師大
Tsung-Ming Huang
Department of Mathematics National Taiwan Normal University, Taiwan
August 28, 2011
師大
Outline
1 Bisection Method
2 Fixed-Point Iteration
3 Newton’s method
4 Error analysis for iterative methods
5 Accelerating convergence
6 Zeros of polynomials and M ¨uller’s method
師大
Bisection Method
Idea
Iff (x) ∈ C[a, b]andf (a)f (b) < 0, then∃ c ∈ (a, b)such that f (c) = 0.
師大
Bisection method algorithm
Given f (x) defined on (a, b), the maximal number of iterations M, and stop criteria δ and ε, this algorithm tries to locate one root of f (x).
Compute u = f (a), v = f (b), and e = b − a If sign(u) = sign(v), then stop
For k = 1, 2, . . . , M
e = e/2, c = a + e, w = f (c) If |e| < δ or |w| < ε, then stop If sign(w) 6= sign(u)
b = c, v = w Else
a = c, u = w End If
End For
師大
Let {cn} be the sequence of numbers produced. The algorithm should stop if one of the following conditions is satisfied.
1 the iteration number k > M ,
2 |ck− ck−1| < δ, or
3 |f (ck)| < ε.
Let [a0, b0], [a1, b1], . . .denote the successive intervals produced by the bisection algorithm. Then
a = a0≤ a1 ≤ a2 ≤ · · · ≤ b0= b
⇒ {an} and {bn} are bounded
⇒ lim
n→∞an and lim
n→∞bn exist
師大
Since
b1− a1 = 1
2(b0− a0) b2− a2 = 1
2(b1− a1) = 1
4(b0− a0) ...
bn− an = 1
2n(b0− a0) hence
n→∞lim bn− lim
n→∞an= lim
n→∞(bn− an) = lim
n→∞
1
2n(b0− a0) = 0.
Therefore
n→∞lim an= lim
n→∞bn≡ z.
Since f is a continuous function, we have that
n→∞lim f (an) = f ( lim
n→∞an) = f (z) and lim
n→∞f (bn) = f ( lim
n→∞bn) = f (z).
師大
On the other hand,
f (an)f (bn) ≤ 0
⇒ lim
n→∞f (an)f (bn) = f2(z) ≤ 0
⇒ f (z) = 0
Therefore, the limit of the sequences {an} and {bn} is a zero of f in [a, b]. Let cn= 12(an+ bn). Then
|z − cn| = lim
n→∞an−1
2(an+ bn)
= 1 2
h
n→∞lim an− bni + 1
2 h
n→∞lim an− ani
≤ max lim
n→∞an− bn ,
lim
n→∞an− an
≤ |bn− an| = 1
2n|b0− a0|.
This proves the following theorem.
師大
Theorem 1
Let{[an, bn]}denote the intervals produced by the bisection algorithm. Then lim
n→∞anand lim
n→∞bnexist, areequal, and represent azerooff (x). If
z = lim
n→∞an= lim
n→∞bn and cn= 1
2(an+ bn), then
|z − cn| ≤ 1
2n(b0− a0) . Remark
{cn}convergestozwith therateofO(2−n).
師大
Example 2
How many steps should be taken to compute a root of f (x) = x3+ 4x2− 10 = 0 on [1, 2] with relative error 10−3? solution: Seek an n such that
|z − cn|
|z| ≤ 10−3 ⇒ |z − cn| ≤ |z| × 10−3. Since z ∈ [1, 2], it is sufficient to show
|z − cn| ≤ 10−3. That is, we solve
2−n(2 − 1) ≤ 10−3 ⇒ −n log102 ≤ −3 which gives n ≥ 10.
師大
Fixed-Point Iteration
Definition 3
xis called afixed pointof a given functionf iff (x) = x.
Root-finding problems and fixed-point problems Find x∗ such that f (x∗) = 0.
Let g(x) = x − f (x). Then g(x∗) = x∗− f (x∗) = x∗.
⇒ x∗is a fixed point for g(x).
Find x∗ such that g(x∗) = x∗. Define f (x) = x − g(x) so that f (x∗) = x∗− g(x∗) = x∗− x∗= 0
⇒ x∗is a zero of f (x).
師大
Example 4
The function g(x) = x2− 2, for −2 ≤ x ≤ 3, has fixed points at x = −1and x = 2 since
g(−1) = (−1)2− 2 = −1 and g(2) = 22− 2 = 2.
