math-skill 2.1.0 → 3.0.0

package/skills/game-theory/original-texts.en.md

@@ -1,131 +0,0 @@
-# Mathematical Sources and Classic Texts
-
-## Zermelo's Theorem (1913)
-
-> Zermelo (1913) proved: in any finite two-player game of perfect information (such as chess), either the first player or the second player has a winning strategy, or both players have a drawing strategy. That is, the game is "determined."
-
-**Core thesis**: For any finite game of perfect information, $\exists$ a winning strategy for the first player $\lor$ $\exists$ a winning strategy for the second player $\lor$ $\exists$ a drawing strategy for both. This proves that no genuine "uncertainty" exists in such games — the source of uncertainty is computational complexity rather than logical structure.
-
-**Key ideas**: The outcome of a perfect-information game is logically predetermined — uncertainty arises from limitations in computational power, not from the game itself. Zermelo's proof employed backward induction, which later became the central tool for subgame perfect equilibrium. Mathematical context: in the same year, Zermelo also introduced the Axiom of Choice (AC) in set theory. Game theory and set theory share the logic that "finite choices can always be exhaustively enumerated."
-
-## von Neumann's Minimax Theorem (1928)
-
-> von Neumann (1928) proved: in a two-player zero-sum game, $\max_{x \in \Delta} \min_{y \in \Delta} \sum_{ij} a_{ij} x_i y_j = \min_{y \in \Delta} \max_{x \in \Delta} \sum_{ij} a_{ij} x_i y_j$. That is, the optimal outcome under a conservative strategy equals the optimal outcome under the opponent's conservative strategy — max-min = min-max.
-
-**Core theorem**: For any two-player zero-sum matrix game $A$, there exist mixed strategies $x^*$ and $y^*$ such that
-
-$$v^* = \max_x \min_y x^\top A y = \min_y \max_x x^\top A y$$
-
-$v^*$ is called the value of the game. The proof utilizes the convex set separation theorem (which later revealed deep connections with the Brouwer fixed-point theorem).
-
-**Key ideas**: The minimax theorem reveals the structural symmetry of zero-sum games — "my maximizing my gain" is equivalent to "the opponent minimizing my gain." The optimal mixed strategy renders the opponent indifferent among all pure strategies. This theorem foreshadowed the duality of linear programming: Dantzig (1947) discovered that LP duality is equivalent to minimax, and von Neumann independently found the same connection.
-
-## Existence of Nash Equilibrium (1950)
-
-> Nash (1950) proved: every finite game (any number of players, non-zero-sum) has at least one mixed-strategy Nash equilibrium. The proof uses the Brouwer fixed-point theorem — constructing a best-response mapping $BR: \Delta \to \Delta$ whose fixed point is a Nash equilibrium.
-
-**Core theorem**: For any finite game $(N, \{S_i\}, \{u_i\})$, $\exists \sigma^* \in \Delta_1 \times \cdots \times \Delta_n$ such that for all $i$ and $s_i \in S_i$,
-
-$$u_i(\sigma^*) \geq u_i(s_i, \sigma^*_{-i})$$
-
-Nash's proof applies the Brouwer fixed-point theorem to the best-response mapping: $F(\sigma) = (BR_1(\sigma_{-1}), \ldots, BR_n(\sigma_{-n}))$, where the fixed point of $F$ is an equilibrium. Subsequent proofs also use the Kakutani fixed-point theorem (for set-valued correspondences).
-
-**Key ideas**: The existence of Nash equilibrium does not depend on the "goodwill" or "cooperative intent" of the players — it is a logical necessity of self-interested behavior in strategic interaction. The Nash equilibrium generalizes the minimax theorem from zero-sum games to non-zero-sum games, and from two-player adversarial settings to multi-player interaction. Mathematical context: the Brouwer fixed-point theorem (1911) is a central result in topology — every continuous map has a fixed point — and game theory and topology converge here.
-
-## Prisoner's Dilemma
-
-> The Prisoner's Dilemma is the most classic example in game theory: two players each have "cooperate" and "defect" strategies. Defection is the dominant strategy, yet (Defect, Defect) is Pareto-dominated by (Cooperate, Cooperate) — individual rationality leads to collective irrationalality.
-
-**Payoff matrix**:
-
-| | Cooperate (C) | Defect (D) |
-|---|---|---|
-| Cooperate (C) | (3, 3) | (0, 5) |
-| Defect (D) | (5, 0) | (1, 1) |
-
-Strictly dominant strategy analysis: for each player, defection strictly dominates cooperation ($5 > 3$ and $1 > 0$). The unique Nash equilibrium (D, D) yields payoff (1, 1), which is Pareto-dominated by (C, C) with payoff (3, 3).
-
-**Key ideas**: The Prisoner's Dilemma reveals the fundamental tension between individual and collective rationality — Nash equilibrium does not guarantee Pareto optimality. This dilemma is ubiquitous in reality: arms races, the tragedy of the commons, price wars, carbon emissions. Axelrod (1984) showed through computer tournaments that in the iterated Prisoner's Dilemma, the "Tit-for-Tat" strategy can sustain cooperation — repeated interaction changes the equilibrium structure.
-
-## Backward Induction & Subgame Perfection
-
-> Backward induction is the central analytical tool for sequential games: starting from terminal nodes and working backward, choosing the optimal action at each step. Selten (1965) defined the subgame perfect equilibrium (SPE): a strategy profile that constitutes a Nash equilibrium in every subgame — eliminating equilibria based on non-credible threats.
-
-**Core concept**: Nash equilibrium permits "non-credible threats" — a player claims they will take a certain action, but when that node is reached, deviation is more profitable. SPE requires the equilibrium to hold in every subgame, thereby eliminating non-credible threats. Backward induction is the standard method for computing SPE.
-
-**Key ideas**: In sequential games, the temporal structure introduces commitment and credibility issues — a "threat" must be rational to be effective. Backward induction decomposes multi-stage decisions into single-stage subproblems, in the spirit of Bellman's principle of optimality. Mathematical context: Zermelo's (1913) theorem used backward induction, and Selten (1965) systematized it into the SPE concept. Selten shared the 1994 Nobel Prize in Economics with Nash and Harsanyi.
-
-## Shapley Value & Cooperative Game Theory (1953)
-
-> Shapley (1953) proposed the unique allocation rule in cooperative games satisfying efficiency, symmetry, dummy player, and additivity — the Shapley value $\phi_i(v)$. It allocates coalition payoffs according to average marginal contributions.
-
-$$\phi_i(v) = \sum_{S \subseteq N \setminus \{i\}} \frac{|S|!(n - |S| - 1)!}{n!} [v(S \cup \{i\}) - v(S)]$$
-
-The Shapley value is the unique allocation satisfying the following four axioms:
-1. **Efficiency**: $\sum_i \phi_i(v) = v(N)$
-2. **Symmetry**: if $i$ and $j$ contribute equally in all coalitions, then $\phi_i = \phi_j$
-3. **Dummy player**: if $i$ contributes zero marginal value to every coalition, then $\phi_i = 0$
-4. **Additivity**: $\phi_i(v + w) = \phi_i(v) + \phi_i(w)$
-
-**Key ideas**: The Shapley value translates "fairness" into mathematical axioms — fairness is not subjective judgment but logical necessity. The central question of cooperative game theory: how should the total payoff generated by a coalition be distributed? The Shapley value provides the unique axiomatically justified answer. The core is another solution concept — no coalition receives less than its standalone value — but the core may be empty.
-
-## Nash Bargaining Solution (1950)
-
-> Nash (1950) proposed the unique bargaining solution satisfying Pareto optimality, symmetry, scale invariance, and independence of irrelevant alternatives. Given a disagreement point $d = (d_1, d_2)$ and a feasible set $S \subseteq \mathbb{R}^2$, the Nash solution maximizes $(u_1 - d_1)(u_2 - d_2)$.
