How can you prove that Newton's method for estimating the cube root converges when $x_0$ is not $0$? Where sequence $\{x_{n}\}$ can be defined with the relation $x_{n+1} = \frac{2}{3}x_n + \frac{1}{3}\frac{z}{x_n^2} $.

I tried to figure out a way to write this as another formula to be proved by induction that it is monotonic and bounded, but I have no idea how to do that. I tried writing out the terms, and cannot figure out a formula for that.

Edit: z is the term of which you are finding the cube root.

  • $\begingroup$ Are you taking initial conditions with the same sign as $z$? (I don't think this matters, but the details are harder if the signs are different.) $\endgroup$ – Ian Apr 3 '17 at 18:13
  • $\begingroup$ I'd like to know for all cases, where signs are the same and signs are different. $\endgroup$ – Lo Ran Apr 3 '17 at 18:17

Sketch: suppose $z \neq 0$, $r=z^{1/3}$. Let $g(x)=x-\frac{x^3-z}{3x^2}$. We have two cases:

If $x>r$ then $x>g(x)>r$. This implies the recursive sequence $g^n(x)$ is bounded and monotone, so it is convergent. The limit of a recursive sequence defined by a continuous mapping must be a fixed point of the mapping, and the only fixed point of $g$ is $r$.

If $x<r$ then it is true that $x<g(x)$ (so that you move toward the root). But in general it is not true that $g(x)<r$, so the sequence is no longer guaranteed to be monotone. So the error may not go to zero monotonically, either. As a workaround, you can show that eventually $g^n(x)>r$, at which point the previous case applies.

To make this work out, you will need to additionally assume that $g^n(x)$ is never exactly zero, which rules out a countable set of initial values. (Thankfully, all of these points, except for $0$ itself, have the opposite sign from $r$, so there is no reason you should ever encounter them.)

  • $\begingroup$ How did you get the formula for g(x)? $\endgroup$ – Lo Ran Apr 3 '17 at 18:29
  • $\begingroup$ @LoRan $g$ is just the function that you iterate for Newton's method for cube roots. In other words Newton's method amounts to $x_{k+1}=g(x_k)$. It is the same as you wrote, just "unsimplified" (and I think you are better off leaving it that way, at least for analytical purposes). $\endgroup$ – Ian Apr 3 '17 at 18:32
  • $\begingroup$ If $z>0$ and $0<x<z^{1/3}$ then $g(x)>z^{1/3}.$ $\endgroup$ – DanielWainfleet Apr 3 '17 at 18:34
  • $\begingroup$ @user254665 Thank you, that means I can combine those cases. $\endgroup$ – Ian Apr 3 '17 at 18:37

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.