Horner

Accelerating Convergence
Zeros of Polynomials
Deation, M ullers Method
Polynomial Approximation
Numerical Analysis and Computing
Lecture Notes #04 Solutions of Equations in One Variable,
Interpolation and Polynomial Approximation Accelerating
Convergence; Zeros of Polynomials; Deation; M ullers Method;
Lagrange Polynomials; Nevilles Method
Joe Mahay,
mahaffy@math.sdsu.edu
Department of Mathematics
Dynamical Systems Group
Computational Sciences Research Center
San Diego State University
San Diego, CA 92182-7720
http://www-rohan.sdsu.edu/jmahay
Spring 2010
Joe Mahay, mahaffy@math.sdsu.edu #4: Solutions of Equations in One Variable (1/58)
Outline
1
Review
Aitkens
2
Method
Steensens Method
2
Fundamentals
Horners Method
3
Deation: Finding All the Zeros of a Polynomial
M ullers Method Finding Complex Roots
4
Fundamentals
Moving Beyond Taylor Polynomials
Lagrange Interpolating Polynomials
Nevilles Method
Review
Aitkens
2
Method
Steensens Method
Introduction
It is rare to have the luxury of quadratic convergence.
(Burden-Faires, p.83)
There are a number of methods for squeezing faster convergence out of
an already computed sequence of numbers.
We here explore one method which seems the have been around since the
beginning of numerical analysis... Aitkens
2
method. It can be used
to accelerate convergence of a sequence that is linearly convergent,
regardless of its origin or application.
A review of modern extrapolation methods can be found in:
Practical Extrapolation Methods: Theory and Applications, Avram
Sidi, Number 10 in Cambridge Monographs on Applied and Compu-
tational Mathematics, Cambridge University Press, June 2003. ISBN:
0-521-66159-5
Review
Aitkens
2
Method
Steensens Method
Recall: Convergence of a Sequence
Denition
Suppose the sequence {p
n
}
n=0
converges to p, with p
n
= p for all
n. If positive constants and exists with
lim
n
|p
n+1
p|
|p
n
p|
=
then {p
n
}
n=0
converges to p of order , with asymptotic error
constant .
Linear convergence means that = 1, and || < 1.
Review
Aitkens
2
Method
Steensens Method
Aitkens
2
Method I/II
Assume {p
n
}
n=0
is a linearly convergent sequence with limit p.
Further, assume we are far out into the tail of the sequence (n
large), and the signs of the successive errors agree, i.e.
sign(p
n
p) = sign(p
n+1
p) = sign(p
n+2
p) = . . .
and that
p
n+2
p
p
n+1
p

p
n+1
p
p
n
p
(the asymptotic limit)
This would indicate
(p
n+1
p)
2
(p
n+2
p)(p
n
p)
p
2
n+1
2p
n+1
p + p
2
p
n+2
p
n
(p
n+2
+ p
n
)p + p
2
We solve for p and get...
Review
Aitkens
2
Method
Steensens Method
Aitkens
2
Method II/II
We solve for p and get...
p
p
n+2
p
n
p
2
n+1
p
n+2
2p
n+1
+ p
n
A little bit of algebraic manipulation put this into the classical
Aitken form:
p
n
= p = p
n

(p
n+1
p
n
)
2
p
n+2
2p
n+1
+ p
n
Aitkens
2
Method is based on the assumption that the p
n
we
compute from p
n+2
, p
n+1
and p
n
is a better approximation to the
real limit p.
The analysis needed to prove this is beyond the scope of this class, see
e.g. Sidis book.
Review
Aitkens
2
Method
Steensens Method
Aitkens
2
Method The Recipe
Given a sequence nite {p
n
}
N
n=0
or innite {q
n
}
n=0
sequence
which converges linearly to some limit.
Dene the new sequences
p
n
= p
n

(p
n+1
p
n
)
2
p
n+2
2p
n+1
+ p
n
, n = 0, 1, . . . , N 2
or
q
n
= q
n

(q
n+1
q
n
)
2
q
n+2
2q
n+1
+ q
n
, n = 0, 1, . . . ,
The numerator is a forward dierence squared, while the
denominator is a second order central dierence.
Review
Aitkens
2
Method
Steensens Method
Aitkens
2
Method Example
Consider the sequence {p
n
}
n=0
, where the sequence is generated by the
xed point iteration p
n+1
= cos(p
n
), p
0
= 0.
Iteration p
n
p
n
0 0.000000000000000 0 .685073357326045
1 1.000000000000000 0.7 28010361467617
2 0 .540302305868140 0.73 3665164585231
3 0 .857553215846393 0.73 6906294340474
4 0 .654289790497779 0.73 8050421371664
5 0.7 93480358742566 0.73 8636096881655
6 0.7 01368773622757 0.73 8876582817136
7 0.7 63959682900654 0.73 8992243027034
8 0.7 22102425026708 0.7390 42511328159
9 0.7 50417761763761 0.7390 65949599941
10 0.73 1404042422510 0.7390 76383318956
11 0.7 44237354900557 0.73908 1177259563
12 0.73 5604740436347 0.73908 3333909684
Note: Bold digits are correct; p

