2.4 error analysis for iterative methodszxu2/acms40390f12/lec-2.4.pdfย ยท newtonโ€™s method as fixed...

9
2.4 Error Analysis for Iterative Methods 1

Upload: others

Post on 03-Aug-2020

1 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: 2.4 Error Analysis for Iterative Methodszxu2/acms40390F12/Lec-2.4.pdfย ยท Newtonโ€™s Method as Fixed Point Problem Solve ๐‘ฅ= r by fixed point method. We write the problem as an equivalent

2.4 Error Analysis for Iterative Methods

1

Page 2: 2.4 Error Analysis for Iterative Methodszxu2/acms40390F12/Lec-2.4.pdfย ยท Newtonโ€™s Method as Fixed Point Problem Solve ๐‘ฅ= r by fixed point method. We write the problem as an equivalent

Definition โ€ข Order of Convergence Suppose {๐‘๐‘›}๐‘›=0

โˆž is a sequence that converges to ๐‘ with ๐‘๐‘› โ‰  ๐‘ for all ๐‘›. If positive constants ๐œ† and ๐›ผ exist with

lim๐‘›โ†’โˆž

|๐‘๐‘›+1 โˆ’ ๐‘|

|๐‘๐‘› โˆ’ ๐‘|๐›ผ= ๐œ†

then {๐‘๐‘›}๐‘›=0โˆž is said to converges to ๐’‘ of order ๐œถ with asymptotic

error constant ๐€. An iterative technique ๐‘๐‘› = ๐‘”(๐‘๐‘›โˆ’1) is said to be of order ๐œถ if the sequence {๐‘๐‘›}๐‘›=0

โˆž converges to the solution ๐‘ = ๐‘”(๐‘) of order ๐œถ. โ€ข Special cases

1. If ๐›ผ = 1 (and ๐œ† < 1), the sequence is linearly convergent 2. If ๐›ผ = 2, the sequence is quadratically convergent 3. If ๐›ผ < 1, the sequence is sub-linearly convergent (undesirable, very slow) 4. If ๐›ผ = 1 and ๐œ† = 0 or 1 < ๐›ผ < 2, the sequence is super-linearly convergent

โ€ข Remark: High order (๐›ผ) โŸน faster convergence (more desirable) ๐œ† is less important than the order (๐›ผ)

2

Page 3: 2.4 Error Analysis for Iterative Methodszxu2/acms40390F12/Lec-2.4.pdfย ยท Newtonโ€™s Method as Fixed Point Problem Solve ๐‘ฅ= r by fixed point method. We write the problem as an equivalent

Linear vs. Quadratic

3

Suppose we have two sequences converging to 0 with:

lim๐‘›โ†’โˆž

|๐‘๐‘›+1|

|๐‘๐‘›|= 0.9, lim

๐‘›โ†’โˆž

|๐‘ž๐‘›+1|

|๐‘ž๐‘›|2= 0.9

Roughly we have: ๐‘๐‘› โ‰ˆ 0.9 ๐‘๐‘›โˆ’1 โ‰ˆ โ‹ฏ โ‰ˆ 0.9๐‘› ๐‘0 ,

๐‘ž๐‘› โ‰ˆ 0.9|๐‘ž๐‘›โˆ’1|2 โ‰ˆ โ‹ฏ โ‰ˆ 0.92๐‘›โˆ’1 ๐‘ž0 , Assume ๐‘0 = ๐‘ž0 = 1

Page 4: 2.4 Error Analysis for Iterative Methodszxu2/acms40390F12/Lec-2.4.pdfย ยท Newtonโ€™s Method as Fixed Point Problem Solve ๐‘ฅ= r by fixed point method. We write the problem as an equivalent

Fixed Point Convergence โ€ข Theorem

Let ๐‘” โˆˆ ๐ถ[๐‘Ž, ๐‘] be such that ๐‘” ๐‘ฅ โˆˆ [๐‘Ž, ๐‘] for all ๐‘ฅ โˆˆ ๐‘Ž, ๐‘ . Suppose ๐‘”โ€ฒ is continuous on (๐‘Ž, ๐‘) and that 0 < ๐‘˜ < 1 exists with |๐‘”โ€ฒ(๐‘ฅ)| โ‰ค ๐‘˜ for all ๐‘ฅ โˆˆ ๐‘Ž, ๐‘ .

If ๐‘”โ€ฒ(๐‘) โ‰  0, then for all number ๐‘0 in [๐‘Ž, ๐‘], the sequence ๐‘๐‘› = ๐‘”(๐‘๐‘›โˆ’1) converges only linearly to the unique fixed point ๐‘ in ๐‘Ž, ๐‘ .

