relieved me from all my duties at home, she took everything on her shoulder and tolerated my out-of-phase sleeping rhythm as well as my changing moods. I also want to say thank you to Prof. Gregor Widholm, the leader of the Institut f r Wiener Klangstil, for his encouragement, tolerance and u understanding, which is really beyond what one can expect. It is mainly he, who is responsible for the agreeable, creative, fruitful, friendly and stimulating working conditions at our institute. I owe gratitude to all my colleagues, too. They certainly had to suffer from my absence and much of my understanding was growing on the ground of interesting and fruitful discussions. Specically I want to mention Helmut K hnelt, who helped a lot to lay u the foundation of this book, by being the co-author of an invited paper on wind instrument acoustics [KK03] that later on served as the seed of this book. As a supervisor of his PhD thesis on the application of the Lattice-

259

4. Applications

Boltzmann method to ute acoustics, I will have an opportunity to make up for his contributions in this eld. Special thanks also to my fellow bass player, PhD student, and violin maker Andrew Brown, for his support in correcting my English. Without his patient help my German accent would certainly be much more obvious, even in writing. In this case it is easy for me to reciprocate for this unpaid work because I can store the big chestnut tree that he has acquired for some years in my garden, until he will shape it to make double basses from this wood. Finally I want to thank my well known fellow researchers, Prof. Murray D. Campbell, and Prof. Giovanni De Poli, who have agreed to act as external referees in my habilitation procedure, not only for their willingness to follow the invitation of the examination committee, but also for the appreciation and encouragement they have expressed towards me in the past when we met at some international conferences.

260

Bibliography

[ABB+ 01] F. Avanzini, B. Bank, G. Borin, G. De Poli, and D. Fontana, F. Rocchesso. Musical instrument modeling: the case of the piano. In MOSART IHP Network, editor, Workshop on Current Research Directions in Computer Music, pages 110, Barcelona, 2001. Audiovisual Inst., Pomper Fabra Univ. http: www.iua.upf.es/mtg/mosart/papers/p18.pdf. R. D. Ayers, R. P. Birkemeier, and L. J. Eliason. Rayleigh wave model for the lip reed: Qualitative aspects. Proceedings of the 137th Meeting of the ASA and 2nd Convention of the EAA Forum Acusticum, Abstract in Acustica, 85:13, 1999. Seiji Adachi. Cfd analysis of air jet deection comparison with nolles measurements. In Roberto Bresin, editor, Proceedings of the SMAC03 (Stockholm Music Acoustics Conference), pages 313316, Stockholm, 2003. KTH Speech, Music and Hearing. Noam Amir, V. Pagneux, and J. Kergomard. A study of wave propagation in varying cross-section waveguides by modal decomposition. part ii. results. Journal of the Acoustical Society of America (JASA), 101(5):25042517, 1997. N. Amir, G. Rosenhouse, and U. Shimony. Reconstructing the bore of brass wind instruments: theory and exper-

[ABE99]

[Ada03]

[APK97]

[ARS94]

261

Bibliography

iment. In Proceedings of the SMAC 93, pages 470475, Stockholm, 1994. Royal Swedish Academy of Music and the authors. [AS65] M. Abramowitz and I. A. Stegun. Handbook of Mathematical Functions. Dover, New York, 1965. Seiji Adachi and Masaaki Sato. Time-domain simulation of sound production in the brass instrument. Journal of the Acoustical Society of America (JASA), 97(6):38503861, 1995. Seiji Adachi and Masaaki Sato. Trumpet sound simulation using a two-dimensional lip vibration model. Journal of the Acoustical Society of America (JASA), 99(2):12001209, 1996. H. L. Atkins. Application of essentially nonoscillatory methods to aeroacoustic ow problems. In NASA Conference Publication 3300, editor, ICASE/LaRC Workshop on Benchmark Problems in Computational Aeroacoustics., 1995. John Backus. The Acoustical Foundations of Music. W. W. Norton and Company, Inc., New York, 1969. 2nd ed. 1977. J.M. Buick, C.L. Buckley, C.A. Greated, and J. Gilbert. Lattice Boltzmann BGK simulation of nonlinear sound waves: the development of a shock front. Journal of Physics A, 33:391728, 2000. A. M. Bruneau, M. Bruneau, Ph. Herzog, and J. Kergomard. Boundary layer attenuation of higher order modes in waveguides. Journal of Sound and Vibration, 119(1):15 27, 1987.

[AS95]

[AS96]

[Atk95]

[Bac69]

[BBGG00]

[BBHK87]

262

Bibliography

[BBKMSI03] Julien Bensa, Stefan Bilbao, Richard Kronland-Martinet, and Julius O. Smith III. The simulation of piano string vibration: From physical models to nite difference schemes and digital waveguides. Journal of the Acoustical Society of America, 14(8/2003):1095 1107, 2003. [BBS01] Andreas Bamberger, Eberhard B nsch, and Kunibert G. a Siebert. Experimental and numerical investigation of edge tones. Technical report, Weierstra-Institut f r Angeu wandte Analysis und Stochastik, Berlin, 2001. Preprint No. 681. G. Borin and G. De Poli. A hysteretic hammer-string interaction model for physical model synthesis. In Proceedings of Nordic Acoustical Meeting, pages 399406, Helsinki, 1996. G. Borin, G. De Poli, and D. Rocchesso. Elimination of delay-free loops in discrete-time models of nonlinear acoustic systems. IEEE Trans. Speech Audio Process., 8(5):597606, 2000. G. Borin, G. De Poli, and A. Sarti. A modular approach to excitor-resonator interaction in physical models synthesis. In ICMA & IPSA, editor, Proceedings International Computer Music Conference [ICMC 1995], pages 4650, Columbus (OH), 1990. G. Borin, G. De Poli, and A. Sarti. Algorithms an structures for synthesis using physical models (music). COMPUTER MUSIC JOURNAL, 16(4):3042, 1992. V. Belevitch. Classical Network Theory. Holden-Day, San Francisco, CA, 1968.

[BDP96]

[BDPR00]

[BDPS90]

[BDPS92]

[Bel68]

263

Bibliography

[Ben60] Arthur H. Benade. On the mathematical theory of woodwind nger holes. Journal of the Acoustical Society of America (JASA), 32(12):15911608, 1960. Arthur H. Benade. Fundamentals of Musical Acoustics. Oxford University Press, New York, 1976. Arthur H. Benade. Equivalent circuits for conical waveguides. Journal of the Acoustical Society of America (JASA), 83:17641769, 1988. A. H. Benade and D. J. Gans. Sound production in wind instruments. Annals of the New York Academy of Science, 155:247263, 1968. A. Barjau and V. Gibiat. Phenomenological models for musical instruments. In Adriano Alippi, editor, Proceedings of ICA 2001 (17th International Congress on Acoustics), Rome, 2001. ICA. Ana Barjau and Vincent Gibiat. Delay lines, nite differences and cellular automata: Three close but different schemes for simulating acoustical propagation in 1d systems. Acta Acustica, 88(4):554566, 2002. Ana Barjau and Vincent Gibiat. Delayed models for simplied musical instruments. Journal of the Acoustical Society of America, 114(7/2003):496 504, 2003. J.M. Buick, C.A. Greated, and D.M. Campbell. Lattice BGK simulation of sound waves. Europhysics Letters, 43(3):23540, 1998. Ana Barjau, Vincent Gibiat, and Delphine Chareyron. Unied modelling of simplied musical instruments in time

[Ben76]

[Ben88]

[BG68]

[BG01]

[BG02]

[BG03]

[BGC98]

[BGC02]

264

Bibliography

domain. In European Acoustics Association, editor, Proceedings of Forum Acusticum Sevilla 2002, pages MUS 06005, Sevilla, 2002. Hirzel Verlag. [BGK54] P. Bhatnagar, E. P. Gross, and M. K. Krook. A model for collision process in gases. Physical Review, 94:51124, 1954. Stefan Damian Bilbao. Wave scattering methods for the numerical integration of partial differential equations. PhD thesis, Stanford University, Department of Electrical Engineering, 2001. A. Barjau, D. H. Keefe, and S. Cardona. Time-domain simulation of acoustical waveguides with arbitrarily spaced discontinuities. Journal of the Acoustical Society of America (JASA), 105(3):19511964, 1999. S. R Bromage, Orlando F. Richards, and D. Murray Campbell. Reproducability and control of the embouchure of an articial mouth for playing brass instruments. In Royal Swedish Academy of Music, editor, Proceedings of the SMAC 03, pages 197200, Stockholm, 2003. Royal Swedish Academy of Music. Gunther Brenner. Numerische simulation komplexer uider transportvorgnge in der verfahrenstechnik, 2002. A. H. Benade and J. H. Smith. Brass wind instruments impulse response measurements. Journal of the Acoustical Society of America (JASA), 70:22, 1981. R. B. Bird, W. E. Steward, and E. N. Lightfoot. Transport Phenomena. Wiley, New York, 1960.

[Bil01]

[BKC99]

[BRC03]

[Bre02]

[BS81]

[BSL60]

265

Bibliography

[Cam99] D. Murray Campbell. Nonlinear dynamics of musical reed and brass wind instruments. Contemporary Physics, 40:415431, 1999. D. Murray Campbell. Brass instruments as we know them today. In Royal Swedish Academy of Music, editor, Proceedings of the SMAC 03, pages 185188, Stockholm, 2003. Royal Swedish Academy of Music,. Gilbert Cullen and D. Murray Campbell. Brass instruments: Linear stability analysis and experiments with an articial mouth. Journal of the Acoustical Society of America (JASA), 86:704724, 2000. Ren Causs and et al. Input impedance of brass musical e e instruments comparison between experiment and numerical models. Journal of the Acoustical Society of America (JASA), 75:241254, 1984. D. Murray Campbell and Clive Greated. The Musicianss Guide to Acoustics. Oxford University Press, New York, 1987. S. C. Chang. The method of space-time conservation element and solution element - a new approach for solving the navier-stokes and euler equations. Journal of Computational Physics, 103:1642, 1995. Jean-Jacques Chattot. Computational Aerodynamics and Fluid Dynamics. Springer, Berlin, 2002. S. Davis. A compact solution to computational acoustics. In ICASE/LaRC Workshop on Benchmark Problems in Computational Aeroacoustics. NASA Conference Publication 3300, 1995.

[Cam03]

[CC00]

[Cea84]

[CG87]

[Cha95]

[Cha02]

[Dav95]

266

Bibliography

[Deq01] Sylvie Dequand. Duct Aeroacoustics: from Technological Applications to the Flute. PhD thesis, Technical University Eindhoven, Eindhoven, 2001. D. dHumi` res, I. Ginzburg, M. Krafczyk, P. Lallemand, e and L.S. Luo. Multiple-relaxation-time lattice Boltzmann models in three dimensions. Phil. Trans. R. Soc. Lond. A, 360:43751, 2002. S. Dequand, S. Hulshoff, H.J. van Kuijk, J.F.H. Willems, and A. Hirschberg. Helmholtz-like resonator self-sustained oscillations, part 2: Detailed ow measurements and numerical simulations. AAIA Journal, 41:416423, 2003. G. De Poli. A tutorial on digital sound synthesis techniques. In C. Roads, editor, The Music Machine, pages 429447. MIT Press, Cambridge, MA, 1991. G. De Poli and A. Piccialli. Waveform transformation techniques for granular synthesis. In DRAKULIC Gordana TOTIC Bojana PRAVICA, Peter, editor, Proceedings of the 13th International World Congress on Acoustics, Yugoslavia 1989, pages 113116, Novi Beograd, 1989. Sava Centar. G. De Poli and A. Piccialli. Pitch-synchronous granular synthesis (music synthesis). In Representations of Musical Signals, pages 187219. MIT Press, Cambridge, MA, 1991. Giovanni De Poli and Davide Rocchesso. Physicallybased musical sound modeling. In sterreichische Computer Gesellschaft, editor, books@ocg.at (Proceedings of the Diderot Forum), pages 7997, Wien, 1999. sterreichische Computer Gesellschaft.

[dGK+ 02]

[DHK+ 03]

[DP91]

[DPP89]

[DPP91]

[DPR99]

267

Bibliography

[EB82] S. J. Elliott and J. M. Bowsher. Regeneration in brass wind instruments. JOURNAL OF SOUND AND VIBRATION, 83(2):181217, 1982. N. Dean Eckhoff. Lectures for Computer Applications in Mechanical Engineering. Kansas State University, online, 2003. Matteo Facchinetti, Xavier Boutillon, and Andrei Constantinescu. Numerical and experimental modal analysis of the reed and pipe of a clarinet. The Journal of the Acoustical Society of America, 113(5):2874 2882, 2003. A. Fettweis. Wave digital lters: Theory and practice. Proceedings of the IEEE, 74:270327, 1986. B. Fabre and A. Hirschberg. Physical modeling of ue instruments: A review of lumped models. ACTA ACUSTICA/ACUSTICA, 86(4):599610, 2000. U. Frisch, B. Hasslacher, and Y. Pomeau. Lattice-gas automata for the navier-stokes equation. Phys. Rev. Letters, 56(14):15051508, 1986. Neville H. Fletcher. Excitation mechanisms in woodwind and brass instruments. Acustica, 43:6372, 1979. Neville H. Fletcher. Autonomous vibration of simple pressure-controlled valves in gas ows. Journal of the Acoustical Society of America (JASA), 93(4):21722180, 1993. Alfred Fettweis and Gunnar Nitsche. Numerical integration of partial differential equations by means of multidimensional wave digital lters. In Proceedings of the International Symposium on Circuits and Systems (ISCAS 90), pages 954 957. IEEE, 1990.

