Non-negative map is locally uniformly continuous on the interior of a cone under hypothesis

55 Views Asked by At

First of all, I'm sorry if I used the wrong tags for this question, please tell me if you want me to change them.

During a reading of a paper, I came across a lemma that is useful for my research (I study the characteristics of a certain function in algebraic geometry) and I tried to prove it without success. Here it is (Lemme 2.7 from B.Lehmann’s paper - "Volume type functions for numerical cycle classes"):

Let $V$ be a finite dimensional $\mathbb Q$-vector space and let $C \subset V$ be a salient full-dimensional closed convex cone (i.e it is a convex cone such that $C - C = V$ and for every $v \in C$, we have $-v \notin C$). Suppose that $f : V \to \mathbb R_{\geq 0}$ is a function satisfying

  1. $f(e) > 0$ for any $e \in C^{\text{int}}$,
  2. there is some constant $c > 0$ so that $f(me) = m^cf(e)$ for any $m \in \mathbb Q_{>0}$ and $e \in C$,
  3. for every $v,e \in C^{\text{int}}$ we have $f(v+e) \geq f(v)$.

Then $f$ is locally uniformly continuous on $C^{\text{int}}$.

All I could observe is that the values taken by elements of the cone C close to $0_V$ must be small by the formula

$$\forall v \in C^{\text{int}}, \forall m \geq 1 : ~~f\left( \dfrac{1}{m}v \right) = \dfrac{1}{m^c} f(v)$$

I tried to formulate a proof in the same style that demonstrates the continuity of a linear application in $0_V$ but I did not have more ideas than that. Any thought is welcome. Thank you!

1

There are 1 best solutions below

0
On BEST ANSWER

This answer is somewhat verbose, it might helpt to just draw the corresponding picture for yourself.

We are assuming that $V=\mathbb{Q}^d$ with the subspace topology inherited from the Euclidean topology on $\mathbb{R}^d$ (the OP clarified in the comments that this is the setting of interest to them). Furthermore, we will assume that $d\geq 2$ (the case $d=1$ is trivial as $C$ would be a half-space and $f(x) = \vert x\vert^m f(e)$ for $x\in C^\mathrm{int}$ and $e=1$ if $C=[0,\infty)$ and $e=-1$ if $C=(-\infty,0]$. Clearly $f$ is locally uniformly continuous). In this answer, we denote by $B(x,r)=\{v\in \mathbb{Q}^d \ : \ \vert v-x\vert <r\}$ the ball in $\mathbb{Q}^d$ of radius $r$ centered at $x$.

We want to prove that the map $f$ is locally uniformly bounded on $C^\mathrm{int}$. The only thing which is of quantitative nature in our assumption is the requirement that $$ f(c x) = c^m f(x) $$ for all $c\in \mathbb{Q}_{>0}$ and all $x\in C$. We would like to make use of this to estimate the difference between close points. Namely, if we want to compare $x,y\in C^\mathrm{int}$, then we would like to make a "zigzag" between $x$ and the ray spanned by $y$. More precisely, we would like to find $\varepsilon(x,y)>0$ such that $$ y-(1-\varepsilon(x,y))x, (1+\varepsilon(x,y))x-y\in C^\mathrm{int}. $$ In that case we have by conditions $1$ and $2$ that $$ \left( 1-\varepsilon(x,y) \right)^m f(x) = f((1-\varepsilon(x,y)x) \leq f(y) \leq f((1+\varepsilon(x,y))x) = \left( 1+\varepsilon(x,y) \right)^m f(x). $$ In particular, this implies $$ \vert f(x)-f(y) \vert \leq \left[ \left( 1+\varepsilon(x,y) \right)^m - \left( 1-\varepsilon(x,y) \right)^m \right] f(x). $$

We need to make sure that both $\varepsilon(x,y)$ are locally uniformly controlled and $f(x)$ too.

Let us fix some $e\in C^\mathrm{int}$ and some $0<\delta <\vert e\vert/8$ such that $B(e,4\delta)\subset C^\mathrm{int}$. In particular, this means that for every $x\in B(e,\delta)$ the cone $C_x = \mathbb{Q}_{>0} B(x,2\delta)$ generated by $B(x,2\delta)$ is contained in $C^\mathrm{int}$. There exists $\theta \in (0,\pi/2)$ such that all $C_x$ contain a cone of the form $$ \{ v\in \mathbb{Q}^d \ : \ 0<\langle v,x \rangle \leq\cos(\theta) \vert x \vert \} \subseteq C_x \subseteq C^\mathrm{int}.$$ In particular, this means that if we were in $\mathbb{R}^d$ (instead of $\mathbb{Q}^d$), then we could just use (highschool) geometry to compute $\varepsilon(x,y)$ for $y\in B(x,\delta)$. Let's do this first.

We can rotate everything to reduce to the case where $x=(\vert x \vert, 0, \dots, 0)\in \mathbb{R}^d$ and $y=(y_1, y_2, 0, \dots, 0)\in \mathbb{R}^d$ and $y_1, y_2\geq 0$ (so really we are in $\mathbb{R}^2$). We are asking where we need to start on the ray generated by $(1,0, \dots, 0)$ to reach $y$ by adding only a vector in $C_{(\vert x \vert,0,\dots, 0)}$. The cone $C_{(\vert x \vert,0,\dots, 0)}$ has an opening angle of $\theta$ around $\mathbb{R}_{\geq 0}(1, 0, \dots, 0)$, thus, starting at $( y_1-y_2 \cot(\theta), 0, \dots, 0)$ would do the job. Similarly, if we start at $y$ and need to reach the ray generated by $(1,0, \dots, 0)$ by adding only a vector in $C_{(\vert x \vert,0,\dots, 0)}$, we can choose $$ (y_1+y_2 \cot(\theta), 0, \dots, 0). $$ We have $$ \vert \vert x \vert - (y_1\pm\cot(\theta)y_2)\vert \leq \vert x-y\vert (1+\cot(\theta)).$$ Here we have used that $0\leq y_1\leq \vert x-y\vert$.

All of those computations assumed that we are in $\mathbb{R}^d$. In order to account for the fact that not all those numbers would be in the reals, we just introduce an additional factor of $2$ (if we'd really care about constants, we could drop it by just approximating suitably and noting that everything is well-behaved). Then we one sees that $$ \varepsilon(x,y) = 2 \vert x-y\vert (1+\cot(\theta)) $$ is an admissible choice (recall that $\theta$ is constant in the choosen neighborhood).

So far we have shown that there exists $\theta\in (0, \pi/4)$ such that for all $x,y\in B(e,2\delta)$ we get $$ \vert f(x)-f(y) \vert \leq \left( \left[ 1+2 \vert x-y\vert (1+\cot(\theta)) \right]^m - \left[ 1-2 \vert x-y\vert (1+\cot(\theta)) \right]^m \right) f(x). $$ Using the same argument as before, comparing $x$ to $e$, we get for all $x,y\in B(e,2\delta)$ $$ \vert f(x)-f(y) \vert \leq \left( \left[ 1+2 \vert x-y\vert (1+\cot(\theta)) \right]^m - \left[ 1-2 \vert x-y\vert (1+\cot(\theta)) \right]^m \right) \left( 1+ 2\delta(1+\cot(\theta)) \right) f(e). $$ This proves that $f$ is locally uniformly continuous on $C^\mathrm{int}$ as there exists a constant $C_{\theta,m}$ depending only on $\theta,m$ such that for all $x,y\in \mathbb{R}^d$ holds $$ \left[ 1+2 \vert x-y\vert (1+\cot(\theta)) \right]^m - \left[ 1-2 \vert x-y\vert (1+\cot(\theta)) \right]^m \leq C_{\theta,m} \vert x-y\vert.$$

Note that this proof gives us a slightly stronger result. Namely, we get that $f$ is locally lipschitz on $C^\mathrm{int}$.