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「Distinguished Lecture Series II: Avi Wigderson, “Expander graphs – constructions and appli」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

On Thursday, Avi Wigderson continued his Distinguished Lecture Series here at UCLA on computational complexity with his second lecture “ Expander Graphs – Constructions and Applications “. As in the previous lecture , he spent some additional time after the talk on an “encore”, which in this case was how lossless expanders could be used to obtain rapidly decodable error-correcting codes .

The talk was largely based on these slides . Avi also has a recent monograph with Hoory and Linial on these topics. (For a brief introduction to expanders, I can also recommend Peter Sarnak’s Notices article . I also mention expanders to some extent in my third Milliman lecture .)

已知结果和反例

Expander graph are sparse graphs with an additional special property. This property can be described in several equivalent ways (combinatorial, probabilistic, or algebraic):

For sake of concreteness, Avi used the spectral gap formulation of expansion:

证明或构造的主线

Definition . A -graph is a graph G on n vertices which is d- regular , i.e. every vertex has d edges connected to it. A -graph is an -graph which has the additional property

for all non-zero vectors whose coordinates sum to zero, where A(G) is the adjacency matrix of G and is the Euclidean norm of v.

阅读时建议盯住的点

A family of graphs of d-regular graphs (thus each is an -graph for some ) forms an expander family if each is a graph for some .

The above definition was a bit technical, but it can be clarified a bit by considering an example of a graph which is definitely not an expander. Consider an -graph G which is disconnected, with the n vertices divided into two disjoint sets A and B of n/2 vertices each, with no edge between A and B. Then if one lets v be the vector which is +1 on A and -1 on B (so that the total sum of coefficients is zero), one can check that , so G is not an -graph for any . More generally,

值得单独记下的条目

• Mixing . A random walk on an expander graph converges very rapidly to the uniform distribution .
• Spectral gap . The second eigenvalue of the adjacency matrix is significantly smaller than the (typical) degree of the graph.
• Start with the corrupted word, and compute all the parity bits. If they all vanish, then we have a codeword and we stop.
• Repeat 2 until all parity bits vanish.

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