師大
Theorem 5 (Existence and uniqueness)
1 If g ∈ C[a, b] such thata ≤ g(x) ≤ bfor all x ∈ [a, b], then g hasa fixed point in [a, b].
2 If, in addition, g0(x)exists in (a, b) and there exists a positive constantM < 1such that|g0(x)| ≤ M < 1for all x ∈ (a, b). Then the fixed point isunique.
師大
Proof
Existence:
If g(a) = a or g(b) = b, then a or b is a fixed point of g and we are done.
Otherwise, it must be g(a) > a and g(b) < b. The function h(x) = g(x) − xis continuous on [a, b], with
h(a) = g(a) − a > 0 and h(b) = g(b) − b < 0.
By the Intermediate Value Theorem, ∃ x∗ ∈ [a, b] such that h(x∗) = 0. That is
g(x∗) − x∗ = 0 ⇒ g(x∗) = x∗. Hence g has a fixed point x∗in [a, b].
師大
Proof
Uniqueness:
Suppose that p 6= q are both fixed points of g in [a, b]. By the Mean-Value theorem, there exists ξ between p and q such that
g0(ξ) = g(p) − g(q)
p − q = p − q p − q = 1.
However, this contradicts to the assumption that
|g0(x)| ≤ M < 1for all x in [a, b]. Therefore the fixed point of g is unique.
師大
Example 6
Show that the following function has a unique fixed point.
g(x) = (x2− 1)/3, x ∈ [−1, 1].
Solution: The Extreme Value Theorem implies that min
x∈[−1,1]g(x) = g(0) = −1 3, max
x∈[−1,1]g(x) = g(±1) = 0.
That is g(x) ∈ [−1, 1], ∀ x ∈ [−1, 1].
Moreover, g is continuous and
|g0(x)| =
2x 3
≤ 2
3, ∀ x ∈ (−1, 1).
By above theorem, g has a unique fixed point in [−1, 1].
師大
Let p be such unique fixed point of g. Then p = g(p) = p2− 1
3 ⇒ p2− 3p − 1 = 0
⇒ p = 1 2(3 −√
13).
師大
Fixed-point iteration or functional iteration
Given a continuous function g, choose an initial point x0 and generate {xk}∞k=0by
xk+1 = g(xk), k ≥ 0.
{xk} may not converge, e.g., g(x) = 3x. However, when the sequence converges, say,
k→∞lim xk= x∗, then, since g is continuous,
g(x∗) = g( lim
k→∞xk) = lim
k→∞g(xk) = lim
k→∞xk+1= x∗. That is, x∗is a fixed point of g.
師大
Fixed-point iteration
Given x0, tolerance T OL, maximum number of iteration M . Set i = 1 and x = g(x0).
While i ≤ M and |x − x0| ≥ T OL Set i = i + 1, x0 = xand x = g(x0).
End While
師大
Example 7 The equation
x3+ 4x2− 10 = 0
has a unique root in [1, 2]. Change the equation to the fixed-point form x = g(x).
(a) x = g1(x) ≡ x − f (x) = x − x3− 4x2+ 10
(b) x = g2(x) = 10x − 4x1/2
x3 = 10 − 4x2 ⇒ x2 = 10
x − 4x ⇒ x = ± 10 x − 4x
1/2
師大
(c) x = g3(x) = 12 10 − x31/2
4x2 = 10 − x3 ⇒ x = ±1
2 10 − x31/2
(d) x = g4(x) =
10 4+x
1/2
x2(x + 4) = 10 ⇒ x = ±
10 4 + x
1/2
(e) x = g5(x) = x −x33x+4x2+8x2−10
x = g5(x) ≡ x − f (x) f0(x)
師大
Results of the fixed-point iteration with initial point x0 = 1.5
師大
Theorem 8 (Fixed-point Theorem)
Let g ∈ [a, b] be such thatg(x) ∈ [a, b]for all x ∈ [a, b]. Suppose that g0exists on (a, b) and that ∃ k with0 < k < 1such that
|g0(x)| ≤ k, ∀ x ∈ (a, b).
Then, for any number x0 in [a, b],
xn= g(xn−1), n ≥ 1, converges to the unique fixed point x in [a, b].
師大
Proof: By the assumptions, a unique fixed point exists in [a, b].
Since g([a, b]) ⊆ [a, b], {xn}∞n=0 is defined and xn∈ [a, b] for all n ≥ 0. Using the Mean Values Theorem and the fact that
|g0(x)| ≤ k, we have
|x − xn| = |g(xn−1) − g(x)| = |g0(ξn)||x − xn−1| ≤ k|x − xn−1|, where ξn∈ (a, b). It follows that
|xn− x| ≤ k|xn−1− x| ≤ k2|xn−2− x| ≤ · · · ≤ kn|x0− x|. (1) Since 0 < k < 1, we have
n→∞lim kn= 0 and
n→∞lim |xn− x| ≤ lim
n→∞kn|x0− x| = 0.
