math-skill 2.1.0 → 3.0.0

package/skills/discrete-combinatorial/original-texts.en.md

@@ -1,184 +0,0 @@
-# Discrete & Combinatorial Thinking — Original Texts & Historical Context
-
----
-
-## 1. Pigeonhole Principle — Dirichlet (1834)
-
-Peter Gustav Lejeune Dirichlet stated the pigeonhole principle in 1834 as the "Schubfachprinzip" (drawer principle), using it for existential proofs in number theory.
-
-> "Wenn man n + 1 Objecte in n Facher distribuiert, so ist mindestens ein Fach mit zwei Objecten besetzt."
-> — Dirichlet, 1834
-
-**Core statement:**
-- Basic form: if n items are placed into m boxes and n > m, then some box contains at least ⌈n/m⌉ items
-- Generalized form: kn+1 items into n boxes → some box contains at least k+1 items
-- Number theory application: for any real number α and integer N, there exists q ≤ N such that |qα − p| < 1/(N+1) (i.e., qα approximates an integer)
-
-**Diophantine approximation:** Partition [0,1) into N+1 equal-length intervals ("boxes"), and consider the N+2 numbers {0, {α}, {2α}, ..., {(N+1)α}} ("items"). The pigeonhole principle guarantees that two must fall into the same interval, yielding |(q₁−q₂)α − ⌊(q₁−q₂)α⌋| < 1/(N+1). This is the standard pigeonhole proof of Dirichlet's approximation theorem.
-
----
-
-## 2. Euler's Generating Functions (1748)
-
-Leonhard Euler pioneered the generating function method in **Introductio in analysin infinitorum** (1748) — transforming sequences into power series, and counting into algebraic operations.
-
-> "If we ask how many ways a given number can be partitioned, we are led to a study of the generating function." — Euler, 1748
-
-**Partition generating function:**
-$$P(x) = \sum_{n=0}^{\infty} p(n) x^n = \prod_{k=1}^{\infty} \frac{1}{1-x^k}$$
-- p(n) = the number of partitions of n into positive integers. p(5) = 7: {5, 4+1, 3+2, 3+1+1, 2+2+1, 2+1+1+1, 1+1+1+1+1}
-- Product form: 1/(1−x^k) = 1+x^k+x^{2k}+... represents "taking 0, 1, 2, ... copies of k"; since each k is independent, the product encodes all combinations
-
-**Pentagonal number theorem (Euler 1750):**
-$$\prod_{k=1}^{\infty}(1-x^k) = \sum_{m=-\infty}^{\infty} (-1)^m x^{m(3m-1)/2}$$
-- Generalized pentagonal numbers e_m = m(3m−1)/2: 0, 1, 2, 5, 7, 12, 15, 22, ...
-- From P(x)·∏(1−x^k) = Σ(−1)^m x^{e_m}, one obtains the recurrence p(n) = p(n−1)+p(n−2)−p(n−5)−p(n−7)+...
-- This is a classic example of a finite recurrence controlling an infinite sequence — the essence of combinatorial thinking
-
-**Hardy-Ramanujan asymptotic formula (1918):**
-$$p(n) \sim \frac{1}{4n\sqrt{3}} \exp\left(\pi\sqrt{\frac{2n}{3}}\right)$$
-
----
-
-## 3. Catalan Numbers
-
-**History:** Eugène Catalan (1838) studied the counting problem of nested parentheses, from which the Catalan numbers derive their name. However, Euler had already studied a related problem (polygon triangulation) in 1751.
-
-**Definition and formula:**
-$$C_n = \frac{1}{n+1}\binom{2n}{n} = \frac{(2n)!}{n!(n+1)!}$$
-- Recurrence: C₀ = 1, C_n = ΣC_i·C_{n−1−i}
-- Generating function: C(x) = ΣC_n x^n = (1−√(1−4x))/(2x), obtained by solving C(x) = 1 + x·C(x)²
-
-**Combinatorial interpretations (60+):**
-- **Binary trees**: the number of ordered binary trees with n nodes = C_n
-- **Dyck paths**: lattice paths from (0,0) to (2n,0), with steps (+1,+1) or (+1,−1), never crossing above y=0
-- **Parenthesizations**: the number of valid nestings of n pairs of parentheses = C_n
-- **Ballot problem**: A receives n+1 votes, B receives n votes; the number of sequences in which A is always ahead of B = C_n (Bertrand 1887)
-- **Polygon triangulation**: the number of distinct triangulations of an (n+2)-gon = C_n
-
-First few terms: C₀=1, C₁=1, C₂=2, C₃=5, C₄=14, C₅=42, C₆=132
-
----
-
-## 4. Ramsey Theory — Ramsey (1930)
-
-Frank P. Ramsey proved in **"On a Problem of Formal Logic"** (1930) that sufficiently large systems necessarily contain some structure — a profound generalization of the pigeonhole principle.
-
-> "In any sufficiently large system, complete disorder is impossible." — Cohen
-
-**Ramsey numbers:**
-$$R(s,t) = \min\{n : \text{every 2-coloring of the complete graph on } n \text{ vertices contains a red } K_s \text{ or a blue } K_t\}$$
-
-**Known values:** R(3,3)=6 (at a party of 6 people, there must be 3 mutual acquaintances or 3 mutual strangers; proof: fix one person, among the remaining 5 at least 3 share the same relationship color by the pigeonhole principle; if any 2 of those 3 share that color they form a monochromatic triangle, otherwise those 3 form a triangle of the other color), R(4,4)=18, R(3,4)=9, R(3,5)=14.
-
-**General bounds:** Erdős 1947 probabilistic method: R(k,k) > 2^{k/2}; upper bound R(k,k) ≤ 4^k. R(5,5) is unknown (43 ≤ R(5,5) ≤ 48). Erdős: "If aliens demanded R(5,5), we should marshal all our computers; if they demanded R(6,6), we should prepare for battle."
-
----
-
-## 5. Graph Theory — Euler (1736)
-
-**Seven Bridges of Königsberg (1736):**
-
-Leonhard Euler, in **"Solutio problematis ad geometriam situs pertinentis"** (1736), solved the Königsberg bridge problem, founding graph theory.
-
-> "The problem... is to find whether one can cross each of the seven bridges exactly once and return to the starting point." — Euler, 1736
-
-**Euler circuit theorem:**
-- A connected graph has an Euler circuit (a closed walk traversing every edge exactly once) if and only if all vertices have even degree
-- The Königsberg graph: the 4 vertices have degrees 3, 3, 3, 5 (all odd), so no Euler circuit exists — the problem has no solution
-
-**Euler's planar formula (1750):**
-$$V - E + F = 2$$
-- For any connected planar graph: the number of vertices V, edges E, and faces F satisfy this formula
-- Corollaries: for a simple connected planar graph E ≤ 3V − 6; for a planar graph containing no K₃, E ≤ 2V − 4
-
-**Matching theory:**
-- König (1931): in a bipartite graph, the maximum matching size = the minimum vertex cover size (König's theorem)
-- Hall's marriage theorem (1935): a bipartite graph G(X,Y) has a perfect matching from X to Y if and only if for every S ⊆ X, |N(S)| ≥ |S|
-
----
-
-## 6. Pólya Enumeration Theorem (1937)
-
-George Pólya, in his 1937 paper, provided a counting method under symmetry — the orbit counting theorem, the meeting point of combinatorial enumeration and group theory.
-
-> "The number of distinct colorings is obtained by averaging over the group action." — Pólya, 1937
-
-**Theorem statement:**
-Let group G act on set X. The number of equivalence classes of k-colorings (number of orbits):
-$$N = \frac{1}{|G|} \sum_{g \in G} k^{c(g)}$$
-where c(g) = the number of cycles of the permutation g. Applications: chemical isomers, bead necklace patterns. Burnside's lemma (1897): N = (1/|G|)Σ|Fix(g)| is a special case of Pólya's theorem. Pólya generalized this to a weighted version: GF × cycle index = symmetric counting.
-
----
-
-## 7. Inclusion-Exclusion Principle
-
-**History:** da Vinci used inclusion-exclusion ideas to compute areas; Sylvester (1882) formalized the principle. Inclusion-exclusion handles overlapping intersections — direct addition double-counts intersections, requiring successive corrections.
-
-Basic form: |A₁∪...∪A_n| = Σ|A_i| − Σ|A_i∩A_j| + ... ± |A₁∩...∩A_n|. Complement counting: "not having property P" = total − "having property P."
-
-**Derangements:** D(n) = n! − C(n,1)(n−1)! + C(n,2)(n−2)! − ... = n!Σ(−1)^k/k! ≈ n!/e as n→∞. Probability form: P(A₁∪...∪A_n) = ΣP(A_i) − ΣP(A_i∩A_j) + ...
-
----
-
-## 8. Recurrence Relations
-
-**Fibonacci sequence:**
-$$F_0 = 0, F_1 = 1, F_n = F_{n-1} + F_{n-2}$$
-- Generating function: F(x) = ΣF_n x^n = x/(1−x−x²)
-- Binet's formula: F_n = (φ^n − ψ^n)/√5, where φ = (1+√5)/2 ≈ 1.618 (the golden ratio)
-
-**Tower of Hanoi:**
-$$T_1 = 1, T_n = 2T_{n-1} + 1$$
-- Solution: T_n = 2^n − 1. A tower of 64 disks requires 2^{64} − 1 ≈ 1.8×10^{19} moves
-
-**Solving recurrences via generating functions:**
-1. Multiply the recurrence by x^n, sum over n → obtain an equation for A(x)
-2. Use initial conditions to solve for A(x) (algebraic or differential equation)
-3. Extract a_n: partial fractions, Taylor expansion, or matching against known power series
-
-**Catalan GF derivation:** C(x) = 1 + x·C(x)² → solve x·C² − C + 1 = 0 → C(x) = (1−√(1−4x))/(2x) → C_n = (1/(n+1))C(2n,n)
-
----
-
-## 9. Combinatorial Optimization
-
-**Minimum spanning tree:**
-- Kruskal (1956): greedily select edges in ascending order of weight, skipping those that form cycles, O(E log E)
-- Prim (1957): expand from a single vertex, each time selecting the minimum-weight connecting edge, O(E log V) (with a priority queue)
-
-**Max-flow min-cut:**
-- Ford-Fulkerson (1956): augmenting path method, incrementally increasing flow until no augmenting path exists
-- Max-flow min-cut theorem: the maximum flow value = the minimum cut capacity (the combinatorial manifestation of LP duality)
-- Edmonds-Karp (1972): BFS selects the shortest augmenting path, guaranteeing O(VE²)
-
-**Matching theory:**
-- Hungarian algorithm (Kuhn 1955): O(V³) maximum weighted bipartite matching
-- Edmonds' blossom algorithm (1965): maximum matching in general graphs, O(V³)
-
----
-
-## 10. The Twelvefold Way
-
-Richard P. Stanley, in **Enumerative Combinatorics Vol. 1**, unified and classified the 12 fundamental counting problems of placing n balls into k boxes (2×2×3: balls labeled/unlabeled × boxes labeled/unlabeled × any/at most one/at least one):
-
-| Balls | Boxes | Condition | Formula |
-|---|---|---|---|
-| distinct | distinct | any | k^n |
-| identical | distinct | any | C(n+k−1, k−1) |
-| distinct | identical | any | Σ S₂(n,i), i=1..k |
-| identical | identical | any | p_k(n) (restricted partitions) |
-| distinct | distinct | at most one ≤1 | P(k,n) = k!/(k−n)! |
-| identical | distinct | at most one ≤1 | C(k,n) |
-| distinct | identical | at most one ≤1 | 1 if n≤k, 0 otherwise |
-| identical | identical | at most one ≤1 | 1 if n≤k, 0 otherwise |
-| distinct | distinct | at least one ≥1 | k!·S₂(n,k) |
-| identical | distinct | at least one ≥1 | C(n−1, k−1) |
-| distinct | identical | at least one ≥1 | S₂(n,k) − S₂(n,k−1) |
-| identical | identical | at least one ≥1 | p_k(n) − p_{k−1}(n) |
-
-**Key insight:** The 2×2×3=12 fundamental counting problems cover the core scenarios in combinatorics. Use EGF for labeled objects, OGF for unlabeled objects. **Stirling numbers:** S₁(n,k) = the number of permutations of n elements with k cycles; S₂(n,k) = the number of ways to partition n elements into k non-empty subsets.
-
----
-
-*This file provides mathematical references and historical context for the discrete-combinatorial skill, covering core results from Euler to modern combinatorics.*

package/skills/discrete-combinatorial/original-texts.md

@@ -1,185 +0,0 @@
-# 离散与组合思想 — 原始文献与历史语境 / Discrete & Combinatorial Thinking — Original Texts & Historical Context
-
----
-
-## 1. Pigeonhole Principle — Dirichlet (1834) / 鸽巢原理 — Dirichlet (1834)
-
-Peter Gustav Lejeune Dirichlet 在 1834 年以"Schubfachprinzip"（抽屉原理）阐述鸽巢原理，用于数论存在性证明。
-
-> "Wenn man n + 1 Objecte in n Facher distribuiert, so ist mindestens ein Fach mit zwei Objecten besetzt."
-> — Dirichlet, 1834
-
-**核心陈述 / Core statement:**
-- 基本形式：若 n 物品放入 m 个盒子且 n > m，则某盒子含至少 ⌈n/m⌉ 个物品
-- 推广形式：kn+1 物品入 n 盒子 → 某盒至少含 k+1 个物品
-- 数论应用：对任意实数 α 和整数 N，存在 q ≤ N 使得 |qα − p| < 1/(N+1)（即 qα 逼近整数）
-
-**Diophantine 逼近 / Diophantine approximation:** 将 [0,1) 分为 N+1 个等长区间（"盒子"），考虑 N+2 个数 {0, {α}, {2α}, ..., {(N+1)α}}（"物品"），鸽巢保证必有两个落入同一区间，得 |(q₁−q₂)α − ⌊(q₁−q₂)α⌋| < 1/(N+1)。这是 Dirichlet 逼近定理的标准鸽巢证明。
-
----
-
-## 2. Euler's Generating Functions (1748) / Euler 的生成函数 (1748)
-
-Leonhard Euler 在 **Introductio in analysin infinitorum** (1748) 中开创生成函数方法——序列变换为幂级数，计数变换为代数运算。
-
-> "If we ask how many ways a given number can be partitioned, we are led to a study of the generating function." — Euler, 1748
-
-**分拆函数生成函数 / Partition GF:**
-$$P(x) = \sum_{n=0}^{\infty} p(n) x^n = \prod_{k=1}^{\infty} \frac{1}{1-x^k}$$
-- p(n) = 将 n 分为正整数之和的分拆数。p(5) = 7：{5, 4+1, 3+2, 3+1+1, 2+2+1, 2+1+1+1, 1+1+1+1+1}
-- 乘积形式：1/(1−x^k) = 1+x^k+x^{2k}+... 表示"取 k 的 0,1,2,... 次"，各 k 独立故乘积编码组合
-
-**五角数定理 / Pentagonal number theorem (Euler 1750):**
-$$\prod_{k=1}^{\infty}(1-x^k) = \sum_{m=-\infty}^{\infty} (-1)^m x^{m(3m-1)/2}$$
-- 广义五角数 e_m = m(3m−1)/2：0, 1, 2, 5, 7, 12, 15, 22, ...
-- 由 P(x)·∏(1−x^k) = Σ(−1)^m x^{e_m}，得递推 p(n) = p(n−1)+p(n−2)−p(n−5)−p(n−7)+...
-- 这是有限递推控制无限序列的经典实例——组合思想的精髓
-
-**Hardy-Ramanujan 渐近公式 (1918):**
-$$p(n) \sim \frac{1}{4n\sqrt{3}} \exp\left(\pi\sqrt{\frac{2n}{3}}\right)$$
-
----
-
-## 3. Catalan Numbers / Catalan 数
-
-**历史 / History:** Eugène Catalan (1838) 研究括号嵌套的计数问题，Catalan 数由此得名。但 Euler 早在 1751 年就研究过类似问题（多边形三角剖分）。
-
-**定义与公式 / Definition and formula:**
-$$C_n = \frac{1}{n+1}\binom{2n}{n} = \frac{(2n)!}{n!(n+1)!}$$
-- 递推：C₀ = 1, C_n = ΣC_i·C_{n−1−i}
-- 生成函数：C(x) = ΣC_n x^n = (1−√(1−4x))/(2x)，由 C(x) = 1 + x·C(x)² 解得
-
-**组合解释（60+种） / Combinatorial interpretations (60+):**
-- **二叉树**：n 个节点的有序二叉树数 = C_n
-- **Dyck 路径**：从 (0,0) 到 (2n,0) 的格点路径，每步 (+1,+1) 或 (+1,−1)，不越 y=0
-- **括号序列**：n 对括号的合法嵌套方式数 = C_n
-- **投票问题 (Ballot problem)**：A 得 n+1 票、B 得 n 票，A 全程领先 B 的序列数 = C_n（Bertrand 1887）
-- **多边形三角剖分**：n+2 边形的不同三角剖分数 = C_n
-
-前几项：C₀=1, C₁=1, C₂=2, C₃=5, C₄=14, C₅=42, C₆=132
-
----
-
-## 4. Ramsey Theory — Ramsey (1930) / Ramsey 理论 — Ramsey (1930)
-
-Frank P. Ramsey 在 **"On a Problem of Formal Logic"** (1930) 中证明：足够大的系统必存在某种结构——鸽巢原理的深刻推广。
-
-> "In any sufficiently large system, complete disorder is impossible." — Cohen
-
-**Ramsey 数 / Ramsey numbers:**
-$$R(s,t) = \min\{n : \text{任意 } n \text{ 阶 2-色完全图中必有红 } K_s \text{ 或蓝 } K_t\}$$
-
-**已知值:** R(3,3)=6（6人聚会必有3人互识或互不识；证明：固定一人，5人中至少3人同色关系（鸽巢），此3人中若2人同色则3人组，否则此3人构成另一色3人组）, R(4,4)=18, R(3,4)=9, R(3,5)=14.
