Some Thoughts on Guaranteed Function Approximation Satisfying Relative Error

Similar documents
Constructing Guaranteed Automatic Numerical Algorithms for U

A Guaranteed, Adaptive, Automatic Algorithm for Univariate Function Minimization

Reliable Error Estimation for Quasi-Monte Carlo Methods

A Deterministic Guaranteed Automatic Algorithm for Univariate 2014 June Function 9 1 Approximatio

Generalizing the Tolerance Function for Guaranteed Algorithms

REAL ANALYSIS II TAKE HOME EXAM. T. Tao s Lecture Notes Set 5

arxiv: v1 [math.ca] 4 Apr 2017

Minimizing the Number of Function Evaluations to Estimate Sobol Indices Using Quasi-Monte Carlo

Local behaviour of Galois representations

Variational inequality formulation of chance-constrained games

Draft. Chapter 2 Approximation and Interpolation. MATH 561 Numerical Analysis. Songting Luo. Department of Mathematics Iowa State University

On principal eigenpair of temporal-joined adjacency matrix for spreading phenomenon Abstract. Keywords: 1 Introduction

Singular integral operators and the Riesz transform

Mathematische Methoden der Unsicherheitsquantifizierung

Asymptotic Analysis 1: Limits and Asymptotic Equality

THEORY OF PROBABILITY VLADIMIR KOBZAR

Mariusz Jurkiewicz, Bogdan Przeradzki EXISTENCE OF SOLUTIONS FOR HIGHER ORDER BVP WITH PARAMETERS VIA CRITICAL POINT THEORY

Approximation in the Zygmund Class

Lecture 15: Quadratic approximation and the second variation formula

APPLIED MATHEMATICS REPORT AMR04/16 FINITE-ORDER WEIGHTS IMPLY TRACTABILITY OF MULTIVARIATE INTEGRATION. I.H. Sloan, X. Wang and H.

arxiv: v2 [math.co] 11 Oct 2016

A Guaranteed Automatic Integration Library for Monte Carlo Simulation

ETIKA V PROFESII PSYCHOLÓGA

NOTES ON SOME EXERCISES OF LECTURE 5, MODULE 2

ADVANCE TOPICS IN ANALYSIS - REAL. 8 September September 2011

p-adic Analysis Compared to Real Lecture 1

Area of left square = Area of right square c 2 + (4 Area of (a, b)-triangle) = a 2 + b 2 + (4 Area of (a, b)-triangle) c 2 = a 2 + b 2.

Elementary factoring algorithms

DS-GA 1002: PREREQUISITES REVIEW SOLUTIONS VLADIMIR KOBZAR

arxiv: v2 [math.ca] 13 May 2015

Erdinç Dündar, Celal Çakan

Absolute Value Information from IBC perspective

P E R E N C O - C H R I S T M A S P A R T Y

CHAPTER 6 : LITERATURE REVIEW

Tempered Distributions

Entropy and Ergodic Theory Notes 21: The entropy rate of a stationary process

MAT 771 FUNCTIONAL ANALYSIS HOMEWORK 3. (1) Let V be the vector space of all bounded or unbounded sequences of complex numbers.

arxiv: v2 [math.ca] 23 Feb 2018

Regularization in Reproducing Kernel Banach Spaces

LECTURE 3: MULTI-LINEAR RESTRICTION THEORY

MATH 1314 Test 2 Review

L E C T U R E 2 1 : M AT R I X T R A N S F O R M AT I O N S. Monday, November 14

INTERPLAY BETWEEN C AND VON NEUMANN

Fourier transform, assorted stuff...

An asymptotic answer to a special case of an open conjecture of Bondy

a P (A) f k(x) = A k g k " g k (x) = ( 1) k x ą k. $ & g k (x) = x k (0, 1) f k, f, g : [0, 8) Ñ R f k (x) ď g(x) k P N x P [0, 8) g(x)dx g(x)dx ă 8

Ground States are generically a periodic orbit

Introduction to Numerical Analysis

Vector Spaces. Vector space, ν, over the field of complex numbers, C, is a set of elements a, b,..., satisfying the following axioms.