Hence, {xn}∞n=0converges to x.
師大
Corollary 9
If g satisfies the hypotheses of above theorem, then
|x − xn| ≤ knmax{x0− a, b − x0} and
|xn− x| ≤ kn
1 − k|x1− x0|, ∀ n ≥ 1.
Proof: From (1),
|xn− x| ≤ kn|x0− x| ≤ knmax{x0− a, b − x0}.
For n ≥ 1, using the Mean Values Theorem,
|xn+1− xn| = |g(xn) − g(xn−1)| ≤ k|xn− xn−1| ≤ · · · ≤ kn|x1− x0|.
師大
Thus, for m > n ≥ 1,
|xm− xn| = |xm− xm−1+ xm−1− · · · + xn+1− xn|
≤ |xm− xm−1| + |xm−1− xm−2| + · · · + |xn+1− xn|
≤ km−1|x1− x0| + km−2|x1− x0| + · · · + kn|x1− x0|
= kn|x1− x0| 1 + k + k2+ · · · + km−n−1 . It implies that
|x − xn| = lim
m→∞|xm− xn| ≤ lim
m→∞kn|x1− x0|
m−n−1
X
j=0
kj
≤ kn|x1− x0|
∞
X
j=0
kj = kn
1 − k|x1− x0|.
師大
Example 10
For previous example,
f (x) = x3+ 4x2− 10 = 0.
For g1(x) = x − x3− 4x2+ 10, we have g1(1) = 6 and g1(2) = −12, so g1([1, 2]) * [1, 2]. Moreover,
g10(x) = 1 − 3x2− 8x ⇒ |g10(x)| ≥ 1 ∀ x ∈ [1, 2]
• DOES NOT guarantee to converge or not
師大
For g3(x) = 12(10 − x3)1/2, ∀ x ∈ [1, 1.5], g30(x) = −3
4x2(10 − x3)−1/2< 0, ∀ x ∈ [1, 1.5], so g3 is strictly decreasing on [1, 1.5] and
1 < 1.28 ≈ g3(1.5) ≤ g3(x) ≤ g3(1) = 1.5, ∀ x ∈ [1, 1.5].
On the other hand,
|g03(x)| ≤ |g03(1.5)| ≈ 0.66, ∀ x ∈ [1, 1.5].
Hence, the sequence is convergent to the fixed point.
師大
For g4(x) =p10/(4 + x), we have r10
6 ≤ g4(x) ≤ r10
5 , ∀ x ∈ [1, 2] ⇒ g4([1, 2]) ⊆ [1, 2]
Moreover,
|g04(x)| =
√ −5
10(4 + x)3/2
≤ 5
√10(5)3/2 < 0.15, ∀ x ∈ [1, 2].
The bound of |g04(x)|is much smaller than the bound of |g30(x)|, which explains the more rapid convergence using g4.
師大
Suppose that f : R → R and f ∈ C2[a, b], i.e., f00exists and is continuous. If f (x∗) = 0and x∗= x + hwhere h is small, then byTaylor’stheorem
0 = f (x∗) = f (x + h)
= f (x) + f0(x)h + 1
2f00(x)h2+ 1
3!f000(x)h3+ · · ·
= f (x) + f0(x)h + O(h2).
Sincehissmall, O(h2)is negligible. It is reasonable to drop O(h2)terms. This implies
f (x) + f0(x)h ≈ 0 and h ≈ −f (x)
f0(x), if f0(x) 6= 0.
Hence
x + h = x − f (x) f0(x) is a better approximation to x∗.
師大
This sets the stage for theNewton-Rapbson’smethod, which starts with an initial approximation x0 and generates the sequence {xn}∞n=0defined by
xn+1 = xn− f (xn) f0(xn).
Since the Taylor’s expansion of f (x) at xkis given by f (x) = f (xk) + f0(xk)(x − xk) +1
2f00(xk)(x − xk)2+ · · · . At xk, one uses thetangent line
y = `(x) = f (xk) + f0(xk)(x − xk)
toapproximate the curveof f (x) and uses the zero of the tangent line to approximate the zero of f (x).
師大
Newton’s Method
Given x0, tolerance T OL, maximum number of iteration M . Set i = 1 and x = x0− f (x0)/f0(x0).
While i ≤ M and |x − x0| ≥ T OL
Set i = i + 1, x0 = xand x = x0− f (x0)/f0(x0).
