A drunkard stands on a long street and flips a coin at every step to go left or right. Where is he after a thousand steps? The instinct that says the pluses and minuses cancel, so he is still near the start, is half right and half badly wrong — his expected position really is 0, but what cancels is direction, not distance.
The real answer: after $n$ steps his typical distance from the origin is about $\sqrt n$ steps — roughly 10 after 100 steps, roughly 100 after 10000. Not $n$ (that would be walking straight), not 0 (that would be cancellation), but the square root. Randomness does not pin you in place; it pushes you away at a rate slower than linear. That is diffusion.
Dimension changes the ending too: a walk on a line or a plane returns to its starting point with probability 1, while in three dimensions the chance is only about 34% (Pólya's theorem). The drunkard always finds his way home; the drunken bird never finds its nest.
$X_i$ are i.i.d. single steps, equal to $+1$ or $-1$ with equal probability; $S_n$ is the position after $n$ steps. Expand the square: $\mathbb{E}[S_n^2]=\sum_i\mathbb{E}[X_i^2]+\sum_{i\ne j}\mathbb{E}[X_iX_j]$. Each term of the first sum equals 1 and there are $n$ of them; every term of the second vanishes, because the steps are independent with mean zero. Hence the typical displacement is $\sqrt{\mathbb{E}[S_n^2]}=\sqrt n$.
Why do the cross terms vanish? Because in the space whose inner product is the expectation of a product, independence (with mean zero) simply is orthogonality. Variances add, and so $\sqrt n$ turns out to be the Pythagorean theorem: lay $n$ mutually perpendicular unit vectors end to end and the resultant has length $\sqrt n$. Independent and perpendicular are the same statement — $\sqrt n$ is Pythagoras echoing inside probability space. Pólya's dimensional threshold is likewise mere arithmetic: the return probability hinges on whether $\sum_n n^{-d/2}$ diverges, and $d=2$ is exactly the dividing line.
In 1905 Einstein used $\langle x^2\rangle=2Dt$ to explain the motion of pollen grains, and from it computed Avogadro's number for the first time — the $\sqrt n$ law became experimental evidence that atoms exist. PageRank is the stationary distribution of a random walk on the web graph; MCMC uses one to explore a posterior; node2vec builds node embeddings from walks on graphs. SGD is a random walk with drift: the gradient is the drift, while minibatch noise accumulates as $\sqrt{\text{steps}}$ — which is where the rule "scale the learning rate with the square root of batch size" comes from.
Let $X_n$ be your bankroll after round $n$ of some game. A martingale has exactly one clause: lay the entire past on the table, and the best prediction of the next bankroll is still whatever it is right now.
It does not say the value stays put — the swings can be violent enough to ruin you. It says the trend is zero: the history contains no extractable directional information. That is the precise content of the word fair.
This settles the doubling system (double after a loss; one win nets you a dollar). It does almost surely win that dollar — at the price of an astronomical loss on the vanishingly unlikely branch of consecutive losses, and the two balance exactly. The optional stopping theorem promotes that remark into a theorem: as long as you cannot foresee the future and your stakes are bounded, no stopping rule can change the expectation.
$\mathcal{F}_n$ formalises everything known up to step $n$ — a $\sigma$-algebra, which is where the measure theory of Day 21 pays off; $\mathbb{E}[\,\cdot\mid\mathcal{F}_n]$ is the best prediction given that information. Replace the equality by $\le$ and you get a supermartingale: once the house takes its cut, your bankroll is exactly that.
A martingale compresses the value judgement "fair" into a single equation, which is what makes it possible to prove theorems about impossibility. One layer deeper: any reasonable process decomposes uniquely into a predictable drift plus a martingale (the Doob decomposition) — the martingale is the pure, unpredictable component of randomness. It is also probability's conservation law: on Day 18 symmetry produced conserved quantities; here the martingale property conserves expectation, and optional stopping is the statement of that law.
The foundational theorem of finance: no arbitrage ⟺ there exists a measure under which discounted prices form a martingale (the risk-neutral measure) — that is the skeleton of Black-Scholes. Azuma–Hoeffding gives concentration bounds for randomised algorithms and online learning. Anytime-valid inference in A/B testing (test martingales, e-values) lets you look at the data whenever you like without inflating false positives, curing exactly the peeking problem from Day 34. In reinforcement learning the TD error is a martingale difference sequence, which is the heart of the convergence proofs.
