<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Gabriel Nagaoka on Medium]]></title>
        <description><![CDATA[Stories by Gabriel Nagaoka on Medium]]></description>
        <link>https://medium.com/@g.n.nagaoka?source=rss-9b71c49ab51d------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*P4jv_4BDCXRtWQKvJhGD4w.jpeg</url>
            <title>Stories by Gabriel Nagaoka on Medium</title>
            <link>https://medium.com/@g.n.nagaoka?source=rss-9b71c49ab51d------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 20:30:45 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@g.n.nagaoka/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[A Intuitivist Approach to Stable Laws]]></title>
            <link>https://medium.com/@g.n.nagaoka/a-intuitivist-approach-to-stable-laws-abd000e28a8c?source=rss-9b71c49ab51d------2</link>
            <guid isPermaLink="false">https://medium.com/p/abd000e28a8c</guid>
            <category><![CDATA[observation]]></category>
            <category><![CDATA[statistics]]></category>
            <category><![CDATA[mathematics-education]]></category>
            <dc:creator><![CDATA[Gabriel Nagaoka]]></dc:creator>
            <pubDate>Sun, 28 Jan 2024 22:45:42 GMT</pubDate>
            <atom:updated>2024-01-29T14:57:04.448Z</atom:updated>
            <content:encoded><![CDATA[<p>Having a fundamentalist view about the hard sciences, I think sometimes the most important aspects of statistics get swept under the carpet to give way for the absolute shallow and lifeless view ultimately sold in almost all on-line courses in &quot;Statistics&quot; and &quot;Machine Learning&quot;. Having said that, I had a thought today that reminded me of one of the most important moments I had in my understanding of statistics which came to me in a discussion with my fellow mathematical physicists friends some good four or five years ago, and it is this precise understanding that I will try to pass on.</p><p>First of all, a disclaimer that this is an intuitivist PoV of the matter, so I do no intend to be formal nor to derive results from first principles. Secondly, the reader should be familiar with basic probability theory in the form of distributions and random variables. This writing is merely a sketch of my understanding of how the <strong>Central Limit Theorem</strong> <strong>is already hinted by seeing convolutions as actions in the &quot;space&quot; parameterized by cumulants</strong>. If you are not familiar with these, don&#39;t worry it will be exposed in the following.</p><p><strong>Introduction</strong></p><p>The main object of study is the simple problem of composing multiple random variables and studying the result. Let me explain what we mean by composing. Say you have two fair dice, and you are playing a game in which a player scores according to the sum of the dice rolls, like in the board game Risk. It is only natural to ask yourself what are the probabilities of a given sum, and how can it be related to the probabilities of the output of the two fair dice. In fact you can inspect that the probability a given sum X is just the sum over all configuration of the dice, say x1 and x2 such that</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*A24cIQL3wkdH6cwfAvmm8A.png" /><figcaption>Eq. 1</figcaption></figure><p>As we have defined x1 and x2 as fair dice, their probability distributions are of course</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*G6WXRHN0OJsr8h6Y4HRB7Q.png" /><figcaption>Eq. 2</figcaption></figure><p>For instance, the probability of scoring 12 is given only by one configuration: x1 = 6 and x2 = 6. As we have 36 possible outputs, the probability of your final score X being 12 is 1/36, in other words P(12) = 1/36.</p><p><strong>Convolutions:</strong></p><p>By repeating this reasoning for all possible values of X you have calculated the convolution of the distributions of the random variables P1(x1) and P2(x2), resulting in the probability distribution of X, P(X). So we see that the convolution process is nothing more than an operation which gobbles up probability distributions and spits out an aggregated distribution related to the addition of these random variables.</p><p>It is common to denote the convolution between two distributions P1(x1) and P2(x2) as</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XTFIroRFnqG3wgGX3Lw_sQ.png" /><figcaption>Eq.3</figcaption></figure><p>As we have reasoned, the process of convolution is simply the answer to the question: what is the probability that the final result of the sum of my initial random variables x1 and x2 are X? The answer, of course is given by summing over all probabilities of P1(x1) and P2(x2) such that the final result is X! In other words:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3lXi0OsJQLKjFKIeX6Nu7g.png" /><figcaption>Eq.4</figcaption></figure><p>where the sum is made over all configurations of xi and the Dirac’s delta distribution introduces the (Eq. 1) constraint.</p><p>Up until now we have been defining the convolution of two distributions, but since the convolution of these distributions is again a distribution, we can generalize the convolution to contain any number of distributions as</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*s7EvXXRhoa9ZpcElYY-MXg.png" /><figcaption>Eq.5</figcaption></figure><p>This expression tells us how to find the probability distribution of the sum of in principle arbitrary random variables as a functional of their distribution probabilities, which is very powerful.