4

Does anybody have a good way to prove the inequality $e^x\leq x+e^{9x^2/16}$ ?

I found this here

Wainwright, M. J. (2019). High-dimensional statistics: A non-asymptotic viewpoint (Vol. 48). Cambridge University Press.

On page 47, the author directly used this inequality

By the way, here are some good proofs of $e^x\leq x+e^{x^2}$. proof of inequality $e^x\le x+e^{x^2}$

I have to mention that the first-order derivative of this function $f(x)=x+e^{9x^2/16}-e^x$ has multiple roots as below.

The first-order derivative: $f'(x)=\frac{9xe^{\frac{9x^2}{16}}}{8}+1-e^x$

Plot of the first-order derivative: the first-order derivative plot

Three roots: 0,0.41490879,0.54479289

By the way, can you find the $\min$ value of $c$ such that the inequality $e^x\leq x+e^{cx^2}$ is satisfied? I tried several values of $c$ less than 9/16, and they work!

  • 1
    Please consider accepting an answer if it helped you. Thanks in advance. –  Dec 10 '22 at 09:20
  • 1
    Using Am-Gm and a good candidate function the refined inequality involves the hyperbolic cosinus or : $$-e^{-x}+2\left(\left(x+e^{-x}\right)e^{\frac{9}{16}x^{2}}\right)^{\frac{1}{2}}>e^{x}$$ for $x>0$ a real number . – Miss and Mister cassoulet char Dec 10 '22 at 12:15
  • Hi Erik Satie, thanks for your idea! But it seems you use a tighter bound, which makes the proof harder. Anyway, your inequality is actually correct. The logic is, equivalently we need to prove $e^x+e^{-x} \leq (x + e^{-x})+e^{9x^2/16}$. Since $(x + e^{-x})+e^{9x^2/16} \geq 2\sqrt{(x+e^{-x})e^{9x^2/16}}$ by Am-Gm, if we could prove that $e^x+e^{-x} \leq 2\sqrt{(x+e^{-x})e^{9x^2/16}}$, then we get the conclusion. – Lee Jelly Dec 10 '22 at 15:23
  • Empy2's answer to find the optimal value of $c$ is fantastic! So the second question is solved! – Lee Jelly Dec 10 '22 at 15:33
  • @LeeJelly For a possible end see https://math.stackexchange.com/questions/331367/cosh-x-inequality .Glad you find it nice . – Miss and Mister cassoulet char Dec 11 '22 at 10:16
  • @LeeJelly Using this link we can show te inequality on (0,1/3) remarking for this interval we have : $$\left(x+e^{-x}\right)e^{\left(\frac{9}{16}-1\right)x^{2}}>1$$ .Now take the logarithm and use derivative . – Miss and Mister cassoulet char Dec 11 '22 at 10:59
  • For the end consider $$f\left(x\right)=e^{\frac{9}{16}x^{2}}$$ and $$f(x)\geq f'\left(a\right)\left(x-a\right)+\frac{f''\left(a\right)}{2}\left(x-a\right)^{2}+f\left(a\right)$$ for $x\ge a$ at the point $a=1/3,1/2,2/3,1$ the proof is rather easy using second derivatives .So now you have a second proof for $x>0$ . – Miss and Mister cassoulet char Dec 12 '22 at 10:04

1 Answers1

3

Remarks: The optimal $c$ is given by $c_0 = \max_{x\ne 0} \frac{\ln(\mathrm{e}^x - x)}{x^2}$. I think that it does not admit a closed form. Here we give a proof of $c = 9/16$.

Letting $x = \frac43 y$, the desired inequality is written as $$\mathrm{e}^{y^2} + \frac43y \ge \mathrm{e}^{4y/3}. \tag{1}$$

Case 1: $y \ge 4/3$

Clearly (1) is true.