[Eck03]

[FBC03]

[Fet86] [FH00]

[FHP86]

[Fle79] [Fle93]

[FN90]

268

Bibliography

[FR90] Neville H. Fletcher and Thomas D. Rossing. The Physics of Musical Instruments, 2nd ed. Addison-Wesley, New York, 1990. Barbara J. Forbes, David B. Sharp, Jonathan A. Kemp, and Aijun Li. Singular system methods in acoustic pulse reectometry. ACTA ACUSTICA/ACUSTICA, 89(5):743753, 2003. C. W. Gear. Numerical Initial Value Problems in Ordinary Differential Equations. Prentice-Hall, New Jersey, 1971. Etienne Guyon, Jean-Pierre Hulin, Luc Petit, and Catalin D. Mitescu. Physical Hydrodynamics. Oxford University Press Inc., New York, 2001. Joel Gilbert. Sound mechanisms of brass instruments, last twenty years results. In European Acoustics Association, editor, Proceedings of Forum Acusticum Sevilla 2002, pages MUS04004IP, Sevilla, 2002. Hirzel Verlag. D. Goldberg. Genetic Algorithms in Search, Optimization and Machine Learning. Addison Wesley Publishing Company, 1988. J. W. Goodrich. Application of a new high order nite difference scheme to acoustic propagation with linearized euler equation. Computational Aero- and Hydro-Acoustics, 147:2934, 1993. Hermann Helmholtz. On the Sensations of Tone as a Physiological Basis for the Theory of Music. Dover, 1954, repr. New York, 1863. F. Q. Hu, M. Y. Hussaini, and J. Manthey. Application of low dissipation and dispersion runge-kutta schemes to

[FSKL03]

[Gea71] [GHPM01]

[Gil02]

[Gol88]

[Goo93]

[Hel63]

[HHM95]

269

Bibliography

benchmark problems in computational aeroacoustics. In ICASE/LaRC Workshop on Benchmark Problems in Computational Aeroacoustics. NASA Conference Publications 3300, 1995. [Hin75] [HL97] J. O. Hinze. Turbulence. McGraw-Hill, New York, 1975. X. He and L.S. Luo. Theory of the lattice boltzmann method: from the boltzmann equation to the lattice boltzmann equation. Physical Review E, 56(6):68116817, 1997. G.C.J. Hofmans. Vortex Sound in Conned ows. PhD thesis, Technical University Eindhoven, Eindhoven, 1998. J. Hardy and Y. Pomeau. Thermo- and hydromechanics for a model uid. J. Mathematical Fluids, 13:10421051, 1972. Thomas H lie and Xavier Rodet. Radiation of a pulsating e portion of a sphere: Application to horn radiation. Acustica, 89(4):565577, 2003. D. Haydock and J.M. Yeomans. Lattice Boltzmann simulations of acoustic streaming. Journal of Physics A, 34:5201 13, 2001. S. Ito, T. Fujisawa, and G. Yagawa. Computational uid analysis of the sound generation mechanism of air-reed instruments. Tech. Rep. Musical Acoustics MA-20, 7:17, 2002. (in japanese). Wilfried Kausel and Paul Anglmayer. A computer program for brass instrument optimization - part i. concept and implementation. In Proceedings of the 137th Meeting of the ASA and 2nd Convention of the EAA Forum Acusticum. DEGA, ASA, EAA, 1999.

[Hof98] [HP72]

[HR03]

[HY01]

[IFY02]

[KA99]

270

Bibliography

[KADA+ 03] I. V. Karlin, S. Ansumali, E. De Angelis, H. C. Ottinger, and S Succi. Entropic lattice boltzmann method for large scale turbulence simulation. J. Fluid Mech (Submitted May 2003), pages 119129, 2003. [Kar03] Matti Karjalainen. Time-domain physical modelling and real-time synthesis using mixed modeling paradigms. In Roberto Bresin, editor, Proceedings of the SMAC03 (Stockholm Music Acoustics Conference 2003), pages 393396, Stockholm, 2003. KTH Speech, Music and Hearing. Wilfried Kausel. An Optimised Architecture for a based Multi-Channel Codec & Filter Device. PhD thesis, Faculty of Electrical Engineering, Univ. of Technology, Vienna, 1999. Wilfried Kausel. Optimization of brasswind instruments and its application in bore reconstruction. Journal of New Music Research, (1 Vol. 30):6982, 2001. Wilfried Kausel. An attempt to use an electrical circuit simulator to better understand the relationship between a brass players intonation and the instruments input impedance. In European Acoustics Association, editor, Revista de Acustica (Proceedings of Forum Acusticum Sevilla), pages MUS06013, Sevilla, 2002. Sociedad Espanola de Acustica. Wilfried Kausel. Bore reconstruction of tubular ducts from acoustic input impedance curve. In IEEE Instrumentation and Measurement Technology Conference, Vail, CO, 2003. IEEE. Wilfried Kausel. Studying lip oscillators of brass instruments: A distributed two dimensional lip model and its electrical equivalent circuit. In Royal Swedish Academy

[Kau99]

[Kau01]

[Kau02]

[Kau03a]

[Kau03b]

271

Bibliography

of Music, editor, Proceedings of the SMAC 03, pages 205 208, Stockholm, 2003. Royal Swedish Academy of Music,. [Kee81] Douglas H. Keefe. Woodwind Tone Hole Acoustics and the Spectrum Transformation Function. PhD thesis, Physics Dept., Case Western Reserve University, Cleveland, OH, 1981. Douglas H. Keefe. Theory of the single woodwind tone hole. Journal of the Acoustical Society of America (JASA), 3:676687, 1982. Douglas H. Keefe. Acoustical wave propagation in cylindrical ducts: Transmission line parameter approximations for isothermal and nonisothermal boundary conditions. Journal of the Acoustical Society of America (JASA), 75(1):5862, 1984. Douglas H. Keefe. Woodwind air column models. Journal of the Acoustical Society of America (JASA), 88(1):3551, 1990. Jonathan A. Kemp. Theoretical and experimental study of wave propagation in brass musical instruments. PhD thesis, University of Edinburgh, Physics Department, Edinburgh, 2002. D. A. Kopriva and J. H. Kolias. A solution of acoustic workship problems by a spectral multidomain method. In ICASE/LaRC Workshop on Benchmark Problems in Computational Aero-Acoustics. NASA Conference Publication 3000, 1995. Wilfried Kausel and Helmut K hnelt. Modelling wave and u uid propagation in wind instruments - methods and applications. In Mirko Cudina, editor, Proceedings of the First

[Kee82]

[Kee84]

[Kee90]

[Kem02]

[KK95]

[KK03]

272

Bibliography

Congress of AAAA, page 14, Portoroz, 2003. Alps Adria Acoustics Association (AAAA). [KL96] J. W. Kim and D. J. Lee. Optimized compact nite difference schemes with maximum resolution. AIAA Journal, 34(5):887893, 1996. D. A. Kopriva. Spectral solution of the time dependent euler gasdynamics equations. In Proc. of the 14th Aero-Acoustics Conference, Paper DGLR/AIAA 92-02036. DGLR/AIAI, 1992. Y. Kagawa, T. Tsuchiya, B. Fujii, and K. Fujioka. Discrete Huygens model approach to sound wave propagation. Journal of Sound and Vibration, 218(3):419444, 1998. Helmut K hnelt. u Simulation Schallfelder in Fl teninstrumenten o Boltzmann-Methode. Masters thesis, Klangstil, University of Music and Vienna, 2002. dreidimensionaler mittels LatticeInstitut f. Wiener performing Arts,

[Kop92]

[KTFF98]

[K h02] u

[K h03] u

Helmut K hnelt. Simulating the Mechanism of Sound u Generation in Flutes using the Lattice Boltzmann Method. In Royal Swedish Academy of Music, editor, Proceedings of the SMAC 03, pages 401404, Stockholm, 2003. Royal Swedish Academy of Music,. D. P. Lockard, K. S. Brentner, and H. L. Atkins. High-accuracy algorithms for computational aeroacoustics. AIAA Journal, 33(2):246251, 1995. S. K. Lele. Compact nite difference schemes with spectral like resolution. Journal of Computational Physics, 103:16 42, 1991.

[LBA95]

[Lel91]

273

Bibliography

[Lig78] [LL00] James Lighthill. Waves in Fluids. Cambridge University Press (repr. 2001), UK, 1978. P. Lallemand and L. S. Lou. Theory of the lattice Boltzmann method: Dispersion, dissipation, isotropy, galilean invariance, and stability. Physical Review E, 61(6):6546 62, 2000. Roe P. L. van Leer B. Lowrie, R. B. A space-time discontinuous galerkin method for the time-accurate numerical solution of hyperbolic conservation laws. In Proc. of the 12th Computational Fluid Dynamics Conference, pages AIAA Paper 951658. AIAA, 1995. Harvard Lomax and Zingg David W. Pulliam, Thomas H. Fundamentals of Computational Fluid Dynamics. Springer, Berlin, 2001. H. Levine and J. Schwinger. On the radiation of sound from an unanged circular pipe. Phys. Rev., 73:383406, 1948. J. Lucero. Dynamics of the two-mass model of the vocal folds: Equilibria, bifurcations, and oscillation region. Journal of the Acoustical Society of America (JASA), 94:3104 3111, 1993. P. O. L thi. Lattice Wave Automata - from Radio Waves to u Fracture Propagation. PhD thesis, Genf, 1998. R. Magalotti, G. Borin, and G. De Poli. A mew kind of dipole source for time-domain simulation of ute-like instruments. In Socit franaise dacoustique, editor, Proceedings of the International Symposium of Musical Acoustics [ISMA 1995], pages 490495, Dourdan (F), 1995. SFA. Philip M. Morse. Vibration and Sound. repr. Acoustical Society of America, 1991, USA, 1936.

[Low95]

[LP01]

[LS48] [Luc93]

[L t98] u [MBDP95]

[Mor36]

274

Bibliography

[MR93] Dan Mapes-Riordan. Horn modeling with conical and cylindrical transmission-line elements. Journal of the Audio Engineering Society, 41(6):471483, 1993. Daniel Noreland. A numerical method for acoustic waves in horns. Acta Acustica, 88(4):576586, 2002. V. Pagneux, Noam Amir, and J. Kergomard. A study of wave propagation in varying cross-section waveguides by modal decomposition. part i. theory and validation. Journal of the Acoustical Society of America (JASA), 100(4):2034 2048, 1996. Xavier Pelorson, Pierre Badin, Pierre Saguet, and Samir El Masri. Numerical simulation of the vocal tract acoustics using the TLM method. In Proceedings of Forum Acusticum Sevilla 2002, pages NUM04004IP, Sevilla, 2002. Hirzel Verlag. H. Pichler, G. Widholm, and W. Winkler. Rechnergesttzte Messung der Parameter von Blechblasinstrumenten. In Proceedings 6th International FASE-CONGRESS 1992, pages 241244, Zrich, 1992. SAS. Y.H. Qian, D. dHumieres, and P. Lallemand. Lattice bgk models for navier-stokes equation. Europhysics Letters, 17(6):47984, 1992. Y.H. Qian and S.A. Orszag. Lattice bgk models for the navier-stokes equation: Nonlinear deviation in compressible regimes. Europhysics Letters, 21(3):2559, 1993. J. W. S. Rayleigh. The Theory of Sound. Dover, 1976, repr. New York, 1877. J. W. S. Rayleigh. The Theory of Sound, Volume Two, 2nd Ed. Dover, 1976, repr. New York, 1894.

[Nor02] [PAK96]

[PBSEM02]

[PWW92]

[QdL92]

[QO93]

[Ray77] [Ray94]

275

Bibliography

[RCG03] Orlando Richards, D. Murray Campbell, and Joel Gilbert. Modelling the lip reed - computational and experimental investigations of the two-mode inward/outward striking behaviour. In Royal Swedish Academy of Music, editor, Proceedings of the SMAC 03, pages 233235, Stockholm, 2003. Royal Swedish Academy of Music. O. Richards, D. M. Campbell, J. Gilbert, and M. A. Neal. Use of experimental studies in determining a two-mass lip model. In European Acoustics Association, editor, Proceedings of Forum Acusticum Sevilla 2002, pages MUS 04005, Sevilla, 2002. Hirzel Verlag. S. W. Rienstra and A. Hirschberg. An Introduction to Acoustics. University of Technology, Eindhoven, Eindhoven, NL, 2002. E. G. Richardson. The Acoustics of Orchestral Instruments and of the Organ. Arnold, London, 1929. H. H. Rosenbrock. An automatic method for nding the greatest or least value of a function. Comp. J., 3:175184, 1960. Paul M. Radavich, Ahmet Selamet, and James M. Novak. A computational approach for owacoustic coupling in closed side branches. Journal of the Acoustical Society of America (JASA), 109(4):13431353, 2001. William Strong and Melville Clark. Synthesis of windinstrument tones. Journal of the Acoustical Society of America (JASA), 41(1):3952, 1967. Gary P. Scavone. Modeling vocal-tract inuence in reed wind instruments. In Roberto Bresin, editor, Proceedings of the SMAC03 (Stockholm Music Acoustics Conference),

[RCGN02]

[RH02]

[Ric29] [Ros60]

[RSN01]

[SC67]

[Sca03]

276

Bibliography

pages 291294, Stockholm, 2003. KTH Speech, Music and Hearing. [Sch77] H. P. Schwefel. Numerische Optimierung von Computermodellen mit der Evolutionsstrategie. Birkh user, Basel, a 1977. R. T. Schumacher. Ab initio calculations of the oscillations of a clarinette. Acustica, 48:7185, 1981. H. P. Schwefel. Evolution and Optimum Seeking. Wiley, New York, 1994. A. Sarti and G. De Poli. Toward nonlinear wave digital lters. IEEE Trans. Sig. Process., 47(6):16541668, 1999. Adachi Seiji. Numerical analysis of an air jet: Toward understanding sounding of air-jet driven instruments. In European Acoustics Association, editor, Proceedings of Forum Acusticum Sevilla 2002, pages MUS04002, Madrid, 2002. Sociedad Espanola de Acustica / Hirzel. David B. Sharp. Acoustic pulse reectometry for the measurement of musical wind instruments. PhD thesis, University of Edinburgh, Physics Department, 1996. Julius O. Smith III. Digital waveguide models for sound synthesis based on musical acoustics. In Mike NEWMAN, editor, Proceedings of the 15th International Congress on Acoustics, Trondheim 1995, pages 489492, Trondheim (N), 1995. ASN. Julius O. Smith III. Nonlinear commuted synthesis of bowed strings. In ICMA & IPSA, editor, Proceedings International Computer Music Conference [ICMC 1997], pages 264267, Thessaloniki (GR), 1997. Aristotle University of Thessaloniki.

