Lecture 1 Linear Variational Problems (Part I). 1. Motivation For those participants wondering why we start a course dedicated to nonlinear problems by.

Slides:



Advertisements
Similar presentations
10.4 Complex Vector Spaces.
Advertisements

5.1 Real Vector Spaces.
Chapter 4 Euclidean Vector Spaces
Introduction to Variational Methods and Applications
Copyright © Cengage Learning. All rights reserved. CHAPTER 5 SEQUENCES, MATHEMATICAL INDUCTION, AND RECURSION SEQUENCES, MATHEMATICAL INDUCTION, AND RECURSION.
INFINITE SEQUENCES AND SERIES
Separating Hyperplanes
Chaper 3 Weak Topologies. Reflexive Space.Separabe Space. Uniform Convex Spaces.
ECE 552 Numerical Circuit Analysis
Function Optimization Newton’s Method. Conjugate Gradients
Tutorial 12 Unconstrained optimization Conjugate gradients.
Lecture 2 Linear Variational Problems (Part II). Conjugate Gradient Algorithms for Linear Variational Problems in Hilbert Spaces 1.Introduction. Synopsis.
1 L-BFGS and Delayed Dynamical Systems Approach for Unconstrained Optimization Xiaohui XIE Supervisor: Dr. Hon Wah TAM.
1 Introduction to Kernels Max Welling October (chapters 1,2,3,4)
Basic Concepts and Definitions Vector and Function Space. A finite or an infinite dimensional linear vector/function space described with set of non-unique.
Optimization Methods One-Dimensional Unconstrained Optimization
Support Vector Machines and Kernel Methods
MADRID LECTURE # 5 Numerical solution of Eikonal equations.
Optical Flow Estimation using Variational Techniques Darya Frolova.
ECE 552 Numerical Circuit Analysis Chapter Six NONLINEAR DC ANALYSIS OR: Solution of Nonlinear Algebraic Equations Copyright © I. Hajj 2012 All rights.
Math for CSLecture 51 Function Optimization. Math for CSLecture 52 There are three main reasons why most problems in robotics, vision, and arguably every.
Maximum likelihood (ML)
CISE-301: Numerical Methods Topic 1: Introduction to Numerical Methods and Taylor Series Lectures 1-4: KFUPM.
5  Systems of Linear Equations: ✦ An Introduction ✦ Unique Solutions ✦ Underdetermined and Overdetermined Systems  Matrices  Multiplication of Matrices.
Quantum One: Lecture 8. Continuously Indexed Basis Sets.
 Limit Cycle Isolated closed trajectories Analysis is, in general, quite difficult. For nonlinear systems, two simple results exist. Limit Cycle
Computational Optimization
1 Chapter 8 The Discrete Fourier Transform 2 Introduction  In Chapters 2 and 3 we discussed the representation of sequences and LTI systems in terms.
Copyright © Cengage Learning. All rights reserved. CHAPTER 11 ANALYSIS OF ALGORITHM EFFICIENCY ANALYSIS OF ALGORITHM EFFICIENCY.
Signal-Space Analysis ENSC 428 – Spring 2008 Reference: Lecture 10 of Gallager.
ME 1202: Linear Algebra & Ordinary Differential Equations (ODEs)
Gram-Schmidt Orthogonalization
CISE-301: Numerical Methods Topic 1: Introduction to Numerical Methods and Taylor Series Lectures 1-4: KFUPM CISE301_Topic1.
CISE301_Topic11 CISE-301: Numerical Methods Topic 1: Introduction to Numerical Methods and Taylor Series Lectures 1-4:
11.2 Series In this section, we will learn about: Various types of series. INFINITE SEQUENCES AND SERIES.
AGC DSP AGC DSP Professor A G Constantinides©1 Hilbert Spaces Linear Transformations and Least Squares: Hilbert Spaces.
1 Chien-Hong Cho ( 卓建宏 ) (National Chung Cheng University)( 中正大學 ) 2010/11/11 in NCKU On the finite difference approximation for hyperbolic blow-up problems.
DE Weak Form Linear System
Chapter 4 Hilbert Space. 4.1 Inner product space.
Survey of Kernel Methods by Jinsan Yang. (c) 2003 SNU Biointelligence Lab. Introduction Support Vector Machines Formulation of SVM Optimization Theorem.
1  The Problem: Consider a two class task with ω 1, ω 2   LINEAR CLASSIFIERS.
1  Problem: Consider a two class task with ω 1, ω 2   LINEAR CLASSIFIERS.
OR Simplex method (algebraic interpretation) Add slack variables( 여유변수 ) to each constraint to convert them to equations. (We may refer it as.
Mathematical Tools of Quantum Mechanics
MA2213 Lecture 2 Interpolation.
Network Systems Lab. Korea Advanced Institute of Science and Technology No.1 Ch. 3 Iterative Method for Nonlinear problems EE692 Parallel and Distribution.
x Examples of Fixed Points Infinite Fixed Points.
Linear Programming Chapter 9. Interior Point Methods  Three major variants  Affine scaling algorithm - easy concept, good performance  Potential.
Comparison Value vs Policy iteration
1 Kernel-class Jan Recap: Feature Spaces non-linear mapping to F 1. high-D space 2. infinite-D countable space : 3. function space (Hilbert.
Summary of the Last Lecture This is our second lecture. In our first lecture, we discussed The vector spaces briefly and proved some basic inequalities.
Network Systems Lab. Korea Advanced Institute of Science and Technology No.1 Maximum Norms & Nonnegative Matrices  Weighted maximum norm e.g.) x1x1 x2x2.
Copyright © Cengage Learning. All rights reserved. Graphs; Equations of Lines; Functions; Variation 3.
§7-4 Lyapunov Direct Method
Complex Variables. Complex Variables Open Disks or Neighborhoods Definition. The set of all points z which satisfy the inequality |z – z0|
Pole Placement and Decoupling by State Feedback
Proving that a Valid Inequality is Facet-defining
Autonomous Cyber-Physical Systems: Dynamical Systems
Polyhedron Here, we derive a representation of polyhedron and see the properties of the generators. We also see how to identify the generators. The results.
Solution of Equations by Iteration
Polyhedron Here, we derive a representation of polyhedron and see the properties of the generators. We also see how to identify the generators. The results.
§2-3 Observability of Linear Dynamical Equations
Preliminary.
Stability Analysis of Linear Systems
Quantum Foundations Lecture 3
Maths for Signals and Systems Linear Algebra in Engineering Lecture 6, Friday 21st October 2016 DR TANIA STATHAKI READER (ASSOCIATE PROFFESOR) IN SIGNAL.
Proving that a Valid Inequality is Facet-defining
Linear Vector Space and Matrix Mechanics
Calculus In Infinite dimensional spaces
1st semester a.y. 2018/2019 – November 22, 2018
Presentation transcript:

Lecture 1 Linear Variational Problems (Part I)

1. Motivation For those participants wondering why we start a course dedicated to nonlinear problems by discussing a particular family of linear problems let us say that: (i) The Newton’s method provides a systematic way to reduce the solution of some nonlinear problems to the solution of a sequence of linear problems.

1. Motivation For those participants wondering why we start a course dedicated to nonlinear problems by discussing a particular family of linear problems let us say that: (i) The Newton’s method provides a systematic way to reduce the solution of some nonlinear problems to the solution of a sequence of linear problems. (ii) During this course, we will show that there are other methods than Newton’s which rely on linear solvers to achieve the solution of nonlinear problems. An example of such methods will be provided by the solution of nonlinear problems in Hilbert spaces by conjugate gradient algorithms.

1. Motivation For those participants wondering why we start a course dedicated to nonlinear problems by discussing a particular family of linear problems let us say that: (i) The Newton’s method provides a systematic way to reduce the solution of some nonlinear problems to the solution of a sequence of linear problems. (ii) During this course, we will show that there are other methods than Newton’s which rely on linear solvers to achieve the solution of nonlinear problems. An example of such methods will be provided by the solution of nonlinear problems in Hilbert spaces by conjugate gradient algorithms. We will start our discussion with the Newton’s method.

2. The Newton’s method in Hilbert spaces Let V be a Hilbert space (real for simplicity); we denote by V’ the dual space of V and by the duality pairing between V and V’. Consider now an operator A (possibly nonlinear) mapping V into V’ and the following equation (E) A(u) = 0. Suppose that operator A is F-differentiable in the neighborhood of a solution u of (E); if v is sufficiently close to u we have then

0 = A(u) = A(v + u – v) = A(v) + A’(v)(u – v) + o(||v – u||), with A’(v) (  L ( V, V’)) the differential of A at v. The abo- ve relation suggest the following algorithm, classically known as the Newton-Raphson method: (1) u 0 is given in V ; then, for n ≥ 0, u n  u n+1 by (2) u n+1 = u n – A’(u n ) –1 A (u n ).

Suppose that A’(u)  Isom(V, V’) and that A’ is locally Lipschitz continuous, then if u 0 is sufficiently close to u we have lim n  +∞ u n = u, the convergence being super-linear. From a ‘practical’ point of view it may be more appropriate to write (2) as (3) A’(u n )(u n+1 – u n ) = – A(u n ), or even better (and equivalently) as: u n+1 – u n  V, (VF) = –,  v  V.

Many remarks are in order, but the only one we will make is that (VF) is a linear variational problem (in the sense of, e.g., J.L. LIONS) since the bilinear functional {v, w}  : V×V  R and the linear functional v  : V  R are both continuous. 3. On a family of linear variational problems in Hilbert spaces: The Lax-Milgram Theorem We are going to focus our discussion on a particular family of linear problems, namely to those problems defined by

considering a triple {V, a, L} verifying the following conditions: (i) V is a real Hilbert space for the scalar (inner) product (.,.) and the associated norm ||. ||. (ii) a: V×V  R is bilinear, continuous and V-elliptic (this last property meaning that there exists  > 0 such that a(v,v) ≥  ||v|| 2,  v  V ). (iii) L: V  R is linear and continuous. Remark 1: We do not assume that the bilinear functional a is symmetric.

