I am learning about local limit theorems. The following example is probably why we don't have a "convergence in density/pmf."

Ex: $X_1,X_2,\ldots$ is a sequence of independent RVs with mean $a$ and variance $\sigma^2$, then the distribution function $F_n(x)$ of the normalized sum $Z_n=\frac{\sum_{i=1}^n (X_i-a)}{\sigma \sqrt n}$ converges to the normal distribution function. However, this does not imply the convergence of density of $p_n(x)$ of $Z_n$ to $\frac{e^{-x^2/2}}{\sqrt{2\pi}}.$

What conditions can be included to ensure convergence of PDFs/PMFs?

  • $\begingroup$ When you say "convergence", you are specifically looking for pointwise convergence? $\endgroup$ – Nate Eldredge Nov 9 '16 at 14:17
  • $\begingroup$ @NateEldredge, I had the same question - what norms are we looking at?. You can also look at convergence (in some norm) of the CDF. $\endgroup$ – usul Nov 9 '16 at 14:21

You should look at the old book

Gnedenko and Kolmogorov: Limit Distributions for Sums of Independent Random Variables.

where they have examples of local limit theorems. I assume many more things were discovered in the meantime. Here is an example of theorem that answers your question.

Theorem. Suppose that $(X_n)_{N\geq 1}$ is a sequence of i.i.d. random variables with common density (PDF) $p(x)$. Denote by $p_n(x)$ the density of $Z_n=X_1+\cdots +X_n$. Assume the following conditions.

  • The random variables $X_n$ are $L^2$, i.e., $$ \sigma^2:=\int_{-\infty}^\infty x^2 p(x) dx<\infty. $$
  • There exists $r\in (1,2]$ and a positive integer $n$ such that $\newcommand{\bR}{\mathbb{R}}$ $p_n\in L^r(\bR)$. (This is guaranteed if e.g. $p(x)$ is Lipschitz.)


$$ \sigma\sqrt{n}p_n(\sigma\sqrt{n}\; x) \to \frac{1}{\sqrt{2\pi}} e^{-\frac{x^2}{2}} $$

uniformly with respect to $x\in\bR$. For a proof and more details see sections 46,47 of the above book.

| cite | improve this answer | |
  • $\begingroup$ Is it possible to obtain the rate of convergence? I.e. how large is the error/remainder if $n$ is large but not infinite? $\endgroup$ – ManUtdBloke Oct 13 at 11:52
  • $\begingroup$ Have you cehcked Terry Tao's blog on this? terrytao.wordpress.com/2015/11/19/… $\endgroup$ – Liviu Nicolaescu Oct 14 at 14:26
  • $\begingroup$ Terry Tao's blog post seems to only discuss convergence in probability, whereas you have specified pointwise/uniform convergence of the pdf. Pointwise/uniform convergence of the pdf is what I am interested in. $\endgroup$ – ManUtdBloke Oct 14 at 15:16
  • $\begingroup$ In Section 47, p. 228, of the above book of Gnedenko and Kolmogorov there are estimates for the rate of convergence provided $$\int_{-\infty}^\infty |x|^k p(x) dx<\infty,$$ for some $k\geq 3. $ $\endgroup$ – Liviu Nicolaescu Oct 18 at 14:20

I don't know if this is an answer to your question: Years before I presented the following result in a lecture about weak convergence: Theorem: Assume that $X_n, n \in N$ are real r.v. with unimodal density $\varphi_n$, $P_0$ a probability measure with unimodal, continuous density $\varphi$, unique maximum value $t^*$ and d.f. $\Phi$, Then if $X_n \to P_0$ in distribution, Then $\liminf_{n \to \infty} \inf_{t \in R} (\varphi_n(t) - \varphi(t)) = 0$ $\lim_{n \to \infty} \sup_{|t-t^*| \geq \delta} |\varphi_n(t) - \varphi(t))| = 0$ for any $\delta > 0$ and $\lim_{n \to \infty} \sup_{t \in R} |\varphi_n(t) - \varphi(t))| = 0$ if any $\varphi_n$ is logarithmically concave.

I don't know where you can find this reasult, but the proof is not too hard. There also is a version for discrete r.v.

The usual simple local theorems are special cases.

| cite | improve this answer | |

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.