Lecture 12: Discrete Laplacian

Similar documents
1 Matrix representations of canonical matrices

APPENDIX A Some Linear Algebra

U.C. Berkeley CS294: Beyond Worst-Case Analysis Luca Trevisan September 5, 2017

Difference Equations

More metrics on cartesian products

3.1 Expectation of Functions of Several Random Variables. )' be a k-dimensional discrete or continuous random vector, with joint PMF p (, E X E X1 E X

Linear Approximation with Regularization and Moving Least Squares

ELASTIC WAVE PROPAGATION IN A CONTINUOUS MEDIUM

Lectures - Week 4 Matrix norms, Conditioning, Vector Spaces, Linear Independence, Spanning sets and Basis, Null space and Range of a Matrix

Lecture 10: May 6, 2013

n α j x j = 0 j=1 has a nontrivial solution. Here A is the n k matrix whose jth column is the vector for all t j=0

Salmon: Lectures on partial differential equations. Consider the general linear, second-order PDE in the form. ,x 2

C/CS/Phy191 Problem Set 3 Solutions Out: Oct 1, 2008., where ( 00. ), so the overall state of the system is ) ( ( ( ( 00 ± 11 ), Φ ± = 1

From Biot-Savart Law to Divergence of B (1)

Numerical Heat and Mass Transfer

Canonical transformations

Chapter 5. Solution of System of Linear Equations. Module No. 6. Solution of Inconsistent and Ill Conditioned Systems

Random Walks on Digraphs

CHAPTER 5 NUMERICAL EVALUATION OF DYNAMIC RESPONSE

CME 302: NUMERICAL LINEAR ALGEBRA FALL 2005/06 LECTURE 13

Transfer Functions. Convenient representation of a linear, dynamic model. A transfer function (TF) relates one input and one output: ( ) system

MMA and GCMMA two methods for nonlinear optimization

Workshop: Approximating energies and wave functions Quantum aspects of physical chemistry

Inner Product. Euclidean Space. Orthonormal Basis. Orthogonal

Lecture 10 Support Vector Machines II

8.1 Arc Length. What is the length of a curve? How can we approximate it? We could do it following the pattern we ve used before

Solutions to Problem Set 6

14 The Postulates of Quantum mechanics

Formal solvers of the RT equation

2 Finite difference basics

PES 1120 Spring 2014, Spendier Lecture 6/Page 1

Week 5: Neural Networks

Report on Image warping

PHYS 705: Classical Mechanics. Calculus of Variations II

Physics 5153 Classical Mechanics. D Alembert s Principle and The Lagrangian-1

Convergence of random processes

Errors for Linear Systems

PHYS 705: Classical Mechanics. Canonical Transformation II

Appendix B. The Finite Difference Scheme

Statistical Mechanics and Combinatorics : Lecture III

NUMERICAL DIFFERENTIATION

DUE: WEDS FEB 21ST 2018

Homework Notes Week 7

The Finite Element Method: A Short Introduction

Norms, Condition Numbers, Eigenvalues and Eigenvectors

Linear Feature Engineering 11

High resolution entropy stable scheme for shallow water equations

Inductance Calculation for Conductors of Arbitrary Shape

For now, let us focus on a specific model of neurons. These are simplified from reality but can achieve remarkable results.

New Method for Solving Poisson Equation. on Irregular Domains

Formulas for the Determinant

Lecture 21: Numerical methods for pricing American type derivatives

Grover s Algorithm + Quantum Zeno Effect + Vaidman

Singular Value Decomposition: Theory and Applications

THE SUMMATION NOTATION Ʃ

DECOUPLING THEORY HW2

princeton univ. F 17 cos 521: Advanced Algorithm Design Lecture 7: LP Duality Lecturer: Matt Weinberg

LECTURE 9 CANONICAL CORRELATION ANALYSIS

Additional Codes using Finite Difference Method. 1 HJB Equation for Consumption-Saving Problem Without Uncertainty

Assortment Optimization under MNL

Hidden Markov Models & The Multivariate Gaussian (10/26/04)

CHAPTER 4. Vector Spaces

j) = 1 (note sigma notation) ii. Continuous random variable (e.g. Normal distribution) 1. density function: f ( x) 0 and f ( x) dx = 1

MTH 263 Practice Test #1 Spring 1999

Georgia Tech PHYS 6124 Mathematical Methods of Physics I

CSCE 790S Background Results

2.3 Nilpotent endomorphisms

MATH 241B FUNCTIONAL ANALYSIS - NOTES EXAMPLES OF C ALGEBRAS

Stanford University CS359G: Graph Partitioning and Expanders Handout 4 Luca Trevisan January 13, 2011

1 GSW Iterative Techniques for y = Ax

Section 8.3 Polar Form of Complex Numbers

NON-CENTRAL 7-POINT FORMULA IN THE METHOD OF LINES FOR PARABOLIC AND BURGERS' EQUATIONS

Expected Value and Variance

LOW BIAS INTEGRATED PATH ESTIMATORS. James M. Calvin

EEE 241: Linear Systems

Lecture 3. Ax x i a i. i i

NP-Completeness : Proofs

Module 3 LOSSY IMAGE COMPRESSION SYSTEMS. Version 2 ECE IIT, Kharagpur

The Quadratic Trigonometric Bézier Curve with Single Shape Parameter

Moments of Inertia. and reminds us of the analogous equation for linear momentum p= mv, which is of the form. The kinetic energy of the body is.

Important Instructions to the Examiners:

Open Systems: Chemical Potential and Partial Molar Quantities Chemical Potential

6.3.4 Modified Euler s method of integration

Supplementary Notes for Chapter 9 Mixture Thermodynamics

STAT 309: MATHEMATICAL COMPUTATIONS I FALL 2018 LECTURE 16

Statistical Inference. 2.3 Summary Statistics Measures of Center and Spread. parameters ( population characteristics )

Robust Norm Equivalencies and Preconditioning

= z 20 z n. (k 20) + 4 z k = 4

Feature Selection: Part 1

Convexity preserving interpolation by splines of arbitrary degree

ANSWERS. Problem 1. and the moment generating function (mgf) by. defined for any real t. Use this to show that E( U) var( U)

Lecture Note 3. Eshelby s Inclusion II

PHYS 705: Classical Mechanics. Newtonian Mechanics

Appendix B. Criterion of Riemann-Stieltjes Integrability

CIVL 8/7117 Chapter 10 - Isoparametric Formulation 42/56

CSci 6974 and ECSE 6966 Math. Tech. for Vision, Graphics and Robotics Lecture 21, April 17, 2006 Estimating A Plane Homography

Lossy Compression. Compromise accuracy of reconstruction for increased compression.

Module 9. Lecture 6. Duality in Assignment Problems

Global Sensitivity. Tuesday 20 th February, 2018

Finding Dense Subgraphs in G(n, 1/2)

Transcription:

Lecture 12: Dscrete Laplacan Scrbe: Tanye Lu Our goal s to come up wth a dscrete verson of Laplacan operator for trangulated surfaces, so that we can use t n practce to solve related problems We are mostly nterested n the standard Posson problem: f g We wll frst ntroduce some basc facts and then talk about dscretzaton By the end, we wll be able to derve a dscretzed lnear system from posson problem and calculate the numercal solutons 1 Facts and Tools Laplacan Operator: Laplacan Operator s a lnear functonal on C (), e : f(x) C () f(x) C () If R n, t can be explctly expressed as a dervatve operator 1 : 2 That s to x 2 say, f we have a regon Ω R n and a functon f(x) C (Ω), then f(x) 2 f(x) x 2 If s a surface, whch s more often the case, one standard defnton of Laplacan operator s: f f, where and are dvergence and gradent operator respectvely However, ths defnton s not stragtforward for dscretzaton Alternatvely, we wll avod ths problem wth Galerkn s approach Before we go on, let s brefly talk about why Laplacan operator s mportant Laplacan operator s used n many mportant partal dfferental equatons, whch are the keys to many mathematcal and physcal models Here are some examples: The heat equaton u t u descrbes the dstrbuton of heat n a gven regon over tme The egenfunctons of (Recall that a matrx s a lnear operator defned n a vector space and has ts egenvectors n the space; smlarly, the Laplacan operator s a lnear operator defned n a functon space, and also has ts egenfunctons n the functon space) are the solutons to the equaton u λu If S s a surface and u C (S), the egenfunctons descrbe the vbraton modes of the surface S Galerkn s Approach Gven a functon defned on, e f : R, ts L 2 dual L f s defned on the functon space L 2 () L f : L 2 () R L f [g] fgda, g L 2 () L 2 () s all the square ntegrable functons on ore rgorously, L 2 () {f : R : f 2 < } The functon g s often called a test functon We wll only deal wth the followng set of test functons n our dscusson: {g C () : g 0} Often, we are nterested n a compact surface wthout boundary, so 1 Note that the sgn of Laplacan operator s nconsstent n dfferent lterature