End While
師大
Three stopping-technique inequalities
(a). |xn− xn−1| < ε, (b). |xn− xn−1|
|xn| < ε, xn6= 0, (c). |f (xn)| < ε.
Note that Newton’s method for solving f (x) = 0 xn+1= xn− f (xn)
f0(xn), for n ≥ 1 is just a special case of functional iteration in which
g(x) = x − f (x) f0(x).
師大
Example 11
The following table shows the convergence behavior of
Newton’s method applied to solving f (x) = x2− 1 = 0. Observe the quadratic convergence rate.
n xn |en| ≡ |1 − xn|
0 2.0 1
1 1.25 0.25
2 1.025 2.5e-2
3 1.0003048780488 3.048780488e-4 4 1.0000000464611 4.64611e-8
5 1.0 0
師大
Theorem 12
Assumef (x∗) = 0, f0(x∗) 6= 0andf (x),f0(x)andf00(x)are continuouson Nε(x∗). Then if x0is chosensufficiently closeto x∗, then
xn+1 = xn− f (xn) f0(xn)
→ x∗.
Proof: Define
g(x) = x − f (x) f0(x). Find an interval [x∗− δ, x∗+ δ]such that
g([x∗− δ, x∗+ δ]) ⊆ [x∗− δ, x∗+ δ]
and
|g0(x)| ≤ k < 1, ∀ x ∈ (x∗− δ, x∗+ δ).
師大
Since f0 is continuous and f0(x∗) 6= 0, it implies that ∃ δ1 > 0 such that f0(x) 6= 0 ∀ x ∈ [x∗− δ1, x∗+ δ1] ⊆ [a, b]. Thus, g is defined and continuous on [x∗− δ1, x∗+ δ1]. Also
g0(x) = 1 − f0(x)f0(x) − f (x)f00(x)
[f0(x)]2 = f (x)f00(x) [f0(x)]2 , for x ∈ [x∗− δ1, x∗+ δ1]. Since f00 is continuous on [a, b], we have g0 is continuous on [x∗− δ1, x∗+ δ1].
By assumption f (x∗) = 0, so
g0(x∗) = f (x∗)f00(x∗)
|f0(x∗)|2 = 0.
Since g0 is continuous on [x∗− δ1, x∗+ δ1]and g0(x∗) = 0, ∃ δ with 0 < δ < δ1 and k ∈ (0, 1) such that
|g0(x)| ≤ k, ∀ x ∈ [x∗− δ, x∗+ δ].
師大
Claim: g([x∗− δ, x∗+ δ]) ⊆ [x∗− δ, x∗+ δ].
If x ∈ [x∗− δ, x∗+ δ], then, by the Mean Value Theorem, ∃ ξ between x and x∗ such that
|g(x) − g(x∗)| = |g0(ξ)||x − x∗|.
It implies that
|g(x) − x∗| = |g(x) − g(x∗)| = |g0(ξ)||x − x∗|
≤ k|x − x∗| < |x − x∗| < δ.
Hence, g([x∗− δ, x∗+ δ]) ⊆ [x∗− δ, x∗+ δ].
By the Fixed-Point Theorem, the sequence {xn}∞n=0defined by xn= g(xn−1) = xn−1− f (xn−1)
f0(xn−1), for n ≥ 1, converges to x∗ for any x0 ∈ [x∗− δ, x∗+ δ].
師大
Example 13
When Newton’s method applied to f (x) = cos x with starting point x0 = 3, which is close to the root π2 of f , it produces x1 = −4.01525, x2= −4.8526, · · · ,which converges to another root −3π2 .
5 4 3 2 1 0 1 2 3 4 5
1. 5 0 1.5
x0 y = c os ( x)
x*
師大
Secant method
Disadvantage of Newton’s method
In many applications, the derivative f0(x)is very expensive to compute, or the function f (x) is not given in an algebraic formula so that f0(x)is not available.
By definition,
f0(xn−1) = lim
x→xn−1
f (x) − f (xn−1) x − xn−1 . Letting x = xn−2, we have
f0(xn−1) ≈ f (xn−2) − f (xn−1)
xn−2− xn−1 = f (xn−1) − f (xn−2) xn−1− xn−2 . Using this approximation for f0(xn−1)in Newton’s formula gives
xn= xn−1−f (xn−1)(xn−1− xn−2) f (xn−1) − f (xn−2) ,
38 / 68
師大
From geometric point of view, we use asecant linethrough xn−1andxn−2 instead of the tangent line to approximate the function at the point xn−1.
