An optimization problem with Newton's method

In summary: No, the rate constant is not 1/4. The rate constant is a coefficient in front of the highest order term in the expansion. In this case, the highest order term is $(\Delta x_k)^2$, so the rate constant is 1/4.Also, please do not include the mathcal O notation in your summary as it is not necessary for understanding the concept of quadratic convergence and rate constant. Your summary should simply state that the rate constant for the second method is 1/4 and how it was derived using an expansion.
  • #1
i_a_n
83
0
Apply Newton's method to $f(x)=(x-2)^4+(x-2)^5$ with initial guess $x_0=3$. We can observe that the sequence converges linearly with rate constant $3/4$. Now apply the iterative mathod $x_{k+1}=x_k-4f(x_k)/f'(x_k)$. This method should converge more rapidly for this problem. But how to prove that the new method converges quadratically and determine the rate constant?also, how to do the first part, that is, how to see linear convergence with rate constant 3/4? And how to prove the second part and find the rate constant?
 
Mathematics news on Phys.org
  • #2
You should be able to demonstrate that Newton's method leads to:

$\displaystyle x_{k+1}=\frac{3}{4}x_{k}+\frac{7}{16}+\frac{3}{16(4x_{k}-5)}$

and the second method leads to:

$\displaystyle x_{k+1}=\frac{7}{4}+\frac{3}{4(4x_{k}-5)}$
 
  • #3
ianchenmu said:
Apply Newton's method to $f(x)=(x-2)^4+(x-2)^5$ with initial guess $x_0=3$. We can observe that the sequence converges linearly with rate constant $3/4$. Now apply the iterative mathod $x_{k+1}=x_k-4f(x_k)/f'(x_k)$. This method should converge more rapidly for this problem. But how to prove that the new method converges quadratically and determine the rate constant?also, how to do the first part, that is, how to see linear convergence with rate constant 3/4? And how to prove the second part and find the rate constant?

Let's define $\Delta x_k$ as the difference between $x_k$ and the actual root.
In your case that means that $\Delta x_k = x_k - 2$.

The method is:

$\qquad x_{k+1} = x_k - \dfrac{f(x_k)}{f'(x_k)}$

It follows that:

$\Delta x_{k+1} = \Delta x_k - \dfrac{f(x_k)}{f'(x_k)}$

$\qquad = \Delta x_k - \dfrac{(x_k-2)^4+(x_k-2)^5}{4(x_k-2)^3+5(x_k-2)^4}$

$\qquad = \Delta x_k - \dfrac{(\Delta x_k)^4+(\Delta x_k)^5}{4(\Delta x_k)^3+5(\Delta x_k)^4}$

$\qquad = \Delta x_k - \dfrac{(\Delta x_k)^4}{4(\Delta x_k)^3} + \mathcal O ((\Delta x_k)^2)$

$\qquad = \Delta x_k - \frac 1 4 \Delta x_k + \mathcal O ((\Delta x_k)^2)$

$\qquad = \frac 3 4 \Delta x_k + \mathcal O ((\Delta x_k)^2)$

To make this approximately zero, we need to multiply the part that is subtraced by 4.
Note that the factor 4 is the multiplicity of the root.
That way we will get (at least) quadratic convergence.

With this factor 4 we get:

$\Delta x_{k+1} = \Delta x_k - 4 \dfrac{f(x_k)}{f'(x_k)}$

$\qquad = \Delta x_k - 4 \dfrac{(\Delta x_k)^4+(\Delta x_k)^5}{4(\Delta x_k)^3+5(\Delta x_k)^4}$

$\qquad = \Delta x_k - \dfrac{1+\Delta x_k}{1+\frac 5 4 \Delta x_k} \Delta x_k$

With a first order expansion of the fraction it follows what the quadratic rate constant will be.
 
  • #4
ILikeSerena said:
Let's define $\Delta x_k$ as the difference between $x_k$ and the actual root.
In your case that means that $\Delta x_k = x_k - 2$.

