24
$\begingroup$

Let $\lambda$ denote the Lebesgue-measure on $\mathbb{R}^n$, and let $C\subset\mathbb{R}^n$ be a convex region.

My question is about $$f(C):=\int_{C} \lambda(C \cap (x + C) ) \mathrm{d} x.$$ How large can $f(C)$ be? Of course, there is the trivial bound $f(C)\leq \lambda(C)^2$ but I would expect more something like $f(C) = O( \lambda(C) )$ at least. This question has probably been answered somewhere, and I suspect that the following might be very well known: for regions $C$ of constant measure, the value of $f(C)$ is maximal if $C$ is an $n$-ball (or something similar). I couldn't find any useful references regarding this question. If someone could help me out with some reference, or even a simple solution, I would be very grateful.

$\endgroup$
6
  • 8
    $\begingroup$ $O(\lambda(C))$ estimate can not hold and it has not even physical meaning, since two expressions are homogeneous with different exponents. $\endgroup$ Oct 2, 2017 at 13:53
  • $\begingroup$ @FedorPetrov Yes, you're right. I completely missed that. It would be nice to know what the extremal case looks like still. $\endgroup$
    – zref
    Oct 2, 2017 at 15:49
  • 4
    $\begingroup$ For $C$ a circle, I calculate $f(C)/\lambda(C)^2=1-3\sqrt{3}/4\pi$, roughly 0.5865 $\endgroup$
    – Matt F.
    Oct 3, 2017 at 10:44
  • 2
    $\begingroup$ For the ball in $R^n$, the corresponding formula seems to be $$\frac{\int_{0}^1 s^{n-1} V_n^{cap}(1,s/2)ds}{V_n^{cap}(1,0)/n}$$ where $V$ is as in math.stackexchange.com/a/162873/210229 $\endgroup$
    – Matt F.
    Oct 3, 2017 at 12:37
  • 2
    $\begingroup$ Looks like the optimal shapes are balls centered at the origin. Indeed, for 3 intervals $I,J,K\subset\mathbb R$ of given lengths, the integral $\int_I|J\cap(x+K)|\,dx$ is maximized when $I,J,K$ are all centered at the origin, so the Steiner symmetrization should do the job. On the other hand, I can easily miss something at 3AM, so it would be nice if someone else could verify (or refute) this claim :-) $\endgroup$
    – fedja
    Oct 4, 2017 at 7:06

3 Answers 3

23
$\begingroup$

The following proposition answers OP's question regarding the upper bound of $$\tau(C) \Doteq f(C)/\lambda^2(C).$$

Let $B_n$ be the closed Euclidean unit ball of $\mathbb{R^n}$ centred at $0$, that is $$B_n = \{ (x_1,\dots, x_n) \in \mathbb{R^n} \, \vert \,\, x_1^2 + \cdots + x_n^2 \le 1\},$$ and let $\tau_n = \tau(B_n)$.

Proposition. The Euclidean unit ball $B_n \subset \mathbb{R}^n$ maximizes $\tau(C)$ among all convex subsets of finite and positive Lebesgue measure, i.e., $\tau(C) \le \tau_n$ holds for any such $C$.

Claim 3 below gives a way to compute efficiently $\tau_n$ for all $n > 0$.

The proof of the proposition follows Fedja's idea. The first part consists in proving that $\tau$ doesn't decrease after [Steiner symmetrization][2]. This is Claim 4 below. The second part uses [1, Theorem 7.1.10] showing that some Euclidean ball centred at $0$ can be obtained as the limit of Steiner symmetrizations of $C$ for any compact convex $C$ of positive Lebesgue measure (equivalently, with non-empty interior).

When $n = 1$, the result is slightly stronger.

Claim 1. Let $C \subset \mathbb{R}$ be a convex subset of finite Lebesgue measure $\lambda(C) > 0$. Then $\tau(C) \le \frac{3}{4}$ and equality holds if and only if $C$ is centred at $0$.