Math 9 Practice Final Exam #1

arxiv: v2 [nlin.ps] 14 Aug 2016

Exotic Crossed Products

Functional Analysis Exercise Class

The Informativeness of k-means for Learning Mixture Models

MAT 128A - Practice Midterm Exam

Efficient Bayesian Multivariate Surface Regression

Public-Key Encryption

C*-algebras - a case study

CS 372: Computational Geometry Lecture 14 Geometric Approximation Algorithms

MATH 360 Final Exam Thursday, December 14, a n 2. a n 1 1

The Riemann Roch theorem for metric graphs

arxiv: v4 [math.ap] 14 Apr 2016

In class midterm Exam - Answer key

Quantifiers For and There exists D

MATH 54 - TOPOLOGY SUMMER 2015 FINAL EXAMINATION. Problem 1

Lecture 11: Clifford algebras

ON DIMENSION-INDEPENDENT RATES OF CONVERGENCE FOR FUNCTION APPROXIMATION WITH GAUSSIAN KERNELS

Hilbert Spaces. Hilbert space is a vector space with some extra structure. We start with formal (axiomatic) definition of a vector space.

1. INTRODUCTION 2. PRELIMINARIES

An Introduction to Analysis of Boolean functions

What s new in high-dimensional integration? designing for applications

Joint distribution optimal transportation for domain adaptation

T i t l e o f t h e w o r k : L a M a r e a Y o k o h a m a. A r t i s t : M a r i a n o P e n s o t t i ( P l a y w r i g h t, D i r e c t o r )

arxiv: v6 [math.nt] 6 Nov 2015

Math 9 Practice Final Exam #2

Entropy and Ergodic Theory Notes 22: The Kolmogorov Sinai entropy of a measure-preserving system

Automatic algorithms

Lecture 2: Homotopy invariance

arxiv: v3 [math.rt] 14 Mar 2014

Lecture 18 Finite Element Methods (FEM): Functional Spaces and Splines. Songting Luo. Department of Mathematics Iowa State University

Station keeping problem

NOTES WEEK 14 DAY 2 SCOT ADAMS

Lecture 17 Methods for System of Linear Equations: Part 2. Songting Luo. Department of Mathematics Iowa State University

Adam Najdecki, Józef Tabor CONDITIONALLY APPROXIMATELY CONVEX FUNCTIONS

Algebra 1 - Chapter 5 Test Review

Beyond Chebyshev technology. Alex Townsend MIT

NOTES WEEK 15 DAY 1 SCOT ADAMS

NONLINEAR EVOLUTION EQUATIONS SHORT NOTES WEEK #1

Multidimensional symbolic dynamics

Characterizing Cycle Partition in 2-Row Bulgarian Solitaire

arxiv: v1 [math.gr] 1 Apr 2019

GENERIC MUCHNIK REDUCIBILITY AND PRESENTATIONS OF FIELDS

Austin Mohr Math 730 Homework. f(x) = y for some x λ Λ

Math 578: Assignment 2

Problem Set 2: Solutions Math 201A: Fall 2016

Math Numerical Analysis Mid-Term Test Solutions

A SPECIAL CLASS OF INFINITE MEASURE-PRESERVING QUADRATIC RATIONAL MAPS RACHEL BAYLESS AND JANE HAWKINS

A TASTE OF COMBINATORIAL REPRESENTATION THEORY. MATH B4900 5/02/2018

Lecture 19 - Covariance, Conditioning

Application of Taylor Models to the Worst-Case Analysis of Stripline Interconnects

Transcription:

Some Thoughts on Guaranteed Function Approximation Satisfying Relative Error Fred J. Hickernell Department of Applied Mathematics, Illinois Institute of Technology hickernell@iit.edu mypages.iit.edu/~hickernell Joint work with Yuhan Ding (IIT) and Henryk Woźniakowski (Columbia U & U Warsaw) October 16, 2014 hickernell@iit.edu Function Approximation w/ Relative Error IIT, 10/16/2014 1 / 10

Function Approximation with Adaptive Linear Splines Given data px 0, y 0 q,..., px n, y n q with y fpx i q for f : r0, 1s Ñ R, Find A n pfq φpy 0,..., y n q : r0, 1s Ñ R such that f A n pfq is small. The linear spline is given by A n pfqpxq : We know that ˆ i f pi ` 1 nxq ` f n ˆi ` 1 n j pnx iq for i n ď x ď i ` 1 n. f A n pfq ď f 2 n 2 for f P W 2, (Clancy et al., 2014). hickernell@iit.edu Function Approximation w/ Relative Error IIT, 10/16/2014 2 / 10

Satisfying Error Tolerances for Balls f A n pfq ď f 2 n 2 for f P W 2,. Absolute error tolerances: the computational cost is bounded by f 2 : mintn : f A n pfq ď ε a @f P B σ u c σ ε a, B σ : tf P W 2, such that f 2 ď σu. Hybrid error tolerances, the computational cost is the same as for the absolute error tolerance: c σ mintn : f A n pfq ď maxpε a, ε r f q @f P B σ u. ε a Relative error tolerances, the computational cost is infinite: mintn : f A n pfq ď ε r f q @f P B σ u hickernell@iit.edu Function Approximation w/ Relative Error IIT, 10/16/2014 3 / 10