The method is:

$\qquad x_{k+1} = x_k - \dfrac{f(x_k)}{f'(x_k)}$

It follows that:

$\Delta x_{k+1} = \Delta x_k - \dfrac{f(x_k)}{f'(x_k)}$

$\qquad = \Delta x_k - \dfrac{(x_k-2)^4+(x_k-2)^5}{4(x_k-2)^3+5(x_k-2)^4}$

$\qquad = \Delta x_k - \dfrac{(\Delta x_k)^4+(\Delta x_k)^5}{4(\Delta x_k)^3+5(\Delta x_k)^4}$

$\qquad = \Delta x_k - \dfrac{(\Delta x_k)^4}{4(\Delta x_k)^3} + \mathcal O ((\Delta x_k)^2)$

$\qquad = \Delta x_k - \frac 1 4 \Delta x_k + \mathcal O ((\Delta x_k)^2)$

$\qquad = \frac 3 4 \Delta x_k + \mathcal O ((\Delta x_k)^2)$

To make this approximately zero, we need to multiply the part that is subtraced by 4.
Note that the factor 4 is the multiplicity of the root.
That way we will get (at least) quadratic convergence.

With this factor 4 we get:

$\Delta x_{k+1} = \Delta x_k - 4 \dfrac{f(x_k)}{f'(x_k)}$

$\qquad = \Delta x_k - 4 \dfrac{(\Delta x_k)^4+(\Delta x_k)^5}{4(\Delta x_k)^3+5(\Delta x_k)^4}$

$\qquad = \Delta x_k - \dfrac{1+\Delta x_k}{1+\frac 5 4 \Delta x_k} \Delta x_k$

With a first order expansion of the fraction it follows what the quadratic rate constant will be.

why
$\qquad = \Delta x_k - \dfrac{(\Delta x_k)^4+(\Delta x_k)^5}{4(\Delta x_k)^3+5(\Delta x_k)^4}$

$\qquad = \Delta x_k - \dfrac{(\Delta x_k)^4}{4(\Delta x_k)^3} + \mathcal O ((\Delta x_k)^2)$?
And where you showed the second method is quadratic convergence and what's the rate constant?
 
  • #5
ianchenmu said:
why
$\qquad = \Delta x_k - \dfrac{(\Delta x_k)^4+(\Delta x_k)^5}{4(\Delta x_k)^3+5(\Delta x_k)^4}$

$\qquad = \Delta x_k - \dfrac{(\Delta x_k)^4}{4(\Delta x_k)^3} + \mathcal O ((\Delta x_k)^2)$?

Do you know what $\mathcal O(y^2)$ means?
If not you can just ignore it and replace each $=$ by $\approx$.When $\Delta x_k$ approaches zero, $(\Delta x_k)^5$ becomes negligible with respect to $(\Delta x_k)^4$.
We can write: $(\Delta x_k)^4+(\Delta x_k)^5 = (\Delta x_k)^4+ \mathcal O((\Delta x_k)^5) = (\Delta x_k)^4( 1 + \mathcal O(\Delta x_k))$.Alternatively, you can do an expansion.

Using $\frac 1 {1+y} = 1 - y + y^2 - ... = 1 - y + \mathcal O(y^2)$

we can expand as follows:

$\qquad \Delta x_k - \dfrac{(\Delta x_k)^4+(\Delta x_k)^5}{4(\Delta x_k)^3+5(\Delta x_k)^4}$

$\qquad = \Delta x_k - \frac 1 4 \dfrac{1+\Delta x_k}{1+ \frac 5 4 \Delta x_k}\Delta x_k$

$\qquad = \Delta x_k - \frac 1 4 (1+\Delta x_k)(1 - \frac 5 4 \Delta x_k + ...)\Delta x_k$

$\qquad = \Delta x_k - \frac 1 4 (1 - \frac 1 4 \Delta x_k + ...)\Delta x_k$

$\qquad = \Delta x_k - \frac 1 4 \Delta x_k + \mathcal O ((\Delta x_k)^2)$
And where you showed the second method is quadratic convergence and what's the rate constant?

Can you make an expansion similar to the one I just did to the following expression (which was the last)?