So, the above claim settles OP's question when $n = 1$. Unsurprisingly, the result generalizes to $\mathbb{R}^n$ in the following way.

Claim 2. Let $C \subset \mathbb{R}^n$ be a parallelotope of positive Lebesgue measure. Then $\tau(C) \le (\frac{3}{4})^n$ and equality holds if and only if $C$ is centred at $0$.

Computing the maximum $\tau_n$ of $\tau$ can be achieved via the formula

Claim 3. We have $$ \tau_n = \frac{3}{2} \frac{\int^{\pi/3}_0 \sin^n(\theta) d\theta}{\int^{\pi/2}_0 \sin^n(\theta) d\theta}.$$ In particular, $\tau_1 = \frac{3}{4}, \, \tau_2 = 1 - \frac{3 \sqrt{3}}{4 \pi}, \tau_3 = \frac{15}{32}, \, \tau_4 = 1 - \frac{9 \sqrt{3}}{8 \pi}, \tau_5 = \frac{159}{512}, \, \tau_6 = 1 - \frac{27 \sqrt{3}}{20 \pi} $.

The above formula can be rephrased as $\tau_n = \frac{3}{2}\frac{W_n'}{W_n}$ where $W_n$ is the $n$-th [Wallis' integral][3] and $W_n'$ is recursively defined by $W'_0 = \frac{\pi}{3}, \,W_1' = \frac{1}{2}$ and $W_n' = - \frac{1}{2n}(\frac{\sqrt{3}}{2})^{n - 1} + \frac{n - 1}{n}W'_{n - 2}$. The latter formula allows an efficient computation of $\tau_n$ for any $n > 0$. As $W_n' \le \frac{\pi}{3}(\frac{\sqrt{3}}{2})^n$ and $W_n \sim \sqrt{\frac{\pi}{2n}}$, the sequence $(\tau_n)_n$ converges to $0$ exponentially fast (we have for instance $10^{-7} < \tau_{91} < 10^{-6}$).

Before proving the claims, let us state some obvious, but important, remarks.

Remark 1. Let $A \in GL_n(\mathbb{R})$ and $d \in \mathbb{R}^n$. Let $C$ be any subset of $\mathbb{R}^n$. Then we have $$f(AC) = \vert \det(A) \vert^2f(C), \quad f(C + d) = \int_C \lambda\left(C \cap (C + d + x) \right)d\lambda(x).$$

In particular the ratio $\tau(C)$ is $GL_n(\mathbb{R})$-invariant but it is not invariant under translation. Indeed, if $C$ is a convex subset of finite positive Lebesgue measure whose interior contains $0$, then the function $f_C(d): d \mapsto f(C + d)$ has a non-empty and bounded support.

The following tightly relates to OP's question.

Question. Let $C \subset \mathbb{R}^n $ be a convex subset of finite positive Lebesgue measure. For which $d$ is $f_C(d)$ maximal?

The following relates $\tau$ to Steiner symmetrization.

Claim 4. Let $C \subset \mathbb{R}^n $ be a convex subset of finite positive Lebesgue measure. Let $H =\{0\} \times \mathbb{R}^{n -1}$ and let $C' \Doteq \text{St}_H(C)$ be the [Steiner symmetrization][2] of $C$ with respect to $H$. Then we have $\tau(C) \le \tau(C')$.

We now turn to the proofs of the above claims.

Proof of Claim 1. Because of Remark 1, we can restrict, without loss of generality, to intervals of the form $[-1, 1] + d$ with $d \ge 0$. A direct computation yields $$ f([-1, 1] + d) = \int_{-1}^1 (2 - \vert d + x \vert)^{+} dx = \left\{\begin{array}{cc} 3 -d^2 & \text{ if } 0 \le d \le 1, \\ \frac{1}{2}(3 - d)^2 & \text{ if } 1 \le d \le 3, \\ 0 & \text{ if } d \ge 3. \end{array}\right. $$ where $g^{+} \Doteq \max(g ,0)$ for $g$ a real-valued function. The result follows.

