- Thread starter
- #1

I found that it has Kappa value equal to 1/2

I am not too sure what happens if when x = 0 is it just a minimum?

But now I am trying to investigate and find the sensitivity of F(x) to errors in x when we use x+ϵ, where ϵ is small.

- Thread starter akerman
- Start date

- Thread starter
- #1

I found that it has Kappa value equal to 1/2

I am not too sure what happens if when x = 0 is it just a minimum?

But now I am trying to investigate and find the sensitivity of F(x) to errors in x when we use x+ϵ, where ϵ is small.

- Admin
- #2

- Mar 5, 2012

- 8,779

Welcome to MHB, akerman!

I found that it has Kappa value equal to 1/2

I am not too sure what happens if when x = 0 is it just a minimum?

But now I am trying to investigate and find the sensitivity of F(x) to errors in x when we use x+ϵ, where ϵ is small.

I am assuming that with a Kappa value you mean the condition number.

If so then we have:

$$\frac{|\Delta y|}{|y|} = \kappa \frac{|\Delta x|}{|x|}$$

where $y=F(x)$, where $\Delta x$ is the error in $x$, and $\Delta y$ is the error in $y$.

In other words: $\kappa$ gives the amplification of the relative error.

At $x=0$, $x$ has an infinite relative error (we're dividing by zero), so $\kappa$ is not defined there.

Since $F(0)=0$, the relative error of $y$ is also infinite.

We can calculate $\kappa$ with:

$$\kappa = \left| \frac{x \cdot F'(x)}{F(x)} \right|$$

This stems from the first order Taylor approximation:

$$F(x + \Delta x) \approx F(x) + \Delta x \cdot F'(x)$$

Since we are using an approximation, the end result is also an approximation.

In particular for points where $F'(x)$ or $F''(x)$ are not defined, the relationship will break down.

This is in particular the case for your function at $x=0$.

- Thread starter
- #3

I still don't get it...Welcome to MHB, akerman!

I am assuming that with a Kappa value you mean the condition number.

If so then we have:

$$\frac{|\Delta y|}{|y|} = \kappa \frac{|\Delta x|}{|x|}$$

where $y=F(x)$, where $\Delta x$ is the error in $x$, and $\Delta y$ is the error in $y$.

In other words: $\kappa$ gives the amplification of the relative error.

At $x=0$, $x$ has an infinite relative error (we're dividing by zero), so $\kappa$ is not defined there.

Since $F(0)=0$, the relative error of $y$ is also infinite.

We can calculate $\kappa$ with:

$$\kappa = \left| \frac{x \cdot F'(x)}{F(x)} \right|$$

This stems from the first order Taylor approximation:

$$F(x + \Delta x) \approx F(x) + \Delta x \cdot F'(x)$$

Since we are using an approximation, the end result is also an approximation.

In particular for points where $F'(x)$ or $F''(x)$ are not defined, the relationship will break down.

This is in particular the case for your function at $x=0$.

So what is the sensitivity of f(x) to errors in x?

And if we consider limit x→0, how many digits can one compute x√ when x is known to an error of 10^−16?

Can you give more detailed explanation.

thanks

- Admin
- #4

- Mar 5, 2012

- 8,779

Since the derivative of $\sqrt x$ is $\frac 1 {2\sqrt x}$, Taylor's approximation gives us:I still don't get it...

So what is the sensitivity of f(x) to errors in x?

$$\sqrt{x+ε} \approx \sqrt{x} + ε \cdot \frac{1}{2\sqrt x}$$

So if the error in $x$ is $ε$, then the error in $√x$ is approximately $ε \cdot \frac{1}{2\sqrt x}$.

The so called absolute sensitivity to errors in x is $\frac{1}{2\sqrt x}$, since an error gets multiplied by this amount.

The relative sensitivity is $\frac 1 2$, since relative errors get multiplied by this amount.

A relative error is the error relative to the value measured. For $x$ this is $ε / x$.

I do not understand your question.And if we consider limit x→0, how many digits can one compute x√ when x is known to an error of 10^−16?

In the limit x→0, √x is simply 0.

However, if x is a regular non-zero value known with an error of $10^{−16}$, then the resultant √x will have an absolute error of $10^{−16} \cdot \frac{1}{2\sqrt x}$ and a relative error of $0.5 \cdot 10^{−16}$.

Where would you like more details?Can you give more detailed explanation.

thanks

- Thread starter
- #5

So having a quesiton such as "how many digits can one compute x√ when x is known to an error of 10^−16?" Can I simply say that absolute error of 10−16⋅12x√ and a relative error of 0.5⋅10^−16?

Also if we have x= x+ ε

Can I specify exactly what the ε is for y =x√ ? Or is it just an assumption that is it a small number?

- Admin
- #6

- Mar 5, 2012

- 8,779

The term "how many digits" is somewhat confusing.The lastest answer is just something I was looking for.

So having a quesiton such as "how many digits can one compute x√ when x is known to an error of 10^−16?" Can I simply say that absolute error of 10−16⋅12x√ and a relative error of 0.5⋅10^−16?

It can typically mean either how many digits behind the decimal point, or it can mean how many significant digits.

Can you clarify which one is intended?

Similarly, when you say "error" do you mean an absolute error or a relative error?

If you have 16

In this case the resultant relative error is $0.5 \cdot 10^{-16}$, meaning you have slightly over 16 significant digits (usually treated as just 16).

What you write about the errors is correct, assuming your initial error is an absolute error.

However, that is apparently not what is being asked, since the question asks "how many digits".

"Exactly" is a strong word.Also if we have x= x+ ε

Can I specify exactly what the ε is for y =x√ ? Or is it just an assumption that is it a small number?

If you want to have the "exact" error in y, you need to calculate $\sqrt{x+ε}-\sqrt x$.

If you are satisfied with the approximate error, you can use the formulas I gave.

- Thread starter
- #7