How the mathematics was forged

Mathematics arrives in textbooks embalmed: definition, theorem, proof, as if it had always existed and someone merely wrote it down. Every tool in this track has a messier and better biography. Each one was forged under the pressure of a problem that would not yield — a debt to settle, a planet to find, a harvest to judge, a war to win, a wire that garbled its messages. Knowing the pressure is not decoration: it tells you what each tool is for, which is the thing the embalmed version hides. This is the track’s story; every thread ends at the door of a module that picks it up with full rigor.

Equations before algebra had letters

Four thousand years ago, Babylonian scribes pressed quadratic problems into clay — a field’s area and perimeter known, its sides wanted — and solved them by recipe, without notation, one worked example at a time. The ninth-century Baghdad scholar al-Khwarizmi collected such recipes into a book on al-jabr (“restoration” — moving a term across the equals sign), giving algebra its name; the Latinization of his own name gave us algorithm. The lesson hiding in the etymology: from the very beginning, this mathematics was procedures for extracting unknowns from measurements — data science’s job description, four millennia early. What was missing for three of those millennia was notation: not until Descartes (1637) fused geometry with symbols — a curve is an equation, a point is a pair of numbers — did shape become computable. Every “vector as a point in feature space” in this track walks through the door Descartes opened.

Calculus: the mathematics of change gets a machine

Newton and Leibniz (1660s–80s, independently, then bitterly) built the machine for motion: the derivative as instantaneous rate, the integral as accumulation, and the fundamental theorem binding them. But the piece this track leans on hardest came from Fermat a generation earlier: his observation that a curve is momentarily flat at its peak — that maxima live where the rate of change vanishes. Set the derivative to zero: with that single move, “find the best” became a computation instead of a search, and every fitted model in this curriculum descends from it. Brook Taylor (1715) added the tool the machine-learning era would prize most: any smooth function is, near a point, approximately polynomial — approximately linear if you’re humble, quadratic if you’re ambitious. Two centuries later that humility became an algorithm: in 1847 Augustin-Louis Cauchy, wrestling with orbital calculations too big to solve directly, proposed simply stepping downhill along the slope — gradient descent, in a four-page note, waiting patiently for computers and the Optimization module.

Probability: mathematics from the gambling table

In 1654 a gambler, the Chevalier de Méré, asked Blaise Pascal an accounting question: two players interrupt a game of chance mid-way — how should the pot be split fairly, given the score? Pascal wrote to Fermat; their correspondence invented expected value and with it the scandalous idea that the unknown future could be calculated. The thread runs fast from there: Jacob Bernoulli proved the law of large numbers (1713) — frequencies converge to probabilities, the license for all estimation; the Reverend Thomas Bayes left behind, unpublished at his death in 1761 (Richard Price sent it to print in 1763), a theorem for reversing conditional probabilities — from evidence back to cause; and Laplace, who reinvented and industrialized all of it, called probability “common sense reduced to calculus” and used it on everything from jury verdicts to the mass of Saturn. When the Probability module derives Bayes’ rule in three lines, remember it took a century and a half of scandal — respectable mathematics did not do uncertainty — before those lines were allowed.

Gauss: error itself becomes mathematics

You know the 1801 Ceres story from the Start Here page — here is its mathematical core. Gauss’s insight was not merely fitting the orbit; it was treating measurement error as a lawful random object. Assume errors are as likely positive as negative, small errors likelier than large, and the average the best guess — and the error law must be the bell curve, from which least squares follows as the optimal procedure. Error, the thing science had treated as shameful failure, became a thing with a distribution — tameable, averageable, priceable. This is the hinge of the whole track: the moment the mathematics of certainty (algebra, calculus) married the mathematics of uncertainty (probability), and statistics was the child. The Probability module’s MLE lesson replays the marriage in modern dress: least squares IS maximum likelihood under Gauss’s error law.