-
-**一般界:** Erdős 1947 概率方法: R(k,k) > 2^{k/2}; 上界 R(k,k) ≤ 4^k. R(5,5)未知(43≤R(5,5)≤48). Erdős："外星人要R(5,5)就动员所有计算机；要R(6,6)就准备战斗。"
-
----
-
-## 5. Graph Theory — Euler (1736) / 图论 — Euler (1736)
-
-**Königsberg 七桥问题 / Seven Bridges of Königsberg (1736):**
-
-Leonhard Euler 在 **"Solutio problematis ad geometriam situs pertinentis"** (1736) 解决柯尼斯堡七桥问题，开创图论。
-
-> "The problem... is to find whether one can cross each of the seven bridges exactly once and return to the starting point." — Euler, 1736
-
-**Euler 环定理 / Euler circuit theorem:**
-- 连通图存在 Euler 环（经过每边恰好一次的闭路径）iff 所有顶点度数为偶数
-- 柯尼斯堡图：4 个顶点度数为 3, 3, 3, 5（均为奇数），故 Euler 环不存在——问题无解
-
-**Euler 平面图公式 / Euler's planar formula (1750):**
-$$V - E + F = 2$$
-- 对任意连通平面图：顶点数 V、边数 E、面数 F 满足此公式
-- 推论：简单连通平面图 E ≤ 3V − 6；不含 K₃ 的平面图 E ≤ 2V − 4
-
-**匹配理论 / Matching theory:**
-- König (1931)：二分图中最大匹配数 = 最小顶点覆盖数（König 定理）
-- Hall 婚姻定理 (1935)：二分图 G(X,Y) 有 X 到 Y 的完美匹配 iff 对任意 S ⊆ X，|N(S)| ≥ |S|
-
----
-
-## 6. Pólya Enumeration Theorem (1937) / Pólya 计数定理 (1937)
-
-George Pólya 在 1937 年的论文中给出对称性下的计数方法——轨道计数定理，组合计数与群论的交汇点。
-
-> "The number of distinct colorings is obtained by averaging over the group action." — Pólya, 1937
-
-**定理陈述 / Theorem statement:**
-设群 G 作用于集合 X，k-着色等价类数（轨道数）:
-$$N = \frac{1}{|G|} \sum_{g \in G} k^{c(g)}$$
-c(g) = 置换 g 的循环数。应用: 化学异构体、珠串图案。Burnside 引理 (1897): N = (1/|G|)Σ|Fix(g)|，是 Pólya 特例。Pólya 推广为带权版本: GF × 循环指数 = 对称计数。
-
----
-
-## 7. Inclusion-Exclusion Principle / 容斥原理
-
-**历史:** da Vinci 用容斥思想计算面积；Sylvester (1882) 形式化。容斥处理交集重叠——直接相加重计交集，需逐层修正。
-
-基本形式: |A₁∪...∪A_n| = Σ|A_i| − Σ|A_i∩A_j| + ... ± |A₁∩...∩A_n|. 补集计数: "不具有P" = 全部 − "具有P".
-
-**错排数:** D(n) = n! − C(n,1)(n−1)! + C(n,2)(n−2)! − ... = n!Σ(−1)^k/k! ≈ n!/e (n→∞). 概率形式: P(A₁∪...∪A_n) = ΣP(A_i) − ΣP(A_i∩A_j) + ...
-
----
-
-## 8. Recurrence Relations / 递推关系
-
-**Fibonacci 数列 / Fibonacci sequence:**
-$$F_0 = 0, F_1 = 1, F_n = F_{n-1} + F_{n-2}$$
-- 生成函数：F(x) = ΣF_n x^n = x/(1−x−x²)
-- Binet 公式：F_n = (φ^n − ψ^n)/√5，φ = (1+√5)/2 ≈ 1.618（黄金比）
-
-**汉诺塔 / Tower of Hanoi:**
-$$T_1 = 1, T_n = 2T_{n-1} + 1$$
-- 解：T_n = 2^n − 1。64 层需 2^{64} − 1 ≈ 1.8×10^{19} 步
-
-**GF 求解递推 / Solving recurrences via GF:**
-1. 递推乘以 x^n，对 n 求和 → A(x) 的方程
-2. 利用初始条件，解出 A(x)（代数/微分方程）
-3. 提取 a_n：部分分式、Taylor 展开、幂级数对照
-
-**Catalan GF 推导:** C(x) = 1 + x·C(x)² → 解 x·C² − C + 1 = 0 → C(x) = (1−√(1−4x))/(2x) → C_n = (1/(n+1))C(2n,n)
-
----
-
-## 9. Combinatorial Optimization / 组合优化
-
-**最小生成树 / Minimum spanning tree:**
-- Kruskal (1956)：按边权升序贪心选取不成环的边，O(E log E)
-- Prim (1957)：从单点扩展，每次选最小权连接边，O(E log V)（优先队列）
-
-**最大流最小割 / Max-flow min-cut:**
-- Ford-Fulkerson (1956)：增广路径法，流量逐步增加直至无增广路径
-- 最大流最小割定理：最大流值 = 最小割容量（LP 对偶性的组合表现）
-- Edmonds-Karp (1972)：BFS 选最短增广路径，保证 O(VE²)
-
-**匹配理论 / Matching theory:**
-- Hungarian 算法 (Kuhn 1955)：O(V³) 最大加权二分匹配
-- Edmonds 带花树算法 (1965)：一般图最大匹配，O(V³)
-
----
-
-## 10. The Twelvefold Way / 十二重道
-
-Richard P. Stanley 在 **Enumerative Combinatorics Vol. 1** 中统一分类 n球入k盒的12种基本计数(2×2×3: 球有标号/无标号 × 盒有标号/无标号 × 任意/至多1/至少1):
-
-| 球 / Balls | 盒子 / Boxes | 条件 / Condition | 公式 / Formula |
-|---|---|---|---|
-| 有标号 distinct | 有标号 distinct | 任意 any | k^n |
-| 无标号 identical | 有标号 distinct | 任意 any | C(n+k−1, k−1) |
-| 有标号 distinct | 无标号 identical | 任意 any | Σ S₂(n,i), i=1..k |
-| 无标号 identical | 无标号 identical | 任意 any | p_k(n)（受限分拆） |
-| 有标号 distinct | 有标号 distinct | 至多一个 ≤1 | P(k,n) = k!/(k−n)! |
-| 无标号 identical | 有标号 distinct | 至多一个 ≤1 | C(k,n) |
-| 有标号 distinct | 无标号 identical | 至多一个 ≤1 | 1 if n≤k, 0 else |
-| 无标号 identical | 无标号 identical | 至多一个 ≤1 | 1 if n≤k, 0 else |
-| 有标号 distinct | 有标号 distinct | 至少一个 ≥1 | k!·S₂(n,k) |
-| 无标号 identical | 有标号 distinct | 至至少一个 ≥1 | C(n−1, k−1) |
-| 有标号 distinct | 无标号 identical | 至少一个 ≥1 | S₂(n,k) − S₂(n,k−1) |
-| 无标号 identical | 无标号 identical | 至少一个 ≥1 | p_k(n) − p_{k−1}(n) |
-
-**关键洞察:** 2×2×3=12种基本计数覆盖组合数学核心场景。有标号用EGF，无标号用OGF。**Stirling数:** S₁(n,k)=n元素k循环排列数; S₂(n,k)=n元素k非空子集数。
-
----
-
-*本文件为离散与组合思想技能的数学参考与历史语境，涵盖从 Euler 到现代组合数学的核心成果。*
-*This file provides mathematical references and historical context for the discrete-combinatorial skill, covering core results from Euler to modern combinatorics.*

package/skills/game-theory/SKILL.en.md

@@ -1,117 +0,0 @@
----
-name: game-theory
-description: |
-  Trigger when optimal strategy depends on others' choices, needing Nash equilibrium computation, zero-sum/non-zero-sum analysis, mechanism design; or designing strategies for multi-agent systems, adversarial training, routing games.
----
-
-# 🎯 Game Theory
-
-> "Your optimal choice depends on others' choices — thinking must be interactive, not just deep."
->
-> — Game Theory, Decision Theory, Mechanism Design
-
-## Core Principle
-
-**Strategic interaction — your best action depends on what others do, and theirs depend on what you do. Nash equilibrium: strategy profile where no player benefits from unilateral deviation.**
-
-Key concepts: **Payoff Matrix**, **Dominant Strategy**, **Mixed Strategy**, **Pareto Optimality**. Classification dimensions: cooperative vs. non-cooperative, zero-sum vs. general-sum, simultaneous vs. sequential.
-
-> **Mathematical Formalization**
->
-> Formal definition: $n$ players, each with strategy set $S_i$ and payoff function $u_i: S_1 \times \cdots \times S_n \to \mathbb{R}$.
->
-> Nash equilibrium: strategy profile $s^* = (s_1^*, \ldots, s_n^*)$ satisfying $u_i(s^*) \geq u_i(s_i, s^*_{-i})$ for all $s_i \in S_i$ and all $i$, where $s^*_{-i}$ denotes the strategies of all players except $i$. Mixed-strategy Nash equilibrium: probability distribution $\sigma^*$ such that $u_i(\sigma^*) \geq u_i(s_i, \sigma^*_{-i})$.
->
-> Dominant strategy: strict $u_i(s_i^*, s_{-i}) > u_i(s_i, s_{-i})$; weak dominance $\geq$ with at least one strict inequality.
->
-> Zero-sum minimax: $\max_x \min_y f(x,y) = \min_y \max_x f(x,y)$.
-
-## GPU-Friendliness (Cross-Cutting Check)
-
-Game-solving maps to GPU with widely varying capability; pass through the eight-dimensional gate in `../../references/gpu-friendly-math.md`:
-
-- **Matrix games / two-player zero-sum (minimax)**: Equivalent to linear programming (LP); payoff matrix operations are GEMM-compatible and batch-parallelizable — **friendly**.
-- **Mixed-strategy equilibrium computation**: Linear systems / indifference equations on the support; small-scale problems are batch-solvable; watch for sparsity and condition number at scale.
-- **Exact large-scale multi-agent equilibrium**: $n$-player Nash equilibrium is PPAD-complete; general case is **intractable** → **anti-pattern**; switch to approximate / learning-based methods (fictitious play, self-play, multi-agent RL, NFSP).
-- **Evolutionary / repeated game simulation**: Population and strategy batch updates are parallelizable; watch for inter-agent communication and synchronization overhead.
-
-Eight-dimensional minimum assessment (formal terms): **Tensorization** — whether players / strategies / rounds admit batched unfolding; **GEMM-mappability** — whether payoff matrices, minimax, LP / linear systems fall into matrix operations; **Complexity** — flag PPAD / combinatorial explosion risks for Nash / mechanism solving; **Memory and KV-Cache** — whether the strategy-state Cartesian product is compressible; **Low-precision stability** — whether equilibrium solving, LP, soft best-response have controlled condition numbers; **Parallelism and communication** — synchronization cost of multi-agent rollouts; **Sparse structure** — whether the interaction graph is structured; **Operator fusion** — whether payoff, mask, and best-response updates can be fused.
-
-> Use in conjunction with `../../references/books/optimization-ml.md` (duality / minimax, game learning algorithms) and `../../references/books/matrix-analysis.md` (matrix games, linear systems).
-
-## When NOT to Use
-
-- **Single-agent decision problems with no interaction** — optimization thinking is more appropriate; no need to model others' strategic responses.
-- **Purely cooperative problems with no conflict of interest** — players' objectives are fully aligned; strategic analysis is unnecessary.
-- **Deterministic problems with no strategic uncertainty** — outcomes are uniquely determined by one's own actions; no reaction from others is involved.
-- **Outcomes determined purely by chance with no strategic choice** — no strategy sets to analyze; game theory does not apply.
-
-## When to Use
-
-- Multiple decision-makers interact; each player's strategy affects others' payoffs; the optimal strategy depends on predicting others' responses.
-- Need to design incentive mechanisms so that self-interested behavior leads to socially optimal outcomes; predict equilibrium states under given rules.
-- Analyze pricing, entry, and exit in competitive markets; benefit allocation and threat points in negotiation / bargaining.
-- **Design strategies and equilibrium objectives for multi-agent systems, adversarial training (GAN / robust learning), and routing games.**
-
-## Method
-
-### Step 1: Identify Players and Strategies
-Define the fundamental elements of the game: players $N=\{1,\ldots,n\}$; strategy sets $S_i$; payoff functions $u_i: S_1\times\cdots\times S_n\to\mathbb{R}$; information structure (symmetric or not? incomplete information? sequential actions? communication possible?). The core task is identifying the "interaction structure" — whose choices affect whose outcomes. Missing a key player is the most common error.
-
-### Step 2: Analyze the Game Type
-First classify: zero-sum / general-sum, simultaneous / sequential, cooperative / non-cooperative, complete / incomplete information. The type determines the method: zero-sum uses minimax / LP; sequential games use backward induction and subgame-perfect equilibrium; cooperative games use the Shapley value to allocate by marginal contribution:
-$$\phi_i(v)=\sum_{S\subseteq N\setminus\{i\}}\frac{|S|!(n-|S|-1)!}{n!}[v(S\cup\{i\})-v(S)]$$
-Misclassifying the type leads directly to applying the wrong equilibrium concept and solution method.
-
-### Step 3: Construct the Payoff Matrix / Function
-For two-player games, use a payoff matrix with cells $(u_1,u_2)$; for $n$-player games, use $u_i(s_1,\ldots,s_n)$. Verify completeness: are any players or strategies missing? Do payoffs accurately reflect preferences? Are there externalities? Payoffs are the foundation of all analysis; inaccurate assessment biases everything downstream.
-
-### Step 4: Search for Dominant Strategies
-Strict dominance: $u_i(s_i^*,s_{-i}) > u_i(s_i,s_{-i})$ for all $s_i\neq s_i^*$ and all $s_{-i}$; weak dominance: $\geq$ with at least one strict inequality. **Iterated elimination of dominated strategies** successively removes inferior strategies to shrink the strategy space. A dominant-strategy equilibrium is the strongest possible conclusion — no player has any incentive to deviate. If none exists, proceed to equilibrium analysis.
-
-### Step 5: Compute Equilibria
-Select the method based on the game type from Step 2. Pure-strategy Nash: for each $s_{-i}$, find the best response $BR_i(s_{-i})$; intersections are equilibria. Mixed strategies: find probability distributions that make others indifferent, $u_j(s_j,\sigma^*_{-j})=u_j(s_j',\sigma^*_{-j})$ for all $s_j,s_j'\in S_j$. Zero-sum matrix games use minimax $\max_x\min_y f=\min_y\max_x f$; sequential games use backward induction; cooperative games output allocation schemes rather than standard Nash. For multiple equilibria, perform Pareto ranking and label optimal vs. inferior equilibria.
-
-### Step 6: Check Equilibrium Stability
-Trembling-hand perfection (stable under small perturbations, excluding equilibria that rely on "opponents never make mistakes"); evolutionarily stable strategy ESS (resists invasion in repeated interactions); off-equilibrium-path incentives (in sequential games, whether there is motivation to honor commitments). Ask: is the equilibrium unique or multiple? Is it robust to small perturbations? Only stable equilibria constitute credible predictions.
-
-### Step 7: Mechanism Design and Improvement
-If no good equilibrium exists, redesign the game: modify the payoff structure (incentives and penalties), add enforceable rules, introduce communication and reputation mechanisms, Vickrey auctions (truthful valuation reporting is a dominant strategy). Mechanism design formalization:
-$$\max\ \text{social welfare}\quad\text{s.t.}\quad\text{incentive compatibility + individual rationality}$$
-The revelation principle: any Bayesian game equilibrium can be implemented by a direct mechanism — truthful type reporting is an equilibrium. Mechanism design is the most practical branch — rather than lamenting non-cooperation, change the rules so that cooperation becomes the self-interested choice.
-
-## Common Errors
-
-| Error | Critique | Correct Approach |
-|---|---|---|
-| Assuming others think the same way you do | Others have different payoff functions / rationality levels; $u_i\neq u_j$ | Explicitly specify each player's payoff function and beliefs |
-| Ignoring mixed strategies | Pure-strategy equilibria may not exist; mixed equilibria always exist (Nash, 1950) | Check whether mixed-strategy analysis is needed |
-| Confusing Nash equilibrium with Pareto optimality | Nash equilibrium can be Pareto-dominated (Prisoner's Dilemma); the two are independent | Separately annotate equilibrium properties and efficiency properties |
-| Ignoring information asymmetry | Incomplete-information equilibria differ fundamentally from complete-information ones | Distinguish Bayesian games from complete-information games |
-| Oversimplifying payoff structure | Omitting key payoff dimensions distorts the equilibrium | Systematically check complete payoffs for all players |
-| Ignoring repeated interaction dynamics | One-shot equilibria may change in repeated games (folk theorem) | Analyze subgame-perfect equilibria |
-| Intractable exact large-scale equilibrium | $n$-player Nash equilibrium is PPAD-complete; exact solutions are infeasible | Pass through the GPU eight-dimensional gate; switch to approximate / learning methods (self-play, MARL) |
-
-## Operating Procedure
-
-When this skill is triggered, the output must include:
-
-1. **[Players]**: All decision-makers and their types (rational / boundedly rational / unknown).
-2. **[Strategy Sets]**: Available strategies for each player; label finite / infinite.
-3. **[Payoff Matrix]**: Payoff values $(u_1,u_2,\ldots)$ for core strategy profiles, or payoff function form.
-4. **[Game Type]**: Zero-sum / non-zero-sum / sequential / cooperative, with corresponding analytical method.
-5. **[Equilibria]**: All applicable equilibria (pure strategy, mixed strategy, subgame-perfect, cooperative allocation, etc.); label multiple equilibria.
-6. **[Stability]**: Trembling-hand perfection, evolutionary stability, off-equilibrium-path incentives.
-7. **[Mechanism Recommendations]**: If equilibria are undesirable, propose mechanism design suggestions (modify payoffs, add rules, introduce reputation, etc.).
-8. **[GPU Feasibility]**: Matrix games / minimax → LP is GEMM-compatible; large-scale multi-agent equilibria may be intractable → approximate / learning methods; pass through the eight-dimensional gate.
-
-**Output must not consist of analysis alone without conclusions.**
-
-## Relations to Other Skills
-
-- **Optimization Thinking**: Nash equilibrium is mutual optimization — each player optimizes their own payoff subject to others' strategies; the equilibrium is the common solution to all optimization problems.
-- **Probability and Statistics**: Mixed strategies select actions via probability distributions; Bayesian games handle type uncertainty probabilistically; equilibria depend on expected payoffs.