The slope of the secant line is
sn−1 = f (xn−1) − f (xn−2) xn−1− xn−2 and the equation is
M (x) = f (xn−1) + sn−1(x − xn−1).
The zero of the secant line x = xn−1−f (xn−1)
sn−1 = xn−1− f (xn−1) xn−1− xn−2 f (xn−1) − f (xn−2) is then used as a new approximate xn.
師大
Secant Method
Given x0, x1, tolerance T OL, maximum number of iteration M . Set i = 2; y0 = f (x0); y1= f (x1);
x = x1− y1(x1− x0)/(y1− y0).
While i ≤ M and |x − x1| ≥ T OL
Set i = i + 1; x0 = x1; y0 = y1; x1= x; y1 = f (x);
x = x1− y1(x1− x0)/(y1− y0).
End While
師大
Method of False Position
1 Choose initial approximations x0 and x1 with f (x0)f (x1) < 0.
2 x2= x1− f (x1)(x1− x0)/(f (x1) − f (x0))
3 Decide which secant line to use to compute x3: If f (x2)f (x1) < 0, then x1and x2 bracket a root, i.e.,
x3= x2− f (x2)(x2− x1)/(f (x2) − f (x1)) Else, x0 and x2 bracket a root, i.e.,
x3= x2− f (x2)(x2− x0)/(f (x2) − f (x0)) End if
師大
Method of False Position
Given x0, x1, tolerance T OL, maximum number of iteration M . Set i = 2; y0 = f (x0); y1= f (x1); x = x1− y1(x1− x0)/(y1− y0).
While i ≤ M and |x − x1| ≥ T OL Set i = i + 1; y = f (x).
If y · y1< 0, then set x0= x1; y0 = y1.
Set x1 = x; y1 = y; x = x1− y1(x1− x0)/(y1− y0).
End While
師大
Error analysis for iterative methods
Definition 14
Let{xn} → x∗. If there are positive constantscandαsuch that
n→∞lim
|xn+1− x∗|
|xn− x∗|α = c,
then we say therate of convergenceis oforderα.
We say that the rate of convergence is
1 linearifα = 1and 0 < c < 1.
2 superlinearif
n→∞lim
|xn+1− x∗|
|xn− x∗| = 0;
3 quadraticifα = 2. 43 / 68
師大
Suppose that {xn}∞n=0and {˜xn}∞n=0are linearly and quadratically convergent to x∗, respectively, with the same constant c = 0.5. For simplicity, suppose that
|xn+1− x∗|
|xn− x∗| ≈ c and |˜xn+1− x∗|
|˜xn− x∗|2 ≈ c.
These imply that
|xn− x∗| ≈ c|xn−1− x∗| ≈ c2|xn−2− x∗| ≈ · · · ≈ cn|x0− x∗|, and
|˜xn− x∗| ≈ c|˜xn−1− x∗|2 ≈ cc|˜xn−2− x∗|22
= c3|˜xn−2− x∗|4
≈ c3c|˜xn−3− x∗|24
= c7|˜xn−3− x∗|8
≈ · · · ≈ c2n−1|˜x0− x∗|2n.
師大
Remark
Quadratically convergent sequences generally converge much more quickly thank those that converge only linearly.
Theorem 15
Let g ∈ C[a, b] with g([a, b]) ⊆ [a, b]. Suppose that g0 is continuous on (a, b) and ∃ k ∈ (0, 1) such that
|g0(x)| ≤ k, ∀ x ∈ (a, b).
Ifg0(x∗) 6= 0, then for any x0∈ [a, b], the sequence xn= g(xn−1), for n ≥ 1
converges onlylinearlyto the unique fixed point x∗ in [a, b].
師大
Proof:
By the Fixed-Point Theorem, the sequence {xn}∞n=0 converges to x∗.
Since g0 exists on (a, b), by the Mean Value Theorem, ∃ ξn between xnand x∗ such that
xn+1− x∗ = g(xn) − g(x∗) = g0(ξn)(xn− x∗).
∵ {xn}∞n=0→ x∗ ⇒ {ξn}∞n=0→ x∗ Since g0 is continuous on (a, b), we have
n→∞lim g0(ξn) = g0(x∗).
Thus,
n→∞lim
|xn+1− x∗|
|xn− x∗| = lim
n→∞|g0(ξn)| = |g0(x∗)|.
Hence, if g0(x∗) 6= 0, fixed-point iteration exhibits linear convergence.