Remark 2: If V is a finite dimensional space, the V- ellipticity of a  the positive definiteness of a (not necessarily true if dim V is infinite). To {V, a, L} we associate the following family of linear (variational) problems: Find u  V such that (LVP) a(u,v) = L(v),  v  V.

Why do we consider such a family of linear problems ? Among the many reasons, let us mention the main ones:

1) Many applications in Mechanics, Physics, Control, Image Processing,…….. lead to such problems once the convenient Hilbert space has been identified (exemples of such situations will be encountered during this course).

Why do we consider such a family of linear problems ? Among the many reasons, let us mention the main ones: 1) Many applications in Mechanics, Physics, Control, Image Processing,…….. lead to such problems once the convenient Hilbert space has been identified (exemples of such situations will be encountered during this course). 2) Let us return to the solution of A(u) = 0 by the Newton’s method and suppose that A = J’ where J is a strongly convex twice differentiable functional (i.e.,   > 0, such that ≥  ||w – v|| 2,  v, w), then the linear problem that we have to solve at each iteration to obtain u n +1 belongs to the (LVP) family *.

Strictly speaking, a problem is (was) variational if it can be viewed as the Euler-Lagrange equation of some problem from the Calculus of Variations. In the particular case of (LVP) this supposes that a is symmetric; indeed, if a is symmetric, then (LVP) is equivalent to Find u  V such that (MP) J(u)  J(v),   V, with J(v) = ½ a(v,v) – L(v). Following the influence of J.L. Lions and G. Stampacchia (mid 1960’s) many authors use the terminology variational equations (and of course inequalities) even when a(.,.) is not symmetric.

The Lax-Milgram Theorem Suppose that the triple {V,a,L} verifies (i)-(iii) then (LVP) has a unique solution. For a proof, see, e.g., Atkinson & Han (Springer). There are several proofs of the LM Theorem, our favorite one relies on the Banach fixed point theorem and can be generalized to variational inequalities. Actually, the LM Theorem can be generalized to complex Hilbert spaces by assuming that a(.,.) is sesquilinear, continuous and verifies Re a(v,v) ≥  ||v|| 2,  v  V.

Among the many comments which can be associated with the Lax-Milgram Theorem, we will focus on the following one, considering its computational implications : The fixed point based proof of the LM Theorem is based is based on the following observations:

Among the many comments which can be associated with the Lax-Milgram Theorem, we will focus on the following one, considering its computational implications : The fixed point based proof of the LM Theorem is based on the following observations: (1) From the Riescz representation Theorem, there exists a unique pair {A, l } such that: A  L (V,V), (Av,w) = a(v,w),  v, w  V,   ||A||, ( l,v) = L(v),  v  V.

Among the many comments which can be associated with the Lax-Milgram Theorem, we will focus on the following one, considering its computational implications : The fixed point based proof of the LM Theorem is based on the following observations: (1) From the Riescz representation Theorem, there exists a unique pair {A, l } such that: A  L (V,V), (Av,w) = a(v,w),  v, w  V,   ||A||, ( l,v) = L(v),  v  V. (2) If 0 < ρ < 2  ||A|| – 2 the mapping v  v – ρ(Av – l ) is a uniformly strict contraction of V.

Among the many comments which can be associated with the Lax-Milgram Theorem, we will focus on the following one, considering its computational implications : The fixed point based proof of the LM Theorem is based on the following observations: (1) From the Riescz representation Theorem, there exists a unique pair {A, l } such that: A  L (V,V), (Av,w) = a(v,w),  v, w  V,   ||A||, ( l,v) = L(v),  v  V. (2) If 0 < ρ < 2  ||A|| – 2 the mapping v  v – ρ(Av – l ) is a uniformly strict contraction of V. (3) (LVP) and Au = l are equivalent.

It follows from observations (1)-(3) that if ρ verifies 0 < ρ < 2  ||A|| – 2 we have geometric convergence of the following algorithm,  u 0  V : (1) u 0 is given in V, n ≥ 0, u n  u n+1 by (2) u n+1 = u n – ρ(Au n – l ). The practical interest of (1)-(2) as written above is limited by the fact that in general A and l are unknown.

A more practical form of (1)-(2) is obtained by replacing relation (2) by the following equivalent one: u n+1  V, (2)’ (u n+1,v) = (u n,v) – ρ[a(u n,v) – L(v)],  v  V. Problem (2)’ is also from the (LVP) family, with the role of a(.,.) played by the V-scalar product (.,.). The fact that ρ is not known in general can be overcome by using various techniques using a sequence { ρ n } n instead of a fixed ρ. If a(.,.) is symmetric, conjugate gradient provides a more efficient alternative to algorithm (1)-(2)’, at little extra cost. Conjugate gradient will be discussed next.