-
-$$\max_{(u_1, u_2) \in S, u_i \geq d_i} (u_1 - d_1)(u_2 - d_2)$$
-
-Nash proved that this solution is the unique allocation satisfying the four axioms. The disagreement point $d$ is each party's payoff when negotiations break down — the more favorable the disagreement point, the more favorable the bargaining outcome.
-
-**Key ideas**: The bargaining problem translates "negotiation" into mathematical optimization — maximizing joint surplus subject to the feasible set and disagreement point constraints. The symmetry and Pareto optimality of the Nash solution make it a natural benchmark for fair negotiation. Mathematical context: this model connects cooperative and non-cooperative game theory — the Nash program asserts that all cooperative game solutions should be derived through non-cooperative game models.
-
-## Evolutionary Game Theory (Maynard Smith, 1973)
-
-> Maynard Smith (1973) proposed the Evolutionarily Stable Strategy (ESS): a strategy $\sigma^*$ is an ESS if for any invading strategy $\sigma \neq \sigma^*$, either $u(\sigma^*, \sigma^*) > u(\sigma, \sigma^*)$, or $u(\sigma^*, \sigma^*) = u(\sigma, \sigma^*)$ and $u(\sigma^*, \sigma) > u(\sigma, \sigma)$.
-
-**Core definition**: ESS is a stronger stability concept than Nash equilibrium — it requires not only "no incentive to deviate" but also "resistance to invasion." Nash equilibrium is a necessary condition for ESS, but ESS adds a stability requirement.
-
-**Key ideas**: Evolutionary game theory extends game theory from "rational choice" to "adaptive dynamics" — players need not be rational, they only need to adjust strategies through repeated interaction. Biological applications: sex-ratio games, Hawk-Dove games, resource competition. Economic applications: formation of market conventions, diffusion of technology standards. Mathematical context: ESS is connected to replicator dynamics — an ESS is a locally stable fixed point of the replicator dynamics.
-
-## Vickrey Auction & Auction Theory (1961)
-
-> Vickrey (1961) proposed the second-price sealed-bid auction: the highest bidder wins but pays the second-highest bid. Bidding one's true valuation is a dominant strategy — a foundational work in mechanism design.
-
-**Core theorem**: In a Vickrey auction, bidding one's true valuation $b_i = v_i$ is a weakly dominant strategy for every bidder. Proof: let $v_i$ be the true valuation and $p$ the second-highest bid. If $v_i > p$, bidding $b_i = v_i$ wins with payoff $v_i - p > 0$; if $b_i > v_i > p$, one still wins with the same payoff; if $b_i < v_i$, one may forgo positive surplus. Hence truthful reporting is optimal.
-
-Vickrey proved that for symmetric risk-neutral bidders, the four standard auctions (English, Dutch, first-price sealed-bid, second-price sealed-bid) yield equal expected revenue — the **Revenue Equivalence Theorem**. Milgrom & Weber (1982) generalized this to affiliated values, finding that the English auction yields the highest revenue.
-
-**Key ideas**: The Vickrey auction is the paradigm of an incentive-compatible mechanism — through clever design of the payment rule, self-interested behavior automatically leads to efficient outcomes. Auction theory applies game theory to market design: how can rules be designed to guide rational agents toward optimal outcomes? Mathematical context: Vickrey shared the 1996 Nobel Prize in Economics.
-
-## Mechanism Design (Hurwicz, Myerson, Maskin — Nobel 2007)
-
-> Mechanism design is the "reverse engineering" of game theory — given a desired outcome, design the rules of the game so that rational agents' self-interested behavior automatically achieves that goal. Hurwicz, Myerson, and Maskin received the 2007 Nobel Prize in Economics for this work.
-
-**Core concepts**:
-- **Incentive Compatibility**: truthful type reporting is an equilibrium strategy — agents have no incentive to misreport
-- **Revelation Principle**: any equilibrium of a Bayesian game can be implemented by a direct mechanism — agents truthfully report types in equilibrium, eliminating the need for complex strategies
-- **Myerson's optimal auction**: under the independent private values model, the mechanism maximizing the seller's expected revenue allocates to the bidder with the highest virtual valuation $v_i - (1 - F_i(v_i))/f_i(v_i)$
-
-$$\text{Virtual valuation: } \varphi_i(v_i) = v_i - \frac{1 - F_i(v_i)}{f_i(v_i)}$$
-
-- **Maskin Monotonicity**: a necessary condition for a social choice function to be Nash-implementable — monotonicity of preferences
-
-**Key ideas**: Mechanism design answers "how to design the rules" rather than "how to act under given rules" — working backward from the desired outcome to the structure of the game. The revelation principle simplifies mechanism design: one need only consider direct mechanisms. Incentive compatibility ensures alignment between rules and incentives — institutional design is not wishful thinking but mathematically verifiable engineering. Mathematical context: Hurwicz (1960s) introduced the concept of incentive compatibility; Myerson (1981) developed optimal auction theory; Maskin (1977/1999) developed Nash implementability theory.
-
-## Signaling Games (Spence, 1973)
-
-> Spence (1973) proposed the signaling model: under asymmetric information, a sender with private information communicates information to a receiver through observable actions (signals). Classic application: education as a signal of ability — high-ability individuals face lower costs of acquiring education, so educational investment conveys information about ability.
-
-**Core model**: The sender has type $t \in \{H, L\}$ (high/low ability), chooses signal $s \in \{e, 0\}$ (educated/uneducated), and the receiver observes $s$ and then chooses response $a$. Separating equilibrium: $H$ chooses $e$, $L$ chooses $0$ — the signal perfectly reveals type. Pooling equilibrium: both types choose the same signal — the signal carries no information.
-
-**Key ideas**: Signaling games reveal the economics of information transmission — signals must have differential costs to convey information (otherwise everyone would imitate). In the Spence model, the social value of education may be entirely signaling value rather than productivity enhancement — this sparked a profound discussion about the nature of education. Spence shared the 2001 Nobel Prize in Economics. Mathematical context: signaling games are a subclass of Bayesian games — type is private information, and strategies are type-dependent.
-
-## Repeated Games & Folk Theorem
-
-> Folk Theorem (informally circulated in the 1970s, systematized by Aumann 1981): in an infinitely repeated game, any individually rational payoff $v_i \geq \min_{s_{-i}} \max_{s_i} u_i(s_i, s_{-i})$ (the minimax payoff) can be sustained as the average payoff of a subgame perfect equilibrium.
-
-**Core theorem**: Let $v = (v_1, \ldots, v_n)$ be a feasible and individually rational payoff vector ($v_i \geq \underline{v}_i$, where $\underline{v}_i$ is $i$'s minimax value). Then, when the discount factor $\delta$ is sufficiently close to 1, $v$ can be realized as the average payoff of a subgame perfect equilibrium of the infinitely repeated game.
-
-Implementation mechanism: punishment strategy — a deviator is "punished" back to the minimax payoff level, so the short-term gain from deviation is offset by the long-term punishment.
-
-**Key ideas**: The Folk Theorem is the central result in repeated games — it explains why cooperation can be sustained in long-run interactions: the short-term gain from deviation is outweighed by the long-term loss from punishment. The theorem also explains why repeated games have infinitely many equilibria — virtually any feasible payoff can be realized. Aumann (2005 Nobel Prize) systematized this theory. Mathematical context: the discount factor $\delta$ approaching 1 makes future punishment sufficiently severe — $\delta \geq \frac{g}{g + l}$, where $g$ is the gain from deviation and $l$ is the punishment loss.

package/skills/game-theory/original-texts.md

@@ -1,131 +0,0 @@
-# 数学出处与经典文献 / Mathematical Sources and Classic Texts
-
-## Zermelo 定理 / Zermelo's Theorem (1913)
-
-> Zermelo（1913）证明：在有限完美信息二人博弈中（如国际象棋），先手或后手必有一方有必胜策略或双方有必和策略。即博弈是"确定的"（determined）。
-
-**核心论点**：对任何有限完美信息博弈，$\exists$ 先手必胜策略 $\lor$ $\exists$ 后手必胜策略 $\lor$ $\exists$ 双方必和策略。证明了此类博弈不存在真正的"不确定性"——不确定性的来源是计算复杂度而非逻辑结构。
-
-**思想**：完美信息博弈的结局在逻辑上是预定的——不确定性来自计算能力的限制而非博弈本身。Zermelo 的证明使用逆向归纳法（backward induction），这在后来成为子博弈完美均衡的核心工具。数学背景：Zermelo 同年也在集合论中提出选择公理（AC），博弈论与集合论共享"有限选择必可穷尽"的逻辑。
-
-## von Neumann Minimax 定理 / von Neumann's Minimax Theorem (1928)
-
-> von Neumann（1928）证明：在二人零和博弈中，$\max_{x \in \Delta} \min_{y \in \Delta} \sum_{ij} a_{ij} x_i y_j = \min_{y \in \Delta} \max_{x \in \Delta} \sum_{ij} a_{ij} x_i y_j$。即保守策略的最优结果等于对手保守策略下的最优结果——max-min = min-max。
-
-**核心定理**：对任何二人零和矩阵博弈 $A$，存在混合策略 $x^*$ 和 $y^*$ 使得
-
-$$v^* = \max_x \min_y x^\top A y = \min_y \max_x x^\top A y$$
-
-$v^*$ 称为博弈的值（value of the game）。证明利用了凸集分离定理（后与 Brouwer 不动点定理有深层联系）。
-
-**思想**：minimax 定理揭示了零和博弈的结构对称性——"我尽可能赢"等价于"对手尽可能防"。最优混合策略使对手对所有纯策略无差异。此定理是线性规划对偶性的先声：Dantzig（1947）发现 LP 对偶与 minimax 等价，von Neumann 独立发现了同样联系。
-
-## Nash 均衡的存在性 / Existence of Nash Equilibrium (1950)
-
-> Nash（1950）证明：每个有限博弈（任意人数、非零和）至少存在一个混合策略纳什均衡。证明使用 Brouwer 不动点定理——构造最优响应映射 $BR: \Delta \to \Delta$，其不动点即为纳什均衡。
-
-**核心定理**：对任何有限博弈 $(N, \{S_i\}, \{u_i\})$，$\exists \sigma^* \in \Delta_1 \times \cdots \times \Delta_n$ 使得对所有 $i$ 和 $s_i \in S_i$，
-
-$$u_i(\sigma^*) \geq u_i(s_i, \sigma^*_{-i})$$
-
-Nash 的证明将 Brouwer 不动点定理应用于最优响应映射：$F(\sigma) = (BR_1(\sigma_{-1}), \ldots, BR_n(\sigma_{-n}))$，$F$ 的不动点即均衡。后续证明也用 Kakutani 不动点定理（对应集值映射）。
-
-**思想**：纳什均衡的存在性不依赖参与者的"善意"或"合作意愿"——它是自私行为在战略互动中的逻辑必然。Nash 均衡推广了 minimax 定理到非零和博弈，从二人对抗到多人互动。数学背景：Brouwer 不动点定理（1911）是拓扑学核心定理——连续映射必有不动点，博弈论与拓扑学在此交汇。
-
-## 囚徒困境 / Prisoner's Dilemma
-
-> 囚徒困境是博弈论最经典的范例：两个参与者各自有"合作"和"背叛"策略。背叛是占优策略，但（背叛, 背叛）是帕累托劣于（合作, 合作）——个体理性导致集体非理性。
-
-**支付矩阵**：
-
-| | 合作 (C) | 背叛 (D) |
-|---|---|---|
-| 合作 (C) | (3, 3) | (0, 5) |
-| 背叛 (D) | (5, 0) | (1, 1) |
-
-严格占优策略分析：对每个参与者，背叛严格优于合作（$5 > 3$ 且 $1 > 0$）。唯一纳什均衡 (D, D) 支付 (1, 1)，帕累托劣于 (C, C) 支付 (3, 3)。
-
-**思想**：囚徒困境揭示了个体理性与集体理性的根本张力——纳什均衡不保证帕累托最优。此困境在现实中无处不在：军备竞赛、公地悲剧、价格战、碳排放。Axelrod（1984）的计算机竞赛表明：在重复囚徒困境中，"一报还一报"（Tit-for-Tat）策略可维持合作——重复互动改变均衡结构。
-
-## 逆向归纳法与子博弈完美 / Backward Induction & Subgame Perfection
-
-> 逆向归纳法（backward induction）是序列博弈的核心分析工具：从终端节点向前回溯，每步选择最优行动。Selten（1965）定义子博弈完美均衡（subgame perfect equilibrium, SPE）：在每个子博弈上都构成纳什均衡的策略组合——排除基于不可信威胁的均衡。
-
-**核心概念**：纳什均衡允许"不可信威胁"——参与者声称将采取某行动，但到达该节点时偏离更有利。SPE 要求均衡在每个子博弈上都成立，因此排除了不可信威胁。逆向归纳法是求解 SPE 的标准方法。
-
-**思想**：序列博弈中，时间结构引入了承诺与可信性问题——"威胁"必须是理性的才有效。逆向归纳法将多阶段决策分解为单阶段子问题，与 Bellman 最优性原理一脉相承。数学背景：Zermelo（1913）的定理使用逆向归纳法，Selten（1965）将其系统化为 SPE 概念。Selten 与 Nash、Harsanyi 共获 1994 年诺贝尔经济学奖。
-
-## Shapley 值与合作博弈 / Shapley Value & Cooperative Game Theory (1953)
-
-> Shapley（1953）提出合作博弈中唯一满足效率、对称性、哑元性和可加性的分配规则——Shapley 值 $\phi_i(v)$。它按边际贡献的平均值分配联盟收益。
-
-$$\phi_i(v) = \sum_{S \subseteq N \setminus \{i\}} \frac{|S|!(n - |S| - 1)!}{n!} [v(S \cup \{i\}) - v(S)]$$
-
-Shapley 值是唯一满足以下四条公理的分配：
-1. **效率性**：$\sum_i \phi_i(v) = v(N)$
-2. **对称性**：若 $i$ 和 $j$ 在所有联盟中贡献相同，则 $\phi_i = \phi_j$
-3. **哑元性**：若 $i$ 对所有联盟无边际贡献，则 $\phi_i = 0$
-4. **可加性**：$\phi_i(v + w) = \phi_i(v) + \phi_i(w)$
-
-**思想**：Shapley 值将"公平"转化为数学公理——公平不是主观判断，而是逻辑必然。合作博弈的核心问题：联盟产生的总收益如何分配？Shapley 值给出唯一满足公理的答案。核心（core）是另一种解概念——任何联盟的分配不低于其自身价值，但核心可能为空集。
-
-## Nash 讨价还价解 / Nash Bargaining Solution (1950)
-
-> Nash（1950）提出满足帕累托最优性、对称性、尺度无关性和独立性缩减的唯二讨价还价解。设威胁点 $d = (d_1, d_2)$，可行集 $S \subseteq \mathbb{R}^2$，Nash 解最大化 $(u_1 - d_1)(u_2 - d_2)$。
-
-$$\max_{(u_1, u_2) \in S, u_i \geq d_i} (u_1 - d_1)(u_2 - d_2)$$
-
-Nash 证明此解是满足四条公理的唯一分配。威胁点 $d$ 是谈判失败时各方的收益——威胁点越有利，谈判结果越有利。
-
-**思想**：讨价还价问题将"谈判"转化为数学优化——在可行集和威胁点约束下最大化联合收益。Nash 解的对称性与帕累托最优性使其成为公平谈判的自然基准。数学背景：此模型将合作博弈与非合作博弈联系起来——Nash 程序（Nash program）主张所有合作博弈解都应通过非合作博弈模型推导。
-
-## 演化博弈论 / Evolutionary Game Theory (Maynard Smith, 1973)
-
-> Maynard Smith（1973）提出演化稳定策略（Evolutionarily Stable Strategy, ESS）：策略 $\sigma^*$ 是 ESS，若对任何入侵策略 $\sigma \neq \sigma^*$，要么 $u(\sigma^*, \sigma^*) > u(\sigma, \sigma^*)$，要么 $u(\sigma^*, \sigma^*) = u(\sigma, \sigma^*)$ 且 $u(\sigma^*, \sigma) > u(\sigma, \sigma)$。
-
-**核心定义**：ESS 是比纳什均衡更强的稳定性概念——不仅要"不被偏离"，还要"不被入侵"。纳什均衡是 ESS 的必要条件，但 ESS 附加了稳定性要求。
-
-**思想**：演化博弈论将博弈论从"理性选择"推广到"适应性动态"——参与者不必理性，只需在重复互动中调整策略。生物学应用：性别比例博弈、鹰鸽博弈（Hawk-Dove）、资源竞争。经济学应用：市场惯例的形成、技术标准的扩散。数学背景：ESS 与复制者动态（replicator dynamics）联系——ESS 是复制者动态的局部稳定点。
-
-## Vickrey 拍卖与拍卖理论 / Vickrey Auction & Auction Theory (1961)
-
-> Vickrey（1961）提出第二价格拍卖（second-price sealed-bid auction）：出价最高者获胜，但支付第二高出价。报真实估值是占优策略——机制设计的奠基之作。
-
-**核心定理**：在 Vickrey 拍卖中，报真实估值 $b_i = v_i$ 是每个投标者的弱占优策略。证明：设真实估值 $v_i$，第二高出价 $p$。若 $v_i > p$，报 $b_i = v_i$ 获胜支付 $p$，收益 $v_i - p > 0$；若 $b_i > v_i > p$ 同样获胜但收益不变；若 $b_i < v_i$ 可能失去正收益。因此如实报告最优。
-
-Vickrey 证明了对称风险中性投标者，四种标准拍卖（英式、荷式、第一价格密封、第二价格密封）的期望收入相等——**收入等价定理（Revenue Equivalence Theorem）**。Milgrom & Weber（1982）推广到关联估值（affiliated values），发现英式拍卖收入最高。
-
-**思想**：Vickrey 拍卖是激励相容（incentive-compatible）机制的典范——通过巧妙设计支付规则，使自利行为自动导向效率最优。拍卖理论将博弈论应用于市场设计：如何通过规则引导理性参与者产生最优结果？数学背景：Vickrey 共获 1996 年诺贝尔经济学奖。
-
-## 机制设计 / Mechanism Design (Hurwicz, Myerson, Maskin — Nobel 2007)
-
-> 机制设计是博弈论的"逆向工程"——给定目标结果，设计博弈规则使理性参与者的自利行为自动实现该目标。Hurwicz、Myerson、Maskin 因此获 2007 年诺贝尔经济学奖。
-
-**核心概念**：
-- **激励相容性（Incentive Compatibility）**：如实报告类型是均衡策略——参与者没有动机谎报
-- **显示原理（Revelation Principle）**：任何贝叶斯博弈的均衡都可由直接机制实现——参与者如实报告类型即达均衡，无需复杂策略
-- **Myerson 最优拍卖**：在独立私人估值模型下，最大化卖家期望收入的最优机制——虚拟估值 $v_i - (1 - F_i(v_i))/f_i(v_i)$ 最高者获胜
-
-$$\text{Virtual valuation: } \varphi_i(v_i) = v_i - \frac{1 - F_i(v_i)}{f_i(v_i)}$$
-
-- **Maskin 可实施性（Maskin Monotonicity）**：社会选择函数可被纳什实施的必要条件——偏好单调性
-
-**思想**：机制设计回答"如何设计规则"而非"给定规则下如何行动"——从博弈的结果反推博弈的结构。显示原理简化了机制设计：只需考虑直接机制。激励相容性保证了规则与动机的一致——制度设计不是空想，而是可被数学验证的工程。数学背景：Hurwicz（1960s）引入激励相容性概念；Myerson（1981）发展最优拍卖理论；Maskin（1977/1999）发展纳什可实施性理论。
-
-## 信号博弈 / Signaling Games (Spence, 1973)
-
-> Spence（1973）提出信号传递模型：在信息不对称下，拥有私人信息的发送者通过可观察的行动（信号）向接收者传递信息。经典应用：教育作为能力信号——高能力者获得教育的成本更低，因此教育投资传递了能力信息。
-
-**核心模型**：发送者有类型 $t \in \{H, L\}$（高/低能力），选择信号 $s \in \{e, 0\}$（受教育/不受教育），接收者观测 $s$ 后选择响应 $a$。分离均衡：$H$ 选 $e$，$L$ 选 $0$——信号完美传递类型。混同均衡：两者都选相同信号——信号无信息量。
-
-**思想**：信号博弈揭示了信息传递的经济学——信号必须有成本差异才能传递信息（否则人人都会模仿）。Spence 模型中，教育的社会价值可能完全是信号价值而非生产力提升——这引发了关于教育本质的深刻讨论。Spence 共获 2001 年诺贝尔经济学奖。数学背景：信号博弈是贝叶斯博弈的子类——类型是私人信息，策略是类型依赖的。
-
-## 重复博弈与 Folk 定理 / Repeated Games & Folk Theorem
-
-> Folk 定理（民间定理，1970s 非正式流传，Aumann 1981 系统化）：在无限重复博弈中，任何个体理性支付 $v_i \geq \min_{s_{-i}} \max_{s_i} u_i(s_i, s_{-i})$（minimax 支付）都可以作为子博弈完美均衡的平均支付被实现。
-
-**核心定理**：设 $v = (v_1, \ldots, v_n)$ 为可行且个体理性的支付向量（$v_i \geq \underline{v}_i$，$\underline{v}_i$ 为 $i$ 的 minimax 值），则在贴现因子 $\delta$ 足够接近 1 时，$v$ 可作为无限重复博弈的子博弈完美均衡的平均支付被实现。
-
-实现机制：惩罚策略（punishment strategy）——偏离者将被"惩罚"回 minimax 支付水平，偏离的短期收益被长期惩罚抵消。
-
-**思想**：Folk 定理是重复博弈的核心结果——它解释了为什么合作在长期互动中可能维持：偏离的短期收益小于长期惩罚的损失。此定理也解释了为什么重复博弈有无数个均衡——几乎任何可行支付都可被实现。Aumann（2005 年诺贝尔奖）系统化了此理论。数学背景：贴现因子 $\delta$ 趋近 1 使得未来惩罚足够重——$\delta \geq \frac{g}{g + l}$，$g$ 为偏离收益，$l$ 为惩罚损失。

package/skills/induction-analogy/SKILL.en.md

@@ -1,145 +0,0 @@
----
-name: induction-analogy
-description: |
-  Trigger when finding patterns from data/experience, specific→general, known→unknown, mathematical induction proofs (weak/strong/structural/transfinite), cross-domain analogical transfer of structures, or borrowing structures from other fields for algorithm/operator design.
