The original version of the central limit theorem (CLT) assumes n<\/em> independently and identically distributed (i.i.d.) random variables X<\/em>1<\/span>, …, Xn<\/span><\/em>, with finite variance. Let Sn<\/span><\/em> =\u00a0X<\/em>1<\/span> + … + Xn<\/span><\/em>. Then the CLT states that<\/p>\n that is, it follows a normal distribution with zero mean and unit variance, as n<\/em> tends to infinity. Here\u00a0\u03bc<\/em>\u00a0 is the expectation of X<\/em>1<\/span>.<\/p>\n Various generalizations have been discovered, including for weakly correlated random variables. Note that the absence of correlation is not enough for the CLT to apply (see counterexamples here<\/a>). Likewise, even in the presence of correlations, the CLT can still be valid under certain conditions.\u00a0 If auto-correlations are decaying fast enough, some results are available, see here<\/a>.\u00a0 The theory is somewhat complicated. Here our goal is to show a simple example to help you understand the mechanics of the CLT in that context. The example involves observations\u00a0X<\/em>1<\/span>, …, Xn<\/span><\/span><\/em> that behave like a simple type of time series: AR(1), also known as autoregressive time series of order one, a well studied process (see section 3.2 in this article<\/a>).<\/p>\n 1. Example<\/strong><\/span><\/p>\n The example in question consists of observations governed by the following time series model: X<\/em>k<\/em>+1<\/span> =\u00a0\u03c1X<\/em>k<\/em><\/span> + Y<\/em>k<\/em>+1<\/span>, with X<\/em>1<\/span> = Y<\/em>1<\/span>, and Y<\/em>1<\/span>, …, Y<\/em>n<\/em><\/span> are i.i.d. with zero mean and unit variance. We assume that |\u03c1<\/em>|\u00a0 <\u00a0 1. It is easy to establish the following:<\/span><\/p>\n Here “~” stands for “asymptotically equal to” as n tends to infinity. Note that the lag-k<\/em> autocorrelation in the time series of observations X<\/em>1<\/span>, …, X<\/em>n<\/em><\/span> is asymptotically equal to\u00a0\u03c1<\/em>^k<\/em> (\u03c1<\/em> at power k<\/em>), so autocorrelations are decaying exponentially fast. Finally, the adjusted CLT (the last formula above) now includes a factor 1 –\u00a0\u03c1<\/em>. If course if\u00a0\u03c1<\/em> = 0, it corresponds to the classic CLT when expected values are zero.<\/span><\/p>\n 1.2. More examples<\/strong><\/p>\n Let X<\/em>1<\/span> be uniform on [0, 1] and X<\/em>k<\/em>+1<\/span> = FRAC(bXk<\/span><\/em>) where b<\/em> is an integer strictly larger than one, and FRAC is the fractional part function<\/a>. Then it is known that Xk<\/span><\/em> also has a uniform distribution on [0, 1], but the Xk<\/span><\/em>‘s are autocorrelated with exponentially decaying lag-k<\/em> autocorrelations equal to 1 \/ b<\/em>^k<\/em>. So I expect that the CLT would apply to this case.\u00a0<\/p>\n Now let\u00a0 X<\/em>1<\/span> be uniform on [0, 1] and X<\/em>k<\/em>+1<\/span> = FRAC(b<\/em>+Xk<\/span><\/em>) where b<\/em> is a positive irrational number. Again, Xk<\/span><\/em> is uniform on [0, 1]. However this time we have strong, long-range autocorrelations, see here<\/a>. I will publish results about this case (as to whether or not CLT still applies) in a future article.<\/p>\n 2. Results based on simulations<\/strong><\/span><\/p>\n The simulation consisted of generating 100,000 time series\u00a0X<\/em>1<\/span>, …, X<\/em>n<\/span>\u00a0<\/em>as in section 1.1,\u00a0<\/em><\/span>with \u03c1\u00a0<\/em><\/span>= 1\/2, each one with\u00a0n<\/em> = 10,000 observations, computing Sn<\/span><\/em>\u00a0for each of them, and standardizing Sn<\/span><\/em> to see if it follows a N<\/em>(0, 1) distribution. The empirical density follows a normal law with zero mean and unit variance very closely, as shown in the figure below. We used uniform variables with zero mean and unity variance to generate the deviates Yk<\/span><\/em>.<\/p>\n Below is one instance (realization) of these simulated time series, featuring the first n<\/em> = 150 observations. The Y-axis represents Xk<\/span><\/em>, the X-axis represents k<\/em>.\u00a0<\/p>\n It behaves quite differently from a white noise due to the auto-correlations.<\/p>\n<\/p>\n To receive a weekly digest of our new articles, subscribe to our newsletter,\u00a0here<\/a>.<\/em><\/span><\/p>\n About the author<\/strong>:\u00a0 Vincent Granville is a data science pioneer, mathematician, book author (Wiley), patent owner, former post-doc at Cambridge University, former VC-funded executive, with 20+ years of corporate experience including CNET, NBC, Visa, Wells Fargo, Microsoft, eBay. Vincent is also self-publisher at\u00a0DataShaping.com<\/a>, and founded and co-founded a few start-ups, including one with a successful exit (Data Science Central acquired by Tech Target).<\/span>\u00a0He recently opened\u00a0Paris Restaurant<\/a>, in Anacortes. You can access Vincent’s articles and books,\u00a0here<\/a>.<\/em><\/span><\/p>\n<\/div>\n Go to Source<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<\/a><\/p>\n
<\/a><\/p>\n
<\/a><\/span><\/p>\n
<\/a><\/p>\n
<\/a><\/p>\n