Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

Let $C\subset \mathbb R^n$ be convex and symmetric about the origin. I am trying to prove that $\gamma(C) \geq \gamma(C+x)$ for any $x\in \mathbb R^n$, where $\gamma$ is the standard Gaussian measure. I tried using the following, which can be deduced from the Prekopa Leindler inequality:

$$\gamma(tA + (1-t)B) \geq t\gamma(A) + (1-t)\gamma(B)$$

for convex sets $A$ and $B$. However, for any of the various choices of $A$ and $B$ I can think of, this does not yield the desired inequality for $C$. Any suggestions for how to precede?

EDIT: The inequality I stated above is wrong; the right side should have the geometric mean. Here is a proof of the correct form of the inequality, using Prekopa Leindler. Let $f(x) = \chi_{tA + (1-t)B} (x) e^{-|x|^2/2}$, $g(x) = \chi_{A} (x) e^{-|x|^2/2}$, $h(x) = \chi_{B} (x) e^{-|x|^2/2}$, where $\chi$ denotes an indicator function. We claim that $f,g,h$ satisfy the hypothesis of the Prekopa Leindler inequality, i.e. $f(tx + (1-t)y) \geq g(x)^t h(y)^{1-t}$. This is obvious if $x\notin A$ $y\notin B$. If $x\in A$ and $y\in B$, then our claim reduces to showing

$$ e^{-|tx + (1-t)y|^2/2} \geq e^{-t|x|^2 - (1-t)|y|^2}$$

which is immediate from convexity of $x\mapsto |x|^2$. Thus Prekopa Leindler gives

$$\int f(x)\, dx \geq \left(\int g(x)\, dx\right)^t \left(\int h(x)\,dx\right)^{1-t}$$

which is equivalent to

$$\gamma(tA + (1-t)B ) \geq \gamma(A)^t \gamma(B)^{1-t} .$$

share|cite|improve this question
Can you explain how your inequality for $\gamma$ follows from Prekopa-Leindler? My understanding is that one gets a geometric mean of $\gamma(A)$ and $\gamma(B)$ (I also commented on did's answer below). – user53153 Jan 2 '13 at 1:12
I have added an explanation. – user15464 Jan 2 '13 at 2:17
Thanks for the addition. In some speculative way, we can guess results about Gaussian measure by looking at results for Lebesgue measure and letting $n\to \infty$. The Lebesgue measure version of the Brunn-Minkowski inequality expresses the concavity of ${\rm volume}^{1/n}$ with respect to Minkowski addition. In the limit $n\to \infty$ we should expect concavity of the logarithm. // To make an even stronger case, think of Brunn-Minkowski as an inequality for generalized mean with $p= 1/n$. When $p=0$ we have the geometric mean. – user53153 Jan 2 '13 at 5:27
up vote 3 down vote accepted

Brunn-Minkowski inequality for gaussian measures states that, for $t$ in $(0,1)$ and $A$ and $B$ both convex, $\gamma(tA+(1-t)B)\geqslant\gamma(A)^t\gamma(B)^{1-t}$. For $t=\frac12$, $A=C+x$ and $B=C-x$, the fact that $C=\frac12(A+B)$ by convexity of $C$, and the fact that $A=B$ by symmetry of $C$ indeed yield $\gamma(C)\geqslant\gamma(C+x)$.

share|cite|improve this answer
Does the Brunn-Minkowski inequality work this way for $\gamma$? I thought it had the geometric mean instead of arithmetic on the right hand side. – user53153 Jan 1 '13 at 23:34
@PavelM You are right, I modified the answer. Thanks. – Did Jan 2 '13 at 0:01

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

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