$\phantom{2}$

Case 2: $y < 4/3$

Using $\mathrm{e}^u \ge 1 + u + \frac12u^2 + \frac16u^3$ for all $u\ge 0$, we have \begin{align*} \mathrm{e}^{y^2} &\ge 1 + y^2 + \frac12y^4 + \frac16 y^6\\[6pt] &\ge 1 + y^2 + \frac12y^4 + \frac16 \cdot \left(\frac12 y^4 - \frac{1}{16}y^2\right) \tag{2}\\[6pt] &= \frac{7}{12}y^4 + \frac{95}{96}y^2 + 1\\[6pt] &\ge \frac{33}{35}\cdot \frac{7}{12}y^4 + \frac{95}{96}y^2 + 1\\[6pt] &= \frac{11}{20}y^4 + \frac{95}{96}y^2 + 1 \end{align*} where in (2) we have used $y^6 + \frac{1}{16}y^2 \ge \frac12 y^4$ (AM-GM).

Thus, it suffices to prove that $$\left(\frac{11}{20}y^4 + \frac{95}{96}y^2 + 1\right) + \frac43y \ge \mathrm{e}^{4y/3}.$$

Since $\frac{11}{20}y^4 + \frac{95}{96}y^2 + \frac43y + 1 > 0$, it suffices to prove that $$\ln\left(\frac{11}{20}y^4 + \frac{95}{96}y^2 + \frac43y + 1\right) \ge \frac{4y}{3}.$$ Denote $\mathrm{LHS} - \mathrm{RHS}$ by $f(y)$. We have $$f'(y) = \frac{-1056y(y - a)(y- 1/2)(y - b)}{792y^4 + 1425y^2 + 1920y + 1440}$$ where $$a = \frac54 - \frac{1}{132}\sqrt{17655} \approx 0.2434, \quad b = \frac54 + \frac{1}{132}\sqrt{17655} \approx 2.2566.$$ Thus, we have $f'(y) < 0$ on $(-\infty, 0)$, and $f'(y) > 0$ on $(0, a)$, and $f'(y) < 0$ on $(a, 1/2)$, and $f'(y) > 0$ on $(1/2, 4/3)$. Also, we have $f(0) = 0$ and $f(1/2) = \ln\frac{1247}{640} - \frac23 > 0$. Thus, $f(y)\ge 0$ for all $y < 4/3$.

We are done.

River Li
  • 37,323
  • Where did the term $\frac{4}{3}y$ come from in the third line of the first display in case 2? – L. F. Dec 11 '22 at 00:09
  • @L.F. I should remove the term $4y/3$. Thanks. – River Li Dec 11 '22 at 01:17
  • @River Li Cool! Your answer is helpful and elegant! I've already accepted your answer! How did you come up with this solution? Only simple calculus&algebra or motivated by some advanced math fields? – Lee Jelly Dec 11 '22 at 19:20
  • @RiverLi Never mind! I finally understand your solution! Use Taylor expansion to transform $e^{x^2}$ into polynomials. Then deal with polynomials. Moreover, Am-Gm is used to transform $x^6$ to $x^4$ and $x^2$, which are low order terms. Anyway, your solution is so elegant! – Lee Jelly Dec 11 '22 at 23:29
  • @LeeJelly Thanks. I first found the Taylor approximation $\mathrm{e}^{y^2} \ge 1 + y^2 + \frac12y^4 + \frac16 y^6$ is enough. If we want to prove $1 + y^2 + \frac12y^4 + \frac16 y^6 + 4y/3 \ge \mathrm{e}^{4y/3}$, the derivative does not admit closed form zeros. So I decrease the degree of the bound $1 + y^2 + \frac12y^4 + \frac16 y^6$ (from six to four). Then, if we want to prove $\frac{7}{12}y^4 + \frac{95}{96}y^2 + 1 + 4y/3 \ge \mathrm{e}^{4y/3}$, the derivative does not admit closed form zeros. Finally, I changed the bound to $\frac{11}{20}y^4 + \frac{95}{96}y^2 + 1$ which works. – River Li Dec 11 '22 at 23:54
  • @LeeJelly Yes, you got it. By the way, this trick may be called shift of extreme points, that is, the extreme points are moved to closed form extreme points. See e.g. my answer for this question: https://math.stackexchange.com/questions/4573486/prove-that-frac-ln-xx-frac1ex-frac12-for-x-0/4575441 – River Li Dec 12 '22 at 00:02