# finding all zeros of a continuous function

Let $$f: \Bbb R \to \Bbb R$$ be continuous. What are efficient algorithms to finding all the zeros in an interval $$[a, b]$$? I am actually only interested in the smallest zero in that interval, if there's an algorithm that can provide me with just that.

• What do you know about the function? Can you efficiently evaluate $f(x)$? Can you efficiently evaluate $f'(x)$ or $f''(x)$? Jul 8, 2022 at 16:47
• Separately, the function $f(x)=x\sin(1/x)$ is continuous on $[0,1]$, for example, but it has infinitely many zeros and you will have trouble finding them all. You need to restrict yourself to a smaller class of functions than just continuity. Jul 8, 2022 at 16:49
• we may even suppose $f$ is differentiable, and that there are finitely many zeros in the interval. But I don't know what $f'$ is explicitly Jul 8, 2022 at 17:00
• Do you know the number of zeros ? Jul 8, 2022 at 23:50
• In that case you can do an incremental search followed by a secant method. Jul 9, 2022 at 3:13

As others have pointed out, you cannot solve this problem in general for all continuous functions. But there are methods that work quite well in practice. One such method is to sample the function at Chebyshev points, and compute the roots of the interpolating polynomial. Monitoring the decay of the Chebyshev coefficients can be used to decide when the function has been sampled adequately enough. For functions that are sufficiently smooth, you can stop when the coefficients are close to the machine epsilon, for example. This method is implemented in the ‘roots’ function of the Chebfun package (https://www.chebfun.org).

Once you have found the Chebyshev coefficients, you form a generalized companion matrix whose eigenvalues are the roots. You could use a Krylov method to find only the smallest root (as in matlab’s eigs command).

• +1, this is what I would've added right now if you hadn't. Throw the function into chebfun, let it approximate with machine accuracy, and then find the roots using linear algebra. Jul 12, 2022 at 22:19
• This is precisely what I was looking for. In fact, after posting this question, I began implementing the method you describe. Jul 13, 2022 at 1:28

The question is ill-formulated, in the sense that you will never be able to find the smallest zero for all continuous functions in $$[a,b]$$, unless you now something more, like for example some derivatives of $$f(x)$$.

For example, for $$x\in[-1,1]$$, let's say that you're looking at the function $$f(x)=\frac{e^{-\frac{x^2}{2s^s}}}{\sqrt{2\pi s^2}}-1$$. For some values of $$s$$, this function has two zeros, arbitrarily close to $$0$$. How do you find them without knowing anything about $$f$$? How do you even know that there are some zeros, if all you can do is sample randomly?

• The objection is incorrect. You may suppose there exists at least one zero, if that's what bothers you (ideally the algorithm tells you if a zero doesn't exist, but it's ok if it doesn't). More importantly -- for a fixed $s$, the function $f$ you present does NOT have 2 roots that are arbitrarily close to zero (such a root would be uniquely $0$, perhaps with multiplicity). It does have 2 zeros, symmetrically located to the left and right of the origin, and that's ok. The algorithm should return the smaller of the two -- the negative root. There's no issue with your example. Jul 9, 2022 at 0:44
• The function I chose is just a centred (and shifted) normal distribution. It will have exactly two roots if $s<\frac1{\sqrt{2\pi}}$. As we take smaller and smaller $s$, the roots will get closer and closer to $0$. We can choose a $s$ so the absolute values of the roots are smaller than any value we want.
– PC1
Jul 9, 2022 at 1:42
• Further, let's have a continuous function $f(x)$ for which $f(a)$ and $f(b)$ are negative. How can you even know that there is at least one zero? How can an algorithm detect that the function is positive over some domain?
– PC1
Jul 9, 2022 at 1:44
• the function you describe is $\Bbb R^2 \to \Bbb R$ (a function of $s$, and $x$, since you allow $s$ to vary), whereas I indicated that the function of interest here is $\Bbb R \to \Bbb R$. Your objection is invalid for say, the standard normal, which would be the appropriate example here. Jul 9, 2022 at 2:22
• I tend to agree with @PC1 - if all we know is that the function is continuous, and it has at least one root in the interval, that's not sufficient for a numerical algorithm to find all its roots in the interval in the general case. If we had some constraints on the derivative of the function that would perhaps allow formulating an algorithm for finding all roots (assuming there is a finite number of them). Jul 9, 2022 at 5:05

Another solution is interval rootfinding, for instance with the interval Newton method. This technique can find all zeros of a given differentiable function while proving automatically at the same time that they are all the zeros in the given interval.

A package to do that is Julia's IntervalRootFinding.jl. All the required derivatives are computed automatically, so you can just provide the function as a formula, e.g., x -> x^2 - 2x.

julia> using IntervalArithmetic, IntervalRootFinding

julia> roots(x -> x^2 - 2x, -10..10)
2-element Vector{Root{Interval{Float64}}}:
Root([-8.17345e-10, 6.8733e-10], :unique)
Root([1.99999, 2.00001], :unique)


The algorithm has identified that the two given intervals (of diameter $$\approx 2\times 10^{-9}$$, a configurable threshold) each contain a unique zero of the function.

Here is an example with the function mentioned in the other answer:

julia> s = 1e-5
1.0e-5

julia> roots(x -> exp(-x^2/2s^2) / sqrt(2pi*s^2) - 1, -10..10)
2-element Vector{Root{Interval{Float64}}}:
Root([4.60292e-05, 4.60314e-05], :unique)
Root([-4.60313e-05, -4.60295e-05], :unique)
$$$$
`
• The function is a black box to me, although from sampling and plotting it (and from the application domain) I am fairly confident that it’s a nice function and is well behaved. Jul 13, 2022 at 1:29
• What do you mean by "black-box" exactly? Just to be sure that automatic differentiation is off the table. Jul 13, 2022 at 9:25
• I can evaluate $f(x)$ for any $x$, but I don’t know what $f$ itself looks like Jul 13, 2022 at 11:39
• But what is in the computer code that you call to evaluate $f$? Library calls? For loops? Code someone else wrote? Is there anything specific that makes it unsuitable for automatic differentiation? Jul 13, 2022 at 11:42
• you have an obscure interface to call $f$. Yes, code someone else wrote, but more importantly, you don't have access to the internal implementation. You can sample, however. For autodiff you would need to define $\nabla f$, but you don't have access to $f$, much less $\nabla f$. Jul 13, 2022 at 13:28