[Sha96]

[SI95]

[SI97]

277

Bibliography

[SI03] Julius O. Smith III. Recent developments in musical sound synthesis based on a physical model. In Royal Swedish Academy of Music, editor, Proceedings of the SMAC 03, pages 35, Stockholm, 2003. Royal Swedish Academy of Music. G. Skordos, P.; Sussman. Comparison between subsonic ow simulation and physical measurements of ue pipes. In Proc. ISMA 1995, pages 7985, Le Normont, France, 1995. J. O. Smith. Physical modeling using digital waveguides (music synthesis) iii. COMPUTER MUSIC JOURNAL, 16(4):7491, 1992. J. O. Smith. Physical modeling update. Computer Music J., 1996(20):4456, 1996. Julius O. Smith. Principles of digital waveguide models of musical instruments. In M. Kahrs and K. Brandenburg, editors, Applications of DSP to Audio & Acoustics, pages 416 466. Kluwer Academic Publishers, 1998. Gary P. Scavone and Julius O. and Smith III. Scattering parameters for the keefe clarinet tonehole model. In Arnold MYERS, editor, Proceedings of the International Symposium of Musical Acoustics [ISMA 1997], pages 433438, St Alban (UK), 1997. Institute of Acoustics. J. Saneyoshi, H. Teramura, and S. Yoshikawa. Feedback oscillations in reed woodwind and brass wind instruments. Acustica, 1987(62):194210, 1987. J. O. Smith and S. Van Duyne. Recent results in piano synthesis via physical modeling. In Socit franaise

[Sko95]

[Smi92]

[Smi96]

[Smi98]

[SSI97]

[STY87]

[SVD95]

278

Bibliography

dacoustique, editor, Proceedings of the International Symposium of Musical Acoustics [ISMA 1995], pages 503510, Dourdan (F), 1995. SFA. [Tho72] [Tho96] P. A. Thompson. Compressible-Fluid Dynamics. McGrawHill, New York, 1972. Jeffrey Paul Thomas. An investigation of the upwind leapfrog method for scalar advection and acoustic/aeroacoustic wave propagation problems. PhD thesis, University of Michigan, Department of Aerospace Engineering, 1996. Michael W. Thompson and William J. Strong. Inclusion of wave steepening in a frequency-domain model of trombone sound production. Journal of the Acoustical Society of America (JASA), 110(1):556562, 2001. Michihisa TSUTAHARA, Kenji SHIKATA, and Takeshi KATAOKA. Direct simulation of edge-tones by the nite difference lattice Boltzmann method. In Proc. of the 16th Computational Fluid Dynamics Symposium, pages A132, 2002. (in japanese). C. K. W. Tam and J. C. Webb. Dispersion-relationpreserving nite differencing schemes for computational acoustics. Journal of Computational Physics, 107:262 281, 1992. Franz J. Vesely. Computational Physics. Einfhrung in die Computative Physik. WUV Universittsverlag, Wien, 1993. C. E. Vilain, X. Pelorson, A. Hirschberg, L. Le Marrec, W. Opt Root, and J. Willems. Contribution to the physical modeling of the lips. inuence of the mechanical boundary conditions. ACTA ACUSTICA/ACUSTICA, 89(5):882887, 2003.

[TS01]

[TSK02]

[TW92]

279

Bibliography

[Wal01] [WC02] Matthew Wall. Galib - genetic algorithm package, 2001. http://lancet.mit.edu/galib-2.4/Overview.html. Maarten van Walstijn and D. Murray Campbell. Largebandwidth measurement of acoustic input impedance of tubular objects. In European Acoustics Association, editor, Proceedings of Forum Acusticum Sevilla 2002, pages MUSGen004, Sevilla, 2002. Hirzel Verlag. Maarten van Walstijn and Murray Campbell. Discrete-time modeling of woodwind instrument bores using wave variables. JASA, Journal of the Acoustical Society of America, 113(01/2003):575 585, 2003. A. G. Webster. Acoustical impedance and the theory of horns and the phonograph. repr. in J. Audio Eng. Soc. 1977, 25(1/2):2428, 1919. J. C. Webster. An electrical method of measuring the intonation of cup-mouthpiece instruments. Journal of the Acoustical Society of America (JASA), 5:902906, 1947. P. Welander. On the temperature jump in a rareed gas. Arkiv fr Fysik, 7(4):50733, 1954. D. Wolf-Gladrow. Lattice-Gas Cellular Automata and Lattice Boltzmann Models: An Introduction. Springer-Verlag, New York, Heidelberg, Berlin, 2000. Gregor Widholm. BIAS Ein Computersystem zur Beurteilung der Qualitt von Blasinstrumenten. Instrumentenbau-Zeitschrift, (7/8):4549, 1993. Gregor Widholm. Brass wind instrument quality measured and evaluated by a new computer system. In Mike NEWMAN, editor, Proceedings of the 15th International

[WC03]

[Web19]

[Web47]

[Wel54] [WG00]

[Wid93]

[Wid95]

280

Bibliography

Congress on Acoustics, Trondheim 1995, pages 517520, Trondheim (N), 1995. ASN. [Wid99] Gregor Widholm. Diagnosis and therapy for brasses. measurement, evaluation and improvement of the acoustic quality of brass wind instruments. In Proceedings of the 137th Meeting of the ASA and 2nd Convention of the EAA Forum Acusticum. DEGA, ASA, EAA, 1999. Andreas Wilde. Virtual Wave Tank. http://www.eas.iis.fraunhofer.de/projects/simaksig/vwt/index de.html. Andreas Wilde. Simulation of the sound wave propagation in turbulent ows using a lattice Boltzmann scheme. In European Acoustics Association, editor, Proceedings of Forum Acusticum Sevilla 2002, pages ELE01008, Sevilla, 2002. Hirzel Verlag. Stephen Wolfram. Mathematica: A System for Doing Mathematics by Compute. Addison-Wesley, 1988. S. Wolfram. Cellular automaton uids: Basic theory. In S. Wolfram, editor, Cellular automata and complexity: collected papers, pages 359408. Addison-Wesley, 1994. Gregor Widholm, Heinrich Pichler, and Thomas Ossmann. BIAS - A Computer Aided Test System for Brass Instruments. AN AUDIO ENGINEERING SOCIETY PREPRINT, (2834), 1989. Valana L. Wells and Rosemary A. Renaut. Computing aerodynamically generated noise. Annual Review Fluid Mechanics, 29:161199, 1997. G. Widholm and W. Winkler. Evaluation of musical instrument quality by computer systems. examples of reali-

[Wil] [Wil02]

[Wol88] [Wol94]

[WPO89]

[WR97]

[WW94]

281

Bibliography

sation. In Proceedings of the SMAC 93, pages 560565, Stockholm, 1994. Royal Swedish Academy of Music. [YM03] Shigeru Yoshikawa and Yoko Muto. Lip-wave generation in horn players and the estimation of lip-tissue elasticity. Acta Acustica, 89(1-2/2003):145 162, 2003. D. W. Zingg, H. Lomax, and Jurgens H. An optimized nite difference scheme for wave propagation problems. AIAA Paper, 93(459), 1993. William E. Zorumski. Generalized radiation impedances and reection coefcients of circular and annular ducts. Journal of the Acoustical Society of America (JASA), 54(6):16671673, 1973. Wei Zhao, Cheng Zhang, H. Frankel, and Luc Mongeau. Computational aeroacoustics of phonation - part I: Computational methods and sound generation mechanisms. Journal of the Acoustical Society of America (JASA), 112:2134 2146, 2002.

[ZLH93]

[Zor73]

[ZZFM02]

282

A.1. Signal Sampling and Reconstruction

Dealing with digitally sampled signals has become so common nowadays that most of the rules and properties which are associated with signal sampling and reconstruction are usually invisible to a typical user. Sound cards of personal computers or compact disc players have already built in all the means to sample and reconstruct audio signals properly as long as some standard procedures are followed. The reason why a theoretical introduction to signal sampling and reconstruction is included in this book is that as soon as the very basic way of using sound recording or data acquisition systems is departed from, a true understanding of all the basic relationships is mandatory. The task of selecting an appropriate windowing function in any of the public domain spectral analysis programs already exceeds a typical users capabilities. The same is true when the program defaults for buffer length, sampling frequency, aliasing lter parameters. . . turn out to be inappropriate. A scientist who will write numerical simulation programs involving digital wave guides or wave digital lters will probably be aware of these fundamentals. Therefore they have been moved to the Appendix of this book as a reference, just in case, for nobody is perfect.

283

L H

1.5 1 0.5

f t 1.5 1 0.5

10

20

30

40

When a continuous analogue signal as shown in Figure A.1(a) is periodically sampled, a corresponding time discrete stream of signal magnitude values (Figure A.2(a), A.2(b)) is created. In other words, a sampled signal is generated by sequentially outputting the amplitudes of a corresponding continuous signal taken at all integer multiples of T , the basic sampling interval. Theoretically, a sampled signal has no beginning and no end just like its analogue counterpart. In practice it will not be possible to process or store signal buffers of innite lengths, so effects caused by truncation of an innite signal buffer have to be taken into account. Multiplying a signal by a so called window function which is non-zero only in a certain nite time interval is normally referred to as windowing. Many useful window functions (Hamming, Hanning, Kaiser, Nuttall. . . ) are known. Figure A.1(b) shows the most simple case of a rectangular window. The effect of this windowing on the results of a subsequent Fourier transformation is generally not negligible and must be studied carefully. Only if the buffer is very long compared to the period of any contained signal component can the effects caused by the multiplication of the analyzed signal with a rectangular window be neglected.

284

The rectangular observation window does not cause any spectral distortion at all if only integer number of periods of signal components are present in the buffer. In this case the basic assumption of the Fourier transform, that the observed signal is strictly periodical and the buffer content represents exactly one of these innite number of periods, is met.

(a) T=1.

Very often those magnitude values composing a sampled signal are quantized in order to be represented by digital values (shown in Figure A.3(a)) introducing a second level of discretization which adds the so called quantization noise (Figure A.3(b)) to the signal. This quantization noise very often cannot be neglected especially when dealing with digital lters, but it can be separated from the spectral effects caused by discretization in the time-domain. The quantization noise caused by the Wave-Audio PCM format (16 bit, linear) is low (1/65536, 96 dB) but the standardized PCM quantization schemes used in telecom systems (North American -law, European Alaw) introduce a very signicant error which dominantly limits signal to total distortion ratio as well as gain tracking and noise gures. In this chapter only the effects caused by discretization in the time-domain are dealt with.

f n T

f n T

(b) T=2.

285

L H

t 5 -1 -2 -3 -4 10 15 20 25 30 qf t 3 2 1

A.1.2. Aliasing

In order to understand the frequency spectra of sampled and reconstructed signals, sampling at a rate equal to the frequency of the sampled signal should be considered. Assuming a phase relationship of /2, the peak of the signal is always sampled. The sampled result appears to be a DC signal and cannot be differentiated from a real DC signal. Figure A.4 illustrates this condition. Normally the sampled signal will not be correlated with the sampling clock. Therefore the phase relationship will not be predictable which even makes the sampled amplitude meaningless. If the signal frequency is now chosen to be slightly below or slightly above the sampling frequency, an interference is generated giving the illusion a much slower signal has been sampled. From the sampled amplitudes it is not possible to differentiate between two frequencies one above and one below the sampling rate providing they are symmetric around the sampling rate (refer to Figure A.5(a) and A.5(b)). In Figure A.6 it can be seen that the so-called Nyquist frequency, where samples alternatingly represent a single positive and a single negative signal value is half the sampling frequency. If the sampling clock happens to have a phase delay of /2 relative to the sine waves phase, then the positive and the negative maximum are sampled. Analyzing the case where the phase delay is a multiple of is left as an exercise for the reader.

286

L H

5 10 15 20 25 30

e t

0.5

-0.5

-1

Figure A.4.: Sampling at the Signal Rate

0.5

0.5

-0.5

-1

-1

L H

L H

f t

L H

t 100 200 300 400 500

f t t 100 200 300 400 500

f t

287

f(t) 1 0.5 t

-10 -0.5 -1

10

20

30

40

Figure A.6.: Sampling at the Nyquist Rate (sampling rate is twice the signal frequency)

The Nyquist frequency is the highest signal frequency which is not transformed into a different frequency by the sampling process itself. Any lower frequency can unambiguously be sampled and reconstructed from a stream of digital signal values.

The described observations can be summarized by a few important relationships which one has to be aware of when sampled data systems are designed. Signal frequency components which are higher than the Nyquist frequency (half of the sampling rate) are aliased or mirrored into the baseband between zero and the Nyquist frequency where those mirror images

288

of higher bands are added to the original base-band signal. Those frequencies are mirrored in a way that they appear in the base-band as having a frequency corresponding to the frequency difference of their real frequency and their closest multiple of the sampling frequency. This distorting spectral effect of sampling signals containing out-ofband components is illustrated in Figure A.7 and Figure A.8. While Figure A.7 shows an arbitrary base-band spectrum and all its mirror images around the sampling frequency and its multiples, Figure A.8 shows the resulting spectrum which is the superposition of all those components. This implies Nyquists law: True reconstruction of a base-band signal from a sampled data stream is not possible unless during sampling precautions have been taken not to distort the base-band signal by mirror images of higher frequency components. That means before sampling a signal for later reconstruction it has to be ensured that only spectral components below the Nyquist frequency are present in the signal before sampling.

f [Hz] 50000 -10 -20 -30 -40 -50 -60 X( ) [dB] images caused by sampling at 64000 Hz 100000 150000 200000 250000

Figure A.7.: Continuous Signal Spectrum and its Mirror Images created by Sampling

289

4 2 f [kHz] 50 -2 -4 100 150 200 250

The spectral components of the sampled signal which are above the Nyquist frequency are not virtual. They are real at least until an actual circuit is used to reconstruct the analogue signal. They originate from the denition of a sampled signal because by denition it represents a contiguous equidistant sequence of delayed Dirac pulses weighted with the amplitude values of the corresponding sample of the related continuous time signal. Therefore the theoretically ideal way of reconstructing an analogue signal from a sampled data stream is to generate this sequence of weighted Dirac pulses and to apply an ideal low-pass lter to cut all the higher mirror images of the base-band i.e. to smoothen the signal until it reects the one which was to be reconstructed.

Finally the correct mathematical formalism describing the above observations and considerations should be looked at in order to justify the generalizations which have been made. As we have said, the sampled signal,

290

-6

X( ) [dB]

which by denition corresponds to a given continuous signal, can be described as a product of the original continuous function f (t) with a sampling function s(t). This is a periodic impulse train, actually the innite sum of ideal unit impulse functions located at all multiples of the sampling period T .

s(t) =

l=

(t lT ),

(A.1)

with (t) being the ideal Dirac impulse function with unity area below the curve and with zero value for t = 0. With this denition the relationship between the spectrum of the continuous time signal and its corresponding sampled signal should be investigated. The spectrum of the continuous time signal can be derived by applying the Fourier transform F

F {x(t)} = X(j) =

x(t)ejt dt.

(A.2)

The sampling function s(t) has a spectrum dened by its Fourier transform (with s = 2/T )

F {s(t)} = S(j) =

l=

s(t)ejt dt =

(t lT ) ejt dt = F {(t lT )} =

(t lT )ejt dt =

=

l=

l= jlT

= s

l=

( ls )A.3) (

l=

For the last transformation no proof is given here. In Figure A.9(a) and

L

l=L

ejlT approaches

291

the sequence of ideal unity impulses described by the nal term as soon as the number of elements in the sum is growing. This result implies that a uniformly spaced impulse train in time, s(t), transforms to a uniformly spaced impulse train in frequency, S(j).

1 0.8 0.6 0.4 0.2 f 2 3 4 5 2 3 4 5 1 0.8 0.6 0.4 0.2 f

Since multiplication in the time-domain is equivalent to convolution in the frequency-domain, we have the relation

F {x(t)s(t)} =

x(t)s(t)ejt dt =

1 1 F {x(t)} F {s(t)} = X(j) S(j) (A.4) = 2 2 with * denoting a linear convolution of X(j) and S(j) in frequency. Considering that S(j) is the uniformly spaced pulse train with a spacing of s it becomes clear that the linear convolution 1 X(j) S(j) = 2 1 = 2

X(j j )S(j )d =

s 2

X(j ljs )

l=

(A.5)

292

does create mirror images of the base-band spectrum for each pulse of s i.e. for each multiple of the sampling frequency as already observed above. S(j) is zero between the pulses and one during the pulse. Nonzero contributions to the integral are X(j j0 ) , X(j j1 ), X(j j2 ), X(j j3 ) . . . with ji being the ith multiple of the sampling frequency 0 . Figure A.10 illustrates the effect of the linear convolution of the spectrum of a band-limited, continuous-time signal with this uniformly spaced pulse train in the frequency-domain. The resulting spectrum is indeed the spectrum of an ideally sampled signal, that is a signal which is zero in between the samples and which contains innitesimal narrow spikes of innite height carrying the whole energy of the signal.

X( )

1

S( )

1

X ( ) * S( )

1

Any other representation of a sampled signal like the one which is created by well known sample and hold circuits has a different spectrum, which naturally can be derived from the basic spectrum multiplied by the transmission function of f.e. a sample and hold term. By ideal low-pass

-2 f s

2 fs

293

-2 f s

2 fs

-2 f c

2 fc

ltering it is possible to eliminate all spectral mirror images except the base-band and to perfectly reconstruct the original continuous time signal, providing the Nyquist theorem was met when the sampled data stream was created. Let us now take a closer look on the ideal Dirac impulse function (t). We said already that an ideal Dirac impulse has unity area below the curve and a zero value for t = 0. The spectrum of one Dirac impulse is dened by its Fourier transform

F {(t)} = (j) =

(t)ejt dt = lim

= lim

1 jt e dt = 2

(A.6)

In the above calculation, the Dirac pulse has been approximated by a rectangular pulse with a height of 1/2 and a width of 2 which satises the unity area condition regardless how small is assumed to be. The result shows that the spectrum of an ideal Dirac impulse is a constant 1, independent of frequency. A Dirac impulse is equivalent to a superposition of an innite number of ideal sine-waves equally representing all frequencies between inf ty and +inf ty and having unity amplitude and zero phase. Let us dene w(t) as an ideal rectangular impulse with amplitude 1 and width 2p. As above we obtain

p

F {w(t)} = W (j) =

w(t)e

jt

dt =

p

ejt dt =

sin(p) j jp e ejp = 2p p

(A.7)

294

This result is very important because in real world signal processing it is not possible to extend a measurement to t = . It is therefore necessary to multiply a steady state input signal with a rectangular window, which is zero for t < 0 and t > t1. If the spectrum of a time-domain window is known, the spectrum of an input signal that has been multiplied by that window can be derived by convoluting the two spectra. A linear transmission system is a system which transforms a linear combination of sine-waves that have arbitrary frequencies into a linear combination of sine-waves of the same frequencies modifying their coefcients and their phases in accordance with a complex function H(f ) its so-called transmission function. If such a linear system H(j) is stimulated by an ideal Dirac impulse with its spectrum (j) = 1 the spectrum of the pulse response is H(j) itself. So, by stimulating a linear system with an ideal Dirac impulse in the time-domain, and measuring or calculating the systems pulse response, and by Fourier transforming that measured result, the frequency-domain transmission function H(j) is obtained.

There is another way to look at the spectrum of a sampled signal. The spectrum is dened by the Fourier transform of the sampled signal i.e. of the product of the continuous time signal with the sampling function. The Fourier transform is dened as a denite integral of its input function multiplied by an exponential time dependent term. We know that the sampling function is zero most of the time so we can replace the integral by the innite sum of the contributions of all signal samples. The integral of one Dirac pulse is 1, which means that the contribution of one sample is the value of the continuous time input function x(t) at the exact position of the sample, a multiple of the sampling period.

295

F {x(t)s(t)} =

x(t)s(t)ejt dt =

l=

=

l=

x(lT )ejlT =

xc [l]z l

(A.8)

In the nal term a transition has been made from the continuous time function x(t) to the table of discrete sampling values xc [n]. Also the abbreviation ejT = z has been introduced. Because of the correspondence between time-domain and frequencydomain, there is a transformation rule saying that a shift by an interval T in the time-domain transforms into a multiplication with ejT in the frequency-domain. Using this formalism it is possible to directly derive signal spectra and transmission functions in the frequency-domain from the signal samples in the time-domain. The product X(z)z1 denotes a sampled data stream which is delayed by one sample compared with x(nT ). The term X(z)z2 indicates a shift by two samples.

A simple example will be given to illustrate how the spectral inuence of a practical signal processing stage can conveniently be derived by means of the z-transformation. In the z-domain, this lter algorithm can be represented by a owchart that is entered by the input signal, which actually represents an equally spaced stream of input samples, and which creates an output signal that again represents a stream of corresponding output samples. The operators used in that owchart are adders, subtractors and two different types of multipliers. Multiplication with constant coefcients indicates a scaling operation applied to all samples of a signal. Multiplication with z1 represents a

296

register delaying its input stream by one sample, thus delivering the sample which was current in the previous clock period. Sampling rate changes are usually indicated by switch symbols representing ideal up-sampling or down-sampling operations. The ideal up-sampling operation does not introduce any spectral changes. Ideal up-sampling means insertion of a certain number of zero value samples between every two samples of the input stream to accommodate the required change in sampling rate. This operation will be covered in Section A.3 on page 303. Figure A.11 shows the z-domain owchart of a simple but important circuit which increases the input sampling rate by a factor of four by repeating each input sample four times. This is different from the ideal up-sampling operation where zero value samples are inserted. It will be shown that this sort of digital sample and hold has some inuence on the up-sampled spectrum, in fact. To calculate the transmission function of the lter part of the circuit, it is necessary to split its operation in an ideal up-sampling part and the lter part, which then can be treated separately. The owchart accordingly contains an ideal up-sampling unit and a summing unit calculating the sum of the current, the previous, the third and the fourth past sample. Three of the four summands are zero, which makes the output to repeat each input sample four times which exactly represents the required function. It should be noted that z = ejT = z 4 in the diagram corresponds to the original sampling rate 1/T of the input signal x while z = ejT is connected to the fourfold faster sampling rate 1/T of the output signal y. The circuit can be described in the z-domain by means of y(z) = x(z)(1 + z 1 + z 2 + z 3 ). The transmission function H(z) results to H(z) = y(z)/x(z) = 1 + z 1 + z 2 + z 3 = Back substitution of z ejT = ej2f /fs (A.11) 1 z 4 . 1 z 1 (A.10) (A.9)

297

z-1 4 x(z) + y(z) z-1 z-1

x(z)

yields the spectral inuence of the ltering operation in the resp. f domain. Figure A.12 shows a plot with normalized frequency axes (sampling rate fs ) and logarithmic amplitude. It is the spectrum of a rst order so called comb lter.

Figure A.12.: Transmission Function of Digital Sample and Hold

0.4

0.6

0.8

298

Repeating samples several times at a new higher sampling rate is in fact an efcient way to multiply the sampling rate. The rst order comb lter which already attenuates unwanted mirror images of the base-band effectively comes for free. Anyhow, the pass-band droop of its transmission function must be considered.

As already said, it is not possible to perfectly reconstruct an analogue signal from a sampled data stream if the original analogue signal has not been perfectly low-pass ltered before it was sampled. It was shown that the spectrum of a sampled signal is a superposition of the original input spectrum with all the mirror images of that spectrum around the sampling frequency and all their multiples.

This means that by sampling a continuous input signal with a spectrum between 0 and innity this spectrum is folded innite times into the baseband, the range between 0 and half the sampling frequency - the so called Nyquist frequency - where all these mirror images are superimposed. This effect has been illustrated once more in Figure A.13. To keep the drawing simple, just the base-band of the resulting spectrum has been taken into account knowing that higher order mirror images will be cut off by the reconstruction lter anyhow. In order to preserve a true image of the interesting part of an input spectrum during sampling, it is required to select a sampling frequency which is at least two times higher than the bandwidth of the so called useful baseband. Additionally it is essential to zero all out-of-band noise and other undesired high frequency signal components by satisfactory low-pass ltering of the signal before sampling. It should be noted once more that the periodically repetitive spectrum of an ideally sampled signal is indeed a true spectrum and a spectrum

299

F N y q u is t

f s a m p le

F N y q u is t

analyzer would show all the mirror images if a more or less ideal weighted pulse train would be connected to its input. This clarication helps to make it clear that there is no difference in sampling a continuous time signal and in re-sampling an already sampled signal. The same folding and replication operation which is applied to the spectrum of the analogue signal can be applied to the spectrum of an already sampled signal. Let us assume that a new sampling frequency fs2 , which is 2 times lower than the original sampling frequency fs1 , is used to downsample an already sampled signal. Let us also assume that the signal was low-pass ltered before it was sampled the rst time in a way to satisfy the Nyquist theorem not only related to the rst sampling frequency but also for the new two times lower sampling frequency. That means the cut-off frequency fc is lower than fs2 /2. The spectrum of the signal after the rst sampling process X1() is periodic in 2fs1 . Down-sampling by a factor of two, which actually means dropping

300

every 2nd signal value, applies the folding and replication process to the spectrum that has been described above. The resulting spectrum X2() obtained this way is shown in Figure 40. It is now periodic in 2fs2 .

