Episode 1's 1/√N and Episode 4's universality were two faces of one theorem ── statistics has been doing renormalization for a hundred years
Something occurred to me after finishing the main series: I wrote Episode 1's \(1/\sqrt N\) and Episode 4's universality as two separate stories. They are two faces of one theorem. Take "add up \(N\) independent variables and divide by \(\sqrt N\)" and regard it as a transformation from distribution to distribution ── it has exactly the same structure as Episode 4's block spins. Doing it twice equals doing it once for four (the composition rule); the Gaussian maps to itself (a fixed point); and the central limit theorem is the universality claim that everything flows there. Better still, the "quirks" of the original distribution ── skewness, kurtosis ── die off cleanly as \(N^{1-k/2}\): these are irrelevant operators. The whole renormalization group appears in a form a high-schooler can compute by hand.
Draw two independent samples from the same distribution (mean 0, variance 1), add them, divide by \(\sqrt2\).
Here \(*\) is convolution (the operation that corresponds to adding). This \(\mathcal{T}\) is a map on the space of probability distributions. Episode 4 said "coarse-graining is a map from model to model" ── same shape. And the composition rule matches too: applying \(\mathcal{T}\) twice equals adding four and dividing by two.
Why divide by \(\sqrt2\)? Variance adds, so two samples give variance 2; dividing by \(\sqrt2\) brings it back to 1. That division is the same rescaling as shrinking the lattice in Episode 4. Coarse-graining always comes as a pair ── squash, then re-measure. The same is true here.
Feed the standard normal \(p(x)=\frac{1}{\sqrt{2\pi}}e^{-x^2/2}\) into \(\mathcal{T}\). A Gaussian convolved with a Gaussian is Gaussian, with variances adding to 2; divide by \(\sqrt2\) and the variance returns to 1 ── exactly the distribution you started with.
Episode 4 defined a fixed point as "a model that maps to itself under coarse-graining." The Gaussian is a fixed point of this transformation. And the central limit theorem can be restated: this fixed point is an attractor for a wide range of starting points.
Here is the computable centrepiece. There are quantities called cumulants \(\kappa_k\) that measure a distribution's quirks (\(\kappa_1\) = mean, \(\kappa_2\) = variance, \(\kappa_3\) behind skewness, \(\kappa_4\) behind kurtosis). Cumulants have exactly one lovely property ── they simply add when you add independent variables.
Adding \(N\) gives \(\kappa_k \to N\kappa_k\). Dividing by \(\sqrt N\) divides the \(k\)-th cumulant by \((\sqrt N)^k\):
$$\kappa_k^{(N)}=\frac{N\kappa_k}{N^{k/2}}=\kappa_k\,N^{\,1-k/2}$$Tabulate it and the renormalization-group classification falls right out.
| \(k\) | quantity | \(N\) dependence | in RG language |
|---|---|---|---|
| 1 | mean | \(N^{+1/2}\) (grows) | relevant (shift it and it diverges) |
| 2 | variance | \(N^{0}\) (fixed) | marginal (it sets the scale) |
| 3 | skewness | \(N^{-1/2}\) (decays) | irrelevant |
| 4 | kurtosis | \(N^{-1}\) (decays) | irrelevant |
| \(k\ge3\) | higher quirks | \(N^{1-k/2}\) | even more irrelevant |
Episode 4 said "relevant directions are few, irrelevant ones overwhelmingly many." Here it is literally one relevant, one marginal, and all the rest irrelevant ── with the decay rates derivable by hand. This is the simplest fully computable instance of the renormalization group.
The information about "what shape the original distribution had" lives in \(\kappa_3,\kappa_4,\kappa_5,\dots\), and all of it dies as \(N^{1-k/2}\) ── so the origin is forgotten. The very phenomenon of Episode 4's water and magnet is happening with dice and coins.
Everything so far quietly assumed the variance is finite (you can only divide by \(\sqrt N\) if variance is defined). What if it isn't?
The standard example is the Cauchy distribution \(p(x)=\frac{1}{\pi(1+x^2)}\). Its tail only falls as \(1/x^2\), so the variance integral diverges. And this distribution ── returns to itself if you add two and divide by \(2\), not by \(\sqrt2\).