師大
Theorem 16
Let x∗ be a fixed point of g and I be an open interval with x∗ ∈ I. Suppose thatg0(x∗) = 0and g00is continuous with
|g00(x)| < M, ∀ x ∈ I.
Then ∃ δ > 0 such that
{xn= g(xn−1)}∞n=1 → x∗ for x0 ∈ [x∗− δ, x∗+ δ]
at leastquadratically. Moreover,
|xn+1− x∗| < M
2 |xn− x∗|2, for sufficiently large n.
師大
Proof:
Since g0(x∗) = 0and g0 is continuous on I, ∃ δ such that [x∗− δ, x∗+ δ] ⊂ Iand
|g0(x)| ≤ k < 1, ∀ x ∈ [x∗− δ, x∗+ δ].
In the proof of the convergence for Newton’s method, we have
{xn}∞n=0⊂ [x∗− δ, x∗+ δ].
Consider the Taylor expansion of g(xn)at x∗ xn+1= g(xn) = g(x∗) + g0(x∗)(xn− x∗) +g00(ξ)
2 (xn− x∗)2
= x∗+g00(ξ)
2 (xn− x∗)2, where ξ lies between xnand x∗.
師大
Since
|g0(x)| ≤ k < 1, ∀ x ∈ [x∗− δ, x∗+ δ]
and
g([x∗− δ, x∗+ δ]) ⊆ [x∗− δ, x∗+ δ], it follows that {xn}∞n=0converges to x∗.
But ξnis between xnand x∗for each n, so {ξn}∞n=0also converges to x∗ and
n→∞lim
|xn+1− x∗|
|xn− x∗|2 = |g00(x∗)|
2 < M 2 .
It implies that {xn}∞n=0 is quadratically convergent to x∗if g00(x∗) 6= 0and
|xn+1− x∗| < M
2 |xn− x∗|2, for sufficiently large n.
師大
For Newton’s method, g(x) = x − f (x)
f0(x) ⇒ g0(x) = 1 − f0(x)
f0(x)+f (x)f00(x)
(f0(x))2 = f (x)f00(x) (f0(x))2 It follows that g0(x∗) = 0. Hence Newton’s method is locally quadratically convergent.
師大
Error Analysis of Secant Method
Reference: D. Kincaid and W. Cheney, ”Numerical analysis”
Let x∗ denote the exact solution of f (x) = 0, ek= xk− x∗ be the error at the k-th step. Then
ek+1 = xk+1− x∗
= xk− f (xk) xk− xk−1
f (xk) − f (xk−1) − x∗
= 1
f (xk) − f (xk−1)[(xk−1− x∗)f (xk) − (xk− x∗)f (xk−1)]
= 1
f (xk) − f (xk−1)(ek−1f (xk) − ekf (xk−1))
= ekek−1
1
ekf (xk) − e1
k−1f (xk−1)
xk− xk−1 · xk− xk−1 f (xk) − f (xk−1)
!
師大
To estimate the numerator
1
ekf (xk)− 1
ek−1f (xk−1)
xk−xk−1 , we apply Taylor’s Theorem
f (xk) = f (x∗+ ek) = f (x∗) + f0(x∗)ek+1
2f00(x∗)e2k+ O(e3k), to get
1 ek
f (xk) = f0(x∗) +1
2f00(x∗)ek+ O(e2k).
Similarly, 1
ek−1f (xk−1) = f0(x∗) +1
2f00(x∗)ek−1+ O(e2k−1).
Hence 1
ekf (xk) − 1
ek−1f (xk−1) ≈ 1
2(ek− ek−1)f00(x∗).
Since xk− xk−1 = ek− ek−1and xk− xk−1
f (x ) − f (x ) → 1 f0(x∗),
師大
we have
ek+1 ≈ ekek−1
1
2(ek− ek−1)f00(x∗) ek− ek−1 · 1
f0(x∗)
!
= 1 2
f00(x∗) f0(x∗)ekek−1
≡ Cekek−1. (2)
To estimate the convergence rate, we assume
|ek+1| ≈ η|ek|α, where η > 0 and α > 0 are constants, i.e.,
|ek+1|
η|ek|α → 1 as k → ∞.
Then |ek| ≈ η|ek−1|αwhich implies |ek−1| ≈ η−1/α|ek|1/α. Hence (2) gives
η|ek|α≈ C|ek|η−1/α|ek|1/α =⇒ C−1η1+α1 ≈ |ek|1−α+α1.
師大
Since |ek| → 0 as k → ∞, and C−1η1+α1 is a nonzero constant, 1 − α + 1
α = 0 =⇒ α = 1 +√ 5
2 ≈ 1.62.