X 1( )

X 2( )

Figure A.14.: Re-Sampling of a Sampled Signal at a Two Times Lower Rate (Decimation) It is not required to do all the low-pass ltering in advance. The requirements of the Nyquist theorem can be met for each sampling or resampling operation separately. Figure 41 shows a case where a sampled signal X1() with a bandwidth close to its Nyquist frequency is digitally low-pass ltered to get X2() in order to cut off spectral components above the Nyquist frequency of the re-sampling process. This properly ltered signal may now undergo the actual down-sampling process (dropping every 2nd sample) and has a resulting spectrum X3() which is now symmetric to and periodic in the new sampling frequency 2fs2 . This ltering and down-sampling process, which is often referred to as decimation, is not limited to a factor of two. As long as the decimation ratio is an integer number, down-sampling can be performed by periodically selecting samples of the input signal to be output at the lower rate, dropping everything in between. The spectrum of the decimated signal will be symmetric to and periodic in the lower sampling frequency 2fs2 regardless how many copies of

-4 f s2

-2 f s 2

2 fs2

4 f s2

301

-2 f s1

-2 f c

2 fc

2 f s1

X 1( )

X 2( )

X 3( )

the base-band have to be inserted between a multiple of the original sampling frequency (n2fs1 ) and its neighbor multiple.

Reducing the sampling rate by a factor which is not an integer number can no longer be accomplished by a simple skipping of signal samples because taking samples in between the sampling points where a sampled signal is dened to be zero would be required in this case. Considering the required spectral periodicities before and after the non-integer sampling rate reduction it can be seen that the base-band would not survive the folding and replication operation undisturbed. Either it is possible to up-sample the signal to an intermediate rate that is an integer multiple of the desired rate or it is necessary to intermediately switch to the continuous time signal before sampling at the output rate can be performed. The continuous time signal is reconstructed from the pulse

302

-4 f s2

-2 f s 2

2 fs2

4 f s2

-2 f s1

-2 f c2

2 f c2

2 f s1

-2 f s1

-2 f c 1

2 fc1

2 f s1

train by ltering and cutting off all higher mirror images of the base-band using an analogue low-pass lter. Because of the overhead involved in such a general re-sampling operation everybody will try to avoid it and to select sampling frequencies with an integer ratio wherever possible.

A sampled signal has been dened as the product of a continuous time signal and a sample function that is zero in between all sample points. The sampling function has non-zero values only inside of an innitesimal narrow area around all multiples of the sampling period. This implies that no change in the sampled signal spectrum can be caused by insertion of additional zero samples where the sampled signal is already zero. A spectrum analyzer could not even detect the new higher sampling rate as long as the inserted extra samples are zero. It still would only see the narrow spikes of weighed -functions at the multiples of the sampling period carrying the whole information contained in the base-band of the signal. Figure A.16 illustrates that it is just the marker for the sampling frequency which is different in both spectra.

X 1( )

X 2( )

-2 f s2

2 f s2

303

-4 f s1

-2 f s1

-2 f c

2 fc

2 f s1

4 f s1

replications of the pass-band are caused by the mirroring effect of the sampling frequency 2fs1 there is one replication of the base-band spectrum between neighbored multiples of the faster sampling frequency 2f s2 (crosshatched in Figure A.16), which is no longer enforced by the sampling laws. It can be ltered out by a digital low-pass lter. The ideal up-sampling process is called interpolation i.e. multiplication of the sampling rate by an integer factor with a subsequent ltering process cutting off former base-band mirror images.

A very simple digital lter with unity coefcients only has already been presented as an example for the z-domain analysis (Figure A.11). The same basic principle will be applied intensively when digital lters are to be analyzed. A digital lter is a circuit generating output samples from a certain number of recent input samples (Finite Impulse Response - FIR) or from a certain number of recent input and output samples (Innite Impulse Response - IIR). The lter topologies of these two types can be compared in Figure A.17(a) and Figure A.17(a). Both types are covered in more detail in the following sections.

x[n] a0 z-1 z-1 z-1 z-1 a1 a2 a3 a4

x[n] z a0 a1

-1

-1

-1

-1

y[n]

z-1

z-1

z-1

z-1

304

a2

a3

a4

b0

b1

b2

b3

b4

+

y[n]

The rst class of so-called Finite Impulse Response (FIR) lters is nonrecursive, meaning it does not have any feedback. So the lter response to a pulse stimulus will decay and the lter cannot output anything but zero after a time dened by the maximum storage depth for the input samples. This is called the number of taps of the lter. The calculation of the nth output sample yn requires just the weighted sum of only a nite number M of past input samples x. This is generally described by

M 1

yn =

k=0

ak xnk

(A.12)

with ak being the M coefcients of the lter. FIR lters are usually formulated in the z-domain, where x(z)z1 , x(z)z 2 , x(z)z 3 , . . . represent signal values x sampled 1, 2, 3, . . . sample periods before the actual one. The corresponding architecture is shown once more in Figure A.18. It can be seen that the circuit of Figure A.11 is already a special case of such an FIR lter with all coefcients ai = 1 and four taps. Its impulse response which is the output sequence of the lter when it is stimulated with a digital Dirac impulse, the sequence {1,0,0,0,0,0,0,. . . } is {1,1,1,1,0,0,0,0,. . . }. As in the example above, it is generally the case that the impulse response of an FIR lter corresponds to the coefcient vector in a unique way. As the only non-zero input sample of a digital -impulse moves from present to past, it is multiplied with all lter coefcients one by one starting with a0 , a1 , a2 . . . thus returning the coefcient series itself at the output y. In other words, the coefcient vector and the rst non-zero part of the impulse response of an FIR lter are identical. This fact helps to design an FIR lter with any desired transmission function in the j-domain. By discrete Fourier transformation of the desired lter spectrum an impulse response can be obtained, which can directly be implemented as coefcient vector of a corresponding FIR lter. In order to get real lter coefcients, the lter magnitudes must be sym-

305

x[n] z a0 a1

-1 -1 -1 -1

metric around fs /2 and antisymmetric in phase, which means that the desired spectrum must be upwards extended with a conjugate complex mirror image of itself. Another view of this fact can be obtained by keeping the basic correspondences between time and frequency-domain in mind. They can be directly derived from the properties of the Fourier transform itself. The timedomain convolution performed by the FIR lter as described in Equation A.12 corresponds to a frequency-domain multiplication with the transmission function H[j] which is the Fourier transform of the time-domain impulse response. FIR lters with a rectangular impulse response are often referred to as comb lters or Sinc lters. Their spectrum is the spectrum of a sampled rectangular pulse which can be veried using the derivations made in Section A.1.7 on page 296. Comb lters are very often used to smooth noisy curves or to downsample data streams. Often they are not recognized as such and rather referred to as running average or running RMS (if the energy content of the signal is to be smoothed or down-sampled). Because of their attractiveness they will be discussed in more detail now.

306

+

y[n]

a2

a3

a4

In the general case the transmission function of a comb lter is H(z) = which is equivalent to H(ejT ) = 1 eDjT 1 ejT

n

1 z D 1 z 1

(A.13)

(A.14)

with D being the so called decimation ratio (the number of non-zero samples in their impulse response) and n being the lter order. Its comb-like magnitude plot looks - at least in the range {0, T } similar to the | sin(x)/x| function which is often referred to as Sinc(x). An example for n = 2 and D = 128 is shown in Figure 44. The zeroes at 1 the decimated sampling frequency fs = DT and their multiples make this D lter suitable as a decimation lter. The resulting data stream can be down-sampled by the factor D. There is much aliasing of out-of-band noise caused by this down-sampling operation because the comb lters low-pass characteristic is by far less than excellent, but narrow frequency regions centered at all multiples of the decimated sampling frequency (the areas inside the notches) stay almost clean. And these are the frequency ranges which would be mirrored into the base-band starting at DC. Looking at the widths of the notches it can be seen that comb lters cannot be used to decimate from a very high sampling rate right away down to a very low one if a at pass-band is required. Depending on the actual base-band width, a comb lter will decimate down to a convenient intermediate rate, where a precision lter will clean the spectrum and atten the pass-band before the signal can be down-sampled to the nal output data rate. In Figure 45 a z-domain circuit topology is shown which creates the transmission characteristics of a 2nd order comb decimation lter. Other

307

Figure A.19.: Spectrum of 2nd Order Comb Filter with D=128

0.02

0.03

0.04

0.05

orders can easily be derived from this example by adapting the number of accumulator stages to the left as well as the number of differential stages to the right.

z-1

z-1

z-D

z-D

308

Y(z

-D

equations can be derived: T1 (z) T2 (z) T3 (z) Y (z) T1 (z)(1 z 1 ) T2 (z)(1 z

1

= = = = = = =

1

(A.15)

T2 (z) D ) )

D

T2 (z)

X(z)/(1 z 1 )2

2

(A.16)

FIR lters of this type do not require any multiplication operation because all their coefcients can be scaled to unity. A rst order Sinc decimation lter can even be implemented using a single accumulator - the differential stage can be omitted if the accumulator is cleared before each accumulation of D samples. Therefore lters of this type are extremely useful where high sampling rates must be reduced in a rst step before high precision ltering can be applied to provide the desired characteristics. Practical implementations will benet from the fact that the decimation switch (the actual reduction of sampling rate) can be moved to the place after the n cascaded accumulators on the left side because only every Dth sample is further processed in the n differential terms on the right side and nally propagated to the output. Comb lters and their efcient implementation are described in [Kau99]. This publication also covers numerical issues when nite precision arithmetic is involved.

309

Another important class of digital lters, the so-called IIR (Innite Impulse Response) or recursive lters, not only processes recent input samples but has also a feedback from its output. New output samples are calculated as weighted sum of a certain number of past input and output samples. Such a feedback loop has the inherent capability to store a state information. It can continue to produce output even when the input signal has already decayed completely. It can even self sustain oscillations when coefcients of past output samples are greater than one.

x[n] a0 z-1 z-1 z-1 z-1

z-1

z-1

z-1

z-1

310

b0

b1

b2

b3

b4

+

y[n]

a1

a2

a3

a4

the lter to generate non-zero output samples as a response to a digital Dirac impulse (a one followed and preceded by an innite number of zeros) even after any nite period of time. Therefore the impulse response theoretically has innite length. In practice the response of even an IIR lter will decay to zero because of the nite signal resolution dened by the number representation and depending on the coefcients a[k] and b[k] at least if the lter is stable. The operation of this class of lters can be generally described by

M 1 M 1

yn =

k=0

ak xnk +

k=0

bk ynk

(A.17)

Equation A.17 can be solved for y and rewritten in the z-domain where x(z)z 1 , x(z)z 2 , x(z)z 3 , . . . represent signal values x sampled 1, 2, 3, . . . sample periods before the actual one and where y(z) stands for the actual output sample, y(z)z1 for the previous output sample, and so on. This way the transmission function of the lter in the z-domain can be calculated. H(z) = a0 + a1 z 1 + a2 z 2 + . . . y(z) = x(z) 1 b0 b1 z 1 b2 z 2 . . . (A.18)

By choosing the orders of the numerator and denominator and by selecting actual values for the coefcient vectors a and b, nearly any type of lter with a great variety of characteristics can be realized. Depending on order and coefcients, the obtained lter can be stable or unstable. Stability is also closely related to other implementation details like nite precision arithmetic, coefcient quantization and signal overload caused by limited number ranges, inappropriate signal scaling and missing wrap around provisions. Careful investigation of stability has to be made in the z-domain or in the j domain which is beyond the scope of this book. Stability analysis methods as well as various design strategies for IIR lters can be taken from a wide variety of literature. Any digital lter design book will cover those topics exhaustively.

311

An IIR lter according to Equation A.17 is uncomputable if b0 = 0 because it would require the actual output sample to compute it. This can indeed be a problem, as Walstijn reported [WC03], when tone holes of woodwinds are to be modelled. Their dimensions are so small that zerodelay loops would result if the sampling rate is not increased to painful levels. In order to circumvent the above mentioned issues and difculties, the discovery of wave digital lters was so widely appreciated.

There is a type of recursive lters which are always computable, referred to as Wave Digital Filters. Their theory has been elaborated by Fettweis in [Fet86]. They are known to have excellent numerical and stability properties in all cases of operation and are therefore worthy of consideration when various transmission functions have to be implemented. Their stability properties derive from the fact that they are closely related to classical passive lter networks, preferably loss-less lters, inserted between resistive terminations. The analogy between a wave digital lter and its corresponding reference lter is based not on the use of voltages and currents as signal parameters, but of so-called wave quantities known from classical circuits, which has given this important class of digital lters its name. The Wave Digital Biquadratic section proposed by Kausel in [Kau99] is an example for an especially well applicable lter block originally presented by Fettweis. It can be arbitrarily scaled without loosing its good properties which makes it a favorable candidate for coefcient optimization. Its topology is shown in Figure A.22. The transmission function in the z-domain is given by H(z) = A2 + A1 z + A0 z 2 B2 + B1 z + z 2 (A.19)

312

y1 z lef

-1

+

z

lef1 = a1 y b1 x

-1

rig

+

lef1

+

a1 a2

rig1

+

-b1 -b2

y ( z ) = x( z )

1 + b1 b2 b1 z b2 z + z 2 1 + a1 a2 a1 z a2 z + z 2

(A.20)

There are no restrictions on the coefcients Ai and Bi , which means they can be optimized freely. As long as the transformation rules are applied correctly in order to derive the actual coefcients ai and bi this second order section enjoys all the stability properties of wave digital lters, including forced-response stability providing proper scaling and number representation.

