Calculus begins with a simple idea: that we can talk about change, and about quantities we approach but never quite reach, without the reasoning falling apart. This section makes that idea rigorous and carries it from the derivatives and gradients of ordinary calculus to the analysis that modern computation rests on. The thread running through it is completeness, the property that a space has no holes. When the terms of a sequence eventually all stay as close to one another as we like, completeness guarantees there is something there for them to close in on. Without it, even an algorithm whose steps settle down may have nothing to converge to.
Within the Compass, this section is the continuous side of the map: smooth change, infinite processes, objects reached only in the limit. Where Section IV (Discrete Mathematics & Algorithms) studies discrete structures made of separate pieces, this section studies what can be approximated as closely as we like but never reached in finitely many steps. It inherits the algebraic vocabulary of Section I (Linear Algebra to Algebraic Foundations), where vector spaces and linear maps become the basic objects of analysis. In turn it supplies the measure and integration that let Section III (Probability & Statistics) speak rigorously about averages, densities, and continuous chance, and the spaces and optimization that much of Section V (Machine Learning) takes place in.
The first half builds analysis on flat ground: optimization, measure and integration, Fourier analysis, metric and topological spaces, and then spaces whose points are functions, with the operators, dual spaces, and spectral theory that act on them. The longer half rebuilds calculus on curved spaces. The surface of a sphere, the group of all rotations, and the shapes that data traces out in high dimensions are not flat, and on them even a derivative has to be rebuilt from the ground up. The result is a language for differentiation that works on any smooth space, and for integration once an orientation is chosen, with a metric added when we want length and volume. One object in that language deserves a name here: a vector field, possibly changing in time, and the flow it generates by carrying every point along it. A flow-matching model learns exactly such a vector field, and it generates data by following the flow.
Infinite processes raise their own problem of trust. A computer can watch the steps of an iterative method get smaller. It cannot show, by running longer, that they are approaching anything at all. That guarantee has to be proved, and completeness is what proofs of this kind rest on. The contraction principle is the model case: if each step shrinks the distance between any two points by a fixed factor less than one, then in a complete space there is exactly one answer, and every run, from any start, converges to it. The geometric half adds a second kind of check. A quantity defined without coordinates gives the same answer in every coordinate system, so the language of manifolds rules out a whole class of mistakes before any computation starts. Some natural next questions are not yet covered here. One is which of the many flows carrying one distribution to another is the cheapest, the subject of optimal transport.