----
-
-# 📈 Induction & Analogy
-
-> "From the specific to the general, from the known to the unknown — induction discovers patterns, analogy transfers experience, but neither equals proof."
->
-> — Pólya, *Mathematics and Plausible Reasoning*; Lakatos, *Proofs and Refutations*
-
-## Core Principle
-
-**Induction discovers general patterns from limited instances — the engine of new knowledge; analogy transfers understanding across domains. Neither is rigorous proof, but both are fundamental to discovering theorems, proposing hypotheses, and generating innovation. Rigor is a tool of verification, not of discovery — Pólya.**
-
-The distinction between induction and analogy: **Induction** moves from N specific cases → a general rule (vertical deepening); **analogy** moves from the structure of domain A → the structure of domain B (lateral transfer); both produce hypotheses rather than theorems and require subsequent verification.
-
-> **Mathematical Formalization**
->
-> The critical distinction between inductive reasoning and inductive proof:
-> - Inductive reasoning: empirical, proposing hypotheses from observations
-> - Inductive proof: logical, mathematical induction is a rigorous deductive proof scheme
-> - Key: inductive reasoning generates conjectures; inductive proof verifies them — the two must not be conflated
->
-> Variants of inductive proof and their applicable scenarios:
-> - **Weak Induction**: Prove P(1) holds + prove ∀k(P(k)→P(k+1)) → conclude ∀n P(n); applicable when P(k+1) depends only on P(k)
-> - **Strong / Complete Induction**: Assume P(1), P(2), …, P(k) all hold, prove P(k+1); applicable when P(k+1) depends on multiple predecessors
-> - **Structural Induction**: Prove P(base) holds + P(composite) follows from P(components); applicable to recursively defined structures (trees, lists, expressions)
-> - **Transfinite Induction**: For a well-ordered set (W, <), prove ∀α∈W(∀β<α P(β) → P(α)) → ∀α∈W P(α); applicable to well-ordered sets indexed beyond the natural numbers
->
-> Formal description of Lakatos's methodology:
-> - **Monster-barring**: Treating a counterexample as an object outside the intended scope of the hypothesis and explicitly excluding it — the simplest form of revision, but one must guard against over-exclusion that renders the hypothesis vacuous
-> - **Lemma-incorporation**: Incorporating the hidden condition exposed by the counterexample into the hypothesis as a new lemma — makes the hypothesis more precise
-> - **Proof-strategy revision**: The counterexample reveals a fundamental flaw in the original proof approach, requiring a different proof framework — the most profound form of revision
->
-> Structural similarity measure for analogies:
-> - Structural similarity = (number of successfully mapped relations + number of successfully mapped components) / (total source relations + total source components)
-> - High validity (homomorphism): most structures have correspondents, key relations are preserved
-> - Medium validity (partial homomorphism): some structures and relations are preserved, some are distorted
-> - Low validity (surface similarity): only appearance or terminology is similar; deep structures differ
-
-## GPU-Friendliness (Cross-Cutting Check)
-
-Induction and analogy are inherently **meta-heuristic** — they generate conjectures and transfer structures, but do not directly prescribe computations. What must actually pass the GPU eight-dimension gate is their **product**: the structure being transferred into algorithm/operator design. If the analogically transferred structure is intractable, it must be adapted.
-
-- **Friendly**: The transferred structure can be expressed as dense GEMM chains or low-rank linear maps — e.g., structural induction ideas mapped to fusable implementations of recursive operators, algebraic homomorphism analogies mapped to low-rank GEMM (dimensions 1/2/4).
-- **Adaptable**: Transferred global/serial structures (naive cohomology computations, unstructured graph traversals, naive RNN-style recurrence) → adapt via block sparsification, low-rank projection, or tropical semiring relaxation (see adaptation techniques in `../../references/gpu-friendly-math.md`).
-- **Anti-pattern**: Adopting a structure solely because it is isomorphic/homomorphic and elegant, while ignoring O(n²) memory, fp64 dependence, or long serial recurrences — "mathematical beauty ≠ tractability"; an analogical conclusion may be mathematically correct yet hardware-infeasible.
-
-Eight-dimension minimum criteria (formal terms): **Tensorization** checks batch-expressibility of the transferred structure; **GEMM-mappability** checks whether the core operator after analogy maps to matrix multiplication; **complexity** checks whether global constructions brought from the source domain explode; **memory and KV-cache** checks whether large tables or long caches are introduced; **low-precision stability** checks whether the analogical structure depends on exact symbolic or high-precision computation; **parallelism and communication** checks whether recurrences or interactions can be chunked; **sparse structure** checks whether sparsity is regular; **operator fusion** checks whether the transferred structure can be realized as a small number of fused ops.
-
-> Used together with `../../references/gpu-friendly-math.md` (eight-dimension gate), `../../references/books/abstract-algebra.md` (homomorphism/isomorphism structures), `../../references/books/algebraic-geometry-rising-sea.md` (cross-domain structure transfer and compression).
-
-## When NOT to Use
-
-- **Rigorous proof is required** — inductive reasoning and analogy can only produce hypotheses, not proofs; however, mathematical induction itself is a proof method.
-- **There is only one or very few samples** — inductive reasoning needs a sufficient number of instances for support.
-- **The two domains are too different** — analogy requires structural similarity, not mere surface resemblance.
-- **The conclusion must have zero error** — inductive reasoning carries uncertainty; analogy carries distortion risk.
-- **The problem can be solved directly by deduction** — there is no need to detour through induction.
-
-## When to Use
-
-- Multiple cases have been observed sharing a common pattern, and one wishes to propose a general hypothesis.
-- Attempting to discover new theorems or new regularities in research.
-- Properties of recursively defined structures (trees, lists, expressions) need to be proved — use structural induction.
-- Propositions involving decomposition (e.g., prime factorization of integers) need to be proved — use strong induction.
-- Arguments over well-ordered sets beyond the natural numbers are needed — use transfinite induction.
-- Cross-disciplinary borrowing — can methods from other fields be applied to one's own problem?
-- **Borrowing structures from other fields for algorithm/operator design** — cross-domain analogical transfer of structures (e.g., algebraic homomorphism → low-rank mapping, sheaf diffusion → block summary), passing the GPU eight-dimension gate before adoption.
-
-## Method
-
-### Step 1: Collect Concrete Cases
-Systematically collect and observe concrete instances: extract key cases from the literature, gather patterns from experimental data, and organize known theorems. Ensure cases cover boundary values, typical values, and extreme values. The diversity and coverage of cases determine the quality of induction — too few or too homogeneous cases will be misleading.
-
-### Step 2: Identify Patterns
-Compare commonalities and differences across cases: what features appear in all (or most) cases? What appears only in some? Are the differences themselves patterned? Be aware that patterns may suddenly break after a number of cases (as in the Borwein integral phenomenon) — focusing only on commonalities can cause one to miss counterexamples.
-
-### Step 3: Formulate a Hypothesis
-Based on the observed pattern, formulate a general hypothesis: **strong hypothesis** ∀n P(n) (holds in all cases); **weak hypothesis** P(n) holds for sufficiently large n (with exceptions); **conditional hypothesis** ∀n∈S P(n) (on a specific subset). Selection principle: try the strong hypothesis first, then weaken it stepwise with counterexamples. The hypothesis must be specific enough to be falsifiable.
-
-### Step 4: Search for Counterexamples
-Actively seek counterexamples that falsify the hypothesis — the hallmark distinguishing science from pseudoscience. If a counterexample is found, revise (weaken or add conditions); if none is found, the hypothesis stands provisionally. Pay special attention to boundary conditions, extreme values, and whether the pattern suddenly fails when parameters change.