Proof of Claim 2. By Remark 1, we can restrict to the convex sets of the form $C = [-1, 1]^n + d$ with $d = (d_1, \dots, d_n) \in \mathbb{R}^n$. We have $f(C) = \int_{[-1, 1]^n} (2 - \vert x_1 + d_1 \vert)^{+} \cdots (2 - \vert x_n + d_n\vert)^{+}dx_1 \cdots dx_n$. It follows from Claim 1, that $f(C) \le \int_{[-1, 1]^n} (2 - \vert x_1 \vert)^{+} \cdots (2 - \vert x_n \vert)^{+}dx_1 \cdots dx_n = 3^n$ and the inequality is strict whenever one of the $d_i$'s is not zero. The result follows.

We denote by $V_n(r)$ the Lebesgue measure of the Euclidean ball of radius $r \ge 0$, that is, $V_n(r) = r^n\frac{\pi^{\frac{n}{2}}}{\Gamma(\frac{n}{2} + 1)}$, and let $\nu_n = V_n(1) = \lambda(B_n)$. Given a real number $h$ such that $-1 \le h \le 1$, we denote by $B_n^{\text{cap}}(h)$ the hyperspherical cap of height $h$ along $x_n$ , that is the set of points $x = (x_1, \dots, x_n) \in B_n$ such that $x_n \ge h$. We also set $V_n^{\text{cap}}(h) = \lambda(B_n^{\text{cap}}(h))$.

Proof of Claim 3. By definition, we have $f(B_n) = \int_{B_n} \lambda(B_n \cap (B_n + x)) d\lambda(x) = 2 \int_{B_n} V_n^{\text{cap}}(\frac{\vert x \vert}{2}) d\lambda(x)$ where $\vert \cdot \vert$ stands for the Euclidean norm of $\mathbb{R}^n$. Changing to [$n$-dimensional spherical coordinates][4], we get $$f(B_n) = 2 \int_{D_n} V_n^{\text{cap}}(\frac{r}{2})r^{n - 1} \sin^{n - 2}(\phi_1)\sin^{n - 3}(\phi_2) \cdots \sin(\phi_{n - 1}) dr d\phi_1 d\phi_2 \cdots d \phi_{n - 1}$$ where $D_n = [0,2\pi] \times [0, \pi] \times \cdots \times [0, \pi] \times [0, 1]$. Using the classical computation of $\nu_n = \lambda(B_n)$, we deduce that $$f(B_n) = 2n\nu_n \int_{[0, 1]} V_n^{\text{cap}}(\frac{r}{2})r^{n - 1}dr = 2^{n + 1}n\nu_n \int_{[0, \frac{1}{2}]} V_n^{\text{cap}}(s)s^{n - 1}ds .$$ Since $V_n^{\text{cap}}(s) = \int_s^1 V_{n - 1}(\sqrt{1 - h^2})dh = \nu_{n - 1} \int_0^{\text{arcos}(s)} \sin^n(\theta)d \theta$, we obtain $$f(B_n) = 2^{n + 1}n\nu_n \nu_{n - 1} \int_{[0, \frac{1}{2}]} J_n(\text{arcos}(s))s^{n - 1}ds$$ where $J_n(\phi) = \int_0^{\phi}\sin^n(\theta)d\theta$. Using the substitution $\theta = \text{arcos}(s)$, we get $$ \begin{eqnarray} \frac{f(B_n)}{2^{n + 1}n\nu_n \nu_{n - 1}} &=& \int_{[\frac{\pi}{3}, \frac{\pi}{2}]} J_n(\theta)\cos(\theta)^{n - 1} \sin(\theta)d\theta \\ &=& \frac{[J_n(\theta) \cos^n(\theta)]^{\frac{\pi}{3}}_{\frac{\pi}{2}} + \int^{\frac{\pi}{2}}_{\frac{\pi}{3}} (\cos(\theta)\sin(\theta))^n d\theta}{n} \\ &=& \frac{2J_n(\frac{\pi}{3}) + \int^{\pi}_{\frac{2\pi}{3}} \sin(\theta)^n d\theta}{2^{n + 1} n} \\ &=& \frac{3J_n(\frac{\pi}{3})}{2^{n + 1} n}. \end{eqnarray} $$ Now the result follows from the identity $\frac{\nu_{n - 1}}{\nu_n} = \frac{1}{2 W_n}$.

