Enrolment options

FMA_3X061_EP — Les grandes notions de l’Analyse

How can an equation that describes only a single instant predict an entire trajectory? Why can a neural network approximate almost any function? What exactly does a quantum computer compute? The answer lies in a small number of great ideas of analysis, those that turn a difficult problem into a simple one.

Whether or not students intend to pursue mathematics further, this core course provides foundations for those who wish to go further, and tools they will encounter everywhere. It unfolds in four parts.

  1. Measuring closeness. To say that two functions are close, one must choose a distance between functions. We will study the topology of normed vector spaces, with emphasis on three notions. Density makes it possible to replace an arbitrary function by a simple one, and it explains why a neural network can approximate any continuous function. Completeness guarantees that a limit exists even before we know it. Compactness ensures that a minimum is attained.
  2. Predicting. Cauchy, a professor at the École polytechnique, laid the rigorous foundations of the theory of differential equations. The Cauchy–Lipschitz theorem (also known as the Picard–Lindelöf theorem) guarantees the existence and uniqueness of solutions, and its modern proof reduces solving the equation to finding a fixed point. We will see that uniqueness may fail when its assumptions are not satisfied, and that a solution may blow up in finite time. Two examples will guide us: the Schrödinger equation, which makes the qubits of quantum computers oscillate, and the gradient flow, whose discretization is used to train neural networks.
  3. Integrating. The Lebesgue integral is studied in parallel in the applied mathematics course. We will revisit it for what it brings to analysis, namely simple theorems for interchanging limits and integrals, and the right function spaces, which are at last complete.
  4. Decomposing. In a Hilbert space, Euclidean geometry remains valid in infinite dimension, and the best approximation is given by a projection. Fourier analysis rests on this principle. It decomposes a signal into elementary waves and turns differentiation into multiplication. We will see two applications: Heisenberg’s uncertainty principle and the quantum Fourier transform. The latter lies at the heart of Shor’s algorithm, which, on a sufficiently large quantum computer, would break most of the public-key cryptosystems in use today.

These ideas cut across disciplines. Interactive web pages will illustrate them.

Guests cannot access this course. Please log in.