-- **Information Theory Thinking**: Information asymmetry and signaling — in Spence signaling games, the information structure directly affects equilibrium existence and properties.
-- **Causal Inference Thinking**: Strategic choices causally affect others' payoffs; causal chains in strategic interaction are bidirectional.
-- **Algorithmic Thinking**: Nash equilibrium computation complexity (PPAD-complete), Lemke-Howson algorithm, evolutionary game simulation.
-- **Modern Mathematics Activation**: `../../references/books/optimization-ml.md` (duality / minimax, game learning algorithms), `../../references/books/matrix-analysis.md` (matrix games, linear algebraic structure of payoff matrices).

package/skills/game-theory/SKILL.md

@@ -1,123 +0,0 @@
----
-name: game-theory
-description: |
-  触发：当策略取决于他人选择、需计算纳什均衡、分析零和/非零和博弈、做机制设计；或为多智能体系统、对抗训练、路由博弈设计策略时调用。
-  English: Trigger when optimal strategy depends on others' choices, needing Nash equilibrium computation, zero-sum/non-zero-sum analysis, mechanism design; or designing strategies for multi-agent systems, adversarial training, routing games.
----
-
-> **语言路由**：若用户消息为英文，请读取并遵循同目录下的 `SKILL.en.md`，按其操作规程以英文输出；中文消息则继续使用本文件。
-
-# 🎯 博弈论思想 / Game Theory
-
-> "你的最优选择取决于他人的选择——思考不仅要深入，还要互动。"
-> "Your optimal choice depends on others' choices — thinking must be interactive, not just deep."
->
-> —— 博弈论、决策论、机制设计 / Game Theory, Decision Theory, Mechanism Design
-
-## 核心原则 / Core Principle
-
-**战略互动——你的最优行动取决于他人做什么，而他人的最优行动取决于你做什么。纳什均衡：无人能从单方面偏离中获益的策略组合。**
-
-**Strategic interaction — your best action depends on what others do, and theirs depend on what you do. Nash equilibrium: strategy profile where no player benefits from unilateral deviation.**
-
-关键概念：**支付矩阵（Payoff Matrix）**、**占优策略（Dominant Strategy）**、**混合策略（Mixed Strategy）**、**帕累托最优性（Pareto Optimality）**。分类维度：合作 vs 非合作、零和 vs 一般和、同时 vs 序列。
-
-> **数学形式化 / Mathematical Formalization**
->
-> 形式化定义：n 个参与者，每人有策略集 $S_i$ 和支付函数 $u_i: S_1 \times \cdots \times S_n \to \mathbb{R}$。
->
-> 纳什均衡：策略组合 $s^* = (s_1^*, \ldots, s_n^*)$ 满足 $u_i(s^*) \geq u_i(s_i, s^*_{-i})$ 对所有 $s_i \in S_i$、所有 $i$ 成立，其中 $s^*_{-i}$ 为除 $i$ 外所有人的策略。混合策略纳什均衡：概率分布 $\sigma^*$ 使 $u_i(\sigma^*) \geq u_i(s_i, \sigma^*_{-i})$。
->
-> 占优策略：严格 $u_i(s_i^*, s_{-i}) > u_i(s_i, s_{-i})$；弱占优 $\geq$ 且至少一处严格。
->
-> 零和 minimax：$\max_x \min_y f(x,y) = \min_y \max_x f(x,y)$。
-
-## GPU 友好性 / GPU-Friendliness（横切检查）
-
-博弈求解映射 GPU 的能力差异极大，过 `../../references/gpu-friendly-math.md` 八维门：
-
-- **矩阵博弈 / 双人零和（minimax）**：等价于线性规划（LP），支付矩阵运算可 GEMM 化、可批量并行——**友好**。
-- **混合策略均衡求解**：支撑集上的线性方程组/无差异方程，小规模可批量解；规模大时注意稀疏性与条件数。
-- **大规模多智能体均衡精确求解**：n 人纳什均衡 PPAD-complete，一般情形**不可算**→ **反模式**，改用近似/学习方法（fictitious play、虚拟自博、多智能体 RL、NFSP）。
-- **演化/重复博弈模拟**：群体与策略批量更新可并行；注意 agent 间通信与同步开销。
-
-八维最低判定（正式术语）：**张量化**看玩家/策略/回合是否能批量展开；**GEMM 可映射**看支付矩阵、minimax、LP/线性方程是否落矩阵运算；**复杂度**标注 Nash/机制求解的 PPAD/组合爆炸风险；**显存与 KV-Cache**看策略-状态笛卡尔积是否可压缩；**低精度稳定**看均衡求解、LP、soft best-response 是否条件数可控；**并行与通信**看多智能体 rollout 的同步成本；**稀疏结构**看交互图是否结构化；**算子融合**看收益、mask、best-response 更新能否融合。
-
-> 配合 `../../references/books/optimization-ml.md`（对偶/minimax、博弈学习算法）与 `../../references/books/matrix-analysis.md`（矩阵博弈、线性方程组）。
-
-## 不适用场景 / When NOT to Use
-