Claim 4 is a specialization of

Fedja's Lemma. Let $$\mathcal{I}(C, D, E) \Doteq \int_C \lambda(D \cap (E + x)) d\lambda(x)$$ where $C$, $D$ and $E$ are convex subsets of $\mathbb{R}^n$ of positive finite Lebesgue measure. Then $\mathcal{I}(C, D, E) \le \mathcal{I}(C', D', E')$, where $X'$ denotes the Steiner symmetrization of $X$ with respect to $H = \{0\} \times \mathbb{R}^{n - 1}$.

The proof of Fedja's Lemma can be reduced to the case $n = 1$. The proof of the case $n = 1$ is eased off by

Lemma 1. Let $g: \mathbb{R} \rightarrow \mathbb{R}^{+} \cup \{0\}$ be an even non-negative function which is concave on its support and non-decreasing on the negative real numbers. Then we have $$\int_I g d\lambda \le \int_{I'} g d\lambda $$ for every convex subset $I$ of $\mathbb{R}$ where $I'$ is the closed interval of measure $\lambda(I)$ centred at $0$.

Proof of Lemma 1. The non-negative number $\int_I g d\lambda$ is the area of convex subset $C(I) \subset \mathbb{R}^2$ located between the $x$-axis and the graph of $g$. By hypothesis, we have $C(I)' \subset C(I') = \int_{I'} g d\lambda$, where $C(I)'$ is Steiner symmetrization of $C(I)$ with respect to $H = \{0\} \times \mathbb{R}$. As $C(I)$ and $C(I)'$ have the same area, the result follows.

We are now in position to prove Fedja's Lemma.

Proof of Fedja's Lemma. Let us assume first that $n = 1$. As $\mathcal{I}(C + x, D, E) = \mathcal{I}(C, D, E + x)$ and $\mathcal{I}(C, D + x, E + x) = \mathcal{I}(C, D, E)$, we can assume, without loss of generality, that $C$ and $D$ are centred at zero. Let us write $E = E’ + \eta$ and let $g(x) \Doteq \lambda(D \cap (E' + x))$. Then we have $$ g(x) = \left\lbrace \begin{array}{ccc} \min(d, e) & \text{ if } \vert x \vert \le \frac{ \vert d - e \vert }{2}, \\ \frac{d + e}{2} - \vert x \vert & \text{ if } \frac{ \vert d - e \vert }{2} \le \vert x \vert \le \frac{d + e}{2}, \\ 0 & \text{ if } \vert x \vert \ge \frac{d + e}{2}, \end{array}\right. $$ where $d = \lambda(D) > 0$ and $e = \lambda(E) > 0$. The function $g$ satisfies the hypothesis of Lemma $1$, thus $$ \mathcal{I}(C, D, E) = \mathcal{I}(C + \eta, D, E’) = \int_{C + \eta} g(x) d\lambda(x) \le \int_{C} g(x) d\lambda(x) = \mathcal{I}(C, D, E’)$$ which proves the result for $n = 1$. Assume now that $n > 1$. We denote by $\lambda_{n - 1}$ the Lebesgue measure on $\mathbb{R}^{n - 1}$ and we identify $H$ with $\mathbb{R}^{n - 1}$. For $\xi \in \mathbb{R^n}$, we write $\xi = (\xi_1, \zeta)$, with $\xi_1 \in \mathbb{R}$ and $\zeta \in H$. Given $A \subset \mathbb{R}^n$ and $\zeta \in H$, we set $A^{\zeta} \Doteq \{ \xi_1 \, \vert \, (\xi_1, \zeta) \in A\} \subset \mathbb{R}$. Let $x = (x_1, z) \in \mathbb{R} \times H$. Observe that $(A^{\zeta})’ = (A’)^{\zeta}$ and $(A + x)^{\zeta} = A^{\zeta - z} + x_1$. By the Fubini-Tonelli Theorem, we have $\lambda_n(D \cap (E + x)) = \int_H \lambda_1(D^{\zeta} \cap (E^{\zeta - z} + x_1)) d \lambda_{n - 1}(\zeta)$. We shall use the abbreviations $dz = d \lambda_{n - 1}(z)$ and $d\zeta = d \lambda_{n - 1}(\zeta)$. Using Fubini-Tonelli’s Theorem and the case $n = 1$, we obtain $$ \begin{eqnarray} \mathcal{I}(C, D, E) & = & \int_{H^2} \mathcal{I}(C^{z}, D^{\zeta}, E^{\zeta - z})d\zeta dz \\ & \le & \int_{H^2} \mathcal{I}((C’)^{z}, (D’)^{\zeta}, (E’)^{\zeta - z}) d\zeta dz \\ & = & \mathcal{I}(C’, D’, E’). \end{eqnarray} $$