Why a Relative Error Tolerance Doesn t Help for Balls Let Ãn be any algorithm using n data. Let ξ ď ζ be the consecutive data sites spaced furthest apart. Define f bump pxq : 1 4pζ ξq2 ` p4x 2ξ 2ζq 2 32 `p4x ξ 3ζq 4x ξ 3ζ p4x 3ξ ζq 4x 3ξ ζ s à n pf bump q 0, f 2 bump 1, f bump pζ ξq2 16 ě 1 16pn ` 1q 2! ) min n : f Ãnpfq ď maxpε a, ε r f q @f P B σ! ) ě min n : σf bump Ãnpσf bump q ď maxpε a, ε r σf bump q " * c σ σ ě min n : 16pn ` 1q 2 ď ε a ε a hickernell@iit.edu Function Approximation w/ Relative Error IIT, 10/16/2014 4 / 10

Satisfying Error Tolerances for Cones Clancy et al. (2014) developed a way of choosing n based on function data to ensure that f A n pfq ď ε a without knowing f 2. Let C τ : tf P W 2, : f 2 ď τ f 1 fp1q ` fp0q u. By noting that for all f P W 2,, f 1 fp1q ` fp0q A n pfq 1 fp1q ` fp0q ď f 2 2n, it may be shown that f 2 ď τ A npfq 1 fp1q ` fp0q 1 τ{p2nq ÐÝ data-based hickernell@iit.edu Function Approximation w/ Relative Error IIT, 10/16/2014 5 / 10

Satisfying Error Tolerances for Cones Clancy et al. (2014) s algorithm chooses n to satisfy f A n pfq ď f 2 n 2 ď τ A npfq 1 fp1q ` fp0q ď ε a looooooooooooooooomooooooooooooooooon 4np2n τq data-based The computational cost for the absolute error tolerance is mintn : f A n pfq ď ε a @f P C τ X B σ u Hybrid error tolerances, the computational cost is unknown: c σ ε a. mintn : f A n pfq ď maxpε a, ε r f q @fc τ P B σ u? What about relative error tolerances (ε a 0)? hickernell@iit.edu Function Approximation w/ Relative Error IIT, 10/16/2014 6 / 10

Bounding Weaker Norms in Terms of Stronger Ones Let g f A 1 pfq, and note that g 1 is continuous. Let ξ be chosen such that g 1 pξq g 1. Also, define ζ ξ ` 1{τ or ζ ξ 1{τ, whichever falls inside r0, 1s. It follows from integration by parts and the triangle inequality Thus 2 g ě gpξq gpζq g1 pxqpx ζq ξ ζ ż ξ ě g 1 pξq ξ ζ 1 2 g 1 pxq dx g 2 pxqpx ζq dx ζ ż ξ ζ g 2 ξ ζ 2 ě g 1 1 τ 1 2 ˆ τ g 1 ˆ 1 τ 2 g 1 1 2τ f ě 1 2 f A 1pfq ě 1 f 1 A 1 pfq 1 τ, hickernell@iit.edu Function Approximation w/ Relative Error IIT, 10/16/2014 7 / 10

Adaptive Hybrid Error Tolerances Now we choose n to satisfy f A n pfq ď f 2 n 2 ď τ A npfq 1 fp1q ` fp0q looooooooooooooooomooooooooooooooooon 4np2n τq data-based ˆ ď max ε a, ε r A n pfq 1 A 1 pfq 1 ˆ ď max ε a, ε r f 1 A 1 pfq 1 τ τ ď maxpε a, ε r f q and get mintn : f A n pfq ď maxpε a, ε r f q @f P C τ X B σ u d ˆ σ min, 1 ε a ε r So either ε a or ε r positive gives bounded computational cost. hickernell@iit.edu Function Approximation w/ Relative Error IIT, 10/16/2014 / 10

Next We Consider W r`1, Use piecewise r th degree polynomials to approximate function. Maybe consider tensor product for d-variate functions. This idea does not work for integration problems. Why? hickernell@iit.edu Function Approximation w/ Relative Error IIT, 10/16/2014 9 / 10

References Clancy, N., Y. Ding, C. Hamilton, F. J. H., and Y. Zhang. 2014. The cost of deterministic, adaptive, automatic algorithms: Cones, not balls, J. Complexity 30, 21 45. hickernell@iit.edu Function Approximation w/ Relative Error IIT, 10/16/2014 10 / 10