Linear algebra: three births and a late coronation

The tool that feels most “designed” for machine learning was born three separate times for three other reasons. As elimination: solving simultaneous equations, from Chinese rod-arithmetic (the Nine Chapters, ~200 BCE) to the systematic method that bears Gauss’s name because he used it on — again — astronomical data. As transformation: Cayley (1858) noticed that the operations themselves (rotate, stretch, shear) could be added and multiplied like numbers — matrices as objects, the Linear Algebra module’s “matrices are verbs” in its original wording. And as spectrum: eigenvalues emerged from physical vibration — Euler’s spinning bodies, Fourier’s heat, later the literal spectra of atoms in quantum mechanics, which is where “spectral theorem” gets its name. The coronation came late: Pearson (1901) found the principal axes of data clouds, Hilbert’s spaces made functions into vectors — and a century later, embeddings made meaning into vectors, and the three births turned out to be one subject. The module teaches them unified; history had no such luxury.

Fisher: statistics grows teeth in a fertilizer plot

Rothamsted agricultural station, 1919. R. A. Fisher, hired to make sense of ninety years of harvest records, faced the question every A/B test still faces: this plot yielded more — luck, or the fertilizer? Out of that mud came modern statistics: likelihood as the principle for estimation (the MLE lesson’s engine), randomized experiments as the design that makes causal claims honest, significance tests, analysis of variance — and Fisher information, the curvature that prices what data can tell you, which you will meet measuring the sharpness of likelihood peaks. Kolmogorov completed the era in 1933 by giving probability proper axioms (measure theory underneath — the reason your densities integrate and your expectations behave), ending two centuries of the subject being brilliant but unlicensed.

Shannon: information becomes a quantity

In 1948, at Bell Labs, Claude Shannon asked the telegraph century’s question properly: what is information, such that a noisy wire can carry it? His answer — surprise is $-\log p$, a source’s information rate is its entropy, channels have exact capacities — created information theory whole, in one paper, with the casual side effect of naming the bit. Two consequences run this curriculum. Practically: every compression format and every error-correcting code in your phone is applied Shannon. Conceptually — and this is the track’s closing twist — his quantities turned out to be the native language of learning: cross-entropy loss is literally the coding cost of your model’s wrong beliefs, and training a language model is compressing the internet. The track’s final module (Information Theory, new door, two lessons down) makes both sentences precise.

The braid

• ~1800 BCEBabylonian scribes solve quadratics by recipe — procedures for unknowns, before notation exists.
• 820sal-Khwarizmi’s al-jabr names algebra; his name becomes “algorithm”.
• 1637Descartes fuses shape and symbol — geometry becomes computable.
• 1654Pascal and Fermat settle a gambler’s dispute — probability is born from a question about fairness.
• 1665–87Newton and Leibniz build calculus; Fermat’s flat-at-the-top insight makes optimization a computation.
• 1715Taylor: every smooth function is locally polynomial — the approximation the whole of deep learning runs on.
• 1763Bayes’ theorem, published posthumously — evidence reversed back to cause.
• 1801–09Gauss finds Ceres: error gets a distribution, least squares gets a justification — statistics is conceived.
• 1847Cauchy, drowning in orbital arithmetic, proposes stepping downhill — gradient descent, 110 years before hardware worthy of it.
• 1858–1901Cayley makes matrices objects; eigenvalues arrive from vibrating physics; Pearson points principal axes at data.
• 1920sFisher, in the fertilizer plots: likelihood, randomized experiments, information — statistics grows teeth.
• 1933Kolmogorov gives probability axioms — two centuries of brilliance, finally licensed.
• 1948Shannon: information is a quantity — entropy, bits, channel capacity, in one paper.
• nowAll five threads braid into machine learning — and into the five modules of this track.

Read the braid backwards and the track’s syllabus assembles itself: Optimization is Fermat–Cauchy (find the best, by stepping downhill); Linear Algebra is elimination–Cayley–spectrum (the language of many numbers at once); Probability & Inference is Pascal–Bayes–Gauss– Fisher (calculated uncertainty); Information Theory is Shannon (surprise, priced). Nothing in the modules ahead was invented for machine learning — and that is precisely their power. Tools forged on planets, harvests and telegraph wires, still cutting three centuries later, are tools whose sharpness does not depend on fashion. When a derivation ahead feels abstract, come back here and find its pressure: somebody once needed exactly that inequality, badly, with money or harvests or messages on the line.

Exercises