Eventually, we can prove the proposition.

Proposition's proof. By [1, Theorem 7.1.10], there is a Euclidean ball $B$ centred at zero which is the limit in the Hausdorff distance topology of Steiner symmetrizations of $C$ with respect to linear hyperplanes. As $\lambda$, $f$ and hence $\tau$ are continuous with respect to the Hausdorff distance topology (we use that $C$ has non-empty interior), we have $\lambda(B) = \lambda(C) > 0$ and we infer from Claim 4 that $\tau(C) \le \tau(B)$. By Remark 1, we have $\tau(B) = \tau_n$, which completes the proof.


[1] S. G. Krantz, H. R. Parks, "The Geometry of Domains in Space", 1999.
[2]: https://en.wikipedia.org/wiki/Symmetrization_methods [3]: https://en.wikipedia.org/wiki/Wallis%27_integrals [4]: https://en.wikipedia.org/wiki/N-sphere

$\endgroup$
4
  • $\begingroup$ I'ved added a post which gives an analytic formula for $\tau_m$ in terms of incomplete beta function. I've also derived a exponential bound which is essentially tight. $\endgroup$
    – dohmatob
    May 10, 2020 at 15:30
  • 1
    $\begingroup$ It seems convexity is not used at all in the proof of "Fedja's Lemma", since in the 1D case, you don't need the sets to be intervals. Also, Fubini-Tonelli still carries over $1 \rightarrow n$. So, I guess compactness + convexity of D and E can be weakened to demanding that $D$ and $E$ be measurable with finite volume. Then, the limitting argument needed to complete the proof of the ultimate result can be replaced with a weaker form of convergence: almost sure convergence to a ball, of Steiner symmetrizations w.r.t random hyperplanes link.springer.com/article/10.1007/s00526-012-0493-4. $\endgroup$
    – dohmatob
    May 28, 2020 at 10:44
  • $\begingroup$ @dohmatob This is a very interesting remark, thank you. I'll check the details. (By the way, you didn't answer my question about tightness: what does "tight" mean for your?) $\endgroup$
    – Luc Guyot
    May 30, 2020 at 12:56
  • $\begingroup$ Fixed (see comment below). For my remark on Steiner symmetrization compact nonconvex sets, actually you get convergence in a nonprobabilistic sense. See Theorem 1.1 of researchgate.net/profile/Aljosa_Volcic/publication/… $\endgroup$
    – dohmatob
    Jun 2, 2020 at 1:37
5
$\begingroup$

How about a short solution to a much general problem, and without convexity assumption (which turns out to be really just a distraction here) ?


So, more generally, for nonempty compact subsets $A$, $B$, and $C$ of $\mathbb R^n$, define $F(A,B,C)$ by

$$ F(A,B,C) := \int_{A}\lambda(B \cap (x + C))dx. $$

In particular, the OP's question is concerned with $f(C) := F(C,C,C)$.

