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「When are eigenvalues stable?」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

I was asked recently (in relation to my recent work with Van Vu on the spectral theory of random matrices) to explain some standard heuristics regarding how the eigenvalues of an matrix A behave under small perturbations. These heuristics can be summarised as follows:

在这类讨论里 I would like to briefly explain why these heuristics are plausible.

已知结果和反例

As any student of linear algebra knows, the spectrum of a matrix A consists of all the complex numbers such that fails to be invertible, thus there exists a non-zero vector v such that is zero. This is a finite set, consisting of at most n points. But there is also an important set containing the spectrum (and which, at times, can be much larger than that spectrum), which is the pseudospectrum of the matrix A. Unlike the spectrum, which is canonically defined, there is more t

The significance of the pseudospectrum is that it describes where the spectrum can go to under small perturbations . Indeed, if lies in the pseudospectrum , so that there exists a unit vector v whose image has magnitude at most , then we see that

证明或构造的主线

and so lies in the spectrum of the perturbation of A. Note that the operator norm of is at most .

Conversely, if does not lie in the pseudospectrum , and is a small perturbation of A (with E having operator norm at most ), then for any unit vector v, one has

阅读时建议盯住的点

by the triangle inequality, and so cannot lie in the spectrum of .

Thus, if the pseudospectrum is tightly clustered around the spectrum, the spectrum is stable under small perturbations; but if the pseudospectrum is widely dispersed, then the spectrum becomes unstable.

值得单独记下的条目

• For normal matrices (and in particular, unitary or self-adjoint matrices), eigenvalues are very stable under small perturbations. For more general matrices, eigenvalues can become unstable if there is pseudospectrum present.

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