Empty red-red-blue triangles

Assume that $|B(\mathcal{C})|>0$ and $\operatorname{disc}(\mathcal{C})>0$, for otherwise the statement is trivially true. Fix an arbitrary blue point $b\in B(\mathcal{C})$. We claim that $b$ contributes at least $\operatorname{disc}(\mathcal{C})$ empty red-red-blue triangles in $\mathcal{C}$. Then the lemma is proved by summing over $b$.

To see this, we draw a ray from $b$ to every red point in $R(\mathcal{C})$. These $|R(\mathcal{C})|$ rays partition $\mathcal{C}$ into $|R(\mathcal{C})|$ sectors. In each sector containing no blue point (later called empty sector), $b$ forms an empty red-red-blue triangle with the two red points on the boundary of the sector. However, there is at most one exception, which is the unique (if exists) sector whose angle at the apex is bigger than $\pi$. Since there are exactly $|B(\mathcal{C})|-1$ blue points other than $b$ in $\mathcal{C}$, the number of empty sectors is bounded from below by $|R(\mathcal{C})|-\bigl{(}|B(\mathcal{C})|-1\bigr{)}=\operatorname{disc}(\mathcal{C})+1$, and so the claim is true. ∎

Fix an arbitrary red point $r\in R$ and an arbitrary blue sector $\mathcal{S}$ with respect to $r$. Assume $\mathcal{S}$ is the sector between rays $\overrightarrow{rr_{1}}$ and $\overrightarrow{rr_{2}}$, where $r_{1},r_{2}\in R$. Pick $b\in\mathcal{S}\cap B$ so that the smaller one of the angles $\angle\bigl{(}\overrightarrow{\mathstrut rb},\overrightarrow{\mathstrut rr_{1}}\bigr{)},\angle\bigl{(}\overrightarrow{\mathstrut rb},\overrightarrow{\mathstrut rr_{2}}\bigr{)}$ is minimized. Note that one of the angles is possibly bigger than $\pi$, but the smaller one cannot. Assume without loss of generality that the minimum is achieved at $\angle\bigl{(}\overrightarrow{\mathstrut rb},\overrightarrow{\mathstrut rr_{1}}\bigr{)}$. It follows that $r,r_{1},b$ form an empty red-red-blue triangle. By summing over all blue sectors with respect to $r$ and all red points $r$, we obtain at least $pn$ empty red-red-blue triangles. Since each such triangle is counted at most twice (from its two red vertices), we obtain $\mathcal{N}_{\mathsf{r}\mathsf{r}\mathsf{b}}(\mathbb{R}^{2})\geq pn/2$. ∎

Lemma 3 implies that we are done if $p>\sqrt{n}$. Assume $p\leq\sqrt{n}$ then, and fix a red point $r_{0}\in R$ with $p(r_{0})=p$. For convenience purpose, we suppose that $n\geq 5$.

Draw a line $\ell$ through $r_{0}$ that bisects $R$. Formally, we fix a line $\ell$ through $r_{0}$ that separates the plane into two closed half-planes $\mathcal{H}_{+},\mathcal{H}_{-}$ with $|R(\mathcal{H}_{+})|=|R(\mathcal{H}_{-})|=\bigl{\lceil}(n+1)/2\bigr{\rceil}\geq 3$. Depending on the parity of $n$, our $\ell$ goes through $0$ or $1$ point of $R$ other than $r_{0}$. Due to the discrete intermediate value theorem, the existence of $\ell$ is obvious. By applying a possible perturbation, we assume further that $\ell$ goes through no point of $B$. Introduce coordinates so that $r_{0}$ is the origin and $\ell$ is the $x$-axis. Choose the direction of the $x$-axis appropriately so that $\mathcal{H}\stackrel{{\scriptstyle\mbox{\tiny def}}}{{=}}\bigl{\{}(x,y)\in\mathbb{R}^{2}:y\geq 0\bigr{\}}$ contains at least a half of the blue points (i.e., $|B(\mathcal{H})|\geq\left\lceil m/2\right\rceil$). Observe that $|B(\mathcal{H})|\geq|R(\mathcal{H})|-1$.

Consider the following algorithm: At the beginning, a ray $\bm{v}$ starts from the positive half of the $x$-axis $\bigl{\{}(x,0):x\geq 0\bigr{\}}$ and rotates counterclockwise around the origin $r_{0}$ to scan through $\mathcal{H}$. This ray will generate a collection of good sectors, where a good sector is a sector $\mathcal{G}$ with $|B(\mathcal{G})|\geq|R(\mathcal{G})|/3$.

• Step $i$:

  After Step $i-1$ (the beginning if $i=1$), when the ray $\bm{v}$ hits the first red point, we label $\bm{v}$ at this moment as $\bm{v}_{i}^{-}$. At the first time when $\bm{v}$ hits some blue point so that the number of blue points is bigger than or equal to a third of the number of red points in the sector between $\bm{v}_{i}^{-}$ and $\bm{v}$, we label $\bm{v}$ at this moment as $\bm{v}_{i}^{+}$. If both $\bm{v}_{i}^{-}$ and $\bm{v}_{i}^{+}$ exist, then $\mathcal{G}_{i}$ is defined as the sector between them. If $\bm{v}$ arrives the negative half of the $x$-axis before one of $\bm{v}_{i}^{-},\bm{v}_{i}^{+}$ appears, then $\mathcal{G}_{i}$ is not generated and the algorithm terminates.