$\Delta x_{k+1} = \Delta x_k - \dfrac{1+\Delta x_k}{1+\frac 5 4 \Delta x_k} \Delta x_k$
 
Last edited:
  • #6
ILikeSerena said:
Do you know what $\mathcal O(y^2)$ means?
If not you can just ignore it and replace each $=$ by $\approx$.When $\Delta x_k$ approaches zero, $(\Delta x_k)^5$ becomes negligible with respect to $(\Delta x_k)^4$.
We can write: $(\Delta x_k)^4+(\Delta x_k)^5 = (\Delta x_k)^4+ \mathcal O((\Delta x_k)^5) = (\Delta x_k)^4( 1 + \mathcal O(\Delta x_k))$.Alternatively, you can do an expansion.

Using $\frac 1 {1+y} = 1 - y + y^2 - ... = 1 - y + \mathcal O(y^2)$

we can expand as follows:

$\qquad \Delta x_k - \dfrac{(\Delta x_k)^4+(\Delta x_k)^5}{4(\Delta x_k)^3+5(\Delta x_k)^4}$

$\qquad = \Delta x_k - \frac 1 4 \dfrac{1+\Delta x_k}{1+ \frac 5 4 \Delta x_k}\Delta x_k$

$\qquad = \Delta x_k - \frac 1 4 (1+\Delta x_k)(1 - \frac 5 4 \Delta x_k + ...)\Delta x_k$

$\qquad = \Delta x_k - \frac 1 4 (1 - \frac 1 4 \Delta x_k + ...)\Delta x_k$

$\qquad = \Delta x_k - \frac 1 4 \Delta x_k + \mathcal O ((\Delta x_k)^2)$

Can you make an expansion similar to the one I just did to the following expression (which was the last)?

$\Delta x_{k+1} = \Delta x_k - \dfrac{1+\Delta x_k}{1+\frac 5 4 \Delta x_k} \Delta x_k$

So in your way, for the second method, is the rate constant 1/4? Since
$\Delta x_{k+1} = \Delta x_k - \dfrac{1+\Delta x_k}{1+\frac 5 4 \Delta x_k} \Delta x_k$
$\qquad = \Delta x_k - \Delta x_k + \frac 1 4 (\Delta x_k)^2 +\mathcal O ((\Delta x_k)^3)$
 
Last edited:
  • #7
ianchenmu said:
So in your way, for the second method, is the rate constant 1/4? Since
$\Delta x_{k+1} = \Delta x_k - \dfrac{1+\Delta x_k}{1+\frac 5 4 \Delta x_k} \Delta x_k$
$\qquad = \Delta x_k - \Delta x_k + \frac 1 4 (\Delta x_k)^2 +\mathcal O ((\Delta x_k)^3)$

Yep! ;)
 
  • #8
ILikeSerena said:
Yep! ;)

Hey I have a similar question may need your help. It's at here:
http://www.mathhelpboards.com/showthread.php?p=15336

You may need to use Taylor series of $f(x_k)$
Thank you!
 
Last edited:

FAQ: An optimization problem with Newton's method

What is Newton's method?

Newton's method is an algorithm used to find the roots of a function or the minima/maxima of an optimization problem. It involves using calculus to iteratively refine an initial guess until a more accurate solution is found.

How does Newton's method work?

Newton's method works by taking an initial guess for the root or minimum/maximum of a function and using the derivative of the function to find the slope of the tangent line at that point. This slope is then used to find a new point on the function that is closer to the actual root or minimum/maximum. This process is repeated until the desired level of accuracy is achieved.

What is an optimization problem?

An optimization problem is a mathematical problem that involves finding the maximum or minimum value of a function, subject to certain constraints. These problems can be solved using various methods, such as Newton's method.

How is Newton's method used for optimization problems?

Newton's method can be used to find the minimum or maximum of an optimization problem by finding the roots of the derivative of the function. This is because the derivative of a function is equal to zero at its minimum or maximum points, and Newton's method is able to find these roots efficiently.

What are the advantages of using Newton's method for optimization problems?

One advantage of using Newton's method for optimization problems is that it can converge to the solution quickly, especially for well-behaved functions. It also allows for more accurate solutions compared to other methods, such as gradient descent. However, it may not always converge to the global minimum or maximum, and it may require knowledge of the second derivative of the function, which can be computationally expensive in some cases.

Similar threads

Replies
7
Views
2K
Replies
21
Views
3K
Replies
1
Views
10K
Replies
9
Views
2K
Replies
1
Views
1K
Replies
15
Views
4K
Replies
1
Views
9K
Back
Top