-
-### Step 5: Attempt a Proof
-For mathematical hypotheses, attempt a rigorous proof; selecting the correct induction variant is key:
-
-- **Weak Induction**: Prove P(1) + ∀k(P(k)→P(k+1)) → ∀n P(n); applicable when P(k+1) depends only on P(k) (summation formulas, simple recurrences); example: 1+2+…+n = n(n+1)/2.
-- **Strong Induction**: Assume P(1)..P(k) all hold and prove P(k+1); applicable when P(k+1) depends on multiple predecessors (decomposition, partition propositions); example: every integer greater than 1 is a product of primes, Fibonacci properties. Key: when P(k+1) must invoke P(j) for j<k, strong induction is required.
-- **Structural Induction**: Prove P(base) + P(composite) follows from P(components); applicable to recursively defined structures (trees, lists, expressions, formulas, program semantics); example: the number of leaf nodes in a binary tree equals the number of internal nodes plus 1. It is the natural extension of mathematical induction to recursive structures.
-- **Transfinite Induction**: For a well-ordered set (W,<), prove ∀α∈W(∀β<α P(β)→P(α)) → ∀α∈W P(α); applicable to well-ordered sets indexed beyond the natural numbers (ordinals, partially ordered sets, topological transitive closures); example: applications of Zorn's lemma, recursive constructions in set theory.
-- **Well-ordering principle equivalence**: Every non-empty subset has a least element ↔ natural number induction ↔ strong induction ↔ transfinite induction on well-ordered sets. Practical implication: choose the most natural formulation.
-
-### Step 6: Analogical Transfer
-Using analogy requires quantitative evaluation rather than relying solely on intuition:
-
-- **Structural similarity measure**: Structural similarity = (number of successfully mapped relations + number of successfully mapped components) / (total source relations + total source components).
-- **Analogy validity grading**: **High validity (homomorphism)** — most structures have correspondents, key relations are preserved, conclusions have high credibility; **Medium validity (partial homomorphism)** — some are preserved and some are distorted, conclusions must be revised before use; **Low validity (surface similarity)** — only appearance/terminology is similar, deep structures differ, conclusions are essentially unreliable.
-- **Systematic verification checklist**: (i) Does every component in the source domain have a correspondent in the target domain? (ii) Does every relation hold in the target domain? (iii) Is every operation defined in the target domain? (iv) Mapping consistency: if a↦a', b↦b', does R(a,b) imply R'(a',b')?
-
-### Step 7: Hypothesis Revision
-Revise the hypothesis based on counterexamples and proof attempts, following Lakatos's methodology: **Monster-barring** — treating the counterexample as outside the intended scope and explicitly excluding it (simplest; guard against over-exclusion leading to vacuity); **Lemma-incorporation** — incorporating the hidden condition exposed by the counterexample as a new lemma (makes the hypothesis more precise); **Proof-strategy revision** — the counterexample reveals a fundamental flaw in the original proof approach, requiring a different proof framework (most profound). After revision, return to Step 4 and iterate until the hypothesis stabilizes.
-
-## Common Errors
-
-| Error | Critique | Correct Approach |
-|-------|----------|-----------------|
-| Induction from insufficient samples | "Patterns" induced from 2–3 cases are extremely unreliable | Sample size must be large enough and cover diverse cases |
-| Treating inductive reasoning as proof | Inductive reasoning only produces hypotheses, not proofs; mathematical induction is a proof method | Distinguish inductive reasoning from inductive proof: the former generates conjectures, the latter verifies them |
-| Confirmation bias | Only seeing cases that support the hypothesis while ignoring counterexamples | Actively search for counterexamples — the most critical step |
-| Surface analogy | Analogy based on surface similarity rather than structural similarity | The basis of analogy is structural isomorphism/homomorphism, not surface resemblance |
-| Overextending analogy | Two domains are not fully isomorphic; analogical conclusions are not fully valid | Quantitatively evaluate analogy strength using structural similarity |
-| Failing to verify after induction | Proposing a hypothesis without testing it | The hypothesis must be tested against counterexamples and/or proved |
-| Using weak induction when strong induction is needed | When P(k+1) depends on multiple predecessors, weak induction cannot complete the inductive step | Identify the dependency structure: if P(k+1) must invoke P(j) for j<k, use strong induction |
-| Evaluating analogy by surface features alone | Surface feature similarity does not imply structural similarity; conclusions may be entirely wrong | Use systematic mapping to compare components, relations, and operations one by one |
-| Conflating inductive reasoning with inductive proof | Inductive reasoning is empirical; inductive proof is deductive — they are fundamentally different in nature | Distinguish clearly: reasoning → hypothesis; proof → theorem |
-| Ignoring pattern disruption (Borwein integrals) | Patterns that hold for the first several cases may suddenly break down; inductive reasoning has blind spots | Check sufficiently many cases; watch for sudden pattern changes at boundaries and under parameter variations |
-| Transferring intractable structures via analogy | Adopting a structure solely because the isomorphism is elegant, while ignoring O(n²) memory or fp64 dependence — "mathematical beauty ≠ tractability" | Run the transferred structure through the GPU eight-dimension gate first; if intractable, adapt (low-rank / block-sparse / relaxation) before adoption |
-
-## Operating Procedure
-
-When this skill is triggered, the output must include:
-
-1. **Case inventory**: `[Case N]: [description] [key features]`, ensuring coverage of boundary and extreme values.
-2. **Pattern identification**: `[Common pattern]: [description] [frequency: N/M] [Difference pattern]: [description]`, noting whether Borwein-type sudden disruption occurs.
-3. **Hypothesis statement**: `[Hypothesis]: [content] [type: strong/weak/conditional]`, trying the strong hypothesis first.
-4. **Counterexample search**: `[Counterexample]: [found/not found]. If found: [description], hypothesis revision: [new statement]`.
-5. **Proof direction**: `[Induction type: weak/strong/structural/transfinite]` with a proof outline; explain why this variant was chosen.
-6. **Analogy mapping**: `[Source domain A] → [Target domain B]: [correspondence] [structural similarity: X/Y] [validity: high/medium/low]`, verifying components, relations, and operations item by item.
-7. **Hypothesis revision**: `[Revision method: monster-barring/lemma-incorporation/strategy revision] [Revised hypothesis]`.
-8. **[GPU feasibility]** (if the transferred structure is used for algorithm/operator design) — run the transferred structure through `../../references/gpu-friendly-math.md` eight-dimension gate, annotate as friendly / retrofittable / unfriendly + adaptation suggestions.
-
-**The output must not present analysis alone without a conclusion; when further execution is needed, state in one sentence at the end what will be done next.**
-
-## Relations to Other Skills
-
-- **Logical deduction**: Induction generates hypotheses; deduction proves them — the two wings of discovery and verification.
-- **Abstraction thinking**: Analogy is essentially identifying the common abstract structure of two domains.
-- **Probability and statistics**: The probabilistic version of inductive reasoning is statistical inference.
-- **Modeling thinking**: The inductive process of discovering patterns from data is also part of modeling.
-- **Algorithmic thinking**: Induction as a correctness proof paradigm — loop invariants are induction hypotheses, and the correctness of recursive programs relies on structural induction.
-- **Discrete / Combinatorial thinking**: Combinatorial induction — inductive arguments on combinatorial objects, such as graph-theoretic induction and inductive recurrences in generating function methods.
-- **Modern mathematics activation**: `../../references/books/abstract-algebra.md` (structural analogy: homomorphism/isomorphism as formalization of analogy), `../../references/books/algebraic-geometry-rising-sea.md` (cross-domain structure transfer: sheaves/restriction maps/Plücker compression).

