Consider a sequence of functions $f_N(x)$ for which I know $\limsup_{N \to \infty} \frac{1}{N} \log f_N(x) = -g(x)$, where $g(x)$ is positive, continuous on $(a, \infty)$, and increases to infinity as $x \to \infty$. Moreover, \begin{equation} g(x) \geq - \log(x)+V(x)+C \hspace{1cm} \text{for } x \geq a, \text{ and } C \text{ a positive constant} \end{equation} where $V(x)$ is a continuous function satisfying \begin{equation} \liminf_{x \to \infty} \frac{V(x)}{\beta \log(x) } >1 \hspace{1cm} \text{for } \beta >1 \end{equation}

I want to show that $\lim_{N \to \infty} \int_b^{\infty} f_N(x) x dx = 0$ for $b >a$.

$\bullet Origin$ of the problem comes from that $f_N$'s are probability measures, ($P(X_n >x)$), satisfying a large deviation principle with good rate function $g(x)$ with the above properties, and I want to show that $\mathbb{E} [ X_n \mathbb{I}\{ X_n > b \} ]$ converges to 0 as $N \to \infty$.

My first question is that, what is the interpretation of $ \liminf_{x \to \infty} \frac{V(x)}{\beta \log(x) } >1$? can we deduce that $V(x) > \beta \log(x)$?

Assuming $V(x) > \beta \log(x)$ for all $x$, here is what I have tried so far:

First note that we have $g(x) \geq \beta' \log(x) + C$ for positive constants $\beta', C$. From $\limsup_{N \to \infty} \frac{1}{N} \log f_N(x) = -g(x)$, given $0<\epsilon <C$, for large enough $N$,we have $f_N(x) \leq e^{-N(g(x) - \epsilon)}$. So, we have: $$ \int_b^{\infty} f_N(x) x dx \leq \int_b^{\infty} e^{-N(g(x) - \epsilon)} x dx \leq e^{-N(C-\epsilon)} \int_b^{\infty} e^{-N \beta' \log(x)} x dx = e^{-N(C-\epsilon)} \frac {b^{2 - N \beta'}}{N \beta' -2} \hspace{5mm} \text{for sufficiently large} N $$ taking the limit, we have the result.

I'm not sure if it is correct or not. Any help is appreciated!


1 Answer 1


Without uniformity on $x$, it is not necessary true. Consider $$f_N(x) = \begin{cases} \exp(-N\cdot g(x)), x < N\\ \text{anything}, x \geq N \end{cases}$$

Then not just $\liminf$, but even $\lim\limits_{N\to \infty} \frac{1}{N} f_N(x) = -g(x)$, but $\int_b^\infty f_N(x)\, dx$ can be made arbitrary for each $N$ independently.

Assuming uniformity in the first limit ($\forall \varepsilon > 0 \exists N_0 \forall x: \frac{1}{N}\log f_N(x) < -g(x) + \varepsilon$) - that's what needed for $f_N(x) \leq e^{-N(g(x) - \epsilon)}$ you are almost correct.

$\liminf\limits_{x \to \infty} h(x) > 1$ means exactly that for some $\varepsilon > 0$ and $x_0$, $h(x) > 1 + \varepsilon$ if $x > x_0$. So you have $g(x) \geq \beta' \log(x) + C$ if $x > x_0$. Then your reasoning shows that $\int_{x_0}^\infty f_N(x)\, dx \to 0$.

We only need to deal with $\int_b^{x_0}f_N(x)\, dx$. But $g(x)$ is bounded from $0$ on $[b, x_0]$ (as continuous positive function on closed segment), so let's say $\inf_{x \in [b, x_0]} g(x) = \delta > 0$, and as for large enough $N$, $F_N(x) \leq \exp(-N(g(x) - \delta / 2)) \leq \exp(-N \delta / 2)$, for large enough $N$, $\int_{b}^{x_0} f_N(X)\, dx \leq \exp(-N \delta / 2) \cdot (x_0 - b) \to 0$.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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