3
$\begingroup$

My questions come from the proof of Theorem 5.14 in section 5.7 of Boucheron, Lugosi, and Massart - Concentration inequalities. My first question can be stated as follows:


Suppose for positive numbers $V,y,\Delta,x,\delta >0$ we have $V \leq y^{-1}\left[2\Delta y^{-1} + \sqrt{2x} + \sqrt{2\delta}\right].$ Then if we define $$y^2 = 2K\epsilon^2\left[\left(\sqrt{x}+\sqrt{\delta}\right)^2 + \epsilon^{-1}K^{-\frac 12}\Delta + \sqrt{\delta\epsilon^{-1}K^{-\frac 12}\Delta} \right] \tag{1}\label{1}$$ for $K >1$ and $\epsilon >0$, then the authors claimed that the inequality for $V$ implies that $\epsilon V \leq K^{-\frac 12}$. To be honest, I have no clue as to the motivation of defining $y$ according to \eqref{1}. If I want to enforce the relation $V \leq K^{-\frac 12}/\epsilon$, I will just set up the relation $$y^{-1}\left[2\Delta y^{-1} + \sqrt{2x} + \sqrt{2\delta}\right] = K^{-\frac 12}/\epsilon,$$ which defines a quadratic equation in the variable $y^{-1}$ and whose solution can be explicitly found out. But this idea does not lead to the suggested expression for $y$….


My second question is again regarding details. Setting $\delta := \left((4\pi)^{-\frac 12} + \sqrt{z}\right)^2$ for fixed $z > 0$. The authors mentioned that by using repeatedly the elementary inequality $2ab \leq \theta a^2 + \theta^{-1}b^2$, we can derive the following upper bound from the identity/definition \eqref{1} (recall that $K >1$ is fixed): $$K^{-\frac 12}y^2 \leq 2K\epsilon^2\left[\epsilon^{-1}\Delta + x + \sqrt{\epsilon^{-1}\Delta x} + \frac{2}{\sqrt{K} - 1}\left(\frac{1}{2\pi} + 2z\right)\right]. \tag{2}\label{2}$$ I tried a lot but still can not figure out the right way to arrive at \eqref{2} from \eqref{1}, so any help is appreciated!

$\endgroup$

1 Answer 1

2
$\begingroup$

If $y^2$ is defined by (1) and $$\epsilon = \frac{168}{97},x= \frac{1168561}{2916},z= \frac{121}{8281},\Delta = \frac{1}{1000},K= \frac{10201}{9604},$$ then the difference between the right-hand side of inequality (2) and its left-hand side is $-2.5248\ldots<0$, so that (2) fails to hold.

(What might be interesting, though, is that, for $y^2,\epsilon,x,z,\Delta,K$ as above, the values of the right-hand side and left-hand side are relatively close to each other: $2635.1\ldots$ and $2637.6\ldots$.)


Also, if $y^2$ is defined by (1), $y>0$, $V=y^{-1}\left(2\Delta y^{-1} + \sqrt{2x} + \sqrt{2\delta}\right)$ and $$K= \frac{602}{61},x= \frac{7936}{57},\delta = \frac{34618}{41},\epsilon = \frac{6}{73},\Delta = \frac{37534}{29},$$ then the ratio of the left-hand side of inequality $$\epsilon V \le K^{-\frac 12} \tag{*}$$ to its right-hand side is $1.2496\ldots>1$, so that (*) fails to hold, too.

$\endgroup$
3
  • $\begingroup$ Thank you professor Pinelis! It's so sad to see errors in highly cited books like this book...Also, may I know do you have a answer to my first question? $\endgroup$
    – Fei Cao
    Commented Jan 5, 2022 at 5:25
  • 1
    $\begingroup$ I am deeply amazed by your ability to construct such non-trivial counter-examples.... Perhaps I need to email the authors about the errors/mistakes they made there. $\endgroup$
    – Fei Cao
    Commented Jan 5, 2022 at 5:53
  • 2
    $\begingroup$ @FeiCao : These counterexamples were obtained by (i) numerical minimization or maximization of the difference or ratio between the two sides of the inequalities, (ii) finding close enough rational approximations to the found minimizing/maximizing values of the variables, and (iii) obtaining the corresponding values of difference and ratio. $\endgroup$ Commented Jan 5, 2022 at 14:15

You must log in to answer this question.

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