We shall proceed in 3 short steps.

  • Step 1. The first step in analyzing $F$ is to realize that we can rewrite it as $F(A,B,C) = G(1_A,1_B,1_C)$, where

\begin{align} G(u,v,w) = \int_{\mathbb R^n}\int_{\mathbb R^n} u(x)w(x-y)v(y)\,dxdy. \end{align}

  • Step 2. Now, by the Riez rearrangement inequality, we know that $$ G(u,v,w) \le G(u^\star,v^\star,w^\star), $$ where $u^\star$ is the Symmetric decreasing rearrangement of the function $u$. On the other hand, it is clear that $$ (1_A)^{\star} = 1_{K(A)}, $$ where $K(A)$ is the centered ball of same volume as $A$. We deduce that $$ F(A,B,C) \le F(K(A),K(B),K(C)), $$

and in particular, $f(C) \le f(K(C))$. The solution to the OP's question then follows from computation for the ball, in the begining of the accepted answer https://mathoverflow.net/a/282941/78539.

$\endgroup$
0
2
$\begingroup$

Analytic formula and an exponential bound

For $r \in [0, 1]$, define $\tau_m(r) \in [0, 1]$ by $$ \tau_m(r) := \frac{1}{\lambda(B_m(r))}\int_{B_m(r)} \frac{\lambda(B_m(1) \cap (B_m(1) + x))}{\lambda(B_m(1))}dx. $$

The case $r=1$ has been analysed in accepted answer. Extending user @Matt's comment, one can establish that (for the core idea, see this post)

$$ \begin{split} \tau_m(r) &= \frac{1}{V_m^{cap}(r, 0)}\int_{0}^r ms^{m-1}V_m^{cap}(1, s/2)ds = \frac{1}{r^mV_m^{cap}(1, 0)}\int_{0}^r ms^{m-1}V_m^{cap}(1, s/2)ds\\ &= \frac{1}{V_m^{cap}(1, 0)}\int_{0}^1 ms^{m-1}V_m^{cap}(1, rs/2)ds, \end{split} $$

where $V_m(r; h)$ is the volume (i.e Lebesgue measure) of the half-lens $\{x \in B_m(r) \mid x_1 \ge h\}$.


Preamble. I propose to obtain an analytic formula for $\tau_m(r)$ in terms of special functions (mostly hypergeometric functions). We will also obtain a good upper bound for $\tau_m(r)$.

The end result will be the following exponential decay of $\tau_m(r)$ as a function on the dimension $m$.

Theorem. For every $r \in [0, 1]$ and large integer $m$, we have the upper bound $$ \begin{split} \tau_m(r) &\lesssim \exp\left(-\frac{mr^2}{8-r^2}\right) + \frac{r(1-r^2/4)^{m/2}}{\sqrt{m\pi}}\\ &\lesssim \exp\left(-\frac{mr^2}{8}\right) + \frac{r(1-r^2/4)^{m/2}}{\sqrt{m\pi}}. \end{split} $$

enter image description here Proof. Now, it is a classical computation that for $h \ge 0$ we have $V_m^{cap}(1, h) = \frac{\lambda(B_{m - 1}(1))}{2}\beta_{1-h^2}\left(\frac{m+1}{2},\frac{1}{2}\right)$, where $\beta_x(a, b) := \int_0^x t^{a-1}(1-t)^{b-1}dt$ defines the incomplete beta function. In particular, $V_m^{cap}(1, 0) = \frac{\lambda(B_{m - 1}(1))}{2}\beta_{1}\left(\frac{m+1}{2},\frac{1}{2}\right)$.