Note that we can also recover the functon f from ts dual L f Fgure 1 gves an ntuton about ths process We can take g to be the square functon and as g approaches a sngle pont when t gets narrow, L f [g] approaches the value of f at that partcular pont Fgure 1: L f f We apply L 2 dual to Laplacan usng the test functons stated above (note that the test functons vansh on the boundary), and use ntegraton by parts: L 2 f [g] g fda boundary terms g fda g fda Notce that we have used Laplacan wthout actually evaluatng t Ths s an example of Galerkn s approach, whch s a class of methods for convertng a contnuous operator problem to a dscrete problem (eg for dscretzng and solvng dfferental equatons) In ths approach we need to decde on a functon space (where functon f come from) and a test functon space (where functon g come from; we can often apply boundary condtons by choosng test functons) We ll see ths method n practce n the next secton 2 Dscretzaton In ths secton, we wll frst pose a dfferent representaton of the Posson problem and then use the tools from the prevous secton to derve a dscretzaton Weak Soluton Consder Posson Equaton f g As stated before, t s hard to drectly derve a good dscretzaton for the equaton Therefore, we seek a dfferent representaton wth the concept of weak soluton The weak solutons for the Posson Equaton are functons f that satsfy the followng set of equatons 2 : φ fda φgda, test functons φ Recall that accordng to Galerkn s method, we need to choose the bass functon for f, g and test functon φ Frst Order Fnte Element Snce we are focusng on the Posson Equaton on a surface, f and g should be functons defned on the surface Therefore we need a set of bass functon on For trangulated surface, the most natural choce of bass functons are the pecewse lnear hat functons h, whch equal one at ther assocated vertex and zero at all other vertces, as shown n Fgure 2 2 The notaton here s dfferent from the last secton Specfcally, f and g are functons we want to dscretze and φ s test functon

Fgure 2: One hat functon per vertex Therefore, f we know the value of f (x) on each vertex, f (v ) a, we can approxmate t wth: X f (x) a h (x) Snce hp a R V Smlarly, we can have (x) are all fxed, we can store f wth only a sngle array ~ g(x) b h (x) Wth the dscretzaton of f and g, usng h as test functons, recall the L2 dual of f and the equatons for weak solutons of Posson Equaton, we can now pose the orgnal Posson problem as a set of V equatons: h gda, {1, 2,, V } h f da For the left hand sde: h f da h fp da h ( a h )da P a h h da Suppose matrx L {L } V V, L h h da Then the left hand sde of the set of P L a 1 P h1 f da L a 2 h2 f da L~a equatons s: P h f da V L V a For the rght hand sde: P h ( b h )da P b h h da h gda Smlarly, suppose matrx A {A } V V, A h h da The rght hand sde of the set of equatons s A~b Now we derve a lnear problem L~a A~b, we only need to calculate the matrces L and A n order to solve ~a Cotan Laplacan We now try to calculate the matrx L by examnng ts element L h h da Snce h are pecewse lnear functons on a trangular face, h s a constant vector on a face, and thus

h h yelds one scaler per face Therefore, to calculate the ntegral above, for each face we multply the scalar h h on that face by the area of the face and then sum the results Let s frst evaluate the gradent Now consder a lnear functon f defned on a trangle so that f(v 1 ) 1, f(v 2 ) f(v 3 ) 0, as shown n Fgure 3a For a lnear functon, we have f(x) f(x 0 ) + f x0 (x x 0 ) Let x 0 v 1, x v 2 and v 3 respectvely, and notce that f les wthn the trangle face, we get: f (v 1 v 3 ) 1 f (v 1 v 2 ) 1 f n 0 Ths yelds: f (v 2 v 3 ) 0 Therefore, f s perpendcular to the edge v 2 v 3, as shown n Fgure 3a (a) Evaluatng gradent (b) Case 1: Same vertex (c) Case 2: Dfferent vertces Fgure 3 Now we know the drecton of the gradent, we wll go on to evaluate ts magntude 1 f (v 1 v 3 ) f l 3 cos( π 2 θ 3) f l 3 sn θ 3 f 1 l 3 sn θ 3 1 h where h s the heght of the trangle correspondng to edge v 2 v 3 Recall that trangle area A 1 2 v 2v 3 h, thus f e 23 2A e 23 s the vector from v 2 to v 3 rotated by a quarter turn n the counter-clockwse drecton Now we have the gradent vectors on a face, we need to take dot products of them to get the scalar assocated wth each face There are two dfferent cases Let s stll look at a sngle trangle face and functons defned on t for now The frst case s when two functons are defned on the same vertex

T f, f dt A f 2 A b h 2 (h cot α+h cot β) 2h (cot α + cot β) 1 2 The second case s when two functons are defned on dfferent vertces (but of the same edge) T f α, f β dt A f α, f β 1 4A e 31, e 12 l 1l 2 cos θ 4A (h/ sn β)(h/ sn α) cos θ 2bh h cos θ 2(h cot α+h cot β) sn α sn β cos θ 2(cos α sn β+cos β sn α) cos θ 2 sn(α+β) cos θ 2 sn θ 1 2 cot θ Now we can apply these results on hat functons {h } by smply summng around each vertex 2h (a) Case 1: Same vertex (b) Case 2: Dfferent vertces Fgure 4: Summng around each vertex h p, h p 1 (cot α + cot β ) 2 h p, h q 1 2 (cot θ 1 + cot θ 2 ) Fnally, we get the cotangent Laplacan matrx L: 1 2 (cot α + cot β ), f L 1 2 (cot α + cot β ), f 0, otherwse (1) means that vertex and vertex are adacent ass atrx For the rght hand sde, we need to calculate the matrx A, whch s often called the mass matrx As t nvolved the product of h and h, the result would be quadratc There are several approaches to deal wth t A straghtforward and consstent approach s to ust do the quadratc ntegral Smlarly to the approach to calculate L, we fst ntegrate on a sngle trangle A trangle { area/6 area/12 f f Then we can sum up around each vertex There are two cases as shown n Fgure 4 (2) A { one rng area 6 f adacent area 12 f (3)

Some propertes of the mass matrx constructed ths way: each row sums to one thrd of the one-rng area of the vertex correspondng to that row; to construct the mass matrx t nvolves only vertex and ts neghbors; t parttons surface area as the weght to assgn to each vertex The matrx can be used for calculatng ntegraton (notce that h 1): f a h a h ( A a h ) 1 T A a Settng a 1 wll gve us the surface area However, there s one drawback of ths approach The mass matrx constructed s not dagonal whch means t s often hard to manpulate t We can turn t nto a dagonal matrx by ntroducng certan approxmatons From the prevous example, we have seen that the mass matrx s actually ntegratng a functon on the surface Therefore, we can try to fnd dfferent ways to do the ntegraton For example, the lumped mass matrx fnds the dual cells of each vertex and approxmate the dagonal of A wth the areas of each cell: a Area(cell ) Fgure 5: Lumped ass atrx wth dual cells Such approxmaton won t make a dfference for smooth functons Intutvely, the functon values of adacent vertces are very close for a smooth functons, so the result won t change a lot f we add them all to the dagonal eanwhle, as the mesh gets more and more refned, we can argue that the result would converge Fgure 6: Barycentrc Lumped ass atrx There are many ways to choose the dual cell One smple soluton s the barycentrc lumped mass matrx We assgn the dual of each face to be the barycentrc of the trangle Therefore, each vertex has a thrd of ts one-rng area assgned to t The resultng cells are lkely to be of rregular shapes One alternatve approach s to take Vorono Cells and use the areas accordngly