Is there a central limit theorem for bounded non identically distributed random variables ? - MathOverflow most recent 30 from http://mathoverflow.net 2013-05-25T05:19:12Z http://mathoverflow.net/feeds/question/29508 http://www.creativecommons.org/licenses/by-nc/2.5/rdf http://mathoverflow.net/questions/29508/is-there-a-central-limit-theorem-for-bounded-non-identically-distributed-random-v Is there a central limit theorem for bounded non identically distributed random variables ? Caroline Fontaine 2010-06-25T13:02:35Z 2010-06-30T14:31:07Z <p>I have a sequence of centered independent random variables $X_i$ that are all bounded by one in absolute value. They are not identically distributed, though. I would like to know if the <strong>central limit theorem</strong> is still true for such a sequence. Putting $S_n= X_1+...+X_n$, do we have $$c_n = P(\ {S_n\over\sigma(S_n)} \in [a,b] ) - {1\over \sqrt{2\pi}}\int_a^b exp(-t2/2) dt \ \rightarrow \ 0\ ?$$ (let's assume $\sigma(S_n)$ goes to infinity with n). I guess it is true but I can't find a reference.</p> <p>Also, what can be said from the rate of convergence of $c_n$ ? Since the $X_i$ are uniformly bounded, does $c_n$ goes to zero exponentially fast ? </p> http://mathoverflow.net/questions/29508/is-there-a-central-limit-theorem-for-bounded-non-identically-distributed-random-v/29515#29515 Answer by Mark Meckes for Is there a central limit theorem for bounded non identically distributed random variables ? Mark Meckes 2010-06-25T13:36:33Z 2010-06-30T14:31:07Z <p>For your first question, the answer is yes, and I don't understand why it isn't better known since all the classical proofs of the central limit theorem generalize easily to that setting. See <a href="http://en.wikipedia.org/wiki/Central_limit_theorem#Lack_of_identical_distribution" rel="nofollow">this section</a> of the Wikipedia page on the central limit theorem.</p> <p><strong>Added:</strong> I overstated slightly, since the classical proofs don't easily generalize to give the most general conditions. But in the OP's setting, as coudy points out, Lindeberg's condition implies that it's enough to have $\sigma(S_n) \to \infty$ (whereas if the $X_i$ were identically distributed we would of course have $\sigma(S_n) = \sqrt{n}\sigma(X)$).</p> <p>For your second question, even under uniform boundedness, $c_n$ only goes to zero like $n^{-1/2}$ in general. See for example <a href="http://www.jstor.org/stable/2044679" rel="nofollow">this paper</a>.</p> http://mathoverflow.net/questions/29508/is-there-a-central-limit-theorem-for-bounded-non-identically-distributed-random-v/29526#29526 Answer by Suresh Venkat for Is there a central limit theorem for bounded non identically distributed random variables ? Suresh Venkat 2010-06-25T16:23:00Z 2010-06-25T16:23:00Z <p>There is an <a href="http://www.amazon.com/Concentration-Measure-Analysis-Randomized-Algorithms/dp/0521884276" rel="nofollow">entire book on concentration inequalities</a> that goes well beyond the iid normal case. Loosely speaking, any collection of variables satisfying some kind of Lipschitz property will have exponential tails that can often lead to a CLT-type form. Azuma's inequality is one example covered in the book. </p> http://mathoverflow.net/questions/29508/is-there-a-central-limit-theorem-for-bounded-non-identically-distributed-random-v/30022#30022 Answer by coudy for Is there a central limit theorem for bounded non identically distributed random variables ? coudy 2010-06-30T08:58:07Z 2010-06-30T08:58:07Z <p><em>Theorem</em> (Billingsley, "probability and measure", example 27.4)</p> <p>Let X_i a sequence of independent, uniformly bounded random variables with zero mean, such that $\sigma(S_n)$ goes to infinity with n. Then $S_n/\sigma(S_n)$ converges in law to the normalized Laplace-Gauss distribution.</p> <p>This follows from the <em>Lindeberg triangular array theorem</em>. As pointed out in the other answers, the convergence can be slow. The <em>Bernstein inequality</em> may be used to bound the tail. Under the assumption of the previous theorem, for all n, we have</p> <p>$$P(S_n>t) \leq exp(-t^2/(\sigma^2(S_n)+Ct/3))$$</p> <p>where C is a bound for the |X_i|'s.</p>