Thus, integrating by parts, we get

$$ \begin{split} V_m^{cap}(1,0)\tau_m(r) &= \int_{0}^1 ms^{m-1}V_m^{cap}(rs/2)ds = \int_{0}^1 \beta_{1-(rs)^2/4}\left(\frac{m+1}{2},\frac{1}{2}\right)ds^m\\ &= \left[s^m\beta_{1-(rs)^2/4}\left(\frac{m+1}{2},\frac{1}{2}\right)\right]_0^1 + r\int_0^1 s^m(1-(rs)^2/4)^{\frac{m-1}{2}}ds \\ &= \beta_{1-(r/2)^2}\left(\frac{m+1}{2},\frac{1}{2}\right) + R_m, \end{split} \tag{1} $$ where $R_m := r\int_0^1 s^m(1-(rs/2)^2)^{\frac{m-1}{2}}ds$. Now, consider the change of variable $s = 2\sqrt{t}$. This gives $s^m(1-s^2/4)^{(m-1)/2}ds=2^mt^{m/2}\cdot(1-t)^{(m-1)/2}\cdot t^{-1/2}dt = 2^mt^{(m-1)/2}(1-t)^{(m-1)/2}dt$. Thus,

$$ \begin{split} R_m &= r\int_0^1 s^m(1-(rs)^2/4)^{\frac{m-1}{2}}ds = \frac{2^m}{r^m}\int_0^{(r/2)^2}t^{(m-1)/2}(1-t)^{(m-1)/2}dt\\ &=: \frac{2^m}{r^m}\beta_{(r/2)^2}\left(\frac{m+1}{2},\frac{m+1}{2}\right). \end{split} \tag{2} $$

Combining (1) and (2) then gives the analytic formula,

Analytic formula. For any $r \in [0, 1]$ and positive integer $m$, it holds that $$ \tau_m(r) = I_{1-(r/2)^2}\left(\frac{m+1}{2},\frac{1}{2}\right) + \frac{(2/r)^m\beta_{(r/2)^2}\left(\frac{m+1}{2},\frac{m+1}{2}\right)}{\beta_1\left(\frac{m+1}{2},\frac{1}{2}\right)}, $$ where $u \mapsto I_u(a, b) := \beta_u(a,b) / \beta_1(a, b)$ is the CDF of $(a,b)$-beta distribution.

Bounding the first term. Let $\delta := (r/2)^2$ and $p = I_{1-\delta}((m+1)/2,1/2)$. It is known since Temme (1992) that for $p \in (0, 1)$ and large $a > 0$, the solution of the equation $p = I(t; a,b)$ is given (approximately) by

$$ t=t_p(a, b) \approx e^{-(1/a)Q_{1-p}(\Gamma(b,1))}, \tag{4} $$

where $Q_{1-p}(\Gamma(b,1))$ is the $1-p$ quantile of the unit-scale gamma distribution with shape parameter $b$. Now by standard concentration results (e.g see Boucheron et al. textbook),

$$ Q_{1-p}(\Gamma(b,1)) \le \log(1/p) + \sqrt{2b\log(1/p)}. \tag{5} $$

In particular, for $a=(m+1)/2$ and $b=1/2$ we get

$$ Q_{1-p}(\Gamma(1/2,1)) \le \log(1/p) + \sqrt{\log(1/p)}. \tag{6} $$

Putting (2), (4), and (6) together and using the basic inequality $1-e^{-z} \le 2z/(2+z)\;\forall z \ge 0$, we see that

$$ \begin{split} \delta &= 1 - t_{p}\left((m+1)/2,1/2\right) \approx 1-e^{-\frac{2Q_{1-p}(\Gamma(1/2,1))}{m+1}} \le 1-e^{-\frac{2}{m+1}\left(\log\left(\frac{1}{p}\right) + \sqrt{\log\left(\frac{1}{p}\right)}\right)} \le \frac{2\alpha}{2+\alpha}, \end{split} $$

where $\alpha = \frac{2}{m+1}(\log(1/p)+\sqrt{\log(1/p)})$. Thus, $\alpha \ge \delta/(1 - \delta/2)$, and this wolfram-alpha query tells us that we must have $p \lesssim \exp(-mr^2/(8-r^2))$. Putting things together then gives

$$ \begin{split} I_{1-(r/2)^2}((m+1)/2,1/2) = p \lesssim \exp(-mr^2/(8-r^2)) \le \exp(-mr^2/8). \end{split} $$