Let $\mathcal{G}_{1},\dots,\mathcal{G}_{k}$ (each containing $r_{0}$) be all good sectors generated by the algorithm. This means that the algorithm terminates in Step $k+1$. Denote by $\mathcal{T}$ the sector between $\bm{v}_{k+1}^{-}$ and the negative half of the $x$-axis. If no red points other than $r_{0}$ exist after $\mathcal{G}_{k}$, then $\mathcal{T}$ denotes the negative half of the $x$-axis (hence $|R(\mathcal{T})|=1$ and $|B(\mathcal{T})|=0$). See Figure 2 for an illustration.

To finish the proof, we need to analyze some detailed properties of our algorithm:

• •

  Each $\bm{v}_{i}^{-}$ comes from a distinct red sector with respect to $r_{0}$, and so

  $$k\leq p\leq\sqrt{n}.$$

  (1)

• •

  The way we defined $\mathcal{G}_{i}$ implies that $\bm{v}_{i}^{-}\in\mathcal{G}_{i},\,r_{0}\in\mathcal{G}_{i}$, and $|B(\mathcal{G}_{i})|=\bigl{\lceil}|R(\mathcal{G}_{i})|/3\bigr{\rceil}$. Then

  $$\operatorname{disc}(\mathcal{G}_{i})=|R(\mathcal{G}_{i})|-|B(\mathcal{G}_{i})|=|R(\mathcal{G}_{i})|-\bigl{\lceil}|R(\mathcal{G}_{i})|/3\bigr{\rceil},$$

  and hence from Lemma 2 and the fact $|R(\mathcal{G}_{i})|\geq 2$ we deduce that

  $$\mathcal{N}_{\mathsf{r}\mathsf{r}\mathsf{b}}(\mathcal{G}_{i})\geq|B(\mathcal{G}_{i})|\cdot\operatorname{disc}(\mathcal{G}_{i})\geq\frac{2}{9}|R(\mathcal{G}_{i})|^{2}.$$

  (2)

  Here we applied the fact that $\left\lceil x/3\right\rceil\cdot\bigl{(}x-\left\lceil x/3\right\rceil\bigr{)}\geq 2x^{2}/9$ holds for any integer $x\geq 2$.

• •

  Observe that the sector $\mathcal{T}$ contains all red points in $R(\mathcal{H})$ that are not in $\mathcal{G}_{1}\cup\dots\cup\mathcal{G}_{k}$, and $|B(\mathcal{T})|<|R(\mathcal{T})|/3$. Reflect $R\cup B$ over the $y$-axis and conduct the same algorithm to generate good sectors $\mathcal{G}_{1}^{\prime},\dots,\mathcal{G}_{k^{\prime}}^{\prime}$. Define $\mathcal{T}^{\prime}$ in the same way (with respect to the new family of good sectors). Then we reflect everything over the $y$-axis again to work on the original picture. See Figure 3 for an illustration of the same example as in Figure 2.

• Similarly, $|B(\mathcal{T}^{\prime})|<|R(\mathcal{T}^{\prime})|/3$. We claim that $\mathcal{T}\cap\mathcal{T}^{\prime}=\{r_{0}\}$. If not, then $\mathcal{T}\cup\mathcal{T}^{\prime}=\mathcal{H}$, and so

  $$|B(\mathcal{H})|\leq|B(\mathcal{T})|+|B(\mathcal{T}^{\prime})|<\frac{|R(\mathcal{T})|}{3}+\frac{|R(\mathcal{T}^{\prime})|}{3}\leq\frac{2|R(\mathcal{H})|}{3}\leq|R(\mathcal{H})|-1,$$

  where in the last inequality we applied $|R(\mathcal{H})|\geq 3$. This contradicts $|B(\mathcal{H})|\geq|R(\mathcal{H})|-1$. It then follows that $|R(\mathcal{T})|+|R(\mathcal{T}^{\prime})|\leq|R(\mathcal{H})|+1$, as $r_{0}$ is counted twice. Assume without loss of generality that $|R(\mathcal{T})|\leq\bigl{(}|R(\mathcal{H})|+1\bigr{)}/2$. Since $|R(\mathcal{H})|\geq 2$, we have $|R(\mathcal{T})|<|R(\mathcal{H})|$, and hence $k\geq 1$. By counting $r_{0}$ to the multiplicity $k+1$, we derive that

  $$|R(\mathcal{G}_{1})|+\dots+|R(\mathcal{G}_{k})|+|R(\mathcal{T})|=|R(\mathcal{H})|+k.$$

  This together with $R(\mathcal{T})\leq\bigl{(}|R(\mathcal{H})|+1\bigr{)}/2,\,k\geq 1$, and $|R(\mathcal{H})|=\bigl{\lceil}(n+1)/2\bigr{\rceil}$ implies that

  $$\sum_{i=1}^{k}|R(\mathcal{G}_{i})|\geq\frac{|R(\mathcal{H})|-1}{2}+k\geq\frac{n}{4}\implies\frac{1}{k}\sum_{i=1}^{k}|R(\mathcal{G}_{i})|\geq\frac{n}{4k}.$$

  (3)

By putting all the ingredients from above together, we conclude that

The proof of Theorem 1 is complete. ∎