7
$\begingroup$

I've read about the bisection method for finding roots of a function in my numerical analysis textbook and one question came to my mind.

Given a relatively complicated function, the chances of finding the exact root (that is, a root that is completely represented in the computer's memory, with all significant figures) are very low. This means that most of the time, we will have a value for which the function is very close to, but not exactly equal to zero.

So what would happen if the function had one root, and another value at which the function gets extremely close to zero, without actually getting there? Would the algorithm fail? And what is the meaning of that eventual "fake" root; is it worth anything?

Thank you.

EDIT: here is a picture showing what I meant

$\endgroup$
  • 3
    $\begingroup$ If you are within the tolerance you have set for the algorithm, $|f(x)|<\epsilon$ then it would count as a root. This could arise if your function is discontinuous at x, but you would probably make sure that isn't the case before you start. $\endgroup$ – Paul Dec 1 '18 at 16:36
  • 1
    $\begingroup$ Are you assuming the function is continuous? $\endgroup$ – PyRulez Dec 1 '18 at 23:10
  • 1
    $\begingroup$ the "fake" root alludes to the fact that the intermediate value theorem is not constructive. $\endgroup$ – Matt A Pelto Dec 2 '18 at 2:50
  • $\begingroup$ The bisection method does NOT "find roots". It finds (preferably, small) intervals where the function changes sign. Ask it to "find a root" of 1/x between -1 and +2, and it will tell you there is one close to 0. That behaviour is NOT a bug in the algorithm, though it may be a bug in the mind of the user who doesn't understand what the algorithm really does! $\endgroup$ – alephzero Dec 2 '18 at 14:26
  • $\begingroup$ Perhaps it is the same bug that is infecting users who insist on putting words in other people's mouths. $\endgroup$ – Matt A Pelto Dec 2 '18 at 21:07
15
$\begingroup$

The bisection method only cares if the function changes sign, so it goes straight past the 'fake' root without noticing.

If the coefficients have a slight error in them, then perhaps the 'fake' root should have been a root.

$\endgroup$
  • 2
    $\begingroup$ This should be the official answer. $\endgroup$ – user150203 Dec 2 '18 at 12:04
9
$\begingroup$

You have to be aware that the bisection method finds a point with a sign change in the values of the numerical evaluation of your function. Due to catastrophic cancellation this can give wide errors even for simple roots. Take for instance the rescaled Wilkinson polynomial $p(x)=\prod_{k=1}^{20}(x-k/10)$ in double floating point precision, after multiplying it out as $x^{20}-21x^{19}+a_{18}x^{18}+...+a_1x+a_0$. Around the root $x=1.9$ the numerical evaluation of a larger sampling of points gives this image

enter image description here

so that depending on the initial interval the bisection method might end at any point between $1.8999$ and $1.9003$


To put this into perspective, the scale for this situation, polynomial evaluation for $|x|\le 2$, is provided by $p(-2)=3.35367..\cdot 10^9$, so that the expected accuracy bound using a machine epsilon of $2⋅10^{-16}$ is indeed about $7⋅10^{-7}$, the observed errors are a little smaller.

$\endgroup$
  • 1
    $\begingroup$ It's strange that you get so apparent cancellation problem with this polynomial using double precision. Mathematica with machine precision handles it pretty well, both using built-in Product function and manual multiplication loop: screenshot. What software did you use to make this plot? $\endgroup$ – Ruslan Dec 1 '18 at 21:39
  • $\begingroup$ I used the monomial form, that is, computed the coefficients and evaluated from there. Using the Horner scheme reduces the band-width by about half. $\endgroup$ – LutzL Dec 1 '18 at 21:46
  • $\begingroup$ Ah, right, I can reproduce it after I Expand the product. $\endgroup$ – Ruslan Dec 1 '18 at 21:48
  • 1
    $\begingroup$ If someone knows so little about numerical analysis that they want to evaluate Wilkinson's polynomial by multiplying it out first, they deserve whatever nonsense answers they get IMO. "Garbage in, garbage out" applies to algorithms as well as data! $\endgroup$ – alephzero Dec 2 '18 at 14:21
  • 4
    $\begingroup$ @alephzero : I think you missed the meaning of the Wilkinson polynomial as an example of a polynomial with a seemingly "nice" root set that behaves "badly" for numerical root-finding algorithms. In general, if one knows the complete factorization of a polynomial, there is no need to apply root-finding algorithms, as one can directly read off the roots. $\endgroup$ – LutzL Dec 2 '18 at 14:25
3
$\begingroup$

As long as you evaluate $f\left(\frac {a+b}2\right)$ to something greater than zero, the method will work fine. It will replace $a$ with $\frac {a+b}2$ as the left end of the interval. The usual termination criterion for bisection and other bracketing approaches is the length of the interval, not the function value. It would keep going, knowing that there is a root somewhere in the interval because the function values at the endpoints have opposite signs.

$\endgroup$
  • $\begingroup$ In my example the computer will think that the very first estimation is a root, while it's actually not, so does it mean, as stated by @Empy2, that the value should have been a root, and that the function actually has 2 roots on the interval? $\endgroup$ – FredV Dec 1 '18 at 17:07
  • 2
    $\begingroup$ If you are using a function value test for a root and the computed value is within it, it is a root to you. That is what you said when you set the test limit. In that case this function has one double and one single root in the interval. As I said, the implementations I have seen test on the length of the interval as that is the uncertainty in the position of the root. As long as the computed value is not exactly $0$ the method will replace one endpoint or the other and continue. Getting a small non-zero value is no problem at all. $\endgroup$ – Ross Millikan Dec 1 '18 at 17:13
3
$\begingroup$

Suppose we are utilizing a computer that has $N$-bit precision ($N\geq4$). If we take any integer $n > N$ and we apply the bisection algorithm to the function $f$ defined on $[0,1]$ by $f(x)=\left(x-\frac12\right)^2\left(x-\frac34\right)-2^{-n}$, then the algorithm will output $x=\frac12$ as the point where $f$ has a root. This is because $f(0)=-\frac3{16}-2^{-n}<0$, $f(1)=\frac1{16}-2^{-n}>0$, and our computer isn't capable of distinguishing $\,f\left(\frac12\right)=-2^{-n}$ from zero.

Here is a plot of our function $f$ if we took $n=10$:enter image description here

Notwithstanding the output from applying the bisection algorithm to $f$ on $[0,1]$, we can see that the only zero of $f$ in $[0,1]$ lies somewhere between $\frac34$ and $1$.


As for the meaning of such a "fake" root, I would say it alludes to the fact that the Intermediate Value Theorem is equivalent to the nonconstructive proposition known as the lesser limited principle of omniscience.

Define a binary sequence $\{a_n\}_{n=1}^\infty$ by \begin{equation}a_n=\begin{cases}0 &\text{ iff either } 2k+3 \text{ is the sum of 3 primes for all } k\leq n \text{ or there exists } k<n \text{ s.t. } a_k=1 \\1&\text{ otherwise.}\end{cases} \end{equation} Define $a=\sum_{n=1}^\infty a_n2^{-n}$, and apply the bisection algorithm to the function $f$ defined on $[0,1]$ by $f(x)=\left(x-\frac12\right)^2\left(x-\frac34\right)-a$. As long as our computation is limited to some finite precision, the algorithm will output $x=\frac12$ as a root of $f$. This output is correct (which I take to mean that it is either identical to or approximately close to a root) if and only if the odd Goldbach conjecture is true.

The way that the binary sequence $\{a_n\}_{n=1}^\infty$ is defined is meant to invoke the limited principle of omniscience, a nonconstructive principle which implies the lesser limited principle of omniscience.


Disclaimer (in response to the valid concerns raised by Euro Micelli): My "answer" is not trying to provide affirmation of the question in the title as I would say the answer to the question posed in the title is "not yes". I will note that even arbitrary precision is still subject to available memory and computation time (as far as I know). I figure we have two sides of the same coin, the bisection method is not constructive and so is the definition of the function $f$ on $[0,1]$. Indeed there are ways to preclude such a false output, and my response has only considered the algorithm underlying the proof of the Intermediate Value Theorem in the most classical and basic setting. I respond to questions on this forum by trying to provide the question asker with some insight and perspective to the best of my knowledge given the contents of their initial post.

$\endgroup$
  • 2
    $\begingroup$ The mild issue I have with this explanation is that it unwittingly sidesteps the question being asked and answers a subtly different one. The proposed function is being pathologically constructed so that the suggested computer cannot avoid the near zero by the use of any conceivable numerical algorithm. Here, we are effectively blaming the algorithm for the limitations of the chosen computer to attack the given problem. Precision limitation is of course a critical consideration for numerical analysis, but this example doesn’t illustrate a limitation of the Bisection method specifically. $\endgroup$ – Euro Micelli Dec 2 '18 at 3:43
  • $\begingroup$ As long as the precision is somehow limited, we can find a positive integer $n$ so that the algorithm stops by outputting the "fake" root at $x=\frac12$ when applied to the function $f(x)=\left(x-\frac12\right)^2\left(x-\frac34\right)-2^{-n}$ on $[0,1]$. Of course for a given function $f$ we can always have better precision so that the algorithm is not duped by some $\hat{x}$ where $f$ is "close" to zero but not identically zero. I am merely noting that for a given precision limitation we can always find a function such that the algorithm fails when applied to the particular function. $\endgroup$ – Matt A Pelto Dec 2 '18 at 5:15

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.