ma2213 lecture 3 approximation. piecewise linear interpolation p. 147 can use nodal basis functions
TRANSCRIPT
MA2213 Lecture 3
Approximation
Piecewise Linear Interpolation p. 147
],[,)( 2112
12 xxx
xx
xxx
1x
],[,)( 3223
23 xxx
xx
xxx
],[,)( 2112
21 xxx
xx
xxx
],[,0)( 213 xxxx
3x
3
2x1
2
1
1
11x
1x
2x
2x
3x
3x
],[,0)( 321 xxxx
],[,)( 3223
32 xxx
xx
xxx
can use Nodal Basis Functions
IntroductionProblem : Find / evaluate a function P that
(i) P belongs to a specified set S of functions, and (ii) P best approximates a function f among the
functions in the set S
Approximates = match, fit, resemble
If S were the set of ALL functions the choice P = f solves the problem – “you can’t get any closer to somewhere than by being there”. If S is not the set of all functions then S must be defined carefully
Furthermore, we must define the approximation criteria used to compare two approximations
Set of FunctionsIn practice S is closed under sums and multiplication
by numbers – this means that it is a vector space
Example: Bases for S = { Polynomials of degree < n } :
Furthermore, S is usually finite dimensional
},...,:)()({ 11RccxbcxPS nj
n
j j
and then S admits a basis })(),...,({ 1 xbxb n
Monomial Basis
Lagrange’s Basis
Newton’s Basis (used with divided differences)
)(),...,(1 xLxL n
12 ,...,,,1 nxxx
)()(),...,)((),(,1 11211 nxxxxxxxxxx
For distinct nodes nxx 1
For possibly repeated nodes nxx 1
Approximation CriteriaLeast Squares p. 178-187
http://en.wikipedia.org/wiki/Carl_Friedrich_Gauss
observed that measurement errors usuallyhave “Gaussian Statistics” and he invented an optimum method to deal with such errors.
In many engineering and scientific problems, data is acquired from measurements
Minimax or Best Approximation p. 159-165Arises in optimal design, game theory as wellas in mathematics of uniform convergence
Least Squares Criteria
Least Squares (over an finite set)
m
k kk xPy1
2))((minimize
Least Squares (over an interval)
b
adxxPx 2))()(f(minimize
Least Squares Over a Finite Set p. 319-333
If
m
k kk xPy1
2))((then we minimize
)()(1
xbcxP j
n
j j
21 1
)(
m
k
n
j kjjk xbcy
by choosing coefficients
ncc ,...,1to satisfy
nixbcyc
m
k
n
j kjjki
,...,1,0)(2
1 1
Least Squares Over a Finite Set
Remark: these are n equations in n variables
Since
ncc ,...,1the coefficients
m
k
n
j kjjki
xbcyc 1 1
2)(
nixbxbcy ki
n
j kjjk
m
k,...,1),()(2
11
satisfy the equations
niyxbcxbxbm
k kkij
n
j
m
k kjki ,...,1,)()()(11 1
Least Squares Equations
Construct matrices
niyxbcxbxbm
k kkij
n
j
m
k kjki ,...,1,)()()(11 1
)()(
)()(
1
111
mnm
n
xbxb
xbxb
B
nc
c
c 1
my
y
y 1
The least squares equations are yBBcB TT The interpolation equations yBc hold if and only if mkyxP kk ,...,1)( , Question When do these equations have solutions ?
Least Squares ExamplesFor
mxxxmxbn 211 ,1,1)(,1
1
1
B 1cc
my
y
y 1
The least squares equations are
m
yyc m
11
so
and the constant function
myyy 21
myymc 11
1111 1)( ccxbc is the least squares approximation or data fit for the data points ),(),...,,(),,( 2211 mm yxyxyx
by a constant function.
Least Squares ExamplesFor
mxxxmxxbxbn 2121 ,2,)(,1)(,2
mx
x
x
B
1
1
1
2
1
2
1
c
cc
my
y
y 1
The least squares equations are
myyy 21
mxxxS 211
The solution gives the least squares data fit
2
1
2
1
21
1
T
T
c
c
SS
Sm22
2212 mxxxS
myyyT 211
mm yxyxyxT 22112
xcc 21 by a polynomial of degree 1
Least Squares MATLAB CODEfunction c = lspf(n,m,x,y)% function c = lspf(n,m,x,y)%% Wayne Lawton 28 August 2007% Least Squares Polynomial Fit% Inputs : n = deg poly + 1, m = # data points, % data arrays x and y of size n x 1% Output : array c of poly coefficients%for i = 1:m for j = 1:n B(i,j) = x(i)^(j-1); endendc = (B'*B)\(B'*y);
Least Squares MATLAB CODE
Least Squares MATLAB CODE
sum of squares error for constant least squares fit
sum of squares error for ‘linear’ least squares fit
1.3250
13.469Question : Why did the ssq error decrease ?
Least Squares Algebraic FormulaThe sum of squares error
)()(.Sq Sum yBcyBc T can be computed, by substituting the value,
yBBBc TT 1)( to obtain
)(.Sq Sum BcyyT
Least Squares MATLAB CODE
ssq error for quadratic ls fit 76.463Question : Why did the error decrease so little ?
Marge, Where are the Least Squares ?
Least Squares Over an IntervalIf
b
adxxPx 2))()(f(then we minimize
)()(1
xbcxP j
n
j j
b
a
n
j jj dxxbcx2
)()(f1
by choosing coefficients
ncc ,...,1to satisfy
nidxxbcxc
b
a
n
j jji
,...,1,02
)()(f1
Least Squares Over an IntervalThese are also n equations in n variables
Since
nidxxbxbcxb
a i
n
j jj ,...,1,)()()(f21
the coefficients
ncc ,...,1satisfy the equations
b
a
n
j jji
dxxbcxc
2)()(f
1
b
a ij
n
j
b
a ji nidxxbxcdxxbxb ,...,1,)()(f)()(1
The ‘ interpolation equation ‘ on the interval is
],[),(f)()(1
baxxxbcxPn
j jj
Least Squares Over an Interval p. 181For S = polynomials of degree
nn xxbxxbxb )(,...,)(,1)( 121
over the interval
n
the equations are
1
0
1
0
11
11,,1,)(f)()(f
1nidxxxdxxbx
ji
c ii
n
j
j
The matrix of coefficients for these equationsis called the Hilbert matrix. It is a well-known example of an ill conditioned matrix and is discussed in Example 6.5.5 on pages 300-301
]1,0[
Legendre Polynomials p. 181-183defined by n
n
n
nn xdx
d
nxP 1
2!
1)( 2 ,1)(0 xP
xxxPxxPxxP 233
25
3212
23
21 )(,)(,)(
satisfy 1
1else ,0)()( dxxPxPji ji
)12/(2)()(1
1 jdxxPxP jj
they are examples of Orthogonal Polynomials
and are useful method to solve the least squared approximation by polynomials p. 183-185
Vandermonde Matrix
1
122
111
1
122
111
1
1
1
det)(),(
1
1
1
det
n
n
n
n
nnn
n
n
xx
xx
xx
xPxP
xx
xx
xx
n
i
i
jjin xxxP
2
1
1
)()(
1
11111 )(),...,()(0)()(n
kknn xxxxxPxPxP
211
222
211
1
1
1
det
nnn
n
n
xx
xx
xx
then use induction on n ),...,( 11 nxx
first expand det by last row to obtain
and
to obtain
Gramm Matrix
nnn
n
gg
gg
G
1
111
Theorem 1. If
dxxbxbgb
a jiij )()(
nibaCbi ,...,1]),,([
is nonsingular.
0],...,[ 1 TnvvvProof
are linearly
independent, then the Gramm matrix defined by
00)(2
1
GvdxxbvvGvb
a
n
j jjT
01
n
j jjbvh
nbb ,...,1lin. ind. and
Derive this equation
Semipositive Definite and Positive DefiniteDefinitions A real n x n matrix is (semi) positivedefinite if for every nonzero vector
Theorem 2. If is both positive semidefinite and symmetric and 0vPvT
nRv
satisfies
0)( vPvT
nRvthen .0vPProof Assume that
)()()( xuvPxuvxF Tu
nRuconstruct the functionand observe that since is positive semidefinite
0)0(),(0)0( dxdF
uuuRxxFF
.0vPvT For every
Since ,PPT
P
P
P
02)0( PvuTdxdFu therefore
nRu .0vP0PvuT for every hence
Gramm MatrixCorollary (of Thm 2). If the Gramm matrix for a
.0v
nibaCbi ,...,1]),,([ is nonsingular then this set of functions is linearly independent.
nTn Rvvv ],...,[ 1
isProof. Assume that the Gramm matrix
0)(2
1
dxxbvvGvb
a
n
j jjT
.01
n
j jjbv
Gnonsingular and that
set of functions
satisfies
It suffices to show that
We observe that
Furthermore, since satisfies the hypothesis ofGtheorem 2, it follows that 0Gv and the proof
is complete.
Questions for Thought and Discussion
Question 1. Is the matrix
2,00 RvMvMvvT Is it true that
semi positive definite, positive definite,symmetric ?
10
01M
Question 2. Give and example of a positive semidefinite matrix that is not positive definite.
Question 3. Find a positive semidefinite and a nonzero vectorMnonsingular matrix
such that
v.0MvvT
Question 4. Give a detailed derivation for all the assertions in the proofs of Theorems 1 and 2.
http://en.wikipedia.org/wiki/Image:Langrange_portrait.jpg
Joseph Louis Lagrange
Jan 1736 – April 1813Giuseppe Lodovico Lagrangia was an Italian-French mathematician and astronomer who made important contributionsto all fields of analysis and number theory and to classical and celestial mechanics as arguably the greatest mathematician of the 18th century. Before the age of 20 he was professor of geometry at the royal artillery school at Turin.
http://en.wikipedia.org/wiki/Isaac_Newton
Isaac Newton
January 1643 – March
1727)was an English physicist, mathematician, astronomer, natural philosopher, and alchemist, regarded by many as the greatest figure in the history of science. His treatisePhilosophiae Naturalis Principia Mathematica, published in 1687, described universal gravitation and the three laws of motion, laying the groundwork for classical mechanics.
http://en.wikipedia.org/wiki/Carl_Friedrich_Gauss
Carl Friedrich Gauss
April 1777 – February 1855 was a German mathematician and scientist of profound genius who contributed significantly to many fields, including numbertheory, analysis, differential geometry, geodesy, magnetism,astronomy and optics. 23, heard about the problem After three months of intense work, he predicted a position for Ceres in Dec 1801 …influential treatment of the method of least squares …minimize the impact of measurement error …under the assumption of normally distributed errors …when asked how he had been able to predict the trajectory of Ceres with such accuracy he replied, "I used logarithms." The questioner then wanted to know how he had been able to look up so many numbers from the tables so quickly. "Look them up?" Gauss responded. "Who needs to look them up? I just calculate them in my head!"
Homework Due Tutorial 2
Question 2. Use Newton’s divided differencemethod to find a polynomial of degree that satisfies
Question 1. Do Problem 8 (a) on page 133.
Question 3. Do Problem 1 on page 330. Please show all details in your computation.
)(2 xP 221)4(,0)1()1( 2
'22 PPP
Question 4. Compute the Gramm Matrix for the three piecewise linear nodal basis functions in the first vufoil if the three nodes are -1, 0, 1 and the interval of integration is [-1,1]
Homework Example
dxxxg )()( 2
1
`1 112
Question 4. Compute the Gramm Matrix for the three piecewise linear nodal basis functions in the first vufoil if the three nodes are -1, 0, 1 and the interval of integration is [-1,1]Solution The Gram matrix will be a 3 x 3 matrix, here is one of its entries
dxxdxxx )1()0()1()(1
`0
0
`1
6
10
0
12
213
31 xx
Questions for Thought and Discussion
Answer 1. The matrix
01
1but 0
1
1
MvMvvv T
is symmetric, however it is not semi positivedefinite and therefore not positive definite.
10
01M
Answer 3. The matrix is semi positive
10
21M
definite and nonsingular and if
1
1v
then 01
1]11[
MvvT and clearly
0Mv