package/skills/induction-analogy/SKILL.md

@@ -1,152 +0,0 @@
----
-name: induction-analogy
-description: |
-  触发：从数据/经验找规律、特殊→一般、已知→未知、数学归纳证明（弱/强/结构/超限）、跨域类比迁移结构、为算法/算子设计借鉴其他领域结构等科研语境时调用。
-  English: Trigger when finding patterns from data/experience, specific→general, known→unknown, mathematical induction proofs (weak/strong/structural/transfinite), cross-domain analogical transfer of structures, or borrowing structures from other fields for algorithm/operator design.
----
-
-> **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。
-
-# 📈 归纳与类比 / Induction & Analogy
-
-> "从特殊到一般，从已知到未知——归纳发现规律，类比迁移经验，两者都不等于证明"
-> "From the specific to the general, from the known to the unknown—induction discovers patterns, analogy transfers experience, but neither equals proof"
->
-> —— Pólya《数学与猜想》、Lakatos《证明与反驳》
-> —— Pólya "Mathematics and Plausible Reasoning", Lakatos "Proofs and Refutations"
-
-## 核心原则 / Core Principle
-
-**归纳是从有限实例中发现一般规律的过程，是创造新知识的引擎；类比是将一个领域的理解迁移到另一个领域的桥梁。两者都不是严格证明，但都是发现定理、提出假说、产生创新的根本方法。严格性是验证的工具，而非发现的工具——Pólya。**
-
-**Induction discovers general patterns from limited instances—the engine of new knowledge; analogy transfers understanding across domains. Neither is rigorous proof, but both are fundamental to discovering theorems, proposing hypotheses, and generating innovation. Rigor is a tool of verification, not of discovery—Pólya.**
-
-归纳与类比的区别：**归纳**从 N 个具体案例 → 一般规律（纵向深入）；**类比**从领域 A 的结构 → 领域 B 的结构（横向迁移）；两者都产生假设而非定理，需后续验证。
-
-> **数学形式化 / Mathematical Formalization**
->
-> 归纳推理与归纳证明的严格区分：
-> - 归纳推理（inductive reasoning）：经验性的，从观察中提出假说
-> - 归纳证明（inductive proof）：逻辑性的，数学归纳法是严格的演绎证明格式
-> - 关键：归纳推理产生猜想，归纳证明验证猜想——两者不可混淆
->
-> 归纳证明的变体及其适用场景：
-> - **弱归纳法（Weak Induction）**：证明 P(1) 成立 + 证明 ∀k(P(k)→P(k+1)) → 得出 ∀n P(n)；适用：P(k+1) 仅依赖 P(k)
-> - **强归纳法（Strong / Complete Induction）**：假设 P(1), P(2), …, P(k) 全部成立，证明 P(k+1) 成立；适用：P(k+1) 依赖多个前驱
-> - **结构归纳法（Structural Induction）**：证明 P(base) 成立 + P(composite) 由 P(components) 推出；适用：递归定义的结构（树、列表、表达式）
-> - **超限归纳法（Transfinite Induction）**：对良序集 (W, <)，证明 ∀α∈W(∀β<α P(β) → P(α)) → ∀α∈W P(α)；适用：索引超越自然数的良序集
->
-> Lakatos 方法论的形式化描述：
-> - **怪物排除（monster-barring）**：将反例归为不属于假说本意范围的对象，在假说中显式排除——最简单的修正，但需警惕过度排除导致假说空洞化
-> - **引理吸收（lemma-incorporation）**：将反例暴露的隐藏条件纳入假说作为新引理——使假说更精确
-> - **证明策略修订（proof-strategy revision）**：反例揭示原证明思路有根本缺陷，需换一种证明框架——最深刻的修正
->
-> 类比的结构性相似度度量：
-> - 结构相似度 = (成功映射的关系数 + 成功映射的组件数) / (源关系总数 + 源组件总数)
-> - 高有效性（同态映射 homomorphism）：大部分结构有对应，关键关系保持
-> - 中有效性（部分同态 partial homomorphism）：部分结构和关系保持，部分失真
-> - 低有效性（表面相似 surface similarity）：仅外观或术语相似，深层结构不同
-
-## GPU 友好性 / GPU-Friendliness（横切检查）
-
-归纳与类比本身偏**元启发式**——它们产生猜想、迁移结构，但不直接规定计算。真正要过 GPU 八维门的是它们的**产物**：被迁移到算法/算子设计中的那个结构。类比迁移来的结构若不可算则需改造。
-
-- **友好**：迁移来的结构能写成稠密 GEMM 链或低秩线性映射——如结构归纳思路落到递归算子的可融合实现、代数同态类比成低秩 GEMM（维度 1/2/4）。
-- **可改造**：迁移来的全局/串行结构（朴素上同调计算、非结构化图遍历、朴素 RNN 式递推）→ 用块稀疏化、低秩投影、热带半环松弛改造（见 `../../references/gpu-friendly-math.md` 改造手法）。
-- **反模式**：仅因结构同构/同态优美就采用，却忽略 O(n²) 显存、fp64 依赖或长串行递推——"数学美 ≠ 可算"，类比结论即便数学正确也未必硬件可行。
-
-八维最低判定（正式术语）：**张量化**检查迁移结构的可批量表达；**GEMM 可映射**检查类比后的核心算子是否落矩阵乘；**复杂度**检查从源领域带来的全局构造是否爆炸；**显存与 KV-Cache**检查是否引入大表或长缓存；**低精度稳定**检查类比结构是否依赖精确符号/高精度；**并行与通信**检查递推或交互能否分块；**稀疏结构**检查稀疏是否规则；**算子融合**检查迁移结构是否可实现为少量 fused ops。
-
-> 配合 `../../references/gpu-friendly-math.md`（八维门）、`../../references/books/abstract-algebra.md`（同态/同构结构）、`../../references/books/algebraic-geometry-rising-sea.md`（跨域结构迁移与压缩）。
-
-## 不适用场景 / When NOT to Use
-
-- **需要严格证明**——归纳推理和类比只能产生假设，不是证明；但数学归纳法本身是证明方法。
-- **样本只有一个或极少**——归纳推理需要足够的实例支撑。
-- **两个领域差异太大**——类比需要结构相似性，而非表面相似。
-- **结论必须零误差**——归纳推理有不确定性，类比有失真风险。
-- **问题可用演绎直接解决**——不必绕道归纳。
-
-## 何时使用 / When to Use
-
-- 观察到多个案例有共同模式，想要提出一般性假说。
-- 在科研中尝试发现新定理或新规律。
-- 需要对递归定义的结构（树、列表、表达式）进行性质证明——使用结构归纳法。
-- 需要证明涉及分解的命题（如整数的素因子分解）——使用强归纳法。
-- 需要对超越自然数的良序集进行论证——使用超限归纳法。
-- 跨学科借鉴——其他领域的方法能否用到我的问题？
