14
Least Square Approximation and Normal Equation 4 th , December 2009 Presented by Kwak, Nam-ju

Least Square Approximation and Normal Equation

  • Upload
    moke

  • View
    89

  • Download
    1

Embed Size (px)

DESCRIPTION

Least Square Approximation and Normal Equation. 4 th , December 2009 Presented by Kwak , Nam- ju. Introduction. Given a data set, we sometimes hope to find a linear function that represents the data the best. We should do our best to minimize the sum of squared error . Introduction. - PowerPoint PPT Presentation

Citation preview

Page 1: Least Square Approximation and Normal Equation

Least Square Approximation and Normal Equation

4th, December 2009Presented by Kwak, Nam-ju

Page 2: Least Square Approximation and Normal Equation

Introduction• Given a data set, we sometimes hope to

find a linear function that represents the data the best.

• We should do our best to minimize the sum of squared error.

Page 3: Least Square Approximation and Normal Equation

Introduction

Page 4: Least Square Approximation and Normal Equation

Normal Equation• Notations

– (vi, wi): an element of the data set– c0+c1v=w: the linear function, supposed to rep-

resent the data set

– Solve the matrix equation to find out the val-ues for c0 and c1.

Page 5: Least Square Approximation and Normal Equation

Normal Equation

• It is expressed again using matrix nota-tions as follows:Ax=d, where each matrix A, x, and d refers to the one of the original expres-sion, in the same order.

Page 6: Least Square Approximation and Normal Equation

Normal Equation• ATAx = ATd is called the normal equation

associated with the least squares prob-lem.

• If ATA has an inverse, then the solution of the normal equation is also a solution of the least squares problem.

Page 7: Least Square Approximation and Normal Equation

Proof• A least squares solution is x such that

rTr=(d-Ax)T(d-Ax) is no larger than(d-Ay)T(d-Ay) for all y’s.

• In other words, we should guarantee that, for all y’s:(d-Ax)T(d-Ax)≤(d-Ay)T(d-Ay).

Page 8: Least Square Approximation and Normal Equation

(d-Ay)T(d-Ay)=((d-Ax)-A(y-x))T((d-Ax)-A(y-x))=((d-Ax)T-(A(y -x))T)((d-Ax)-A(y-x))=(d-Ax)T(d-Ax)-(A(y-x))T(d-Ax) -(d-Ax)TA(y-x)+(A(y-x))T(A(y-x))≥(d-Ax)T(d-Ax)-2((y-x)TAT(d-Ax).

Proofy=x+(y-x)

If this term is 0, the inequality(d-Ax)T(d-Ax)≤(d-Ay)T(d-Ay)Always holds.

Page 9: Least Square Approximation and Normal Equation

Proof• Let’s make ((y-x)TAT(d-Ax)=0.• In general y-x≠0. Therefore AT(d-Ax)=0.

AT(d-Ax)=ATd-ATAx=0ATAx = ATd

• Now, we have the desired result here.• If ATA is invertible, x can be solved.

Page 10: Least Square Approximation and Normal Equation

Example

Page 11: Least Square Approximation and Normal Equation

Further Considerations• With the form of

c0+c1f(v)=walso can be treated using a matrix, which is somewhat transformed from the original one, as follows:

Page 12: Least Square Approximation and Normal Equation

Further Considerations• We can also estimate the curve represent-

ing the data set.– c0+c1v+c2v2=w

Page 14: Least Square Approximation and Normal Equation

Questions and Answers• Any Question?