3
$\begingroup$

Let:

$$S_n = \frac{1}{n} \sum_{i=0}^n \mathrm{e}^{n f(i/n)} g(i/n)$$

$$I_n = \int_0^1 \mathrm{e}^{nf(x)}g(x)\mathrm{d}x$$

where $f(x)$ and $g(x)$ are smooth functions (assume they are differentiable up to any order required), $g(x) > 0$, and $n$ is a positive integer.

Note that the sum looks like a Riemann sum, but not quite, because there is an $n$ in the exponent.

Prove or disprove, that $R_n = |(S_n - c I_n) / S_n| \rightarrow 0$ as $n \rightarrow \infty$, for some constant $c$ that may depend on $f,g$ and/or its derivatives, but does not depend on $n$. If true, find the value of $c$ and characterize how fast $R_n$ decays as $n$ increases; is it exponentially fast?

$\endgroup$
10
  • 1
    $\begingroup$ Only small neighborhoods near the points where $f$ attains its maximal value really matter. If the maximum is attained at some inner point, you are fine with $c=1$. If it is attained only at endpoints, then the computation of $c$ is a bit more complicated. The question about the speed of convergence is more delicate. It depends on many factors but is rarely exponential (though it may be; say, if $g(x)=1$, $f(x)=x-x^2$). With finite smoothness you'll normally get just a power decay. $\endgroup$
    – fedja
    Commented Jul 13, 2017 at 21:03
  • $\begingroup$ @fedja I have some very specific functions in mind, but I do not want to meddle this post in too much details. What kind of techniques are available so that I prove these things? Are there some general conditions under which $R_n$ decays exponentially? Oh, you can assume that the max of $f(x)$ occurs in the interior. $\endgroup$
    – valle
    Commented Jul 13, 2017 at 22:54
  • $\begingroup$ Unfortunately, the exponential decay is a very rare case. I'd rather prefer to see the functions you have in mind and try to figure out what rate I can guarantee for them than to try to find the weakest assumptions under which the decay is exponential. $\endgroup$
    – fedja
    Commented Jul 13, 2017 at 23:07
  • $\begingroup$ There is one general thing I can say though: if $f,g$ are analytic in a neighborhood of $[0,1]$ and $f$ attains its global maximum only at interior points, then you are fine. This condition can be relaxed a bit, but, alas, not by much. $\endgroup$
    – fedja
    Commented Jul 13, 2017 at 23:50
  • 1
    $\begingroup$ Yes, I can. But will it be enough for you? If it will, I'll post a detailed explanation today or tomorrow. The trickery consists of the Poisson summation formula and shifting the contour to bound the Fourier transform. It is the second part that requires analyticity. I haven't attempted a proof that analyticity is necessary for the exponential decay in all cases, but it is possible to show that no weaker "natural" assumption will suffice. However, if your functions are, indeed, analytic, then why should you care, and if they are not, the chance that there is some non-generic effect is slim.. $\endgroup$
    – fedja
    Commented Jul 14, 2017 at 0:33

2 Answers 2

3
$\begingroup$

The main observation:

Let $f,g$ be analytic in the disk $\{|z|\le 2\delta\}$ and real on the interval $(-2\delta,2\delta)$. Assume that $f(0)=0$ and $f(x)<0$ for $0<|x|<2\delta$. Let $\psi$ be any $C^2$-smooth function on $\mathbb R$ such that $\psi\equiv 1$ on $(-\delta,\delta)$ and $\operatorname{supp}\psi\subset(-2\delta,2\delta)$. Let $n$ be a large integer and let $\Lambda$ be any arithmetic progression with step $1/n$. Put $F_n(x)=\psi(x)g(x)e^{nf(x)}$. Then $$ \frac 1n \sum_{x\in\Lambda}F_n(x)-\int_{\mathbb R}F_n(x)\,dx=O(e^{-cn}) $$ with some $c>0$.

Proof:

By the Poisson summation formula, the absolute value of the left hand side is bounded by $ \sum_{y\in n\mathbb Z\setminus\{0\}}|\widehat F_n(y)| $ where $\widehat F_n(y)=\int_{\mathbb R}F_n(x)e^{-2\pi i yx}\,dx$. Thus, our aim is to estimate the Fourier transform of $F_n$. Since we want to bound an infinite sum, it will be more convenient to get a uniform bound for the Fourier transform of $F_n''$ outside $[-n,n]$, which will give us an extra factor $1/y^2$ when passing to $F_n$ itself. Notice that $F_n''=G_ne^{nf}$ where $G_n$ is continuous, supported inside $(-2\delta,2\delta)$ and analytic in $\{|z|<\delta\}$. Also $|G_n|\le Cn^2$.

Since $f$ attains its strict maximum at the origin and is analytic in a neighborhood, there exist $a>0$, $k\in\mathbb N$ such that $f(z)=-2az^{2k}+\text{higher order terms}$ and the first term dominates in some disk $\{|z|\le 2\Delta\}$. In particular, $|f(z)|\le 3a|z|^{2k}$ when $|z|\le 2\Delta$ and $\Re f(z)\le -a|z|^{2k}$ when $|z|\le 2\Delta$ and $|\Im z|\le \frac \pi{8k}|\Re z|$, say. Now take the integral defining the Fourier transform of $F_n''$ with $y\ge n$ and shift the contour of integration to the one going from $-\infty$ to $-\Delta$ to $-\Delta-\frac\pi{8k}\Delta i$ to $\Delta-\frac\pi{8k}\Delta i$ to $\Delta$ to $+\infty$.