This result implies that C−1η1+α1 → 1 and η → C1+αα = f00(x∗)
2f0(x∗)
0.62
. In summary, we have shown that
|ek+1| = η|ek|α, α ≈ 1.62, that is, therate of convergenceissuperlinear.
Rate of convergence
secantmethod: superlinear Newton’smethod: quadratic bisectionmethod: linear
師大
Each iteration of method requires
secant method: one function evaluation
Newton’s method: two function evaluation, namely, f (xk) and f0(xk).
⇒ two steps of secant method are comparable to one step of Newton’s method. Thus
|ek+2| ≈ η|ek+1|α ≈ η1+α|ek|3+
√ 5
2 ≈ η1+α|ek|2.62.
⇒ secant method is more efficient than Newton’s method.
Remark
Two steps of secant method would require a little more work than one step of Newton’s method.
師大
Aitken’s ∆2 method
Accelerate the convergence of a sequence that islinearly convergent.
Suppose {yn}∞n=0is a linearly convergent sequence with limit y. Construct a sequence {ˆyn}∞n=0that converges more rapidly to y than {yn}∞n=0.
For n sufficiently large, yn+1− y
yn− y ≈ yn+2− y yn+1− y. Then
(yn+1− y)2≈ (yn+2− y)(yn− y), so
y2n+1− 2yn+1y + y2 ≈ yn+2yn− (yn+2+ yn)y + y2
師大
and
(yn+2+ yn− 2yn+1)y ≈ yn+2yn− y2n+1. Solving for y gives
y ≈ yn+2yn− yn+12 yn+2− 2yn+1+ yn
= ynyn+2− 2ynyn+1+ y2n− yn2+ 2ynyn+1− y2n+1 yn+2− 2yn+1+ yn
= yn(yn+2− 2yn+1+ yn) − (yn+1− yn)2 (yn+2− yn+1) − (yn+1− yn)
= yn− (yn+1− yn)2
(yn+2− yn+1) − (yn+1− yn). Aitken’s ∆2 method
ˆ
yn= yn− (yn+1− yn)2
(yn+2− yn+1) − (yn+1− yn). (3)
師大
Example 17
The sequence {yn= cos(1/n)}∞n=1converges linearly to y = 1.
n yn yˆn
1 0.54030 0.96178 2 0.87758 0.98213 3 0.94496 0.98979 4 0.96891 0.99342 5 0.98007 0.99541 6 0.98614
7 0.98981
{ˆyn}∞n=1converges more rapidly to y = 1 than {yn}∞n=1.
師大
Definition 18
For a given sequence {yn}∞n=0, the forward difference ∆ynis defined by
∆yn= yn+1− yn, for n ≥ 0.
Higher powers of ∆ are defined recursively by
∆kyn= ∆(∆k−1yn), for k ≥ 2.
The definition implies that
∆2yn= ∆(yn+1− yn) = ∆yn+1− ∆yn= (yn+2− yn+1) − (yn+1− yn).
So the formula for ˆynin (3) can be written as ˆ
yn= yn−(∆yn)2
∆2yn
, for n ≥ 0.
師大
Theorem 19
Suppose {yn}∞n=0 → ylinearlyand
n→∞lim
yn+1− y yn− y < 1.
Then {ˆyn}∞n=0→ y faster than {yn}∞n=0in the sense that
n→∞lim ˆ yn− y yn− y = 0.
Aitken’s ∆2method constructs the terms in order:
y0, y1 = g(y0), y2= g(y1), yˆ0= {∆2}(y0), y3 = g(y2), ˆ
y1 = {∆2}(y1), . . . .
⇒ Assume |ˆy0− y| < |y2− y|
師大
Steffensen’s method constructs the terms in order:
y0(0)≡ y0, y1(0) = g(y(0)0 ), y2(0)= g(y1(0)), y0(1)= {∆2}(y0(0)), y(1)1 = g(y(1)0 ), y2(1)= g(y(1)1 ), . . . . Steffensen’s method (To find a solution of y = g(y))
Given y0, tolerance T ol, max. number of iteration M . Set i = 1.
While i ≤ M
Set y1= g(y0); y2= g(y1); y = y0− (y1− y0)2/(y2− 2y1+ y0).
If |y − y0| < T ol, then STOP.
Set i = i + 1; y0 = y.
End While Theorem 20
Suppose x = g(x) has solution x∗ with g0(x∗) 6= 1. If ∃ δ > 0 such that g ∈ C3[x∗− δ, x∗+ δ], then Steffensen’s method gives quadraticconvergence for any x0 ∈ [x∗− δ, x∗+ δ].
