8
$\begingroup$

Assume that $f$ is a probability density on $x \in (0,1)$, I want to obtain a bound of the following form (if it is possible at all): $$ \int_0^1 \frac{|f'|^2}{f} \,\mathrm{d} x \geq C\,\int_0^1 |f-1|^2\, \mathrm{d} x$$ for some universal constant $C > 0$. My strategy is as follows: by the log-Sobolev inequality, we have $$\int_0^1 \frac{|f'|^2}{f} \,\mathrm{d} x \geq C\,\int_0^1 f\,\ln f \,\mathrm{d} x.$$ Thus, I am expecting that a $L^p$-type (with $p=2$) Csiszar–Kullback-Pinsker inequality of the form $$\int_0^1 f\,\ln f \,\mathrm{d} x \geq C\,\int_0^1 |f-1|^2\, \mathrm{d} x \quad (*)$$ can hold. However, I searched very hard in the literature and the only relevant material I can found is this $L^p$-type Csiszar–Kullback-Pinsker inequality, which is not enough to deduce $(*)$ as the infimum of $\psi''(s) = 1/s$ over $s \in (0,\infty)$ is not strictly greater than zero (here $\psi(s) = s\,\ln(s) - s + 1$). Thus, it seems that unless we have certain (pointwise) boundedness assumptions imposed on $f$, there is no hope for the validity of $(*)$. I get stuck at this point I would appreciate any ideas which can help me out.

$\endgroup$
0

1 Answer 1

13
$\begingroup$

First of all such a Csiszar–Kullback-Pinsker inequality or whatever cannot possibly be true since $x^2$ explodes faster than $x\log x$ so you can make a local adjustment so that the right-hand side is infinite while the left-hand side is not. Yet, your original inequality is true and here is the proof:

By the Cauchy-Schwarz inequality we have $$\int_0^1 \frac{|f'(x)|^2}{f(x)}dx \int_0^1 f(x)dx \ge \left(\int_0^1 |f'(x)|dx\right)^2.$$

Since $\int_0^1 f(x)dx = 1$, we just have to prove that the right-hands side of this is at least $C\int_0^1 |f(x)-1|^2dx$. I will assume that since we are talking about $f'(x)$, the function $f$ is at least continuous. Since we also have $\int_0^1 f(x)dx = 1$, by the intermediate value theorem there must exist $x_0\in (0, 1)$ such that $f(x_0) = 1$. Applying the fundamental theorem of calculus, we can see that for all $x\in (0, 1)$ we have

$$|f(x) -1| = |f(x) - f(x_0)| \le \int_0^1 |f'(x)|dx.$$

Squaring this and integrating over $(0, 1)$ we get

$$\int_0^1 |f(x)-1|^2dx \le \left(\int_0^1 |f'(x)|dx\right)^2,$$

that is the required inequality holds with $C = 1$. I'm pretty sure $C$ can be optimized with this method, but I will leave it to someone else.

$\endgroup$
1
  • $\begingroup$ Sorry to bother, could you help me understand how the log-Sobolev inequality is being applied to get $\int_0^1 \frac{|f'|^2}{f} \,\mathrm{d} x \geq C\,\int_0^1 f\,\ln f \,\mathrm{d} x$? I wrote $\int_0^1 \frac{|f'|^2}{f} \,\mathrm{d} x=\int_0^1 f((\ln |f|)')^2\,\mathrm{d} x$ and compared with what appears on Wikipedia but got nowhere $\endgroup$
    – Dabed
    Apr 18, 2023 at 6:01

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.