</p><p>You may ask now, why do we care about it? And the answer is because whenever we look at nature or any statistical system, and we want to measure some quantity of these systems, we often come to find that our<br>measurements follow some statistical distribution, which can be seen as the sum of various small statistical distributions (or random variables). For instance, if we want to describe the number of orders of a given day that an e-commerce receives, it is only reasonable that we assume there are various factors that contribute to it, for example the weather, the day of the week, the price of the products and etc. In this sense, if we know the distributions for all these factors, we can derive the distribution of their composition via (Eq.5)</p><p><strong>Cumulants</strong></p><p>Now that we have talked about convolutions, we turn our attention to the so-called characteristic functions. As the name suggests, these functions serve the purpose of characterizing the distribution functions. By the use<br>of characteristic functions, one is able to completely describe a probability distribution as a set of parameters called cumulants. The characteristic function of a distribution is simply its Fourier transform</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-76umai0FDW_7UUY3vKa9A.png" /><figcaption>Eq.6</figcaption></figure><p>which can be rotated back to the original distribution with the inverse transform</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*e96mde3r0J1cgexeyOnSMA.png" /><figcaption>Eq.7</figcaption></figure><p>The cumulants cn are defined as</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yWUYZ-wx-RrVoN4kje6cSQ.png" /><figcaption>Eq.8</figcaption></figure><p>This definition may seem arbitrary, but it is quite the opposite. As we are trying to study the convolution of distributions, it is only natural to look for so-called “covariant quantities”, that is, quantities that transform “nicely” under convolutions. In other words, given the cumulants of two original distributions, the cumulants of their convolution is simply the sum of the original cumulants!</p><p><em>The punchline:</em> any given distribution P (x) can be uniquely identified by its cumulants cn, meaning there is an array of numbers (c1, c2, · · · , cn) that is in one-to-one correspondence with the distribution P (x). Whenever we take the convolution of two distributions, say, P1(x1) and P2(x2), the resulting distribution P (X) = [P1 ∗ P2](X) has its own cumulants given by the sum of the cumulants of P1(x1) and P2(x2),</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*emit2xZ5simJr0-z23m8bQ.png" /><figcaption>Eq.9</figcaption></figure><p>where Cn are the cumulants of P(X), and c(i) the cumulants of Pi(xi). In other words, there is an explicit vector space structure given by the cumulants under the convolution operation! Such that any distribution<br>can be identified with a point in an n-dimensional Real vector space and their convolution can be identified with their vectorial sum.</p><p><strong>Central Limit Theorem (CLT)</strong></p><p>What does the Central Limit Theorem has to do with any of this? Well, the most practical consequence of the CLT is that the limit of a large number of convolutions, any distribution convoluted with itself converges to a stable law. In the case where the cumulants are finite and well defined, it means that once you have a distribution convoluted with itself many times, it must collapse to a Gaussian distribution. To see that, we first introduce the normalized cumulants λn which are nothing more than an dimensionless version of the cumulants cn. We must recall that the probability distributions live inside the integrals, so that they are indeed dimensionful quantities. To define λn, we first recall the good old variance σ, which is also a dimensionful quantity. From the cumulants and σ, we construct the dimensionless quantities</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F0IWjKYMhRFn-mFEVQG8tA.png" /><figcaption>Eq.10</figcaption></figure><p>As this is just a scaling, one can also use λn as the identifying the distributions in a vector space.</p><p>Now comes the beautiful result: if you start with a given distribution, which is completely defined by its λn, and you convolute it with itself N times, the normalized cumulants get mapped as</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EoYHJyHkj8R6oiDWSwMaFA.png" /><figcaption>Eq.11</figcaption></figure><p>which means they are scaled by a factor of N^(1−2n/2), so that for n ∈ {1, 2} as N → ∞ the cumulants increase, and for n &gt; 2 the cumulants tend to zero. Now you ask yourself, what is the distribution that has λ1 and λ2 as its only non-vanishing parameters? In other words, what is the distribution that is completely specified by its mean and its standard deviation? Well, its a Gaussian!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/440/1*k6T2gIhnbyNcbdTHDDUtng.png" /><figcaption>Flow of three distinct distributions to the Gaussian plane (c1,c2,0,0,…)</figcaption></figure><p>What I find exciting about these statistical &quot;comments&quot; is that they hold in generality for distributions that fall to zero faster than any inverse polynomial, and the convolutions that lead to the annihilation of higher order cumulants have no information about the distributions that we start from. In fact, Eq.11 defines a flow equation, indicating the sub-manifold of (c1,c2,0,0,…) of the space of cumulants as a fixed point, which is exactly the space of all gaussian distributions.</p><p>A more detailed derivation of these results in a similar language can be found in the book <em>Theory of Financial Risk and Derivative Pricing: From Statistical Physics to Risk Management.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=abd000e28a8c" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>