11
needs p
13
, and p
12
additionally needs p
14
.
Review
Aitkens
2
Method
Steensens Method
Faster Convergence for Aitken-Sequences
Theorem (Convergence of Aitken-
2
-Sequences)
Suppose {p
n
}
n=0
is a sequence that converges linearly to the limit
p, and for n large enough we have (p
n
p)(p
n+1
p) > 0. Then
the Aitken-accelerated sequence { p
n
}
n=0
converges fast to p in the
sense that
lim
n
_
p
n
p
p
n
p
_
= 0.
We can combine Aitkens method with xed-point iteration in
order to get a xed-point iteration on steroids.
Review
Aitkens
2
Method
Steensens Method
Steensens Method: Fixed-Point Iteration on Steroids
Suppose we have a xed point iteration:
p
0
, p
1
= g(p
0
), p
2
= g(p
1
), . . .
Once we have p
0
, p
1
and p
2
, we can compute
p
0
= p
0

(p
1
p
0
)
2
p
2
2p
1
+ p
0
At this point we restart the xed point iteration with p
0
= p
0
,
e.g.
p
3
= p
0
, p
4
= g(p
3
), p
5
= g(p
4
),
and compute
p
3
= p
3

(p
4
p
3
)
2
p
5
2p
4
+ p
3
Review
Aitkens
2
Method
Steensens Method
Steensens Method: The Quadratic, g-g-A, Waltz! Quadratic Convergence
Algorithm: Steensens Method
Input: Initial approximation p
0
; tolerance TOL; maximum
number of iterations N
0
.
Output: Approximate solution p, or failure message.
1. Set i = 1
2. While i N
0
do 3--6
3
Set
p
1
= g(p
0
), p
2
= g(p
1
),
p = p
0
(p
1
p
0
)
2
/(p
2
2p
1
+ p
0
)
4. If |p p
0
| < TOL then
4a. output p
4b. stop program
5. Set i = i + 1
6. Set p
0
= p
7. Output: Failure after N
0
iterations.
Review
Aitkens
2
Method
Steensens Method
Steensens Method: Potential Breakage
3
If at some point p
2
2p
1
+ p
0
= 0 (which appears in the
denominator), then we stop and select the current value of p
2
as our approximate answer.
Both Newtons and Steensens methods give quadratic
convergence. In Newtons method we compute one function value
and one derivative in each iteration. In Steensens method we
have two function evaluations and a more complicated algebraic
expression in each iteration, but no derivative. It looks like we
got something for (almost) nothing.
Review
Aitkens
2
Method
Steensens Method
Steensens Method: Potential Breakage
3
If at some point p
2
2p
1
+ p
0
= 0 (which appears in the
denominator), then we stop and select the current value of p
2
as our approximate answer.
Both Newtons and Steensens methods give quadratic
convergence. In Newtons method we compute one function value
and one derivative in each iteration. In Steensens method we
have two function evaluations and a more complicated algebraic
expression in each iteration, but no derivative. It looks like we
got something for (almost) nothing. However, in order the
guarantee quadratic convergence for Steensens method, the xed
point function g must be 3 times continuously dierentiable, e.g.
f C
3
[a, b], (see theorem-2.14 in Burden-Faires). Newtons
method only requires f C
2
[a, b] (BF Theorem-2.5).
Review
Aitkens
2
Method
Steensens Method
Steensens Method: Example 1 of 2
Below we compare a Fixed Point iteration, Newtons Method, and
Steensens Method for solving:
f (x) = x
3
+ 4x
2
10 = 0
or alternately,
p
n+1
= g(p
n
) =
10
p
n
+ 4
Review
Aitkens
2
Method
Steensens Method
Fixed Point Iteration
i p
n
g(p
n
)
0 1.50000 1.34840
1 1.34840 1.36738
2 1.36738 1.36496
3 1.36496 1.3652
4 1.36526 1.36523
5 1.36523 1.36523
Review
Aitkens
2
Method
Steensens Method
Newtons Method
i x
n
f (x
n
)
0 1.50000 1.51600e-01
1 1.36495 -3.11226e-04
2 1.36523 -1.35587e-09
Review
Aitkens
2
Method
Steensens Method
Newtons Method
i x
n
f (x
n
)
0 1.50000 1.51600e-01
1 1.36495 -3.11226e-04
2 1.36523 -1.35587e-09
Steensens Method
i p
n
p
1
p
2
p |p p
2
|
0 1.50000 1.34840 1.36738 1.36527 3.96903e-05
1 1.36527 1.36523 1.36523 1.36523 2.80531e-12
Fundamentals
Horners Method
Denition: Degree of a Polynomial
A polynomial of degree n has the form
P(x) = a
n
x
n
+ a
n1
x
n1
+ + a
1
x + a
0
, a
n
= 0
where the a
i
s are constants (either real, or complex) called the
coecients of P.
Why look at polynomials? Well be looking at the problem
P(x) = 0 (i.e. f (x) = 0 for a special class of functions.)
Polynomials are the basis for many approximation methods, hence
being able to solve polynomial equations fast is valuable.
Wed like to use Newtons method, so we need to compute P(x)
and P
(x) as eciently as possible.