師大
Zeros of polynomials and M ¨ uller’s method
• Horner’s method:
Let
P (x) = a0+ a1x + a2x2+ · · · + an−1xn−1+ anxn
= a0+ x (a1+ x (a2+ · · · + x (an−1+ anx) · · · )) . If
bn = an,
bk = ak+ bk+1x0, for k = n − 1, n − 2, . . . , 1, 0, then
b0= a0+ b1x0= a0+ (a1+ b2x0) x0 = · · · = P (x0).
Consider
Q(x) = b1+ b2x + · · · + bnxn−1.
師大
Then
b0+ (x − x0)Q(x) = b0+ (x − x0) b1+ b2x + · · · + bnxn−1
= (b0− b1x0) + (b1− b2x0)x + · · · + (bn−1− bnx0)xn−1+ bnxn
= a0+ a1x + · · · + anxn= P (x).
Differentiating P (x) with respect to x gives
P0(x) = Q(x) + (x − x0)Q0(x) and P0(x0) = Q(x0).
Use Newton-Raphson method to find an approximate zero of P (x):
xk+1= xk−P (xk)
Q(xk), ∀ k = 0, 1, 2, . . . . Similarly, let
cn = bn= an,
ck = bk+ ck+1xk, for k = n − 1, n − 2, . . . , 1, then c1 = Q(xk).
師大
Horner’s method(Evaluate y = P (x0)and z = P0(x0)) Set y = an; z = an.
For j = n − 1, n − 2, . . . , 1 Set y = aj+ yx0; z = y + zx0. End for
Set y = a0+ yx0. If xN is an approximate zero of P , then
P (x) = (x − xN)Q(x) + b0 = (x − xN)Q(x) + P (xN)
≈ (x − xN)Q(x) ≡ (x − ˆx1)Q1(x).
So x − ˆx1is an approximate factor of P (x) and we can find a second approximate zero of P by applying Newton’s method to Q1(x). The procedure is called deflation.
師大
• M ¨uller’s method for complex root:
Theorem 21
If z = a + ib is a complex zero of multiplicity m of P (x) with real coefficients, then ¯z = a − biis also a zero of multiplicity m of P (x)and (x2− 2ax + a2+ b2)m is a factor of P (x).
Secant method: Given p0 and p1, determine p2 as the intersection of the x-axis with the line through (p0, f (p0))and (p1, f (p1)).
M ¨uller’s method: Given p0, p1
and p2, determine p3 by the intersection of the x-axis with the parabola through (p0, f (p0)), (p1, f (p1))and (p2, f (p2)).
師大
Let
P (x) = a(x − p2)2+ b(x − p2) + c
that passes through (p0, f (p0)), (p1, f (p1))and (p2, f (p2)). Then f (p0) = a(p0− p2)2+ b(p0− p2) + c,
f (p1) = a(p1− p2)2+ b(p1− p2) + c, f (p2) = a(p2− p2)2+ b(p2− p2) + c = c.
It implies that c = f (p2),
b = (p0− p2)2[f (p1) − f (p2)] − (p1− p2)2[f (p0) − f (p2)]
(p0− p2)(p1− p2)(p0− p1) , a = (p1− p2) [f (p0) − f (p2)] − (p0− p2) [f (p1) − f (p2)]
(p0− p2)(p1− p2)(p0− p1) .
師大
To determine p3, a zero of P , we apply the quadratic formula to P (x) = 0and get
p3− p2 = 2c b ±√
b2− 4ac. Choose
p3 = p2+ 2c b + sgn(b)√
b2− 4ac
such that the denominator will be largest and result in p3 selected as the closest zero of P to p2.
師大
M ¨uller’s method (Find a solution of f (x) = 0)
Given p0, p1, p2; tolerance T OL; maximum number of iterations M Set h1 = p1− p0; h2 = p2− p1;
δ1= (f (p1) − f (p0))/h1; δ2 = (f (p2) − f (p1))/h2; d = (δ2− δ1)/(h2+ h1); i = 3.
While i ≤ M
Set b = δ2+ h2d; D =pb2− 4f (p2)d.
If |b − D| < |b + D|, then set E = b + D else set E = b − D.
Set h = −2f (p2)/E; p = p2+ h.
If |h| < T OL, then STOP.
Set p0= p1; p1 = p2; p2 = p; h1= p1− p0; h2= p2− p1; δ1= (f (p1) − f (p0))/h1; δ2 = (f (p2) − f (p1))/h2; d = (δ2− δ1)/(h2+ h1); i = i + 1.
End while