โ€ข Proof: ๐‘๐‘›+1 โˆ’ ๐‘ = ๐‘” ๐‘๐‘› โˆ’ ๐‘” ๐‘ = ๐‘”โ€ฒ ๐œ‰๐‘› ๐‘๐‘› โˆ’ ๐‘ , ๐œ‰๐‘› โˆˆ (๐‘๐‘›, ๐‘)

Since {๐‘๐‘›}๐‘›=0โˆž converges to ๐‘, {๐œ‰๐‘›}๐‘›=0

โˆž converges to ๐‘.

Since ๐‘”โ€ฒ is continuous, lim๐‘›โ†’โˆž

๐‘”โ€ฒ(๐œ‰๐‘›) = ๐‘”โ€ฒ(๐‘)

lim๐‘›โ†’โˆž

|๐‘๐‘›+1โˆ’๐‘|

|๐‘๐‘›โˆ’๐‘|= lim

๐‘›โ†’โˆž๐‘”โ€ฒ ๐œ‰๐‘› = |๐‘”โ€ฒ(๐‘)| โŸน linear convergence

4

Page 5: 2.4 Error Analysis for Iterative Methodszxu2/acms40390F12/Lec-2.4.pdfย ยท Newtonโ€™s Method as Fixed Point Problem Solve ๐‘ฅ= r by fixed point method. We write the problem as an equivalent

Speed up Convergence of Fixed Point Iteration

โ€ข If we look for faster convergence methods, we must have ๐‘”โ€ฒ ๐‘ = 0

โ€ข Theorem Let ๐‘ be a solution of ๐‘ฅ = ๐‘” ๐‘ฅ . Suppose ๐‘”โ€ฒ ๐‘ = 0 and ๐‘”โ€ฒโ€ฒ is continuous with ๐‘”โ€ฒโ€ฒ ๐‘ฅ < ๐‘€ on an open interval ๐ผ containing ๐‘. Then there exists a ๐›ฟ > 0 such that for ๐‘0 โˆˆ ๐‘ โˆ’ ๐›ฟ, ๐‘ + ๐›ฟ , the sequence defined by ๐‘๐‘›+1 =๐‘” ๐‘๐‘› , when ๐‘› โ‰ฅ 0, converges at least quadratically to ๐‘. For sufficiently large ๐‘›

๐‘๐‘›+1 โˆ’ ๐‘ <๐‘€

2|๐‘๐‘› โˆ’ ๐‘|2

Remark:

Look for quadratically convergent fixed point methods which ๐‘” ๐‘ = ๐‘ and ๐‘”โ€ฒ ๐‘ = 0.

5

Page 6: 2.4 Error Analysis for Iterative Methodszxu2/acms40390F12/Lec-2.4.pdfย ยท Newtonโ€™s Method as Fixed Point Problem Solve ๐‘ฅ= r by fixed point method. We write the problem as an equivalent

Newtonโ€™s Method as Fixed Point Problem

Solve ๐‘“ ๐‘ฅ = 0 by fixed point method. We write the problem as an equivalent fixed point problem:

๐‘” ๐‘ฅ = ๐‘ฅ โˆ’ ๐‘“ ๐‘ฅ solve: ๐‘ฅ = ๐‘”(๐‘ฅ) ๐‘” ๐‘ฅ = ๐‘ฅ โˆ’ ๐›ผ๐‘“ ๐‘ฅ solve ๐‘ฅ = ๐‘” ๐‘ฅ , ๐›ผ is a constant

๐‘” ๐‘ฅ = ๐‘ฅ โˆ’ ๐œ™ ๐‘ฅ ๐‘“ ๐‘ฅ solve ๐‘ฅ = ๐‘” ๐‘ฅ , ๐œ™ ๐‘ฅ is differentiable

Newtonโ€™s method is derived by the last form:

Find differentiable ๐œ™ ๐‘ฅ with ๐‘”โ€ฒ ๐‘ = 0 when ๐‘“ ๐‘ = 0.

๐‘”โ€ฒ ๐‘ฅ =๐‘‘

๐‘‘๐‘ฅ๐‘ฅ โˆ’ ๐œ™ ๐‘ฅ ๐‘“ ๐‘ฅ = 1 โˆ’ ๐œ™โ€ฒ๐‘“ โˆ’ ๐œ™๐‘“โ€ฒ

Use ๐‘”โ€ฒ ๐‘ = 0 when ๐‘“ ๐‘ = 0 ๐‘”โ€ฒ ๐‘ = 1 โˆ’ ๐œ™โ€ฒ ๐‘ โˆ™ 0 โˆ’ ๐œ™ ๐‘ ๐‘“โ€ฒ ๐‘ = 0

๐œ™ ๐‘ = 1/๐‘“โ€ฒ(๐‘)

This gives Newtonโ€™s method

๐‘๐‘›+1 = ๐‘” ๐‘๐‘› = ๐‘๐‘› โˆ’๐‘“(๐‘๐‘›)

๐‘“โ€ฒ(๐‘๐‘›)

6

Page 7: 2.4 Error Analysis for Iterative Methodszxu2/acms40390F12/Lec-2.4.pdfย ยท Newtonโ€™s Method as Fixed Point Problem Solve ๐‘ฅ= r by fixed point method. We write the problem as an equivalent

Multiple Roots

โ€ข How to modify Newtonโ€™s method when ๐‘“โ€ฒ ๐‘ = 0. Here ๐‘ is the root of ๐‘“ ๐‘ฅ = 0.