Now just notice that $e^{nf(z)}e^{-2\pi i zy}$ is bounded by $e^{-cn}$ everywhere on the new contour. If is obvious for all pieces except the bottom horizontal one because $\Re f$ is negative and separated from $0$ there. However on the bottom horizontal piece we have $$ \Re[nf(z)-2\pi i yz]\le n|f(z)|-2\pi \frac\pi{8k} y\Delta \\ \le \left[3a(2\Delta)^{2k}-2\pi \frac\pi{8k}\Delta\right]n $$ for $y\ge n$ and we can always make $\Delta>0$ smaller, if needed, to ensure that the expression in the brackets is negative. The pre-factor $G_n$ is $O(n^2)$ everywhere on the contour, so it does not affect this estimate too much.

The rest should be more or less clear (take an appropriate partition of unity and blah-blah-blah) but feel free to ask questions if you meet any difficulty :-) .

$\endgroup$
12
  • $\begingroup$ Thanks! I have to digest this. Any chance of looking into the case where the max of $f$ occurs at $0$ or $1$? In this case we probably need a proportionality constant between the sum and the integral that is $\ne 1$. $\endgroup$
    – valle
    Commented Jul 14, 2017 at 15:24
  • $\begingroup$ @becko Indeed. I corrected. I suspect there are more stupid misprints though :-) $\endgroup$
    – fedja
    Commented Jul 14, 2017 at 15:43
  • $\begingroup$ I do not understand how you bound the integrand in the new contour, from $-\infty$ to $-\Delta$, and from $\Delta$ to $\infty$. $\endgroup$
    – valle
    Commented Jul 15, 2017 at 11:41
  • 1
    $\begingroup$ @becko A negative continuous function on a compact set is separated from $0$, isn't it? $\endgroup$
    – fedja
    Commented Jul 15, 2017 at 19:49
  • 1
    $\begingroup$ @becko Yes, we can. Let's say $f$ attains its maximum at $0$ (and $0$ only). If $f'(0)=0$, then it is still $1$. Otherwise it is the same as for $f(x)=f'(0)x$, $g(x)=1$, as M.Dus heas already observed. However the convergence can be slow now. If not a secret, what are you really after with all that? $\endgroup$
    – fedja
    Commented Jul 16, 2017 at 11:41
2
$\begingroup$

If I'm not mistaken, take $f:x\mapsto x$ and $g:x\mapsto 1$. Then, $(S_n-I_n)/S_n=\frac{\frac{\mathrm{e}^n}{n}\frac{\mathrm{1}}{\mathrm{e}-1}-\frac{1}{n}\frac{1}{\mathrm{e}-1}+\frac{1}{n}}{\frac{1}{n}\frac{\mathrm{\mathrm{e}^{n+1}}-1}{\mathrm{e}-1}}$, which converges to $\frac{1}{e}$.

$\endgroup$
9
  • $\begingroup$ You are correct. I modified the question. $I_n$ needs to be multiplied by a constant so that it approaches $S_n$ (I think). Please don't delete this answer because it is a good example. And sorry for modifying the question. $\endgroup$
    – valle
    Commented Jul 13, 2017 at 15:28
  • $\begingroup$ Hum, I think that should be true for polynomials, but I'm not sure. $\endgroup$
    – M. Dus
    Commented Jul 13, 2017 at 16:02
  • $\begingroup$ Anyway, here is a counter-example (with a non-polynomial function): Take $f(x)=\mathrm{e}^{x}=g(x)$. Then, $S_n=\frac{1}{n}\sum\mathrm{e}^{n\mathrm{e}^{k/n}}\mathrm{e}^{k/n}$, so $S_n\geq \mathrm{e}^n$. However, $I_n=\frac{1}{n}(\mathrm{e}^{n\mathrm{e}}-\mathrm{e}^n)$, so that $I_n/S_n$. Then, even with a constant $c$, you get $(S_n-cI_n)/S_n$ converges to $1$. $\endgroup$
    – M. Dus
    Commented Jul 13, 2017 at 16:06
  • 1
    $\begingroup$ Yep. Unfortunately, when the maximum is at an endpoint, the exponential decay of $R_n$ is out of question more often than not (say, for $f(x)=-x^2$ and $g(x)=1$ you still have $c=1$ but the convergence is no better than the typical convergence of the left Riemann sum to the integral ($R_n=O(1/n)$ or something like that). That's why I didn't try to address this case in the hope that the OP doesn't really need it. $\endgroup$
    – fedja
    Commented Jul 14, 2017 at 15:26
  • 1
    $\begingroup$ @becko The problem with the endpoint is that you'll have to start the contour right there and won't be able to bypass the maximum on the ascent and the descent like I did. That will result in the usual Laplace asymptotics along the imaginary axis as the main term, and that one is just a series in inverse powers. $\endgroup$
    – fedja
    Commented Jul 14, 2017 at 15:37

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .