Take the 2-minute tour ×
MathOverflow is a question and answer site for professional mathematicians. It's 100% free, no registration required.

$\newcommand{\bsV}{\boldsymbol{V}}$ $\newcommand{\bsE}{\boldsymbol{E}}$ $\newcommand{\bR}{\mathbb{R}}$ Suppose that $\bsV$ is an $N$-dimensional real Euclidean space. Denote by $\newcommand{\eA}{\mathscr{A}}$ $\eA$ the space of symmetric positive semidefinite operators $A:\bsV\to \bsV$. To each $A\in \eA$ we can associate in a canonical fashion a centered Gaussian measure $\gamma_A$ on $\bsV$ which is concentrated on $(\ker A)^\perp$. For example, if $A$ is nondegenerate, then

$$ \gamma_A(dv)= \frac{1}{\sqrt{\det 2\pi A}} e^{-\frac{1}{2} (A^{-1}v,v)}dv, $$

while if $A=0$, then $\gamma_0$ is the Dirac measure concentrated at the origin.

Fix a locally Lipschitz function $f:\bsV\to\bR$ which is positively homogeneous of degree $\alpha \geq 2$. For any $A\in \eA$ we denote by $\bsE_A(f)$ the expectation of $f$ with respect to the probability measure $\gamma_A$ on $\bsV$. Consider the function

$$ \eA\ni A\mapsto \bsE_A(f)\in \bR. $$

This function is continuous and positively homogeneous of degree $\frac{\alpha}{2}$, i.e.,

$$ \bsE_{tA}(f)=t^{\frac{\alpha}{2}} \bsE_A(f),\;\;\forall t>0,\;\;A\in\eA. $$

I am interested in its modulus of uniform continuity on the ball

$$\eA_1:=\bigl\lbrace A\in\eA;\;\;\Vert A\Vert\leq 1\bigr\rbrace, $$

I was able to prove that on this ball the above function is Holder continuous, with Holder exponent $\frac{1}{2N+3}$. This suffices for the applications I have in mind, but I strongly suspect that it is far from optimal. I believe that the Holder exponent $\frac{1}{2}$ is uniformly optimal in the following sense: there exist $C, r>0$ so that for any $A, B\in\eA_1$ satisfying

$$\Vert A- B\Vert \leq r, $$

we have

$$\bigl\vert \bsE_A(f)-\bsE_B(f)\bigr\vert\leq C\Vert A-B\Vert^{\frac{1}{2}}. \tag{1} $$

Remark. To see that the exponent $\frac{1}{2}$ is the best one can hope for consider the case $\bsV=\bR^2$, $f(x,y)=|xy|$ and $\newcommand{\ve}{{\varepsilon}}$ and the Gaussian measures

$$ \gamma_{A_\ve}=\frac{1}{2\pi\ve} e^{-\frac{1}{2\ve^2}x^2-\frac{1}{2}y^2} |dxdy| $$

Then $\Vert A_\ve-A_0\Vert =\ve^2$,

$$\bsE_{A_0}(f)=0,\;\; \bsE_{A_\ve}(f)=\left(\int_{\bR}|x|e^{-\frac{1}{2}x^2} |dx|\right)^2 \ve. $$

My question is the following: have you encountered continuity results of this sort, and if so, can you indicate some references that deal with this? Thanks!

share|improve this question
Shouldn't this follow from the operator square root being Holder continuous? i.e., $\lVert\sqrt{A}-\sqrt{B}\rVert\le C\lVert A-B\rVert^{1/2}$. –  George Lowther May 14 '13 at 15:43
This argument will do the trick for nonsingular $A$'s. In this situation $f$ need not be Lipschitz. For singular $A$'s things are trickier. The measure $\gamma_A$ is the product between the Dirac measure on $\ker A$ and nonsingular Gaussian on $(\ker A)^\perp$. –  Liviu Nicolaescu May 14 '13 at 17:15
@Liviu: The measure $\gamma_A$ is always the same as the distribution of $\sqrt{A}X$ (singular or not), where $X$ has the standard Gaussian distribution. So, no need to treat the singular case any differently. –  George Lowther May 14 '13 at 17:22
@Duh! Thank you very much. It is embarrassingly obvious now. –  Liviu Nicolaescu May 14 '13 at 17:35
@Liviu: I'll post it as an answer in a few moments –  George Lowther May 14 '13 at 18:19

1 Answer 1

up vote 6 down vote accepted

This problem reduces quickly to Holder continuity of the operator square root. That is, there exists a $C > 0$ such that $$ \begin{align} \lVert\sqrt{A}-\sqrt{B}\rVert\le C\lVert A-B\rVert^{1/2}&&{\rm(1)} \end{align} $$ for any positive semidefinite operators $A,B$.[1]

Assuming (1), the proof of continuity with Holder exponent $1/2$ as mentioned in the question is quite direct. If $X$ is an $\mathbb{R}^N$-valued standard normal random variable, then $\sqrt{A}X$ has distribution $\gamma_A$. So, if $f$ has Lipschitz constant $K$ on the unit ball then, writing $\hat X=X/\lVert X\rVert$, $$ \begin{align} \left\lvert\mathbb{E}_A(f)-\mathbb{E}_B(f)\right\rvert &=\left\lvert\mathbb{E}[f(\sqrt{A}X)-f(\sqrt{B}X)]\right\rvert\cr &=\left\lvert\mathbb{E}[\lVert X\rVert^\alpha(f(\sqrt{A}\hat{X})-f(\sqrt{B}\hat{X}))]\right\rvert\cr &\le K\mathbb{E}[\lVert X\rVert^\alpha]\lVert\sqrt{A}-\sqrt{B}\rVert\cr &\le CK\mathbb{E}[\lVert X\rVert^\alpha]\lVert A-B\rVert^{1/2}, \end{align} $$ as required.

[1] Holder continuity of the square root looks quite obvious, so I would expect it to be standard. I don't have a reference for this though, but I can give a proof now using the Taylor expansion of the square root (maybe there is a quicker proof). It is tempting to suggest that it holds with $C=1$ as in the 1-dimensional case, but this is possibly rather too strong when $A$ and $B$ do not commute.

Multiplying through by a scalar, we can assume that $\lVert A\rVert$ and $\lVert B\rVert$ are bounded by $1$. Then, write $A=1-X$, $B=1-Y$ for positive semidefinite operators $X,Y$ with $\lVert X\rVert,\lVert Y\rVert\le1$. By Taylor expansion, $$ \sqrt{1-X}=1-\sum_{n=1}^\infty a_n X^n $$ where $$ \begin{align} a_n&=\frac12\prod_{k=2}^n\left(1-\frac3{2k}\right) \le\frac12\prod_{k=2}^n\exp\left(-\frac3{2k}\right)\cr &\le\exp\left(-\frac32\log n\right) =n^{-3/2}. \end{align} $$ Then, $$ \sqrt{A}-\sqrt{B}=\sum_{n=1}^\infty a_n(Y^n-X^n). $$ As $X,Y$ have norm bounded by 1, the term $Y^n-X^n$ has norm bounded by $n\lVert Y-X\rVert=n\lVert A-B\rVert$. As $X,Y$ are also positive semidefinite, $Y^n-X^n$ is bounded by 1 in norm. So, $$ \begin{align} \lVert\sqrt{A}-\sqrt{B}\rVert&\le\sum_{n=1}^\infty n^{-3/2}\min\left(n\lVert A-B\rVert,1\right)\cr &=\lVert A-B\rVert\sum_{n\le\lVert A-B\rVert^{-1}}n^{-1/2}+\sum_{n > \lVert A-B\rVert^{-1}}n^{-3/2}. \end{align} $$ As $\sum_{n\le x}n^{-1/2}$ and $\sum_{n > x}n^{-3/2}$ are bounded by fixed multiples of $\sqrt x$ and $1/\sqrt x$ respectively (for $x\ge1$), both terms on the right hand side of the inequality above are bounded by a multiple of $\lVert A-B\rVert^{1/2}$. This gives (1) as required. Note that the constant $C$ does not depend on the dimension $N$, and (1) also holds in infinite dimensions.

share|improve this answer
Beautiful! Thanks again. –  Liviu Nicolaescu May 14 '13 at 21:15

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.