Fundamentals
Horners Method
Fundamentals
Theorem (The Fundamental Theorem of Algebra)
If P(x) is a polynomial of degree n 1 with real or complex
coecients, then P(x) = 0 has at least one (possibly complex)
root.
The proof is surprisingly(?) dicult and requires understanding of
complex analysis... We leave it as an exercise for the motivated
student!
Fundamentals
Horners Method
Key Consequences of the Fundamental Theorem of Algebra 1 of 2
Corollary
If P(x) is a polynomial of degree n 1 with real or complex
coecients then there exists unique constants x
1
, x
2
, . . . , x
k
(possibly complex) and unique positive integers m
1
, m
2
, . . . , m
k
such that
k
i =1
m
i
= n and
P(x) = a
n
(x x
1
)
m
1
(x x
2
)
m
2
(x x
k
)
m
k
The collection of zeros is unique.
m
i
are the multiplicities of the individual zeros.
A polynomial of degree n has exactly n zeros, counting
multiplicity.
Fundamentals
Horners Method
Key Consequences of the Fundamental Theorem of Algebra 2 of 2
Corollary
Let P(x) and Q(x) be polynomials of degree at most n. If x
1
, x
2
,
. . . , x
k
, with k > n are distinct numbers with P(x
i
) = Q(x
i
) for
i = 1, 2, . . . , k, then P(x) = Q(x) for all values of x.
If two polynomials of degree n agree at at least (n +1) points,
then they must be the same.
Fundamentals
Horners Method
Horners Method: Evaluating Polynomials Quickly 1 of 4
Let
P(x) = a
n
x
n
+ a
n1
x
n1
+ + a
1
x + a
0
We want an ecient method to solve
P(x) = 0.
In 1819, William Horner developed an ecient algorithm with n
multiplications and n additions to evaluate P(x
0
).
Technique is called Horners Method or Synthetic Division.
Fundamentals
Horners Method
Assume
P(x) = a
n
x
n
+ a
n1
x
n1
+ + a
1
x + a
0
If we are looking to evaluate P(x
0
) for any x
0
, let
b
n
= a
n
, b
k
= a
k
+ b
k+1
x
0
, k = (n 1), (n 2), . . . , 1, 0,
then b
0
= P(x
0
). We have only used n multiplications and n
additions.
At the same time we have computed the decomposition
P(x) = (x x
0
)Q(x) + b
0
,
where
Q(x) =
n1
k=1
b
k+1
x
k
.
If b
0
= 0, then x
0
is a root of P(x).
Fundamentals
Horners Method
Huh?!? Where did the expression come from? Consider
P(x) = a
n
x
n
+ a
n1
x
n1
+ + a
1
x + a
0
= (a
n
x
n1
+ a
n1
x
n2
+ + a
1
)x + a
0
= ((a
n
x
n2
+ a
n1
x
n3
+ )x + a
1
)x + a
0
= (. . . ((
. .
n1
a
n
x + a
n1
. .
b
n1
)x + )x + a
1
)x + a
0
Horners method is simply the computation of this parenthesized
expression from the inside-out...
Fundamentals
Horners Method
Now, if we need to compute P
(x
0
) we have
P
(x)
x=x
0
= (x x
0
)Q
(x) + Q(x)
x=x
0
= Q(x
0
)
Which we can compute (again using Horners method) in (n 1)
multiplications and (n 1) additions.
Proof? We really ought to prove that Horners method works. It
basically boils down to lots of algebra which shows that the
coecients of P(x) and (x x
0
)Q(x) + b
0
are the same...
A couple of examples may be more instructive...
Fundamentals
Horners Method
Example#1: Horners Method
For P(x) = x
4
x
3
+ x
2
+ x 1, compute P(5):
x
0
= 5 a
4
= 1 a
3
= 1 a
2
= 1 a
1
= 1 a
0
= 1
b
4
x
0
b
3
x
0
b
2
x
0
b
1
x
0
b
4
= 1
Fundamentals
Horners Method
For P(x) = x
4
x
3
+ x
2
x
0
= 5 a
4
= 1 a
3
= 1 a
2
= 1 a
1
= 1 a
0
= 1
b
4
x
0
= 5 b
3
x
0
b
2
x
0
b
1
x
0
b
4
= 1 b
3
= 4
Fundamentals
Horners Method
For P(x) = x
4
x
3
+ x
2
x
0
= 5 a
4
= 1 a
3
= 1 a
2
= 1 a
1
= 1 a
0
= 1
b
4
x
0
= 5 b
3
x
0
= 20 b
2
x
0
b
1
x
0
b
4
= 1 b
3
= 4 b
2
= 21
Fundamentals
Horners Method
For P(x) = x
4
x
3
+ x
2
x
0
= 5 a
4
= 1 a
3
= 1 a
2
= 1 a
1
= 1 a
0
= 1
b
4
x
0
= 5 b
3
x
0
= 20 b
2
x
0
= 105 b
1
x
0
b
4
= 1 b
3
= 4 b
2
= 21 b
1
= 106
Fundamentals
Horners Method
For P(x) = x
4
x
3
+ x
2
x
0
= 5 a
4
= 1 a
3
= 1 a
2
= 1 a
1
= 1 a
0
= 1
b
4
x
0
= 5 b
3
x
0
= 20 b
2
x
0
= 105 b
1
x
0
= 530
b
4
= 1 b
3
= 4 b
2
= 21 b
1
= 106 b
0
= 529
Hence, P(5) = 529, and
P(x) = (x 5)(x
3
+ 4x
2
+ 21x + 106) + 529
Similarly we get P
(5) = Q(5) = 436

x
0
= 5 a
3
= 1 a
2
= 4 a
1
= 21 a
0
= 106
b
3
x
0
= 5 b
2
x
0
= 45 b
1
x
0
= 330
b
3
= 1 b
2
= 9 b
1
= 66 b
0
= 436
Fundamentals
Horners Method
Algorithm: Horners Method
Algorithm: Horners Method
Input: Degree n; coecients a
0
, a
1
, . . . , a
n
; x
0
Output: y = P(x
0
), z = P
(x
0
).
1. Set y = a
n
, z = a
n
2. For j = (n 1), (n 2), . . . , 1
Set y = x
0
y + a
j
, z = x
0
z + y
3. Set y = x
0
y + a
0
4. Output (y, z)
5. End program
Deation Finding All the Zeros of a Polynomial
If we are solving our current favorite problem
P(x) = 0, P(x) a polynomial of degree n
and we are using Horners method of computing P(x
i
) and P
(x
i
),
then after N iterations, x
N
is an approximation to one of the roots
of P(x) = 0.
We have
P(x) = (x x
N
)Q(x) + b
0
, b
0
0
Let r
1
= x
N
be the rst root, and Q
1
(x) = Q(x).
We can now nd a second root by applying Newtons method to
Q
1
(x).
Deation Finding All the Zeros of a Polynomial
After some number of iterations of Newtons method we have
Q
1
(x) = (x r
2
)Q
2
(x) + b
(2)
0
, b
(2)
0
0
If P(x) is an n
th
-degree polynomial with n real roots, we can apply
this procedure (n 2) times to nd (n 2) approximate zeros of
P(x): r
1
, r
2
, . . . , r
n2
, and a quadratic factor Q
n2
(x).
At this point we can solve Q
n2
(x) = 0 using the quadratic
formula, and we have n roots of P(x) = 0.
This procedure is called Deation.
Quality of Deation
Now, the big question is are the approximate roots r
1
, r
2
, . . . ,
r
n
good approximations of the roots of P(x)???
Unfortunately, sometimes, no.
In each step we solve the equation to some tolerance, i.e.
|b
(k)
0
| < tol
Even though we may solve to a tight tolerance (10
8
), the errors
accumulate and the inaccuracies increase iteration-by-iteration...
Question: Is deation therefore useless???
Improving the Accuracy of Deation
The problem with deation is that the zeros of Q
k
(x) are not good
representatives of the zeros of P(x), especially for high ks.
As k increases, the quality of the root r
k
decreases.
Maybe there is a way to get all the zeros with the same quality?
Improving the Accuracy of Deation
The problem with deation is that the zeros of Q
k
(x) are not good
representatives of the zeros of P(x), especially for high ks.
As k increases, the quality of the root r
k
decreases.
Maybe there is a way to get all the zeros with the same quality?
The idea is quite simple... in each step of deation, instead of just
accepting r
k
as a root of P(x), we re-run Newtons method on the
full polynomial P(x), with r
k
as the starting point a couple of
Newton iterations should quickly converge to the root of the full
polynomial.
Improved Deation Algorithm Outline
Algorithm Outline: Improved Deation
1. Apply Newtons method to P(x) r
1
, Q
1
(x).
2. For k = 2, 3, . . . , (n 2) do 3--4
3. Apply Newtons method to Q
k1
r
k
, Q
k
(x).
4. Apply Newtons method to P(x) with r
k
as the initial point
r
k
Apply Horners method to Q
k1
(x) with x =r
k
Q
k
(x)
5. Use the quadratic formula on Q
n2
(x) to get the two remaining
roots.
Note: Inside Newtons method, the evaluations of polynomi-
als and their derivatives are also performed using Horners
method.
Deation & Improvement Wilkinson Polynomials
The Wilkinson Polynomials
P
W
n
(x) =
n
k=1
(x k)
have the roots {1, 2, . . . , n}, but provide surprisingly tough
numerical root-nding problems. (Additional details in Math 543 .)
In the next few slides we show the results of Deation and
Improved Deation applied to Wilkinson polynomials of degree 9,
10, 12, and 13.
Deation & Improvement P
W
9
(x) and P
W
10
(x)
0 2 4 6 8 10
10
16
10
14
10
12
10
10
10
8
Relative Error of the Computed Roots
Deflation
Improved
0 2 4 6 8 10
10
14
10
12
10
10
10
8
Deflation
Improved
Figure: [Left] The result of the two algorithms on the Wilkinson polynomial of degree
9; in this case the roots are computed so that |b
(k)
0
| < 10
6
. [Right] The result of
the two algorithms on the Wilkinson polynomial of degree 10; in this case the roots are
computed so that |b
(k)
0
| < 10
6
. In both cases the lower line corresponds to improved
deation and we see that we get an improvement in the relative error of several orders
of magnitude.
Deation & Improvement P
W
12
(x) and P
W
13
(x)
0 2 4 6 8 10 12
10
16
10
14
10
12
10
10
10
8
10
6
Deflation
Improved
0 2 4 6 8 10 12 14
10
16
10
14
10
12
10
10
10
8
10
6
10
4
Deflation
Improved
Figure: [Left] The result of the two algorithms on the Wilkinson polynomial of degree
12; in this case the roots are computed so that |b
(k)
0
| < 10
4
. [Right] The result of
the two algorithms on the Wilkinson polynomial of degree 13; in this case the roots are
computed so that |b
(k)
0
| < 10
3
. In both cases the lower line corresponds to improved
deation and we see that we get an improvement in the relative error of several orders
of magnitude.
What About Complex Roots???
One interesting / annoying feature of polynomials with real
coecients is that they may have complex roots, e.g.
P(x) = x
2
+ 1 has the roots {i , i }. Where by denition
i =
1.
If the initial approximation given to Newtons method is real, all
the successive iterates will be real... which means we will not nd
complex roots.
One way to overcome this is to start with a complex initial
approximation and do all the computations in complex arithmetic.
Another solution is M ullers Method...
M ullers Method
M ullers method is an extension of the Secant method...
Recall that the secant method uses two points x
k
and x
k1
and
the function values in those two points f (x
k
) and f (x
k1
). The
zero-crossing of the linear interpolant (the secant line) is used as
the next iterate x
k+1
.
M ullers method takes the next logical step: it uses three points:
x
k
, x
k1
and x
k2
, the function values in those points f (x
k
),
f (x
k1
) and f (x
k2
); a second degree polynomial tting these
three points is found, and its zero-crossing is the next iterate x
k+1
.
Next slide: f (x) = x
4
3x
3
1, x
k2
= 1.5, x
k1
= 2.5, x
k
= 3.5.
M ullers Method Illustration f (x) = x
4
3x
3
1
0 1 2 3 4
0
20
M ullers Method Fitting the Quadratic Polynomial
We consider the quadratic polynomial
m(x) = a(x x
k
)
2
+ b(x x
k
) + c
at the three tting points we get
f (x
k2
) = a(x
k2
x
k
)
2
+ b(x
k2
x
k
) + c
f (x
k1
) = a(x
k1
x
k
)
2
+ b(x
k1
x
k
) + c
f (x
k
) = c
We can solve for a, b, and c:
a =
(x
k1
x
k
)(f (x
k2
) f (x
k
)) (x
k2
x
k
)(f (x
k1
) f (x
k
))
(x
k2
x
k
)(x
k1
x
k
)(x
k2
x
k1
)
b =
(x
k2
x
k
)
2
(f (x
k1
) f (x
k
)) (x
k1
x
k
)
2
(f (x
k2
) f (x
k
))
(x
k2
x
k
)(x
k1
x
k
)(x
k2
x
k1
)
c = f (x
k
)
M ullers Method Identifying the Zero
We now have a quadratic equation for (x x
k
) which gives us two
possibilities for x
k+1
:
x
k+1
x
k
=
2c
b
b
2
4ac
In M ullers method we select
x
k+1
= x
k

