Mathematics as a Modeling Language

Machine-learning mathematics is easiest to learn when equations are treated as compressed arguments. A vector says which quantities belong together. A probability distribution says which outcomes are possible and how uncertain they are. A loss function states what counts as a mistake. A gradient describes a local direction of change. A generalization bound connects what happened on sampled data to what might happen beyond it.

The symbols are not the point by themselves. The point is to make claims precise enough that another person can inspect the assumptions, reproduce the calculation, and challenge the conclusion.

Mathematics fits the physical world far better than anyone can explain. Eugene Wigner called that fit “a wonderful gift which we neither understand nor deserve” in a 1960 essay. A gift is not a guarantee. Nothing in the notation promises that the chosen equation matches the system in front of you.

The rest of this lesson keeps that warning concrete rather than decorative. A launcher was lost 37 seconds after ignition to a conversion nobody had tested. A missile battery's tracking equation drifted 687 metres purely as a function of uptime. A flu model fitted from 50 million candidate search terms then ran high in 100 weeks out of 108. An admissions table and its own departments disagree about the sign of the effect. In none of the four did anyone get the algebra wrong.

A fraud model illustrates why one branch of mathematics is rarely enough. Six questions arrive at once, and two of them — what a score means, and whether the calculation survives being carried out — are worth seeing with published measurements attached rather than taken on trust.

This course emphasizes operational understanding over memorized theorems. You will still encounter equations. But each equation will be tied to a geometric picture, a modeling decision, or a diagnostic. The course does not attempt to replace a full degree in linear algebra, probability, statistics, or optimization. It selects the parts that repeatedly govern modern ML systems and develops them deeply enough for technical judgment.

By the end, you should be able to read an unfamiliar method and ask better questions. What space does it operate in? What uncertainty does it represent? What objective does it optimize? What evidence supports the conclusion? What numerical approximation makes it possible?

Most mathematical questions in ML can be located under one primary lens, even though real systems often combine several. Geometry represents examples, directions, distances, projections, and transformations. Change describes sensitivity, derivatives, gradients, and optimization trajectories. Uncertainty models random variables, distributions, conditional beliefs, and predictions. Evidence connects samples to estimates, uncertainty, and generalization claims. Information measures uncertainty reduction, coding cost, divergence, and shared signal. Computation determines whether the intended calculation is stable, tractable, and reproducible.

Computation is the lens readers most often demote to an implementation detail. So it is worth seeing a case where it decided the outcome on its own.

A Patriot missile battery at Dhahran, Saudi Arabia failed to intercept an incoming Scud on 25 February 1991. The Scud struck an Army barracks and killed 28 Americans. The U.S. General Accounting Office reported on the failure on 4 February 1992, and its finding opens: “On February 25, 1991, a Patriot missile defense system operating at Dhahran, Saudi Arabia, during Operation Desert Storm failed to track and intercept an incoming Scud.”

The tracking mathematics was never in dispute. Time was multiplied by a 24-bit fixed-point representation of 0.1, a value with no exact finite binary expansion. The stored constant was therefore slightly wrong, and the resulting error grew in proportion to how long the battery had been switched on. The report tabulates the drift: 0.0275 seconds and a 55-metre shift of the range gate after 8 hours, 0.0687 seconds and 137 metres after 20 hours, and 0.3433 seconds and 687 metres after the roughly 100 hours Alpha Battery had been running. Robert Skeel reproduced the same 0.0275 s and 0.3433 s errors in SIAM News in July 1992, and identified them as a relative error of 2^-20.

No one made an algebraic mistake. Register width and uptime did the rest. Only the computation lens asks about either.

Suppose a team says, “Customers with similar behavior should receive similar recommendations.” The sentence leaves at least four choices unresolved. What represents a customer? How is similarity measured? Which behavior window counts? And what makes one recommendation better than another?

Translating that into mathematics might represent each customer by a vector, compare vectors using cosine similarity, and optimize a ranking loss. Each choice narrows the claim. Each also creates a failure mode. Cosine similarity ignores magnitude, recent behavior may dominate, and ranking labels may reflect exposure bias.

Mathematization is therefore not decoration. It is the act of choosing what will be preserved, what will be ignored, and which errors the system will pay for.

A formula in a design document or model review earns the same five questions every time. What is the claim, stated in the real world without symbols? What are the objects — vectors, matrices, random variables, functions, samples, constraints? What assumptions connect them: independence, linearity, smoothness, stationarity, distributional form? What is actually computed — the estimate, loss, gradient, distance, probability, or bound the system produces? And what could disconfirm it: which counterexample, diagnostic, slice, perturbation, or held-out test?

The third question is the one that gets waved through, so here is what waving it through has cost. The maiden flight of Ariane 5, Flight 501, was lost on 4 June 1996, 37 seconds after the start of the main engine ignition sequence. The ESA/CNES inquiry board traced the loss to an unprotected data conversion: a 64-bit floating-point horizontal-bias value, BH, forced into a 16-bit signed integer. The code doing the conversion was inertial-platform alignment code. It served no purpose at all after lift-off.

The assumption was not hidden by accident. An analysis had identified seven variables at risk of an Operand Error. Protection was added to four. Three, including BH, were left unprotected in order to keep the SRI computer inside an 80% maximum workload target. The report records the reasoning that stood in for a test: “The reason for the three remaining variables, including the one denoting horizontal bias, being unprotected was that further reasoning indicated that they were either physically limited or that there was a large margin of safety, a reasoning which in the case of the variable BH turned out to be faulty.”

The claim, the objects and the computed quantity were all in order. What failed was step three. An assumption connecting a physical quantity to a 16-bit range was argued rather than checked — and written down as such.

A learner can manipulate symbols, explain a mechanism, or evaluate whether the mechanism matches a real problem. Symbol fluency reads and rearranges expressions; it is useful and easy to mistake for understanding. Mechanism fluency connects terms to model behaviour and predicts the effect of changing a quantity. Modeling judgment checks assumptions against the data-generating process and chooses diagnostics that could falsify the claim.

The third level is the goal. The cheapest demonstration of why is a table where the arithmetic is trivial and the answer still changes sign.

Berkeley's graduate admissions for fall 1973 are that table. 8,442 men applied and 44% were admitted. 4,321 women applied and 35% were admitted. The aggregate is easy to compute and hard to argue with, and a 1975 paper in Science opened its abstract with it: “Examination of aggregate data on graduate admissions to the University of California, Berkeley, for fall 1973 shows a clear but misleading pattern of bias against female applicants.”

Split the same applicants by department and the pattern comes apart. In the six largest departments — the 4,526-applicant table still distributed as R's UCBAdmissions dataset — women had the higher admission rate in four of the six. The unstandardised six-department totals read 45% for men, 1,198 of 2,691, and 30% for women, 557 of 1,835. Standardise each sex's rate to those departments' combined applicant pool and it becomes 39% for men against 43% for women. The paper's own pooled analysis reports “a small but statistically significant bias in favor of women” once it corrects for “the tendency of women to apply to graduate departments that are more difficult for applicants of either sex to enter”.

Symbol fluency computes every one of those percentages correctly. Mechanism fluency explains what an admission rate is and how standardising changes it. Only modeling judgment asks which population the rate should be computed over. That question, not the division, decides the sign of the answer. 44% against 35% and 39% against 43% describe the same applicants.