Distributions that return to themselves under this scaling are called stable distributions, and they form a family of fixed points labelled by \(\alpha\) (the tail weight). \(\alpha=2\) is the Gaussian, \(\alpha=1\) the Cauchy. And what decides which fixed point you land on is only how the tail falls ── the shape near the middle is irrelevant, in every sense. In Episode 4's words:
stable distribution = fixed point, domain of attraction = universality class, tail exponent \(\alpha\) = the relevant parameter
| Starting distribution | Tail | Divide by | Flows to |
|---|---|---|---|
| uniform, binomial, exponential, … (anything with finite variance) | light | \(\sqrt N\) | Gaussian (\(\alpha=2\)) |
| Cauchy | \(1/x^{2}\) | \(N\) | Cauchy (\(\alpha=1\)) |
| Pareto type (tail index 1.5) | \(1/x^{2.5}\) | \(N^{2/3}\) | the \(\alpha=1.5\) stable law |
The Gaussian looks "obvious" only because most quantities around us have finite variance. In heavy-tailed worlds ── earthquake magnitudes, city populations, financial crashes ── a different fixed point rules. That is why "just take the average and relax" fails there.
The figure below actually convolves and rescales the distribution you pick, over and over (computed in your browser). The slider is the number added, \(N=2^k\).
Pick uniform, coin or exponential and within a few steps they all lie on top of the Gaussian (grey dashed) ── utterly different starting points, identical destination. The coin flip (±1) is especially good: watch a Gaussian rise out of two spikes. And pick Cauchy and it never becomes Gaussian. The shape doesn't change from beginning to end ── it is already sitting on a different fixed point.
One more correspondence the main series failed to use.
In 1922 Fisher introduced sufficient statistics. A function \(T(x)\) of the data is "sufficient" if, once you know \(T\), the raw data carries no further information about the parameter.
That is exactly what Episode 1 did ── temperature and pressure are sufficient statistics for the microstate. Translated into statistical language, "throw it away and the answer doesn't change" becomes "reduce to a sufficient statistic."
Jaynes (1957) argued: subject to the constraint that only the mean energy is known, choose the distribution that is maximally ignorant about everything else ── i.e. maximise entropy ── and the Gibbs distribution \(e^{-E/kT}\) comes out automatically.
So statistical mechanics can be derived as a procedure of inference, adding no new physical law. When Episode 1 said "entropy = the amount of information discarded," that was not a metaphor but the central claim of this lineage. Sufficient statistics, maximum entropy and exponential families are three faces of one mathematical structure.
Established: that the central limit theorem can be formulated as convergence to a fixed point of a transformation on distribution space (the probabilistic view of the renormalization group, Jona-Lasinio 1975 and others); the cumulant scaling \(\kappa_k^{(N)}=\kappa_k N^{1-k/2}\) (elementary and exact); that stable distributions are the fixed points of \(N^{1/\alpha}\) scaling and that domains of attraction are determined by the tail exponent (the Gnedenko–Kolmogorov classification, rigorous); Fisher's sufficient statistics and the factorisation theorem; Jaynes' derivation of the Gibbs distribution from maximum entropy; and that at a critical point the diverging correlation length breaks the CLT's premise so fluctuations become non-Gaussian ── all established mathematics and physics.
But do not over-identify: (1) The fixed points here correspond, in field-theory terms, to the Gaussian (free / mean-field) fixed point. Interacting fixed points like the Wilson–Fisher one ── the star of Episode 4, and where anomalous dimensions come from ── do not fit inside this simple CLT frame. The order matters: "the CLT is the simplest example of RG," not "RG is a restatement of the CLT." (2) The cumulant classification assumes the distribution is smooth enough that cumulants exist. (3) "Sufficient statistic = coarse-graining" is a correspondence once you fix the question (parameter estimation); change the question and what you should keep changes too (the general theory is Bonus ④).
Treat "add \(N\) independent variables and divide by \(\sqrt N\)" as a transformation on distribution space and the whole renormalization group appears ── the Gaussian is a fixed point, the central limit theorem is the claim that it attracts, and since the \(k\)-th cumulant scales as \(\kappa_k N^{1-k/2}\), the mean is relevant, the variance marginal, and skewness and everything above it irrelevant. Episode 4's classification, derivable by hand.
There is more than one fixed point. With infinite variance you divide by \(N^{1/\alpha}\) and a family of stable fixed points appears (\(\alpha=2\) Gaussian, \(\alpha=1\) Cauchy). What decides where you land is only the fall-off of the tail ── the middle of the distribution is irrelevant. That is a universality class. And when the correlation length diverges at a critical point, the CLT's premise breaks, which is exactly why non-trivial exponents like \(\beta\approx0.326\) appear. Episode 1 (where the CLT works) and Episode 4 (where it breaks) are joined by a single line.
Print / make a PDF: ⌘+P (Ctrl+P on Windows). On screen, pick a starting distribution with the buttons and raise the number added with the slider to watch it flow to a fixed point. Only Cauchy fails to reach the Gaussian. "See the answer" opens each solution.