2c
b + sign(b)
b
2
4ac
we are maximizing the (absolute) size of the denominator, hence
we select the root closest to x
k
.
Note that if b
2
4ac < 0 then we automatically get complex roots.
M ullers Method Algorithm
Algorithm: M ullers Method
Input: x
0
, x
1
, x
2
; tolerance tol ; max iterations N
0
Output: Approximate solution p, or failure message.
1. Set h
1
= (x
1
x
0
), h
2
= (x
2
x
1
),
1
= [f (x
1
) f (x
0
)]/h
1
,
2
= [f (x
2
) f (x
1
)]/h
2
, d = (
2

1
)/(h
2
+ h
1
), j = 3.
2. While j N
0
do 3--7
3. b =
2
+ h
2
d, D =
_
b
2
4f (x
2
)d complex?
4. If |b D| < |b + D| then set E = b + D else set E = b D
5. Set h = 2f (x
2
)/E, p = x
2
+ h
6. If |h| <tol then output p; stop program
7. Set x
0
= x
1
, x
1
= x
2
, x
2
= p, h
1
= (x
1
x
0
),
h
2
= (x
2
x
1
),
1
= [f (x
1
) f (x
0
)]/h
1
,
2
= [f (x
2
) f (x
1
)]/h
2
,
d = (
2