-- **为算法/算子设计借鉴其他领域的结构**——跨域类比迁移结构（如代数同态→低秩映射、层扩散→块摘要），迁移后过 GPU 八维门再采用。
-
-## 方法流程 / Method
-
-### 第一步：收集具体案例 / Collect Concrete Cases
-系统收集与观察具体实例：阅读文献提取关键案例、收集实验数据中的模式、整理已知定理；确保案例覆盖边界值、典型值与极端值。案例的多样性与覆盖度决定归纳质量——太少或太单一会误导。
-
-### 第二步：识别模式 / Identify Patterns
-比较案例的共同点与差异：什么特征在所有（或大多数）案例中出现？什么只出现在部分？差异是否有规律？注意模式可能在若干案例后突然破坏（如 Borwein 积分现象）——只看共同点会忽视反例。
-
-### 第三步：提出假说 / Formulate a Hypothesis
-基于观察到的模式提出一般性假说：**强假说** ∀n P(n)（所有情况成立）；**弱假说** P(n) 对足够大的 n 成立（存在例外）；**条件假说** ∀n∈S P(n)（特定子集）。选择原则：优先尝试强假说，再用反例逐级弱化。假说必须足够明确以至于可被推翻。
-
-### 第四步：搜索反例 / Search for Counterexamples
-主动寻找推翻假说的反例——区分科学与伪科学的关键。找到反例则修改（弱化或加条件）；找不到则假说暂时成立。特别关注边界条件、极端值、参数变化时模式是否突然失效。
-
-### 第五步：尝试证明 / Attempt a Proof
-对数学假说尝试严格证明，选择正确的归纳变体是关键：
-
-- **弱归纳法（Weak Induction）**：证明 P(1) + ∀k(P(k)→P(k+1)) → ∀n P(n)；适用 P(k+1) 仅依赖 P(k)（求和公式、简单递推）；示例 1+2+…+n = n(n+1)/2。
-- **强归纳法（Strong Induction）**：假设 P(1)..P(k) 全部成立证 P(k+1)；适用 P(k+1) 依赖多个前驱（分解、拆分命题）；示例：每个大于 1 的整数都是素数乘积、斐波那契性质。关键：P(k+1) 需引用 P(j)(j<k) 时必须用强归纳。
-- **结构归纳法（Structural Induction）**：证 P(base) + P(composite) 由 P(components) 推出；适用递归定义结构（树、列表、表达式、公式、程序语义）；示例：二叉树叶节点数 = 内部节点数 + 1。它是数学归纳法在递归结构上的自然推广。
-- **超限归纳法（Transfinite Induction）**：对良序集 (W,<) 证 ∀α∈W(∀β<α P(β)→P(α)) → ∀α∈W P(α)；适用索引超越自然数的良序集（序数、偏序集、拓扑传递闭包）；示例：Zorn 引理应用、集合论递归构造。
-- **良序原理等价性**：每个非空子集有最小元 ↔ 自然数归纳法 ↔ 强归纳法 ↔ 良序集上的超限归纳。实践意义：选最自然的形式。
-
-### 第六步：类比迁移 / Analogical Transfer
-使用类比必须定量评估而非仅凭直觉：
-
-- **结构性相似度度量**：结构相似度 = (成功映射的关系数 + 成功映射的组件数) / (源关系总数 + 源组件总数)。
-- **类比有效性分级**：**高有效性（同态映射 homomorphism）**——大部分结构有对应、关键关系保持，结论可信度高；**中有效性（部分同态）**——部分保持部分失真，结论需修正后使用；**低有效性（表面相似）**——仅外观/术语相似，深层结构不同，结论基本不可信。
-- **系统性检验清单**：① 源领域每个组件，目标领域是否有对应？② 每个关系在目标领域是否成立？③ 每个操作在目标领域是否有定义？④ 映射一致性：若 a↦a', b↦b'，则 R(a,b) 是否蕴含 R'(a',b')？
-
-### 第七步：假说修正 / Hypothesis Revision
-据反例与证明尝试结果修正假说，遵循 Lakatos 方法论：**怪物排除（monster-barring）**——将反例归为不属于假说本意范围并显式排除（最简单，警惕过度排除致空洞化）；**引理吸收（lemma-incorporation）**——将反例暴露的隐藏条件纳入为新引理（使假说更精确）；**证明策略修订（proof-strategy revision）**——反例揭示原证明思路根本缺陷，换证明框架（最深刻）。修正后回到第四步，迭代直至假说稳定。
-
-## 常见错误 / Common Errors
-
-| 错误 / Error | 批评 / Critique | 正确做法 / Correct Approach |
-|---|---|---|
-| 样本不足就归纳 | 从 2-3 个案例归纳的"规律"极不可靠 | 样本量要足够大，且覆盖各种情况 |
-| 把归纳推理当证明 | 归纳推理只产生假设，不是证明；数学归纳法才是证明 | 区分归纳推理与归纳证明，前者产猜想后者验证 |
-| 确认偏差 | 只看到支持假说的案例，忽视反例 | 主动寻找反例——最关键的一步 |
-| 表面类比 | 基于表面相似而非结构相似的类比 | 类比的依据是结构同构/同态，不是表面相似 |
-| 类比过度 | 两个领域不完全同构，类比结论不完全成立 | 用结构性相似度量化评估类比强度 |
-| 归纳后不验证 | 提出假说后不去检验 | 假说必须经过反例检验和/或证明 |
-| 需强归纳却用弱归纳 | P(k+1) 依赖多个前驱时，弱归纳无法完成归纳步骤 | 识别依赖结构：若 P(k+1) 需引用 P(j)(j<k)，用强归纳 |
-| 仅凭表面特征评估类比 | 表面特征相似不意味着结构相似，结论可能完全错误 | 用系统映射逐一比较组件、关系、操作 |
-| 混淆归纳推理与归纳证明 | 归纳推理是经验性的，归纳证明是演绎性的——性质完全不同 | 明确区分：推理→假说，证明→定理 |
-| 忽略模式破坏（Borwein积分） | 模式在前若干案例中成立后可能突然失效，归纳推理有盲区 | 检查足够多案例，关注边界与参数变化时的模式突变 |
-| 类比迁移来不可算的结构 | 仅因结构同构优美就采用，忽略 O(n²) 显存或 fp64 依赖——"数学美 ≠ 可算" | 迁移来的结构先过 GPU 八维门，不可算则改造（低秩/块稀疏/松弛）后再采用 |
-
-## 操作规程 / Operating Procedure
-
-当本 skill 被触发时，输出必须包含：
-
-1. **案例清单**：`[案例N]: [描述] [关键特征]`，确保覆盖边界值与极端值。
-2. **模式识别**：`[共同模式]: [描述] [出现频率: N/M] [差异模式]: [描述]`，注意是否有 Borwein 型突然破坏。
-3. **假说表述**：`[假说]: [内容] [类型: 强/弱/条件]`，优先尝试强假说。
-4. **反例搜索**：`[反例]: [找到/未找到]。如找到: [描述]，假说修正: [新表述]`。
-5. **证明方向**：`[归纳类型: 弱/强/结构/超限]` 并给出证明思路；说明为何选择该变体。
-6. **类比映射**：`[源领域A] → [目标领域B]: [对应关系] [结构相似度: X/Y] [有效性: 高/中/低]`，逐项检验组件、关系、操作。
-7. **假说修正**：`[修正方式: 怪物排除/引理吸收/策略修订] [修正后假说]`。
-8. **[GPU 可行性]**（若迁移结构用于算法/算子设计）— 迁移来的结构过 `../../references/gpu-friendly-math.md` 八维门，标注友好/可改造/不友好 + 改造建议。
-
-**输出不得只给分析而无结论；需要继续执行时，在结尾用一句话说明接下来将做什么。**
-
-## 与其他 skill 的关系 / Relations to Other Skills
-
-- **逻辑演绎**：归纳产生假说，演绎证明假说——发现与验证的双翼。
-- **抽象化思想**：类比本质上是识别两个领域的共同抽象结构。
-- **概率与统计**：归纳推理的概率版本就是统计推断。
-- **建模思想**：从数据中发现规律的归纳过程也是建模的一部分。
-- **算法思想**：归纳作为正确性证明范式——循环不变量即归纳假设，递归程序正确性依赖结构归纳。
-- **离散/组合思想**：组合归纳——对组合对象的归纳论证，如图论归纳、生成函数方法中的归纳递推。
-- **现代数学激活**：`../../references/books/abstract-algebra.md`（结构类比：同态/同构作为类比的形式化）、`../../references/books/algebraic-geometry-rising-sea.md`（跨域结构迁移：层/限制映射/Plücker 压缩）。