-- **单人决策问题，与他人无互动**——优化思想更适用，无需考虑他人策略响应。
-- **纯合作问题，无利益冲突**——参与者目标完全一致，无需战略分析。
-- **确定性问题，无战略不确定性**——结果由自身行动唯一决定，不涉及他人反应。
-- **纯粹运气决定的结果，各方无策略选择**——无策略集可分析，博弈论不适用。
-
-## 何时使用 / When to Use
-
-- 多个决策者相互影响，各自策略影响他人收益；最佳策略取决于他人响应，需预判他人行动。
-- 需设计激励机制，使自利行为导向社会最优结果；预测给定规则下的均衡稳定状态。
-- 分析竞争市场中的定价、进入、退出；谈判/讨价还价中的利益分配与威胁点。
-- **为多智能体系统、对抗训练（GAN/鲁棒学习）、路由博弈设计策略与均衡目标**。
-
-## 方法流程 / Method
-
-### 第一步：识别参与者与策略
-定义博弈基本要素：参与者 $N=\{1,\ldots,n\}$；策略集 $S_i$；支付函数 $u_i: S_1\times\cdots\times S_n\to\mathbb{R}$；信息结构（是否对称？是否不完全信息？行动是否序列？能否通信？）。核心是识别"互动结构"——谁的选择影响谁的结果，遗漏关键参与者是最常见错误。
-
-### 第二步：分析博弈类型
-先判定零和/一般和、同时/序列、合作/非合作、完全信息/不完全信息。类型决定方法：零和用 minimax/LP；序列博弈用逆向归纳与子博弈完美均衡；合作博弈用 Shapley 值按边际贡献分配：
-$$\phi_i(v)=\sum_{S\subseteq N\setminus\{i\}}\frac{|S|!(n-|S|-1)!}{n!}[v(S\cup\{i\})-v(S)]$$
-误判类型会直接导致均衡概念和求解方法错用。
-
-### 第三步：构建支付矩阵/函数
-2 人博弈用支付矩阵，单元格 $(u_1,u_2)$；n 人博弈用 $u_i(s_1,\ldots,s_n)$。验证完备性：是否遗漏参与者或策略？支付是否准确反映偏好？是否存在外部性？支付是分析根基，评估不准则后续全偏。
-
-### 第四步：寻找占优策略
-严格占优 $u_i(s_i^*,s_{-i}) > u_i(s_i,s_{-i})$ 对所有 $s_i\neq s_i^*$、所有 $s_{-i}$；弱占优 $\geq$ 且至少一处严格；**迭代消去劣策略**依次删去劣策略缩小空间。存在占优策略均衡是最强结论——无人有任何偏离动机；否则进入均衡分析。
-
-### 第五步：计算均衡
-按第二步的博弈类型选方法。纯策略纳什：对每个 $s_{-i}$ 找最优响应 $BR_i(s_{-i})$，交叉点即均衡。混合策略：求使他人无差异的概率分布 $u_j(s_j,\sigma^*_{-j})=u_j(s_j',\sigma^*_{-j})$ 对所有 $s_j,s_j'\in S_j$。零和矩阵博弈可走 minimax $\max_x\min_y f=\min_y\max_x f$；序列博弈用逆向归纳；合作博弈输出分配方案而非普通纳什。多重均衡做帕累托排序，标注最优与劣均衡。
-
-### 第六步：检查均衡稳定性
-颤抖手完美（微小失误下仍稳定，排除依赖"对手绝不犯错"的均衡）；演化稳定策略 ESS（重复互动中抵御入侵）；偏离均衡路径激励（序列博弈中是否有动机遵守承诺）。追问：均衡唯一还是多重？对微小扰动是否鲁棒？稳定均衡才是可信预测。
-
-### 第七步：设计机制与改进
-若不存在好的均衡，重新设计博弈：改变支付结构（奖惩）、增加可执行规则、引入通信与声誉、Vickrey 拍卖（报真实估值为占优策略）。机制设计形式化：
-$$\max\ \text{social welfare}\quad\text{s.t.}\quad\text{incentive compatibility + individual rationality}$$
-显示原理：任何贝叶斯博弈均衡都可由直接机制实现——如实报告类型即均衡。机制设计是最实用分支——与其抱怨人不合作，不如改变规则让合作成为自利选择。
-
-## 常见错误 / Common Errors
-
-| 错误 / Error | 批评 / Critique | 正确做法 / Correct Approach |
-|---|---|---|
-| 假设他人像自己一样思考 | 他人支付函数/理性水平不同，$u_i\neq u_j$ | 明确各参与者支付函数与信念 |
-| 忽略混合策略 | 纯策略均衡可能不存在，混合均衡总存在（Nash, 1950） | 检查是否需混合策略分析 |
-| 混淆纳什均衡与帕累托最优 | 纳什均衡可帕累托劣（囚徒困境），两者独立 | 分别标注均衡性与效率性 |
-| 忽略信息不对称 | 不完全信息均衡与完全信息截然不同 | 区分贝叶斯博弈与完全信息博弈 |
-| 过度简化支付结构 | 遗漏关键支付维度使均衡失真 | 系统检查所有参与者完整支付 |
-| 忽略重复互动动态 | 单次均衡在重复博弈中可能改变（folk theorem） | 分析子博弈完美均衡 |
-| 大规模均衡精确求解不可算 | n 人纳什均衡 PPAD-complete，精确解不可行 | 过 GPU 八维门，改用近似/学习（虚拟自博、MARL） |
-
-## 操作规程 / Operating Procedure
-
-当本 skill 被触发时，输出必须包含：
-
-1. **[参与者]**：所有决策者及其类型（理性/有限理性/未知）。
-2. **[策略集]**：每个参与者的可选策略，标注有限/无限。
-3. **[支付矩阵]**：核心策略组合的支付值 $(u_1,u_2,\ldots)$ 或支付函数形式。
-4. **[博弈类型]**：零和/非零和/序列/合作，对应分析方法。
-5. **[均衡]**：所有适用均衡（纯策略、混合策略、子博弈完美、合作分配等），标注多重均衡。
-6. **[稳定性]**：颤抖手完美性、演化稳定性、偏离均衡路径激励。
-7. **[机制建议]**：若均衡不理想，提出机制设计建议（改变支付、增加规则、引入声誉等）。
-8. **[GPU 可行性]**：矩阵博弈/minimax→LP 可 GEMM；大规模多智能体均衡是否不可算→近似/学习方法，过八维门。
-
-**输出不得只给分析而无结论。**
-
-## 与其他 skill 的关系 / Relations to Other Skills
-
-- **优化思想**：纳什均衡是互相优化——每人在他人策略约束下优化自己支付，均衡是所有优化问题的共同解。
-- **概率与统计**：混合策略以概率分布选行动，贝叶斯博弈用概率处理类型不确定性，均衡依赖期望支付。
-- **信息论思想**：信息不对称与信号传递——Spence 信号博弈中信息结构直接影响均衡存在与性质。
-- **因果推断思想**：策略选择因果影响他人收益，战略互动中的因果链是双向的。
-- **算法思想**：纳什均衡计算复杂度（PPAD-complete）、Lemke-Howson 算法、演化博弈模拟。
-- **现代数学激活**：`../../references/books/optimization-ml.md`（对偶/minimax、博弈学习算法）、`../../references/books/matrix-analysis.md`（矩阵博弈、支付矩阵的线性代数结构）。