โ€ข Definition: Multiplicity of a Root A solution ๐‘ of ๐‘“ ๐‘ฅ = 0 is a zero of multiplicity ๐‘š of ๐‘“ if for ๐‘ฅ โ‰  ๐‘, we can write ๐‘“ ๐‘ฅ = ๐‘ฅ โˆ’ ๐‘ ๐‘š๐‘ž ๐‘ฅ , where lim๐‘ฅโ†’๐‘ ๐‘ž(๐‘ฅ) โ‰  0.

โ€ข Theorem ๐‘“ โˆˆ ๐ถ1[๐‘Ž, ๐‘] has a simple zero at ๐‘ in (๐‘Ž, ๐‘) if and only if ๐‘“ ๐‘ = 0, but ๐‘“โ€ฒ(๐‘) โ‰  0.

โ€ข Theorem The function ๐‘“ โˆˆ ๐ถ๐‘š[๐‘Ž, ๐‘] has a zero of multiplicity ๐‘š at point ๐‘ in (๐‘Ž, ๐‘) if and only if 0 = ๐‘“ ๐‘ = ๐‘“โ€ฒ ๐‘ = ๐‘“โ€ฒโ€ฒ ๐‘ = โ‹ฏ = ๐‘“ ๐‘šโˆ’1 ๐‘ , but ๐‘“ ๐‘š (๐‘) โ‰  0

7

Page 8: 2.4 Error Analysis for Iterative Methodszxu2/acms40390F12/Lec-2.4.pdfย ยท Newtonโ€™s Method as Fixed Point Problem Solve ๐‘ฅ= r by fixed point method. We write the problem as an equivalent

Newtonโ€™s Method for Zeroes of Higher Multiplicity (๐‘š > 1)

Define the new function ๐œ‡ ๐‘ฅ =๐‘“(๐‘ฅ)

๐‘“โ€ฒ(๐‘ฅ).

Write ๐‘“ ๐‘ฅ = ๐‘ฅ โˆ’ ๐‘ ๐‘š๐‘ž(๐‘ฅ), hence

๐œ‡ ๐‘ฅ =๐‘“(๐‘ฅ)

๐‘“โ€ฒ(๐‘ฅ)= ๐‘ฅ โˆ’ ๐‘

๐‘ž ๐‘ฅ

๐‘š๐‘ž ๐‘ฅ + ๐‘ฅ โˆ’ ๐‘ ๐‘žโ€ฒ ๐‘ฅ

Note that ๐‘ is a simple zero of ๐œ‡ ๐‘ฅ .

โ€ข Apply Newtonโ€™s method to ๐œ‡ ๐‘ฅ to give:

๐‘ฅ = ๐‘” ๐‘ฅ = ๐‘ฅ โˆ’๐œ‡ ๐‘ฅ

๐œ‡โ€ฒ ๐‘ฅ

= ๐‘ฅ โˆ’๐‘“ ๐‘ฅ ๐‘“โ€ฒ ๐‘ฅ

๐‘“โ€ฒ ๐‘ฅ 2 โˆ’ ๐‘“ ๐‘ฅ ๐‘“โ€ฒโ€ฒ ๐‘ฅ

โ€ข Quadratic convergence: ๐‘๐‘›+1 = ๐‘๐‘› โˆ’๐‘“ ๐‘๐‘› ๐‘“โ€ฒ ๐‘๐‘›

๐‘“โ€ฒ ๐‘๐‘›2โˆ’๐‘“ ๐‘๐‘› ๐‘“โ€ฒโ€ฒ ๐‘๐‘›

8

Page 9: 2.4 Error Analysis for Iterative Methodszxu2/acms40390F12/Lec-2.4.pdfย ยท Newtonโ€™s Method as Fixed Point Problem Solve ๐‘ฅ= r by fixed point method. We write the problem as an equivalent

Drawbacks:

โ€ข Compute ๐‘“โ€ฒโ€ฒ(๐‘ฅ) is expensive

โ€ข Iteration formula is more complicated โ€“ more expensive to compute

โ€ข Roundoff errors in denominator โ€“ both ๐‘“โ€ฒ(๐‘ฅ) and ๐‘“(๐‘ฅ) approach zero.

9