Shrink the steps of a random walk and pack them ever more densely, approaching continuous time. The decisive question is: as the time step $\Delta t\to0$, how small should each spatial step be?
Choose $\Delta t$ and the variance collapses to zero, leaving a motionless straight line; choose a constant and the variance explodes. The only surviving scale is $\sqrt{\Delta t}$ — the mirror image of the $\sqrt n$ law.
The consequence is startling: the path is continuous everywhere and differentiable nowhere. The reason is self-similarity — magnify any small piece (time ×4, space ×2) and you see the same jaggedness, so it never straightens into a tangent the way a smooth curve does (a direct relative of the fractals of Day 17). This is also the source of "$dW$ squared is of the same order as $dt$": displacement squared scales like time, so displacement is of higher order in smallness than time itself.
The axioms: it starts at 0; increments over non-overlapping intervals are independent; each increment is normal with mean 0 and variance equal to the length of the interval; paths are continuous. Variance $=t-s$ is just the formal way of writing the $\sqrt{\Delta t}$ scale. And $(dW)^2=dt$ is the core of Itô calculus: second-order terms are no longer negligible, so the chain rule acquires an extra term $\tfrac12 f''\,dt$.
Here the normal distribution is not an assumption but a forced conclusion: whether the single step is a coin, a die, or any law with finite variance, the rescaled limit converges to the same Brownian motion (Donsker's invariance principle). Microscopic detail is forgotten entirely and one object remains — this is mathematics' deepest species of beauty: universality. The second beauty is its equivalence with the heat equation (Feynman–Kac): probability and analysis translate into each other here.
Financial assets are modelled by geometric Brownian motion, and Black-Scholes follows from Itô's lemma; in physics it is the noise term of the Langevin equation. The forward process of a diffusion model is precisely the stepwise injection of Brownian noise, and the reverse denoiser learns the score — today's image and video generation rests on the mathematics of Einstein's 1905 paper. The continuous-time approximation of SGD (a stochastic differential equation) is used to explain flat minima and generalisation.
Take a $1000\times1000$ matrix, fill every entry independently with a random number (keeping it symmetric), and compute its thousand eigenvalues. You would expect a mess. The opposite happens: plotted as a histogram they fall almost perfectly under a semicircular curve; draw fresh random numbers and the shape is the same semicircle again.
More counterintuitive still is the repulsion between eigenvalues: independently scattered random points clump together, whereas eigenvalues spread out almost evenly, as though they carried like charges.
At the level of a single entry there is total disorder; at the level of the whole spectrum there is rigid order. Dimension does not pile randomness into a bigger mess — it irons it flat.
Wigner's semicircle law: for an $N\times N$ real symmetric matrix with independent entries of mean 0 and variance $\sigma^2$, divide the eigenvalues by $\sigma\sqrt N$ and their empirical distribution converges to the formula above as $N\to\infty$. That $\sqrt N$ scaling is the $\sqrt n$ law once more — multiplying by a vector adds $N$ random terms within a row, whose magnitude is exactly $\sqrt N$. The counterpart in data analysis is the Marchenko–Pastur law.
This is the prettiest display of concentration in high dimension: the fluctuation of a macroscopic function of $N^2$ random numbers — the spectral distribution — dies away like $1/N$, and the random becomes determinate. The term for it is self-averaging. Then in 1972, over tea at Princeton, Montgomery and Dyson noticed that the spacings between zeros of the Riemann $\zeta$ function match the eigenvalue spacings of random matrices (the GUE) exactly — the deepest open problem in number theory shares a statistical law with quantum chaos, and half a century later nobody knows why.
PCA and covariance denoising: Marchenko–Pastur supplies the baseline for what the spectrum looks like if it is all noise, so only eigenvalues past the upper edge can be real signal (used for portfolios and gene expression). Initialising a neural network is in essence controlling the spectral radius of the weight matrices — the Xavier/He variance formulas exist so that signal neither explodes nor dies across depth; the spectrum of the Hessian explains the wildly unequal curvature along different directions of the training landscape. In engineering there are also MIMO channel capacity and the phase-transition thresholds of compressed sensing.