1
)/(h
2
+ h
1
), j = j + 1
8. output M ullers Method failed after N
0
iterations.
Now We Know Everything About Solving f (x) = 0 !?
Lets recap... Things to remember...
The relation between root nding (f (x) = 0) and xed point
(g(x) = x).
Key algorithms for root nding: Bisection, Secant Method, and
Newtons Method. Know what they are (the updates), how to
start (one or two points? bracketing or not bracketing the root?),
can the method break, can breakage be xed? Convergence
properties.
Also, know the mechanics of the Regula Falsi method, and
understand why it can run into trouble.
Fixed point iteration: Under what conditions do FP-iteration
converge for all starting values in the interval?
Recap, continued...
Basic error analysis: order , asymptotic error constant .
Which one has the most impact on convergence? Convergence
rate for general xed point iterations?
Multiplicity of zeros: What does it mean? How do we use this
knowledge to help Newtons method when were looking for a
zero of high multiplicity?
Convergence acceleration: Aitkens
2
-method. Steensens
Method.
Zeros of polynomials: Horners method, Deation (with
improvement), M ullers method.
Fundamentals
Nevilles Method
New Favorite Problem:
Interpolation and Polynomial Approximation
Fundamentals
Nevilles Method
Weierstrass Approximation Theorem
The following theorem is the basis for polynomial approximation:
Theorem (Weierstrass Approximation Theorem)
Suppose f C[a, b]. Then > 0 a polynomial P(x) :
|f (x) P(x)| < , x [a, b].
Note: The bound is uniform, i.e. valid for all x in the interval.
Note: The theorem says nothing about how to nd the polyno-
mial, or about its order.
Fundamentals
Nevilles Method
Illustrated: Weierstrass Approximation Theorem
0 2 4 6 8 10
1
0.5
0
0.5
1
1.5
2
2.5
f
f+
f
Figure: Weierstrass approximation Theorem guarantees that we (maybe with sub-
stantial work) can nd a polynomial which ts into the tube around the function
f , no matter how thin we make the tube.
Fundamentals
Nevilles Method
Candidates: the Taylor Polynomials???
Natural Question:
Are our old friends, the Taylor Polynomials, good candidates
for polynomial interpolation?
Answer:
No. The Taylor expansion works very hard to be accurate in
the neighborhood of one point. But we want to t data at
many points (in an extended interval).
[Next slide: The approximation is great near the expansion point
x
0
= 0, but get progressively worse at we get further away from the
point, even for the higher degree approximations.]
Fundamentals
Nevilles Method
Taylor Approximation of e
x
on the Interval [0, 3]
0 0.5 1 1.5 2 2.5 3
0
5
10
15
20
e^x
P_0(x)
P_1(x)
P_2(x)
P_3(x)
P_4(x)
P_5(x)
Fundamentals
Nevilles Method
Taylor Approximation of f (x) = 1/x about x = 1
Let
f (x) =
1
x
The Taylor expansion about x
0
= 1 is
P
n
(x) =
n
k=0
(1)
k
(x 1)
k
.
Note: f (3) = 1/3, but P
n
(3) satises:
n 0 1 2 3 4 5 6
P
n
(3) 1 -1 3 -5 11 -21 43
The Taylors series only converges |x 1| < 1 by the ratio test (a
geometric series). Thus, not valid for x = 3.
Fundamentals
Nevilles Method
Lookahead: Polynomial Approximation
Clearly, Taylor polynomials are not well suited for approximating a
function over an extended interval.
We are going to look at the following:
Lagrange polynomials Nevilles Method. [This Lecture]
Newtons divided dierences.
Hermite interpolation.
Cubic splines Piecewise polynomial approximation.
(Parametric curves)
(Bezier curves used in e.g. computer graphics)
Fundamentals
Nevilles Method
Interpolation: Lagrange Polynomials
Idea: Instead of working hard at one point, we will prescribe
a number of points through which the polynomial must pass.
As warm-up we will dene a function that passes through the
points (x
0
, f (x
0
)) and (x
1
, f (x
1
)). First, lets dene
L
0
(x) =
x x
1
x
0
x
1
, L
1
(x) =
x x
0
x
1
x
0
,
and then dene the interpolating polynomial
P(x) = L
0
(x)f (x
0
) + L
1
(x)f (x
1
),
then P(x
0
) = f (x
0
), and P(x
1
) = f (x
1
).
P(x) is the unique linear polynomial passing through
(x
0
, f (x
0
)) and (x
1
, f (x
1
)).
Fundamentals
Nevilles Method
An n-degree polynomial passing through n + 1 points
We are going to construct a polynomial passing through the points
(x
0
, f (x
0
)), (x
1
, f (x
1
)), (x
2
, f (x
2
)), . . . , (x
N
, f (x
n
)).
We dene L
n,k
(x), the Lagrange coecients:
L
n,k
(x) =
n
i=0, i=k
x x
i
x
k
x
i
=
x x
0
x
k
x
0