Bounding the second term. It is a well-known result that

$$ \beta_z(a,b) \equiv a^{-1}z^a{}_2F_1(a,1-b,a+1;z), $$

where $z \mapsto {}_2F_1(u, v, w; z)$ is the hypergeometric function. Thus, one computes

$$ \begin{split} (2/r)^m\beta_{(r/2)^2}\left(\frac{m+1}{2},\frac{m+1}{2}\right) &= \frac{r}{m+1}{}_2F_1\left(\frac{m+1}{2},\frac{1-m}{2},\frac{m+3}{2},\frac{r^2}{4}\right) \\ &= \frac{r}{m+1}{}_2F_1\left(\frac{1-m}{2},\frac{m+1}{2},\frac{m+3}{2},\frac{r^2}{4}\right)\\ &\approx \frac{r}{m+1}{}_2F_1\left(\frac{1-m}{2},\frac{m+1}{2},\frac{m+1}{2},\frac{r^2}{4}\right)\\ &= \frac{r}{m+1}(1-r^2/4)^{(m-1)/2}, \end{split} $$

where we have used the powerful identity ${}_2F_1(u, v, v; z) \equiv (1 - z)^{-u}$. On the other hand, Stirling's formula gives

$$ \beta_1\left(\frac{m+1}{2},\frac{1}{2}\right) \approx \Gamma(1/2)\left(\frac{m+1}{2}\right)^{-1/2} = \sqrt{\frac{\pi}{m+1}}. $$

Thus,

$$ \frac{(2/r)^m\beta_{(r/2)^2}\left(\frac{m+1}{2},\frac{m+1}{2}\right)}{\beta_1\left(\frac{m+1}{2},\frac{1}{2}\right)} \approx \frac{r(1-(r/2)^2)^{(m-1)/2}}{\sqrt{\pi(m+1)}} $$

Putting every together then completes the proof of the theorem. $\quad\quad\quad\Box$

$\endgroup$
7
  • $\begingroup$ @LucGuyot There were a bunch of typos in what I wrote, leading to the inconsistent inequalities you mentioned. Fixed. Let me know if you still have any trouble following. $\endgroup$
    – dohmatob
    May 13, 2020 at 7:51
  • $\begingroup$ I was hasty in simplifying (6) to $\ldots \le 2\log(1/p)$, which is only true for $p \le 1/e$. Actually, I can do without this simplification and still get the $\exp(-mr^2/(8-r^2))$ term of my bound. See update. $\endgroup$
    – dohmatob
    May 13, 2020 at 10:13
  • $\begingroup$ Sorry, I wasn't very careful in the calculations. I meant $1-e^{-z} \le 2z / (2+z)$ for all $z \ge 0$. Also, indeed the next inequalities were reversed. Everything should be fine now. Thanks for spotting the errors. $\endgroup$
    – dohmatob
    May 16, 2020 at 14:19
  • $\begingroup$ Because you are not using the big-O nor the small-o notation, and because there are at least two approximations in your computations, I would expect a constant factor $C$ (as close to $1$ as one would desire) to appear on the right side of the theorem's inequality. You also claim that your upper bound is tight. In which sense? Consider $r = 1$. We have $e^{-1/7} \approx 0.8668 > 0.8660 \approx \frac{\sqrt{3}}{2}$. So $(\frac{\sqrt{3}}{2})^m$ would be a better asymptotic upper bound of $\tau_m(1)$. $\endgroup$
    – Luc Guyot
    May 21, 2020 at 13:12
  • $\begingroup$ Ah, the $\le $ in my theorem was meant to be be implicitly read as $\lesssim$. I'll make it more explicit, if that helps. BTW, glad to know that my bounds match (upto within dust) the quantity you computed using a completely different strategy. $\endgroup$
    – dohmatob
    May 21, 2020 at 14:22

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.