313

The basic operations performed in a digital lter are addition and subtraction as well as scaling of samples using constant coefcients. Additionally registers are required to store past signal values or past output values. It is clear that scaling of signal values that involves multiplication operations is the most expensive computation in a digital lter. Coefcient values of one would therefore be highly desirable. An example for a non-recursive lter with unit coefcients has already been given (refer to Figure A.11). Because of the rectangular pulse response its transmission function is the discrete Fourier transform of a rectangular window. Although this lter has a good stop-band attenuation at already low lter orders, there is a signicant pass-band droop which can not be avoided (shown in Figure A.12). The simple circuit of Figure A.11 which repeats each input sample four times to increase the sampling rate from 8 kHz to 32 kHz can be used as a simple interpolator. Unfortunately, very often it is desired to combine an extremely at pass-band response with a sharp cutoff region and good stop-band attenuation. This rules out the exclusive use of Sinc lters. On the other hand Sinc lters are much cheaper and can operate at much higher sampling rates. So clearly the optimum combination in a real-time signal or sound processing system is a multi-rate approach with a Sinc (comb) lter as primary decimation stage and another lter running close to the target rate, providing the desired lter characteristics and compensating for the droop introduced by the decimator. In sound synthesis tasks the opposite strategy can be followed. The biggest and most accurate part of the simulation and synthesis task can often be made at a sampling rate which is much lower than the nal sampling rate of the sound output device. A comb interpolator can be used to interface between the working data stream and the output data stream.

314

B. Interpolation Tables

B.1. Forward Coefcients

Deriv: 1, Order: 1 Numerator: (1) Denominator: (x)

fk fk+1 1 1

fk fk+1 3 4 1 2

fk+2 1 1

Deriv: 1, Order: 3

1, x, x2 Numerator: Denominator: 6 x, x2 , 2 x3

315

B. Interpolation Tables

fk fk+1 11 18 2 5 1 3 fk+2 9 4 3 fk+3 2 1 1

Deriv: 1, Order: 4

1, x, x2 , x3 Numerator: Denominator: 12 x, 12 x2 , 4 x3 , 6 x4

fk+3 16 56 14 4

fk+4 3 11 3 1

Deriv: 1, Order: 5

1, x, x2 , x3 , x4 Numerator: Denominator: 60 x, 12 x2 , 8 x3 , 6 x4 , 24 x5

fk fk+1 fk+2 fk+3 fk+4 137 300 300 200 75 45 154 214 156 61 17 71 118 98 41 3 14 26 24 11 1 5 10 10 5

fk+5 12 10 7 2 1

316

Deriv: 1, Order: 6

1, x, x2 , x3 , x4 , x5 Numerator: Denominator: 60 x, 180 x2 , 16 x3 , 36 x4 , 48 x5 , 120 x6

fk fk+1 fk+2 fk+3 fk+4 fk+5 fk+6 147 360 450 400 225 72 10 812 3132 5265 5080 2970 972 137 49 232 461 496 307 104 15 35 186 411 484 321 114 17 7 40 95 120 85 32 5 1 6 15 20 15 6 1

fk 1

fk+1 2

fk+2 1

fk fk+1 2 5 1 3

fk+2 4 3

fk+3 1 1

317

B. Interpolation Tables

Deriv: 2, Order: 4 Numerator: 1, x, x2 Denominator: 12 x2 , 2 x3 , 2 x4

fk+4 11 3 1

fk+3 56 14 4

Deriv: 2, Order: 5

1, x, x2 , x3 Numerator: Denominator: 12 x2 , 4 x3 , 2 x4 , 6 x5

fk+5 10 7 2 1

Deriv: 2, Order: 6

1, x, x2 , x3 , x4 Numerator: Denominator: 180 x2 , 8 x3 , 12 x4 , 12 x5 , 24 x6 fk+1 fk+2 fk+3 fk+4 fk+5 fk+6 fk 812 3132 5265 5080 2970 972 137 49 232 461 496 307 104 15 35 186 411 484 321 114 17 7 40 95 120 85 32 5 1 6 15 20 15 6 1

318

Deriv: 3, Order: 3 Numerator: (1) Denominator: x3

fk fk+1 1 3

fk+2 3

fk+3 1

fk fk+1 5 18 1 4

fk+2 24 6

fk+3 14 4

fk+4 3 1

Deriv: 3, Order: 5

1, x, x2 Numerator: Denominator: 4 x3 , x4 , 2 x5

fk fk+1 17 71 3 14 1 5

fk+4 41 11 5

fk+5 7 2 1

319

B. Interpolation Tables

Deriv: 3, Order: 6

1, x, x2 , x3 Numerator: Denominator: 8 x3 , 6 x4 , 4 x5 , 6 x6

fk fk+1 fk+2 fk+3 fk+4 fk+5 fk+6 49 232 461 496 307 104 15 35 186 411 484 321 114 17 7 40 95 120 85 32 5 1 6 15 20 15 6 1

fk 1

fk+1 4

fk+2 6

fk+3 4

fk+4 1

fk fk+1 3 14 1 5

fk+2 26 10

fk+3 24 10

fk+4 11 5

fk+5 2 1

320

Deriv: 4, Order: 6 Numerator: 1, x, x2 Denominator: 6 x4 , 2 x5 , 2 x6

fk 35 7 1

321

B. Interpolation Tables

Deriv: 1, Order: 1 Numerator: (1) Denominator: (x)

fk1 1

fk 1

fk2 1 1

fk1 4 2

fk 3 1

Deriv: 1, Order: 3

1, x, x2 Numerator: Denominator: 6 x, x2 , 2 x3

fk3 2 1 1

fk2 9 4 3

fk1 18 5 3

fk 11 2 1

322

Deriv: 1, Order: 4 Numerator: 1, x, x2 , x3 Denominator: 12 x, 12 x2 , 4 x3 , 6 x4

fk4 3 11 3 1

fk3 16 56 14 4

fk2 36 114 24 6

fk1 fk 48 25 104 35 18 5 4 1

fk5 12 10 7 2 1

fk4 75 61 41 11 5

Deriv: 1, Order: 6

1, x, x2 , x3 , x4 , x5 Numerator: Denominator: 60 x, 180 x2 , 16 x3 , 36 x4 , 48 x5 , 120 x6

323

B. Interpolation Tables

fk6 10 137 15 17 5 1 fk5 fk4 fk3 fk2 fk1 fk 72 225 400 450 360 147 972 2970 5080 5265 3132 812 104 307 496 461 232 49 114 321 484 411 186 35 32 85 120 95 40 7 6 15 20 15 6 1

fk2 1

fk1 2

fk 1

fk3 1 1

fk2 4 3

fk1 5 3

fk 2 1

Deriv: 2, Order: 4

1, x, x2 Numerator: Denominator: 12 x2 , 2 x3 , 2 x4

324

fk4 11 3 1 fk3 56 14 4 fk2 114 24 6 fk1 fk 104 35 18 5 4 1

Deriv: 2, Order: 5

1, x, x2 , x3 Numerator: Denominator: 12 x2 , 4 x3 , 2 x4 , 6 x5

fk5 10 7 2 1

fk4 61 41 11 5

fk1 fk 154 45 71 17 14 3 5 1

Deriv: 2, Order: 6

1, x, x2 , x3 , x4 Numerator: Denominator: 180 x2 , 8 x3 , 12 x4 , 12 x5 , 24 x6

fk6 137 15 17 5 1

fk5 fk4 fk3 fk2 fk1 fk 972 2970 5080 5265 3132 812 104 307 496 461 232 49 114 321 484 411 186 35 32 85 120 95 40 7 6 15 20 15 6 1

325

B. Interpolation Tables

Deriv: 3, Order: 3 Numerator: (1) Denominator: x3

fk3 1

fk2 3

fk1 3

fk 1

fk4 3 1

fk3 14 4

fk2 24 6

fk1 18 4

fk 5 1

Deriv: 3, Order: 5

1, x, x2 Numerator: Denominator: 4 x3 , x4 , 2 x5

fk5 7 2 1

fk4 41 11 5

fk3 98 24 10

fk2 118 26 10

fk1 71 14 5

fk 17 3 1

326

Deriv: 3, Order: 6

1, x, x2 , x3 Numerator: Denominator: 8 x3 , 6 x4 , 4 x5 , 6 x6

fk6 15 17 5 1

fk4 1

fk3 4

fk2 6

fk1 4

fk 1

fk5 2 1

fk4 11 5

fk3 24 10

fk2 26 10

fk1 14 5

fk 3 1

327

B. Interpolation Tables

Deriv: 4, Order: 6 Numerator: 1, x, x2 Denominator: 6 x4 , 2 x5 , 2 x6

fk6 17 5 1

fk1 fk 186 35 40 7 6 1

328

fk1 1 1

fk fk+1 0 1 2 1

Deriv: 1, Order: 4

1, x, x2 , x3 Numerator: Denominator: 12 x, 12 x2 , 4 x3 , 6 x4

fk2 1 1 1 1

fk1 8 16 2 4

fk fk+1 0 8 30 16 0 2 6 4

fk+2 1 1 1 1

329

B. Interpolation Tables

fk3 1 2 1 1 1 1 fk2 9 27 8 12 4 6 fk1 45 270 13 39 5 15 fk fk+1 0 45 490 270 0 13 56 39 0 5 20 15 fk+2 9 27 8 12 4 6 fk+3 1 2 1 1 1 1

Deriv: 1, Order: 8

1, x, x2 , x3 , x4 , x5 , x6 , x7 Numerator: Denominator: 840 x, 5040 x2, 480 x3 , 1440 x4 , . . . . . . 144 x5 , 480 x6 , 1440 x7 , 5040 x8

fk4 3 9 7 7 1 1 1 1

fk3 32 128 72 96 9 12 6 8

fk+3 32 128 72 96 9 12 6 8

fk+4 3 9 7 7 1 1 1 1

Deriv: 1, Order: 10

1, x, x2 , x3 , x4 , x5 , x6 , x7 , x8 , x9 Num.: Denom.: 2520 x, 25200 x2, 60480 x3 , 90720 x4, 6912 x5, . . . . . . 28800 x6 , 17280 x7, 15120 x8, 80640 x9, 362880 x10

330

fk5 2 8 205 82 13 13 5 1 1 1 fk4 fk3 fk2 fk1 fk 25 150 600 2100 0 125 1000 6000 42000 73766 2522 14607 52428 70098 0 1261 9738 52428 140196 192654 152 783 1872 1938 0 190 1305 4680 9690 12276 52 207 408 378 0 13 69 204 378 462 8 27 48 42 0 10 45 120 210 252

fk fk+1 fk+2 fk+3 fk+4 fk+5 0 2100 600 150 25 2 73766 42000 6000 1000 125 8 0 70098 52428 14607 2522 205 192654 140196 52428 9738 1261 82 0 1938 1872 783 152 13 12276 9690 4680 1305 190 13 0 378 408 207 52 5 462 378 204 69 13 1 0 42 48 27 8 1 252 210 120 45 10 1

Deriv: 1, Order: 12

1, x, x2 , x3 , x4 , x5 , x6 , x7 , x8 , x9 , x10 , x1 1 Num.: Denom.: 27720 x, 831600 x2, 604800 x3, 2721600 x4, . . . . . . 290304 x5, 7257600 x6, 345600 x7, 1814400 x8, . . . . . . 161280 x9, 4354560 x10, 7257600 x11, 39916800 x12

331

B. Interpolation Tables

fk6 5 50 479 479 139 695 31 31 1 5 1 1 fk5 72 864 6840 8208 1936 11616 410 492 12 72 10 12 fk4 495 7425 46296 69444 12500 93750 2404 3606 60 450 44 66 fk3 2200 44000 198760 397520 48176 481760 7550 15100 164 1640 110 220 fk+2 7425 222750 603315 1809945 101559 1523385 13275 39825 261 3915 165 495 fk2 7425 222750 603315 1809945 101559 1523385 13275 39825 261 3915 165 495 fk+3 2200 44000 198760 397520 48176 481760 7550 15100 164 1640 110 220 fk1 23760 1425600 764208 4585248 99744 2992320 11652 69912 216 6480 132 792 fk+4 495 7425 46296 69444 12500 93750 2404 3606 60 450 44 66 fk 0 2480478 0 6222216 0 3735732 0 84084 0 7644 0 924 fk+6 5 50 479 479 139 695 31 31 1 5 1 1

fk+1 23760 1425600 764208 4585248 99744 2992320 11652 69912 216 6480 132 792

fk1 1

fk fk+1 2 1

332

Deriv: 2, Order: 4 Numerator: 1, x, x2 Denominator: 12 x2 , 2 x3 , 2 x4

fk2 1 1 1

fk1 16 2 4

fk fk+1 30 16 0 2 6 4

fk+2 1 1 1

fk3 2 1 1 1 1

fk2 27 8 12 4 6

fk1 270 13 39 5 15

fk+2 27 8 12 4 6

fk+3 2 1 1 1 1

Deriv: 2, Order: 8

1, x, x2 , x3 , x4 , x5 , x6 Num.: Denom.: 5040 x2 , 240 x3 , 480 x4 , 36 x5 , 96 x6 , 240 x7 , 720 x8

333

B. Interpolation Tables

fk4 9 7 7 1 1 1 1 fk3 128 72 96 9 12 6 8 fk2 1008 338 676 26 52 14 28 fk1 8064 488 1952 29 116 14 56 fk 14350 0 2730 0 150 0 70 fk+1 8064 488 1952 29 116 14 56 fk+2 1008 338 676 26 52 14 28 fk+3 128 72 96 9 12 6 8 fk+4 9 7 7 1 1 1 1

Deriv: 2, Order: 10

1, x, x2 , x3 , x4 , x5 , x6 , x7 , x8 Numerator: Denominator: 25200 x2, 30240 x3, 30240 x4 , 1728 x5, . . . . . . 5760 x6 , 2880 x7, 2160 x8 , 10080 x9, 40320 x10 fk4 fk3 fk2 fk1 fk 125 1000 6000 42000 73766 2522 14607 52428 70098 0 1261 9738 52428 140196 192654 152 783 1872 1938 0 190 1305 4680 9690 12276 52 207 408 378 0 13 69 204 378 462 8 27 48 42 0 10 45 120 210 252

fk5 8 205 82 13 13 5 1 1 1

fk+1 fk+2 fk+3 fk+4 fk+5 fk 73766 42000 6000 1000 125 8 0 70098 52428 14607 2522 205 192654 140196 52428 9738 1261 82 0 1938 1872 783 152 13 12276 9690 4680 1305 190 13 0 378 408 207 52 5 462 378 204 69 13 1 0 42 48 27 8 1 252 210 120 45 10 1

334

Deriv: 3, Order: 4 Numerator: (1, x) Denominator: 2 x3 , x4

fk+2 1 1

fk2 1 1

fk1 2 4

fk 0 6

fk+1 2 4

Deriv: 3, Order: 6

1, x, x2 , x3 Numerator: Denominator: 8 x3 , 6 x4 , 4 x5 , 6 x6

fk3 1 1 1 1

fk2 8 12 4 6

fk1 13 39 5 15

fk fk+1 0 13 56 39 0 5 20 15

fk+2 8 12 4 6

fk+3 1 1 1 1

Deriv: 3, Order: 8

1, x, x2 , x3 , x4 , x5 Numerator: Denominator: 240 x3 , 240 x4 , 12 x5 , 24 x6 , 48 x7 , 120 x8

fk4 7 7 1 1 1 1

fk3 72 96 9 12 6 8

fk2 fk1 fk fk+1 fk+2 338 488 0 488 338 676 1952 2730 1952 676 26 29 0 29 26 52 116 150 116 52 14 14 0 14 14 28 56 70 56 28

fk+3 72 96 9 12 6 8

fk+4 7 7 1 1 1 1

335

B. Interpolation Tables

Deriv: 3, Order: 10 Numerator: 1, x, x2 , x3 , x4 , x5 , x6 , x7 Denominator: 30240 x3, 15120 x4, 576 x5 , 1440 x6, . . .

. . . 576 x7 , 360 x8 , 1440 x9 , 5040 x10 fk4 fk3 fk2 fk1 fk 2522 14607 52428 70098 0 1261 9738 52428 140196 192654 152 783 1872 1938 0 190 1305 4680 9690 12276 52 207 408 378 0 13 69 204 378 462 8 27 48 42 0 10 45 120 210 252

fk5 205 82 13 13 5 1 1 1

fk fk+1 fk+2 fk+3 fk+4 fk+5 0 70098 52428 14607 2522 205 192654 140196 52428 9738 1261 82 0 1938 1872 783 152 13 12276 9690 4680 1305 190 13 0 378 408 207 52 5 462 378 204 69 13 1 0 42 48 27 8 1 252 210 120 45 10 1

fk2 1

fk1 4

fk 6

fk+1 4

fk+2 1

336

Deriv: 4, Order: 6 Numerator: 1, x, x2 Denominator: 6 x4 , 2 x5 , 2 x6

fk3 1 1 1

fk2 12 4 6

fk1 39 5 15

fk fk+1 56 39 0 5 20 15

fk+2 12 4 6

fk+3 1 1 1

Deriv: 4, Order: 8

1, x, x2 , x3 , x4 Numerator: Denominator: 240 x4 , 6 x5 , 8 x6 , 12 x7 , 24 x8

fk4 7 1 1 1 1

fk3 96 9 12 6 8

fk2 676 26 52 14 28

fk1 fk fk+1 fk+2 1952 2730 1952 676 29 0 29 26 116 150 116 52 14 0 14 14 56 70 56 28

fk+3 96 9 12 6 8

fk+4 7 1 1 1 1

337

B. Interpolation Tables

20

40

60

80

100

1st order

-2

-6

-4

-8

20 -2

40

60

80

100

2nd order

-4

-6

4th order

-8

6th order

-10

8th order

-12

12th order

10th order

338

1

20 -1

40

60

80

100

nd

order

-2

-3

-4

-5

6th order

-6

20

40

60

80

100

-2

2nd order

-4

-6

4th order

-8

6th order

-10

-12

339

B. Interpolation Tables

1

20

40

60

80

100

-1

3rd order

-2

4th order

-3

5th order

-4

6th order

20

40

60

80

100

-2

4th order

-4

6th order

-6

-8

8th order

-10

10th order

340

20

40

60

80

100

4th order

-1

5th order

-2

6th order

-3

20

40

60

80

100

-2

4th order

-4

-6

6th order

-8

8th order

341

B. Interpolation Tables

342

Index

A-contractive, 111 A-stable, 111 Abramowitz, 119 accuracy, 84 acoustic eld, 36 Adachi, 184, 185, 196, 247 Adams method order=2, 111 Adams-Bashforth method order=2, 110, 111 order=3, 110 predictor, 114 Adams-Moulton corrector, 114 method order=2, 110, 111 order=3, 111 order=4, 114 adaptor, 167 ADI method, 106 adiabatic state transitions, 34 advection, 80

aero-acoustics, 24, 62, 66, 80, 97 aero-dynamics, 66 air column, 34, 36, 136 jet, 36 resonance, 187 aliasing, 84 alternating direction implicit, 106 Amir, 145, 147, 172, 176 amplication matrix, 91, 92 analytical solutions, 32 approximation, 55, 56, 58 curve, 67 error, 75, 87, 88, 90, 112 function, 67 order, 82 polynomial, 66 polynomial, general, 68 quality, 70 articial lips, 185, 194, 207, 210 mouth, 185 atmospheric conditions, 33

343

Index

mean values, 33 quiescent pressure, 32 atoms, 40 attenuation, 65, 96 Avanzini, 148 axial-symmetry, 34, 135 Ayers, 214 back substitution, 98100, 104 Backus, 16 backward difference, 56, 68, 70, 113 operator, 55 order=1, 127 order=2, 110, 111, 151 travelling pulse, 175, 176 waves, 50, 137, 178 traversal, 101 bafe, 141 Bamberger, 248 Banachiewicz, 99 banded matrices, 101, 103 Barjau, 148 basic lip models, 201 bassoon, 36 Benade, 16 Bensa, 148 Berkeley, 186 Bernoulli, 15 pressure, 185, 188, 190, 199 Bessel function, 142, 143 BGK-Model, 42 Bhatnagar, 42 BIAS, 225 Bilbao, 158, 160 bilinear transform, 152, 161 binomial coefcients, 55 block iterative methods, 106 Boltzmann distribution function, 40 equation, 28, 31, 39, 41, 123, 134 bore information, 172 reconstruction, 136, 172 Borin, 148, 160 boundary condition lossy, 143 conditions, 32, 33, 64, 78, 80, 131 element method, 51 nodes, 157 reections, 80 scheme, 82, 109 value problems, 54, 121 Bowsher, 188 brass wind instruments, 34, 184 Bruneau, 143 buoyancy, 35 buzzing lip, 36, 158 CAA, 133 calibration, 226 Campbell, 16, 185, 207 capillary, 228

344

Index

Causs , 141 e central average, 56, 60, 63, 69 difference, 69, 85, 88 operator, 56, 60, 62 operator, order=2, 58 order=2, 108 order=4, 81 chain matrix, 141 change rate, 21, 27 Chattot, 37 chemical reactions, 27 chirp, 150, 226 Cholesky, 99 circuit simulator, 91, 182 clarinet, 247 collision accidents, 39 model, 42 simulated, 40 term, 41, 42 comb lter, 84 common-force connection, 167 common-velocity connection, 167 compatibility law, 37 complex plane, 96 complex variable theory, 153 compressibility, 36 computational advanced concepts, 186 aero-acoustics, 53, 133, 249 algorithm, 157 analysis, 247 complexity, 54 efciency, 61, 130, 182 uid-dynamics, 72, 80, 183 methodology, 44 methods, 181, 224 physics, 43, 186 resources, 252 stencils, 47, 122 computer optimization, 229 concert hall, 149 conservation differential form, 19 energy, 19, 26 equations, 22 integral form, 23, 26, 28, 47, 80 laws, 33 mass, 19, 36 momentum, 19, 23, 37 conservative locally, 160 consistency, 84, 90, 153 continuity position, slope, curvature, 76 convection, 24 phenomena, 33 speed, 29, 32 velocity, 42 convective deformation, 22, 27 uid transport, 33 translation, 19 transport, 27

345

Index

convergence, 88, 90, 104, 153 rate, 105 speed, 105 convolution, 148, 149 cross-sectional area, 137 Crout, 99, 100 crystalloidal lattice, 40 curl, 38, 39 curvature, 75 cut-off frequency, 141, 147 cylindrical coordinates, 34 duct, 137, 142 slices, 139 D3Q15, 124 D3Q19, 124 damping, 65, 79 deformation, 21, 25 rate, 25 delay-free loops, 170 denominator coefcient tables, 70 density constant, 37 distribution, 124 variations, 33 DePoli, 148, 160 Dequand, 251 derivative, 56, 67 higher order, 55 in time, 56, 60, 62 spatial, 24, 60, 63 difference, 55 central, 56 forward, 56 higher order, 5557 quotient, 57, 66 differencing, 54, 56, 62 dynamical switching, 80 explicit, 57 forward, 67 unsymmetrical, 82 upwind, 80 differential cube, 21 equation hyperbolic, 80 non-linear, order=1, 120 ordinary, 109 ordinary, order=1, 109 ordinary, order=2, 109 form, 22 passivity, 160 quotient, 5355, 57, 66, 87 differentiate, 67 differentiation product, 20 diffusion, 22 constant, 28 digital wave guide theory, 160 Dirac pulse, 150, 172 discontinuity, 137, 138, 144 discrete Fourier transform, 136, 148

346

Index

lattice nodes, 41 velocities, 41 discrete-time signals, 153 discretization, 5456, 80, 86 Boltzmann-Equation, 31 spatial, 109 dispersion, 65, 96 dissipation volume changes, 25 distorted wave number, 88 distribution function, 30, 31, 39, 41 divergence, 20, 21, 27, 3739 momentum equation, 34 double reed, 36 downward differences, 83 DuFort, 86 DuFort-Frankel scheme, 86 dusty environment, 25 Eckhoff, 72 efciency, 136 eigenfunction, 142 eigenproblem Laplacian, 142 eigenvalue, 91, 92, 104, 105 Elliott, 16, 188 energy, 40 density internal, 27 kinetic, 27 equation, 27 exchange, 26 external mechanical, 33 internal, 26, 31 leakage, 164 sound, 32 enthalpy density, 27 specic, 28 equation Boltzmann, 123 continuity, 22 diffusion, 37, 93 hyperbolic, wave, 60 implicit, 63, 74, 75, 77, 79 Laplace, 106 linear convection, 59, 60, 62, 65, 95 momentum, 23, 38, 39 of movement, 31, 161 of state, 28 ordinary differential, 92, 166 parabolic, 37 Poisson, 106, 107 thermal diffusion, 86 transport, 21, 41 wave, 32, 34 equidistant breakpoints, 73 equilibrium, 25, 35 local, 25 prole, 41 state, 31 equivalent circuit, 182 error amplication, 91

347

Index

estimate, 90 estimation, 56 order, 86 term, 85 Euler, 15 corrector, 118 equation, 23, 35 explicit method, 110, 111 implicit method, 110, 111 predictor, 118 Euler-Cauchy improved method, 113 predictor, 116 scheme, 92 explicit method, 61, 110 most accurate scheme, 111 extrapolation, 54, 83, 113 formula, 83 order=1, 83 scheme, 83 Fabre, 183 Facchinetti, 247 Fast Fourier Transform, 148 feedback loop, 79, 136 Fettweis, 160 lter kernel, 148, 149 nite difference method, 44, 55 difference scheme, 155 difference theory, 158 element method, 48, 80 element, clarinet, 247 impulse response lters, 305 volume, 20 volume method, 46, 80 volume methods, 80 FIR lter, 305 rst derivative spline approximation, 74 rst order scheme, 96 Fletcher, 16, 188 ow discontinuities, 23 steady, 22 ue pipes, 34 uid compressible, 19 deformation rate, 25 density, 40 dynamical applications, 120 dynamics, 123 element, 27 friction-less, 23 homogeneous, 59 incompressible, 38 incompressible, viscous, 36 mesoscopic view, 28, 29 microscopic view, 29 molecules, 29 Newtonian, 25 particle displacement, 33 stagnant, 33 stress tensor, 24 uniform, 33 velocity, 30, 40

348

Index

velocity eld, 37 viscosity, 35 viscous, 19 ute, 34, 36, 134, 136 simulation, 252 force, 23 convective, 24 densities, 23 density, 25 external, 24, 27, 33 gravitational, 24, 35 viscous, 24 force-node, 164 force-waves, 162 forward difference, 68, 97 operator, 55, 58, 60, 62 order=2, 72 quotient, order=1, 92 extrapolation, 83 travelling pulse, 175 waves, 50, 137, 178 traversal, 101 Fourier coefcients, 80 components, 87, 93 spatial, 93 decomposition, 80 spatial, 93 error analysis, 87 series, 85 fourth order accuracy, 75 Frankel, 86 free air, 32, 35 friction element, 164 Frisch, 40 Galerkin, 49 GAlib, 230 gases, 35 Gauss distribution, 149 divergence theorem, 21 integral theorem, 22 Gauss-Seidel, 105 method, 108 relaxation, 105 Gaussian elimination, 98, 100, 105 Gear, 119 Gilbert, 16, 207 global continuity, 75 passivity, 160 glottis, 249 Goldberg, 230 gradient density, 31 pressure, 24, 37 speed, 21, 35 gravity, 27, 29, 33 Greated, 16 grid discretization, 54 equidistant, 54 hierarchical, 54

349

Index

increasing resolution, 66 regular, 54 spatial, multiple, 54 Gross, 42 group delay, 136 guided ows, 35 H lie, 145 e Hankel functions, 145 Hardy, 40 harmonic oscillator, 92 Hasslacher, 40 heat conduction, 27, 28, 33, 35 conduction coefcient, 27 conversion, 27 ux, 28 release, 27 release rate, 27 specic, 28 Heisenberg, 39 Helmholtz, 15 helmholtz resonator, 251 Hermit, 73 higher order moments, 39, 40 scheme, 108 Hirschberg, 183, 211 Hofmans, 251 Huygens, 158 principle, 155 hybrid approaches, 171 ideal gas constant, 32 impedance matrix, 144 implicit difference schemes, 97, 101 interpolation schemes, 101 methods, 110 non-linear schemes, 119 third-order scheme, 111 impulse invariant transformation, 151 innite bafe, 145 ange, 144 initial value problems, 54 instantaneous power, 164 Institut f. Wiener Klangstil, 184 integer particle counts, 41 integral form, 19, 45, 46, 51 laws, 23 integrating mass density, 29 particle velocities, 30 velocity space, 30 interaction ow and sound, 66 interpolation, 54, 66 function, 67, 69 harmonic, 80 Hermitian, 73 Lagrange formula, 72 order, 69, 81 stencil, 84

350

Index

intonation, 136 inverse problem, 172 inward striking, 187 irrotational part, 38 isobar, 27 isotropy, 123 iteration, 54 alternating direction, 107 benchmark, 108 iterative method, 38, 103 Jacobi, 104, 105 method, 105, 108 Jacobi radius, 105 jet deection, 247 K hnelt, 123, 184 u Karjalainen, 170 Keefe, 16, 177 Kemp, 137, 144, 145, 151 Kergomard, 147 kinetic energy, 165 Knudsen number, 128 Krook, 42 labium, 36 Lagrange, 15, 72, 73 laminar ow region, 35 Laplace transformation, 165 Laplacian operator, 39, 141 lattice Boltzmann, 31, 41, 123, 158, 248 gas cellular automata, 40 node, 40 velocities, 123 layer peeling algorithm, 173 leapfrog predictor, 118 scheme, 85, 87, 109111 Lees scheme, 111 Leibnitz, 53, 66 Levine, 141 LGCA, 40, 41 Lighthill, 133 linear acoustics, 33 convection, 109 equation system, 54, 56, 63, 97 linearization, 119 density, 33 pressure, 33 lip model, 182 resonance, 187 Lomax, 72 losses numerical, 65 thermo-viscous, 32 viscous, 35 low-pass characteristic, 79 LU-decomposition, 64, 99, 100, 104 lumped model, 207

351

Index

Mach number, 33, 128 higher, 35 macroscopic, 19, 28 continuum, 28, 29 properties, 40 Magalotti, 148 Mapes-Riordan, 192 mapping frequency, 153 mass, 161 density, 2830, 32, 42 ow, 23, 31 Mathematica, 64 matrix 5-diagonal, 103 7-diagonal, 103 9-diagonal, 103 banded, 78, 101 coefcient, 62, 70 coefcient tables, 70 coefcients, 97 column swapping, 98 determinant, 100 diagonal, 62, 100 diagonally dominated, 103, 105 inverse, 100 inversion, 99, 107 main diagonal, 97, 99, 104 notation, 63 Pad formula, 77 e recursive approach, 101 row swapping, 98 singular, 103 sparse, 62 square, 78 stiff, 103 transformation, 107 triangle, 98 triangular, 97, 99, 105 tridiagonal, 78, 79, 101, 107 upper triangular, 99 maximum relative error, 87 Maxwell local distribution, 42 Maxwell-Boltzmann distribution, 31, 41 mechanical work, 27 medium incompressible, 37 transmission, 32 mesh computational, 46, 80 irregular, 54 mesoscopic calculation volumes, 29 spatial scale, 29 method implicit, 61 microscopic, 40 view, 39 Milne, 112 corrector, 118 fourth order scheme, 111 MIT, 231 MLS-signal, 150, 226

352

Index

modal conversion, 147 decomposition, 34, 80, 135, 141 wave number, 142 complex, 143 mode conversion, 137 mode conversion, 144 molecular level, 29 molecules, 29, 40 momentum change, 24 external mechanical, 33 Morse, 16 mouthpiece, 36, 141 multi-mass system, 205 multi-modal impedance matrix, 143 propagation, 147 propagation, lossy, 143 multi-step methods, 109 multiplicator coefcient tables, 70 musical acoustics, 34, 35, 181 musical instruments, 32, 36, 147, 149 Muto, 182, 214 Nabla operator, 22 Navier-Stokes equations, 19, 3234, 123 rst equation, 22, 36 second equation, 23, 24, 37 third equation, 26, 28 Newton, 53, 66, 68, 72 Newton-Gregory, 68 non-linear coupling, 36 systems, 119 non-linear problems, 90 non-linearities, 34 Noreland, 171 numerator coefcient tables, 70 numerical t, 80 precision, 103, 104 solutions, 32 Nyquist frequency, 152, 153 theorem, 84, 152 open boundary, 157 optimization, 18, 136 ordering operator, 107 oscillator, 136 outward striking, 187, 194 Pad e formula, 49, 74, 75, 77, 88 Pad formula, 97 e Pagneux, 145, 147 parabola order=2, 67 partial differential equations, 54, 55, 93

353

Index

parabolic, 57 particle, 40 collisions, 31 density, 30, 31 distribution, 30 ow, 31 velocities, 30 particle collisions, 42 Pelorson, 158 perfect gases, 28 perfect scheme, 96 periodic function, 87 periodical system, 79 perturbation variables, 33 phase-space, 30, 41, 123 physical model, 135, 149 piston radiator, 141 pivoting, 101, 103 complete, 99 partial, 99 plane wave model, 141 playable tones, 136 Poisson-Equation, 37 polynomial, 85 coefcients, 69 expression, 66 Pomeau, 40 port-resistance, 163 potential energy, 165 potential ow, 38 power-waves, 163 predictor-corrector methods, 112 stability, 114 pressure, 28, 39, 40 amplitude, 33 disturbances, 35 nite variations, 37 variations, 32 probability density, 39, 40 product matrix, vector, 24 projection equation, 143 projection method, 37, 38 propagation long distance, 25, 35 plane wave, 34 Pulliam, 72 pulsating sphere, 145 pulse response, 149 Radavich, 250 radiation impedance, 141, 145 multi-modal, 144 multi-mode, 137 random number generator, 40 random particle collisions, 41 rate-of-strain tensor, 25 Rayleigh, 15, 18, 141 rectangular cross-section, 34 rectangular ducts, 137 reeds clarinet, 35 lip, 34 oboe, 35 reection-free port, 170

354

Index

regime of oscillation, 185 relative error, 87 relaxation, 35, 124 constant, 42 time, 42 required resolution, 87 resonator, 35, 36, 136 response, 136 Reynolds transport theorem, 19, 21 Richards, 207 Richardson, 16 rigid walls, 157 robustness, 160 Rodet, 145 Rosenbrock, 232 Rossing, 16 roundoff error, 99 Runge-Kutta, 117 algorithm, 90 for second order ODE, 118 order=2, 117 order=4, 117 order=4, example, 118 sampled data stream, 149 sampled data systems, 151 sampling points per wave length, 87 Saneyoshi, 188 Sarti, 160 scalar magnitudes, 28 scattering, 173 approaches, 159 expressions, 167 formulations, 160 rule, 156 theory, 158 Scavone, 148, 193 second derivative spline approximation, 75 sensitivity, 76 Sharp, 151, 172 shear stress, 32 viscosity, 140 shift operator, 83 shock waves, 34 side branch resonators, 249 signal sinusoidal, 59 signal owcharts, 164 simplications, 32, 41 simulation particles, 40 acoustics, 31 aero-acoustical, 66 aerodynamical, 35 explicit formula, 58, 60 uid propagation, 31 framework, 18 method, 41 programs, modern, 54 startup, 109 unstable, 60 update time step, 109 singing voice, 34

355

Index

Smith Julius O., 148, 160, 182 solenoidal part, 38 solids, 25 solutions analytical, 35 numerical, 35 SOR algorithm, 108 method, 105, 106 sound energy,radiated, 36 far-eld propagation, 64 generation, 36 pressure, 136 pressure level, 32, 33 pressures, high, 34 production, 184 propagation, 32 propagation, simulating, 157 source, non-reecting, 172 speed, 33, 37 synthesis, 136 timbre, 136 velocity, 33 sounding pitch, 187 source density, 20 sources chemical, 27 external, 22, 27 mechanical, 27 thermal, 27 sparse matrix, 103, 107 spatial lattice, 123 spectral radius, 105 spectral transformation rules, 93 speed change, 23 spherical sector, 140 wave fronts, 147 wavelets, 155 SPICE, 91, 182, 186, 215 spline, 74, 75 cubic, 76 implementation, 77 implicit system, 79 spring, 161 spurious artifacts, 96 reections, 65 square lattice, 40 stability, 76, 88, 153, 160 analysis, 198 condition, 91, 115 predictor-corrector, 114 standing waves, 80 Stanford, 160, 182 state variables, 149 statistics, 29 steady state, 59, 79 stencil, 87, 108, 109 Sterling, 69 stiff systems, 43 stimulus, 36, 59 signal, 149

356

Index

stochastic simplications, 39 Stokes hypothesis, 25 strain, 21 stress, 21, 22 tensor, 26 viscous, 24, 27 Strong, 16, 136, 220 Strouhal number, 248 successive over-relaxation method, 105 superposition wavelets, 156 surface integral, 21, 26 surface waves, 183 sweep, 226 swinging-door model, 196 symmetry, 40, 123 system borders, 26 denition, 26 tabulated functions, 54 Taylor analysis, 82 expansion, 111, 120 series, 74 temperature, 28, 32, 42 global, 31 tensor stress, 2628 unit, 25 termination impedance, 139 thermal conductivity, 35 diffusion problem, 57 thermo-viscous loss, 139 thermodynamics rst theorem, 26 Thomas algorithm, 102, 106, 107 Thompson, 136, 220 three body problem, 29 time derivative, 109 marching, 109 stepping adaptive, 54 methods, 108 time derivative mass conservation, 33 time-domain methods, 148 time-shift operator, 151 TLM-method, 157 tone holes, 36 total change rate, 22 density, 30 energy, 27 energy density, 31 enthalpy, 27 mass, 29, 30 momentum, 22 work, 27 transformation rules, 154 transmission

357

Index

function, 149 line, 194 line matrix, 157 line mesh, 155 matrix, 139 transport coefcient, 22 transverse model, 188, 196 trapezoidal, 111 integration, 161 rule, 113, 155 scheme, 120 trial function, 80 triangulation, 99 tridiagonal matrices, 101 truncation error, 86, 96, 103 Tschebyscheff acceleration, 106 ultra-sound, 25 undersampling, 84 unit circle, 96, 104, 153 unit-delay, 164 update algorithm, 96 form, 91, 93 upwind principle, 97 upwind sample, 96 variational problems, 80 vector dyadic product, 23 eld, 28 eld,velocity, 27 inner product, 24 product, 23 space, 39 stream function, 38 velocity, 29 distribution, 38 divergence, 37 molecular, 30 no divergence, 36 node, 164 potential, 38 space, 30, 123 transitions, 41 ventilation ducts, 35 Vesely, 119 Vilain, 211 Virtual Wave Tank, 157 viscosity, 25, 33 air, 32 dynamic, 25 kinematic, 35 volume, 25, 35 viscous damping, 35 effects, 32 loss, 139, 157 stress tensor, 32 vocal folds, 34 tract, 158, 249 volume difference, 21 element, 21, 27

358

Index

nite, 21 integral, 21, 41 material, 19 velocity, 136, 137 viscosity, 25 Von Neumann, 62 analysis, 9396, 114 vortex density method, 38 shedding, 36 vorticity, 38 Wall, 231 wall vibrating, 64 Walstijn, 246 warping frequency, 154 wave digital lter, 160 digital lter theory, 160 digital lter, MD, 160 dispersion, 88 equation, 2D, 159 guide, conical, 139 guide, cylindrical, 139 guides, narrow, 32 number, 87 propagation, 32, 59 propagation speed, 65, 88 quantities, 160 steepening, 34, 136, 218 WDF-elements, 164 Webster, 15 weighting coefcients, 80 Welander, 42 white noise, 150, 226 wind instruments, 32, 34 oscillator, 34 windy environmental conditions, 35 Yoshikawa, 182, 214 z-domain, 151 z-transformation, 151 zero divergence condition, 38 Zhao, 249 Zorumski, 144

359