x x
k1
x
k
x
k1
x x
k+1
x
k
x
k+1

x x
n1
x
k
x
n
,
which have the properties
L
n,k
(x
k
) = 1; L
n,k
(x
i
) = 0, i = k.
Fundamentals
Nevilles Method
Example of L
n,k
(x)
0 1 2 3 4 5 6
-0.5
0
0.5
1
This is L
6,3
(x), for the points x
i
= i , i = 0, . . . , 6.
Fundamentals
Nevilles Method
The n
th
Lagrange Interpolating Polynomial
We use L
n,k
(x), k = 0, . . . , n as building blocks for the Lagrange
interpolating polynomial:
P(x) =
n
k=0
f (x
k
)L
n,k
(x),
which has the property
P(x
i
) = f (x
i
), i = 0, . . . , n.
This is the unique polynomial passing through the points
(x
i
, f (x
i
)), i = 0, . . . , n.
Fundamentals
Nevilles Method
Error bound for the Lagrange interpolating polynomial
Suppose x
i
, i = 0, . . . , n are distinct numbers in the interval [a, b],
and f C
n+1
[a, b]. Then x [a, b] (x) (a, b) so that:
f (x) = P
Lagrange
(x) +
f
(n+1)
((x))
(n + 1)!
n
i =0
(x x
i
),
where P
Lagrange
(x) is the n
th
Lagrange interpolating polynomial.
Compare with the error formula for Taylor polynomials
f (x) = P
Taylor
(x) +
f
(n+1)
((x))
(n + 1)!
(x x
0
)
n+1
,
Problem: Applying the error term may be dicult...
The error formula is important as Lagrange polynomials are used
for numerical dierentiation and integration methods.
Fundamentals
Nevilles Method
The Lagrange and Taylor Error Terms
Just to get a feeling for the non-constant part of the error terms in
the Lagrange and Taylor approximations, we plot those parts on
the interval [0, 4] with interpolation points x
i
= i , i = 0, 1, . . . , 4:
0 0.5 1 1.5 2 2.5 3 3.5 4
4
3
2
1
0
1
2
3
4
0 0.5 1 1.5 2 2.5 3 3.5 4
0
200
400
600
800
1000
1200
Figure: [Left] The non-constant error terms for the Lagrange interpolation oscillates in the interval [4, 4]
(and takes the value zero at the node point x
k
), and [Right] the non-constant error term for the Taylor
extrapolation grows in the interval [0, 1024].
Fundamentals
Nevilles Method
Practical Problems
Applying (estimating) the error term is dicult.
The degree of the polynomial needed for some desired accuracy is
not known until after cumbersome calculations checking the
error term.
If we want to increase the degree of the polynomial (to e.g.
n + 1) the previous calculations are not of any help...
Building block for a x: Let f be a function dened at x
0
, . . . , x
n
,
and suppose that m
1
, m
2
, . . . , m
k
are k (< n) distinct integers,
with 0 m
i
n i . The Lagrange polynomial that agrees with
f (x) the k points x
m
1
, x
m
2
, . . . , x
m
k
, is denoted P
m
1
,m
2
,...,m
k
(x).
Note: {m
1
, m
2
, . . . , m
k
} {0, 1, . . . , n}.
Fundamentals
Nevilles Method
Increasing the degree of the Lagrange Interpolation
Theorem
Let f be dened at x
0
, x
1
, . . . , x
k
, and x
i
and x
j
be two distinct
points in this set, then
P(x) =
(x x
j
)P
0,...,j 1,j +1,...,k
(x) (x x
i
)P
0,...,i 1,i +1,...,k
(x)
x
i
x
j
is the k
th
Lagrange polynomial that interpolates f at the k + 1
points x
0
, . . . , x
k
.
Fundamentals
Nevilles Method
Recursive Generation of Higher Degree Lagrange Interpolating Polynomials
x
0
P
0
x
1
P
1
P
0,1
x
2
P
2
P
1,2
P
0,1,2
x
3
P
3
P
2,3
P
1,2,3
P
0,1,2,3
x
4
P
4
P
3,4
P
2,3,4
P
1,2,3,4
P
0,1,2,3,4
Fundamentals
Nevilles Method
Nevilles Method
The notation in the previous table gets cumbersome... We
introduce the notation Q
Last,Degree
= P
LastDegree,. . . ,Last
, the table
becomes:
x
0
Q
0,0
x
1
Q
1,0
Q
1,1
x
2
Q
2,0
Q
2,1
Q
2,2
x
3
Q
3,0
Q
3,1
Q
3,2
Q
3,3
x
4
Q
4,0
Q
4,1
Q
4,2
Q
4,3
Q
4,4
Compare with the old notation:
x
0
P
0
x
1
P
1
P
0,1
x
2
P
2
P
1,2
P
0,1,2
x
3
P
3
P
2,3
P
1,2,3
P
0,1,2,3
x
4
P
4
P
3,4
P
2,3,4
P
1,2,3,4
P
0,1,2,3,4
Fundamentals
Nevilles Method
Algorithm: Nevilles Method Iterated Interpolation
Algorithm: Nevilles Method
To evaluate the polynomial that interpolates the n + 1 points
(x
i
, f (x
i
)), i = 0, . . . , n at the point x:
1. Initialize Q
i ,0
= f (x
i
).
2.
FOR i = 1 : n
FOR j = 1 : i
Q
i ,j
=
(x x
i j
)Q
i ,j 1
(x x
i
)Q
i 1,j 1
x
i
x
i j
END
END
3. Output the Q-table.

Horner

Загружено:

Сведения о документе

Исходное описание:

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Horner

Загружено:

Авторское право:

Доступные форматы

Accelerating Convergence

12 0.73 5604740436347 0.73908 3333909684

Note: Bold digits are correct; p

(x) as eciently as possible.

(5) = Q(5) = 436

Вам также может понравиться