Is the following conjecture true or false:

Given any five coplanar points, we can always draw at least one pair of non-intersecting circles coplanar with the points, such that two of the given points are diameter endpoints of one circle, and another two of the given points are diameter endpoints of the other circle.
Tangent circles are considered to be non-intersecting. Coincident circles are considered to be intersecting.


Example 1

Another example:

Example 2

I cannot find a counter-example, nor can I prove the conjecture. I made a generator of five (pseudo)random points.

(If, instead, we were given four points, the conjecture would not be true: for example, if the four points were the vertices of an equilateral triangle plus the centre, then we could not draw a pair of non-intersecting circles.)

Context: I was thinking about this question about random points in a disk. Staring at various sets of five points, I came up with this conjecture.

Edit: The generalized conjecture ($2n+1$ points and $n$ circles) is not true.

Edit 2: I have asked, and answered, a similar question. Maybe it might provide ideas for this question.

Edit 3 Posted on MO.

  • 1
    $\begingroup$ I'm saying that this kind of problem is decidable, so a computer can automatically find a proof (or a counterexample), perhaps even in reasonable time. I imagine it'd be a lot of case analysis, though, so it might not be human readable. I might try. $\endgroup$ Nov 19 at 10:13
  • 1
    $\begingroup$ I wrote a Mathematica notebook computing this expression but the cylindrical decomposition is too slow for Wolfram Cloud's free plan. $\endgroup$ Nov 19 at 11:27
  • 1
    $\begingroup$ Just noticed it's also wrong, it looks at the sum of squared distances instead of the square of the sum of distances $\endgroup$ Nov 19 at 11:35
  • 2
    $\begingroup$ The following idea can be promising. If some three of the given points belong to a straight-line then any distinct circles with the diameters spanned by these points are nonintersecting. Otherwise the points are in general position and by the happy ending problem we can choose four points which are vertices of the convex quadrilateral. $\endgroup$ Nov 20 at 21:58
  • 3
    $\begingroup$ If the bounty wraps up without getting an answer, I recommend posting this at mathoverflow. I think it would be well-received! $\endgroup$ Nov 25 at 20:20

7 Answers 7


A counterexample is given by the following five points: $$(0,0),(1,0), \Big(-\frac{64867}{77629},\frac{3389}{60094}\Big), \Big(\frac{5981}{56176},\frac{32211}{34172}\Big), \Big(\frac{5925}{117812},-\frac{103221}{116516}\Big).$$

Calculations and the way this example was found are presented in this pdf image of a Mathematica notebook. In particular, it is noted there that a certain "defect" of this example, which makes it a counterexample, is very small, about $3\times10^{-6}$. So, here one can have here two "very slightly intersecting" circles, their intersection certainly invisible to the eye if both circles are fully shown.

  • $\begingroup$ Brilliant! Here is the desmos graph I used to check all fifteen cases. $\endgroup$
    – Dan
    2 days ago
  • $\begingroup$ After some local random hand searching near your counterexample I found (0,0), (1,0),(-0.8357,0.0564), (0.1065,0.9430), (0.0503,-0.8861) with a defect of 5.5e-7. $\endgroup$ 17 hours ago
  • $\begingroup$ @LewisBaxter : Thank you for your comment. $\endgroup$ 16 hours ago

Here is a way to turn it from geometry into algebra without square-roots. I don't have a solution either.
Given four points $(a,b),(c,d)$ and $(e,f),(g,h)$, we are looking for a point $(x,y)$ on two circles, so two dot products are zero: $$(x-a,y-b).(x-c,y-d)=0\\ (x-e,y-f).(x-g,y-h)=0$$ The difference between these is linear so $(x,y)$ lies on a straight line and can be parameterised as $$(x,y)=\left(\frac{eg-ac+t}{e+g-a-c},\frac{fh-bd-t}{f+h-b-d}\right)$$ where $t$ is free. The two equations at the start become a single quadratic in $t$, and the point $(x,y)$ exists if the discriminant is nonnegative.
The numerator I get for the determinant is
$$-4(eg-ac+fh-bd)^2+((a+c)(f+h)-(e+g)(b+d))^2 +4(e+g-a-c)(eg(e+g)-ac(a+c)) +4(f+h-b-d)(fh(f+h)-bd(b+d))\lt0$$ There are fifteen sets of four points paired up. Is it possible to show not all fifteen discriminants can be positive?


This is just some ideas, enter image description here

Consider the two points such that the distance between them is maximum, In the diagram, I call those two points $C$ and $B$, we will work in the complex plane with the origin as $O$, the midpoint of $CB$.

There is a special blue circle inside my diagram, centered at $O$ is such that if it had a radius of $1$, then circle $BC$ will have a radius of $\sqrt2$. (Scale accordingly)

This blue circle has the property of if any two points lie within this circle, then the circle formed by these two points necessarily do not intersect with circle $CB$. Proof: Consider any two points inside or on the unit circle call $x,y$ Then the point on the circle formed by $xy$ furthest away from the origin has magnitude: (1) $|(x+y)/2+i(x-y)/2|$ where i is the imaginary unit. Then bound using triangle inequality:

$(1)\le(0.5)(|x||1+i|+|y||1-i|)\le 0.5(2\sqrt2)=\sqrt2$

With this, we see that it is necessary that at most 1 point can lie within the blue circle.

Then I tried to bash 3 cases,

(1) Two points lie outside of circle $CB$.

(2) One point lie outside of circle $CB$ and one between circle $CB$ and the blue circle

(3) Two points lie between circle $CB$ and the blue circle.

I guess a way to approach the rest is to complex bash the magnitudes.

Consider any four complex points, $m,n,q,r$ . Then the circles formed by $mn$ and $qr$ necessarily intersect if

(Distance beteween the centers)< (The sum of the individual radiuses)

I know that this claim does not work when a circle is contained inside another, that is why I chose $CB$ to be the largest as I conjecture that after the argument above, there should be no circle inside circle configuration.

Partial progress, will edit if there are any more progress.


EDIT 2: One excellent counterexample has been provided by another anwser, the four points of which are nearly a square, resulting in the "very slightly intersecting". I ever tried to find one starting from a square but failed. It turns out my final success was the idea "the example that is 'closest' to an counterexample is the four points of a square together with its center".

EDIT: The following "proof" is wrong, so don't care much about it :P

Now I think the example that is "closest" to an counterexample is the four points of a square together with its center. In this case the circles intersect with each other, or are tangent. (However tangent circles are considered to be non-intersecting.)

Given five points, there are ten circles. We can imagine how difficult it is to let them intersect with each other. It is not hard to deal with the first four points, but the fifth point will introduce four more circles and we should consider each of the four intersecting or not with the original six circles. (You may understand better by trying to find an example where only one pair of circles are allowed to intersect with each other.)

This is not a proof but just provide some ideas.

Denote the five points as $A,B,C,D,E$. Suppose no three of them are collinear, otherwise the conclusion holds immediately.

First we pick $A,B$ such that they have the minimum distance. Let $P$ be the midpoint of $AB$.
Among $C,D,E$ assume $C$ is the farthest one to $P$. Then we can draw some circles.
In the following figure, the red circles means $D,E$ must be outside them, since $AB$ is the shortest, the green circles means $D,E$ must be inside it, since $C$ is the one farthest from $P$, and the annulus with dashed lines means that $D$ must be inside it if we want the circle with diameter $CD$ intersects with the one with diamter $AB$.

Now we can suppose both $D,E$ are in the annulus. We can observe that only two small regions are left for them. In fact, the actual regions such that circle($CD$ or $CE$) and circle($AB$) intersects is much smaller, and one of them is approximately drawn in black.
If $D,E$ are in the same regions (e.g. both in the black region) then circle($DE$) do not intersect with circle $AB$. If $D,E$ are in different regions then circle($DE$) will cover circle($AB$).

An illustration

New contributor
ImbalanceDream is a new contributor to this site. Take care in asking for clarification, commenting, and answering. Check out our Code of Conduct.

Some Ideas , which might lead to Improvements , variations & ETC.

[[ I am not sure whether this will be fruitful , I am throwing some Ideas here. ]]

When we have 2 Points $A=(A_x,A_y)$ & $(B=B_x,B_y)$ in the Cartesian Plane , the Center of the Circle 1 with Diameter $AB$ is $[A+B]/2$.

Likewise for $C$ & $D$ : Center of Circle 2 is $[C+D/2$

We can get the radius like $|[A-B]/2|$ & $|[C-D]/2|$

The Circles will not intersect if 1 contains 2 wholly or if 1 is very far away from 2. Other-wise , there will be some intersection.

1 contains 2 : Distance between Centers is less than Difference between radii
1 is very far from 2 : Distance between Centers is more than Sum of radii

Intersecting Case :
Difference between radii is less than Distance between Centers Sum of radii is more than Distance between Centers


In other words , formally :
Distance between Centers : $[ A + B ] / 2 - [ C + D ] / 2 \tag{1}$
Sum of radii : $[ A - B ] / 2 + [ C - D ] / 2 \tag{2}$
Difference between radii : $[ A - B ] / 2 - [ C - D ] / 2 \tag{3}$

Necessary Condition that Intersection occurs :
$| [ A - B ] / 2 - [ C - D ] / 2 | \le ||[ A + B ] / 2| - |[ C + D ] / 2|| \le |[ A - B ] / 2 + [ C - D ] / 2| \tag{4}$

That looks complicated , but the English language version given earlier is quite simple.
We could even write it like this :
$| rad(AB) - rad(CD) | \le DIST(AB,CD) \le | rad(AB) + rad(CD) | \tag{5}$

Now , we can use 4 out of the 5 Points $A,B,C,D,E$ "cyclically" & that relation must hold if there is some intersection.
We will have 15 ( Combinatoric Calculation ) such relations , counting with uniqueness.

IDEA 1 : What we want to Prove is : that relation can hold for some , but not all 15.

IDEA 2 : Alternately : It is a Contradiction when all 15 relations hold.

IDEA 3 : We could try to add up all those relations to get "cyclical" relation which we want to claim is false.
$\sum | rad(AB) - rad(CD) | \sum \le DIST(AB,CD) \le \sum | rad(AB) + rad(CD) | \tag{6}$

Summation is over the "cyclical" choice of 4 Points out of all 5 Points.

[[ I am not sure whether this will be fruitful , I am throwing some Ideas here. ]]


I believe I found a proof, which relies in the following two statements:

  1. Given a set of five non-collinear points on the Euclidean plane, it is always possible to find a convex subset of 4 points.

  2. Given a convex subset of 4 points on the Euclidean plane, there are always two opposite edges the midpoint of which are separated by a distance equal or larger than the average of their lengths.

From these statements it is possible to do a proof by construction.

The remaining case with three or more collinear points has trivial solution by choosing any two pairs of points on the line and constructing two circles, which will necessarily kiss at the common point.

Sketch proof of theorem 1: Consider three arbitrary non-collinear points on the plane, (p,q,r). Add a fourth non-collinear point, s. If s lies outside the (p,q,r) triangle we found a convex quadrilateral (p,q,r,s). Otherwise, we have decomposed the (p,q,r) triangle into three triangles, (p,q,s), (q,r,s) and (p,r,s). Add a fifth point t. If t lies outside the original (p,q,r) triangle, we found the convex quadrilateral (p,q,r,t). Otherwise, t HAS TO lie inside one of the smaller triangles, say (p,q,s), which in turn means it lies outside both (q,r,s) and (p,r,s), from which we can build either convex quadrilateral (q,r,s,t) or (p,r,s,t).

  • $\begingroup$ Quote: "... the common point". The OP states that we must choose four points. $\endgroup$
    – Dan
    Nov 27 at 21:35
  • 4
    $\begingroup$ Statement 2 is false; consider the points $(\pm 1, 0), (0,\sqrt{3}), (0,-\epsilon)$. $\endgroup$ Nov 28 at 10:11
  • $\begingroup$ @RavenClawPrefect You are correct. I have come to the conclusion there is only one family of configurations in which all pairs of circles overlap, similar to OPs second example, and in all those cases it is possible to fully include one disk inside another. $\endgroup$ Nov 28 at 13:42
  • $\begingroup$ mathoverflow.net/a/459369/8133 $\endgroup$ 2 days ago

HINT.- A method to solve this problem with the help of a computer

Five planar points $P_i=(x_i,y_i);\space 1\le i\le5$, been given, the circle having $\overline{P_iP_j}$ as diameter has equation $$x^2+y^2-(x_i+x_j)x-(y_i+y_j)y+x_ix_j+y_iy_j=0$$ It follows that if the given condition is true, then, among the ten possible circles as desired, there are at least two circles either, disjoints or tangents. Since the resultant of two equations $$x^2+y^2-(x_i+x_j)x-(y_i+y_j)y+x_ix_j+y_iy_j=0\quad(1)\\x^2+y^2-(x_h+x_k)x-(y_h+y_k)y+x_hx_k+y_hy_k=0\quad(2) $$ is of the second degree so we need a system of a pair of the ten circles such that the discriminant of their resultant be negative or null (which correspond to disjoint or tangent circles). This is easy to do with fixed five points but for variable ones we rather can use a little program for computer because the algorithm is tedious.

  • $\begingroup$ mathoverflow.net/a/459369/8133 $\endgroup$ 2 days ago
  • $\begingroup$ @Noah Schweber: Tanks very much for your reference: Keeping into account that I have written: It follows that if the given condition is true, we can see that I did not discarded the conjecture would be false. With your counterexample, a simple calculation will give all discriminant positive (Sorry for bad English without the Google translator) $\endgroup$
    – Piquito
    2 days ago

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .