陶哲轩博客写数学问题时，通常先把对象定义清楚，再给直觉、反例和证明轮廓。把「Milliman Lecture II: Additive combinatorics and random matrices」改写成可阅读的中文笔记，重点是：问题在问什么、已知到哪一步、下一步最容易走偏在哪。原站广告、分享条和导航已去掉。

问题在问什么

This is my second Milliman lecture , in which I talk about recent applications of ideas from additive combinatorics (and in particular, from the inverse Littlewood-Offord problem) to the theory of discrete random matrices . In many areas of physics, chemistry, and computer science, one often has to study large matrices A, which for sake of discussion we shall take to be square, thus is an matrix for some large integer n, and the are real or complex numbers. In some cases A wi

Two particularly fundamental examples of discrete ensembles are

已知结果和反例

The Bernoulli and sparse Bernoulli ensembles arise naturally in computer science and numerical analysis, as they form a simple model for simulating the effect of numerical roundoff error (or other types of digital error) on a large matrix. Continuous ensembles such as the Gaussian ensembles, in contrast, are natural models for matrices in the analog world, and in particular in physics and chemistry. [For reasons that are still somewhat mysterious, these ensembles, or more pre

If A is drawn randomly from one of the above matrix ensembles, then we have a very explicit understanding of how each of the coefficients of the matrix A behaves. But in practice, we want to study more “global” properties of the matrix A which involve rather complicated interactions of all the coefficients together. For instance, we could be interested in the following (closely related) questions:

证明或构造的主线

As any student of linear algebra knows, these questions can be answered satisfactorily if one knows the eigenvalues (counting multiplicity) and singular values of the matrix A. (As the matrix A is not self-adjoint , the eigenvalues can be complex-valued even if the coefficients of A are real-valued; however, the singular values are always non-negative reals, because is self- adjoint and positive semi-definite .) For instance:

So, one of the fundamental problems in the theory of random matrices is to understand how the eigenvalues and singular values of a random matrix A are distributed. (More generally, it is of interest to study the eigenvalues and singular values of A+B, where A is drawn from a standard random matrix ensemble, and B is a fixed deterministic matrix, but for simplicity 下面会 not discuss this case here.) But how does one get a handle on these numbers?

阅读时建议盯住的点

The direct approach of working with the characteristic equation (or ) looks very unpromising; one is asking to find the roots of a large degree polynomial, most of whose coefficients depend in a hopelessly complicated way on the coefficients on A.

In the special cases of the Gaussian orthogonal and unitary ensembles, there is a massive amount of algebraic structure coming from the action of O(n) and U(n) that allows one to explicitly compute various multidimensional integrals, and this approach actually works! One gets a very explicit and useful explicit formula for the joint eigenvalue distribution (first worked out by Ginibre, I believe) this way. But for more general ensembles, such as the Bernoulli ensemble, such a

值得单独记下的条目

• The complex Gaussian random matrix ensemble , in which the are distributed according to a complex normal distribution . This ensemble has the feature of being invariant under the unitary group U(n).
• The Bernoulli ensemble , in which each is distributed independently and uniformly in the set , thus A is a random matrix of signs.
• The lazy (or sparse) Bernoulli ensemble , in which each is independently equal to -1 or +1 with probability p/2, and equal to 0 with probability 1-p, for some fixed , thus A is a sparse matrix of signs of expected density p.
• Dynamics . Given a typical vector , what happens to the iterates in the limit ?
• Expansion and contraction . Given a non-zero vector x, how does the norm of Ax compare with the norm of x? What is the largest ratio ? The smallest ratio? The average ratio?
• Invertibility . Is the equation solvable for every vector b? Do small fluctuations in b always cause small fluctuations in x, or can they cause large fluctuations? (In other words, is the invertibility problem stable?)
• The ratio has a maximal value of , a minimal value of , and a root mean square value of if the orientation of x is selected uniformly at random.
• The matrix A is invertible if and only if all eigenvalues are non-zero, or equivalently if is positive.

阅读和落地时建议先做的 5 件事

1. 用自己的语言重写定义和结论，不看原文能不能说清对象是什么。
2. 找一个最小反例或边界情形，确认假设少一条会怎样。
3. 把证明拆成可独立检验的引理，每步只保留一个新想法。
4. 若涉及计算或形式化，先写可复现的小例子，再谈一般情形。
5. 记下尚未解决的缺口：缺估计、缺构造，还是缺正确的范畴。

和智能体、形式化工具怎么接

龙虾PRO做 OpenClaw 落地时，数学笔记最有用的部分往往是「可检验的步骤」：定义、反例、引理边界。智能体适合帮忙展开计算和检索，不适合代替你决定哪条假设能扔。

本文侧重全链路风控方法论。落地时请用自身业务单据做回放验证，不要把示例阈值直接当生产策略。 相关：风控体检 · 方案资源