math-skill 2.1.0 → 3.0.0

package/skills/induction-analogy/original-texts.en.md

@@ -1,140 +0,0 @@
-# Mathematical Sources and Classic Texts
-
-## Mathematical Induction (Pascal, 1654)
-
-**Principle**:
-> If P(1) holds and P(n) -> P(n+1), then P(n) holds for all positive integers n.
-
-**Philosophical implication**: Deriving the "infinite" from the "finite" -- requiring only a base case and an inductive step.
-
-**Classic examples**: 1+2+...+n = n(n+1)/2; Fibonacci properties; proofs in graph theory
-
-## Strong (Complete) Induction
-
-> If P(1)...P(k) all holding implies P(k+1), then P(n) holds for all positive integers n.
-
-**Applicable scenarios**: When P(k+1) depends on many or even all preceding propositions, not just P(k).
-
-**Examples**: Unique factorization theorem; Fibonacci recurrence F(n)=F(n-1)+F(n-2); Fundamental Theorem of Arithmetic
-
-## Structural Induction
-
-> For recursively defined structures (lists, trees, expressions): prove the property holds for base elements and is preserved under construction steps, then it holds for all instances.
-
-**Applications in computer science**: Compiler correctness; type system safety (well-typed programs do not crash); data structure invariants (red-black tree balance); algorithm termination proofs
-
-**Examples**: |L1 ++ L2| = |L1| + |L2|; tree height <= number of nodes
-
-## Transfinite Induction (Cantor Ordinals)
-
-> If P(gamma) for all gamma < beta implies P(beta), then P(alpha) holds for all ordinals alpha.
-
-**Significance**: Induction extends beyond the natural numbers -- it applies to any well-ordered set.
-
-**Key concepts**: Cantor ordinals omega, omega+1, omega*2, omega^2, omega^omega...; equivalence with Zorn's lemma; application: proving that every vector space has a basis
-
-## Well-Ordering Principle
-
-> Every nonempty subset of the natural numbers has a least element; logically equivalent to mathematical induction.
-
-**Equivalence proofs**: Well-ordering implies induction: the set of counterexamples has a least element m, so P(m-1) holds while P(m) does not -- contradiction. Induction implies well-ordering: use induction to prove that every nonempty subset contains a least element.
-
-**Generalization**: The Well-Ordering Theorem (Zermelo 1904) -- every set can be well-ordered; equivalent to the Axiom of Choice.
-
-## Polya's *Mathematics and Plausible Reasoning* (1954)
-
-> "Induction, analogy, specialization, generalization -- principal methods of discovering mathematical truths."
-
-**Inductive pattern**: 1. Observe cases -> 2. Discover patterns -> 3. Formulate conjectures -> 4. Verify (prove or disprove)
-
-## Mathematical Status of Analogy
-
-> "Analogy is a tool for discovery, not for proof."
-
-**Classic analogies**: Sound waves -> wave theory of light; Planetary orbits -> Bohr atomic model; Water flow -> electric current (voltage analogous to water pressure)
-
-## Euler and the Basel Problem (1735)
-
-> Euler discovered Sum(1/n^2) = pi^2/6 by inductive reasoning -- the most celebrated inductive discovery in mathematics.
-
-**Inductive process**: Computed partial sums approximately 1.6449 -> recognized pi^2/6 -> drew an analogy between the infinite product of sin(x) and the root-coefficient relations of polynomials -> rigorous proof. Polya specifically used this to illustrate "plausible reasoning."
-
-## Fermi Estimation
-
-> Solve unknowns via analogy and order-of-magnitude estimation.
-
-**Classic problem**: "How many piano tuners are there in Chicago?" Population ~3 million -> ~1 million households -> 1/20 own pianos -> ~50,000 pianos -> tuned once per year, 2 hours each -> tuner works 2000 hours/year -> approximately 50 tuners.
-
-**Value**: Reasonable analogical estimates yield order-of-magnitude correct results.
-
-## Mill's Five Methods (1843)
-
-> Mill systematized inductive reasoning in *A System of Logic*.
-
-1. **Method of Agreement**: Multiple instances share only one common circumstance -> that circumstance is likely the cause
-2. **Method of Difference**: Two instances differ in only one circumstance -> that difference is likely the cause
-3. **Joint Method**: Combining agreement and difference to strengthen credibility
-4. **Method of Concomitant Variations**: A phenomenon varies as a certain circumstance varies -> causal relationship
-5. **Method of Residues**: After subtracting the effects of known causes, remaining effects are explained by remaining causes
-
-## Ramanujan: Intuitive Induction (1887-1920)
-
-> Ramanujan discovered formulas via astonishing pattern recognition, often without formal proof -- a genius of intuitive induction.
-
-**Examples**: Series for 1/pi extrapolated from a few terms to general form; mock theta functions (pattern described, proof delayed by decades). Hardy: "His intuition was almost supernatural, yet logical rigor was often lacking."
-
-**Lesson**: Inductive discovery can far outpace proof -- the tension between intuition and rigor is a driving force of mathematical progress.
-
-## Borwein Integrals: A Cautionary Tale (2000s)
-
-> A pattern holds for many terms then suddenly breaks -- a profound warning about inductive reasoning.
-
-**Phenomenon**: Integral of sin(x)/x dx = pi/2; adding sin(x/3) still gives pi/2; ... continues through sin(x/13); but upon adding sin(x/15) it suddenly deviates! Mathematical root cause: the Patel-Vitali condition.
-
-**Lesson**: The first N cases may hold while case N+1 breaks the pattern. Inductive conclusions require theoretical support and cannot rely solely on empirical evidence.
-
-## Lakatos's *Proofs and Refutations* (1963/1976)
-
-> Mathematical discovery evolves through counterexamples, monster-barring, and lemma-incorporation -- not a linear "conjecture -> proof" process.
-
-**Core concepts**: Monster-barring (modifying definitions to exclude counterexamples); Lemma-incorporation (adding the conditions of counterexamples to the premises); Concept-stretching (counterexamples driving the expansion of concepts)
-
-**Example**: The proof history of Euler's polyhedron formula V-E+F=2 -- from Cauchy's naive proof to successive corrections for holes and tunnels.
-
-## Hume's Problem of Induction (1739)
-
-> "The sun has risen every day in the past; this does not logically guarantee it will rise tomorrow."
-
-**Fundamental flaw**: Deriving the infinite from the finite always carries uncertainty.
-
-**Scientific method's response**: Induction produces hypotheses (not theorems) -> falsifiable (Popper) -> continuously tested -> counterexamples lead to revision or abandonment
-
-## Carnap's Inductive Logic (1950s)
-
-> Carnap attempted to formalize and quantify inductive reasoning -- a logical foundation for inductive probability.
-
-**Core idea**: Confirmation function c(h,e) -- the degree to which evidence e confirms hypothesis h. Inductive logic should have rigorous rules; prior probabilities should be based on logical symmetry rather than subjective belief.
-
-**Limitations**: The system is highly sensitive to the linguistic framework -- different predicate sets yield different degrees of confirmation; Goodman's "grue paradox" further reveals the difficulties.
-
-## Problem of Underdetermination
-
-> Multiple analogies or theories can equally fit the same data -- data underdetermines theory choice.
-
-**Mathematical analogy**: Finitely many data points can be fit by infinitely many functions; multiple inductive hypotheses may be simultaneously consistent with experience; theory choice requires additional criteria: simplicity, explanatory power, predictive power.
-
-**Philosophical roots**: The Quine-Duhem thesis -- any hypothesis can be shielded from falsification by adjusting auxiliary hypotheses.
-
-## Machine Learning as Induction
-
-> Machine learning is the engineering of inductive reasoning -- inferring general rules from data.
-
-**Statistical learning theory**: PAC learning (Valiant 1984) -- outputting a hypothesis with low error with high probability; generalization bounds -- the quantitative relationship between sample complexity and hypothesis class complexity; VC dimension -- a mathematical measure of "inductive capacity."
-
-**Contrast with mathematical induction**: Mathematical induction = deterministic reasoning over an infinite domain; statistical induction = probabilistic reasoning from finite samples; generalization = the inductive leap
-
-## Power and Limits of Analogy
-
-**Power**: Transferring knowledge across domains (one of the most important sources of innovation); converting the unknown into the known; generating new ideas and hypotheses
-
-**Limits**: Analogy is not proof -- conclusions require independent verification; surface similarity does not imply structural similarity; every analogy has a breaking point -- two domains can never be perfectly isomorphic

package/skills/induction-analogy/original-texts.md

@@ -1,154 +0,0 @@
-# 数学出处与经典文献 / Mathematical Sources and Classic Texts
-
-## 数学归纳法 / Mathematical Induction (Pascal, 1654)
-
-**原理**：
-> 如果 P(1) 成立，且 P(n) → P(n+1)，则 P(n) 对所有正整数 n 成立。
-> If P(1) holds and P(n) → P(n+1), then P(n) holds for all positive integers n.
-
-**哲学含义**：从"有限"推导"无限"——只需基础与递推两步 / Deriving "infinite" from "finite" — just base case and inductive step.
-
-**经典案例**：1+2+...+n = n(n+1)/2；Fibonacci 性质；图论证明
-
-## 强归纳法 / Strong (Complete) Induction
-
-> 若 P(1)...P(k) 全部成立可推出 P(k+1)，则 P(n) 对所有正整数成立。
-> If P(1)...P(k) all holding implies P(k+1), then P(n) holds for all n.
-
-**适用场景**：P(k+1) 依赖多个甚至全部前命题，而非仅 P(k) / When P(k+1) depends on many or all preceding propositions, not just P(k).
-
-**案例**：唯一分解定理 / Unique factorization；Fibonacci 递推 F(n)=F(n-1)+F(n-2)；算术基本定理
-
-## 结构归纳法 / Structural Induction
-
-> 对递归定义结构（列表、树、表达式），证明性质对基础元素成立且在递推步骤中保持，则对所有实例成立。
-> For recursively defined structures: prove property holds for base elements and is preserved under construction steps → holds for all instances.
-
-**计算机科学应用**：编译器正确性；类型系统安全性（良类型程序不崩溃）；数据结构不变量（红黑树平衡）；算法终止性证明
-
-**案例**：|L₁ ++ L₂| = |L₁| + |L₂|；树高度 ≤ 节点数
-
-## 超限归纳法 / Transfinite Induction (Cantor Ordinals)
-
-> 若对所有 γ < β，P(γ) 成立可推出 P(β) 成立，则 P(α) 对所有序数 α 成立。
-> If P(γ) for all γ < β implies P(β), then P(α) holds for all ordinals α.
-
-**意义**：归纳超越自然数——对任意良序集成立 / Induction extends beyond naturals — applies to any well-ordered set.
-
-**关键概念**：Cantor 序数 ω, ω+1, ω·2, ω², ω^ω...；Zorn 引理等价性；应用：证明每个向量空间都有基
-
-## 良序原理 / Well-Ordering Principle
-
-> 自然数的每个非空子集有最小元素，与数学归纳法逻辑等价。
-> Every nonempty subset of naturals has a least element; logically equivalent to mathematical induction.
-
-**等价论证**：良序→归纳：反例集有最小元 m，则 P(m-1) 成立而 P(m) 不成立，矛盾。归纳→良序：用归纳法证每个非空子集含最小元。
-
-**推广**：良序定理（Zermelo 1904）——任何集合可良序化，与选择公理等价。
-
-## Pólya《数学与猜想》/ *Mathematics and Plausible Reasoning* (1954)
-
-> "归纳、类比、特例化、一般化——发现数学真理的主要方法。"
-> "Induction, analogy, specialization, generalization—principal methods of discovering mathematical truths."
-
-**归纳模式**：1. 观察案例 → 2. 发现模式 → 3. 提出猜想 → 4. 验证（证明或证伪）
-
-## 类比的数学地位 / Mathematical Status of Analogy
-
-> "类比是发现的工具，不是证明的工具。" / "Analogy is a tool for discovery, not for proof."
-
-**经典类比**：声波→光波动理论；行星轨道→Bohr 原子模型；水流→电流（电压≈水压）
-
-## Euler 与 Basel 问题 / Euler and the Basel Problem (1735)
-
-> Euler 通过归纳发现 ∑1/n² = π²/6 ——数学史上最著名的归纳发现。
-> Euler discovered ∑1/n² = π²/6 by inductive reasoning — the most celebrated inductive discovery in mathematics.
-
-**归纳过程**：计算部分和 ≈1.6449 → 识别 π²/6 → 类比 sin(x) 无穷乘积与多项式根系数关系 → 严格证明。Pólya 专门以此说明"合情推理"。
-
-## Fermi 方法 / Fermi Estimation
-
-> 通过类比和数量级估计解决未知问题 / Solve unknowns via analogy and order-of-magnitude estimation.
-
-**经典问题**："芝加哥有多少钢琴调音师？"：人口~300万→~100万户→1/20有钢琴→~5万台→每年调1次×2小时→调音师2000小时/年→≈50名调音师
-
-**价值**：合理的类比估计可得数量级正确答案 / Reasonable analogy yields order-of-magnitude correct results.
-
-## Mill 五种归纳方法 / Mill's Five Methods (1843)
-
-> Mill 在《逻辑体系》中系统化归纳推理 / Mill systematized inductive reasoning in *A System of Logic*.
-
-1. **一致法 / Agreement**：多场合仅一共同情况→该情况可能为原因
-2. **差异法 / Difference**：两场合仅一情况不同→该差异可能为原因
-3. **联合法 / Joint Method**：一致+差异结合，增强可信度
-4. **共变法 / Concomitant Variations**：现象随某情况变化而变化→因果关系
-5. **残差法 / Residue**：扣除已知原因效应后，剩余效应由剩余原因解释
-
-## Ramanujan：直觉归纳 / Ramanujan: Intuitive Induction (1887–1920)
-
-> Ramanujan 以惊人模式识别发现公式，常无形式证明——直觉归纳天才。
-> Ramanujan discovered formulas via astonishing pattern recognition, often without formal proof — genius of intuitive induction.
-
-**案例**：1/π 级数从少数项推广到一般形式；mock theta 函数（模式描述了，证明延迟数十年）。Hardy："他的直觉近乎超自然，但逻辑常缺。"
-
-**教训**：归纳发现可远超前于证明——直觉与严格性的张力是数学发展的动力。
-
-## Borwein 积分：归纳警醒 / Borwein Integrals: Cautionary Tale (2000s)
-
-> 模式持续很久然后突然断裂——归纳推理的深刻警示。
-> A pattern holds for many terms then suddenly breaks — a profound warning about inductive reasoning.
-
-**现象**：∫sin(x)/x dx = π/2；加入 sin(x/3) 仍 = π/2；...持续到 sin(x/13)；但加入 sin(x/15) 时突然偏离！数学根源：Patel-Vital 条件。
-
-**教训**：前 N 个案例成立，第 N+1 个可能打破模式。归纳结论需理论支撑，不能仅依赖经验 / First N cases may hold while case N+1 breaks. Inductive conclusions need theoretical support.
-
-## Lakatos《证明与反驳》/ *Proofs and Refutations* (1963/1976)
-
-> 数学发现不是线性"猜想→证明"，而是通过反驳、怪物排除、引理嵌入不断演进。
-> Mathematical discovery evolves through counterexamples, monster-barring, and lemma-incorporation — not linear "conjecture→proof".
-
-**核心概念**：怪物排除 / Monster-barring（修改定义排除反例）；引理嵌入 / Lemma-incorporation（将反例条件加入前提）；概念拉伸 / Concept-stretching（反例推动概念扩展）
-
-**案例**：多面体 Euler 公式 V-E+F=2 的证明生成——从 Cauchy 朴素证明到对空洞、隧道的逐步修正。
-
-## Hume 归纳问题 / Hume's Problem of Induction (1739)
-
-> "过去太阳每天升起，不能逻辑保证明天也会升起。"
-> "The sun has risen every day in the past; this does not logically guarantee it will rise tomorrow."
-
-**本质缺陷**：从有限到无限永远存在不确定性 / Deriving infinite from finite always carries uncertainty.
-
-**科学方法解决**：归纳产假说（非定理）→ 可证伪（Popper）→ 持续检验 → 反例则修正/放弃
-
-## Carnap 归纳逻辑 / Carnap's Inductive Logic (1950s)
-
-> Carnap 试图形式化量化归纳推理——为归纳概率建立逻辑基础。
-> Carnap attempted to formalize and quantify inductive reasoning — a logical foundation for inductive probability.
-
-**核心**：确证函数 c(h,e)——证据 e 对假说 h 的确证度 / Confirmation function c(h,e): degree of confirmation. 归纳逻辑应有严格规则；先验概率基于逻辑对称性而非主观信念。
-
-**局限**：系统对语言框架高度敏感——不同谓词集导致不同确证度；Goodman "绿蓝悖论"(grue paradox) 进一步揭示困难。
-
-## 不充分决定 / Problem of Underdetermination
-
-> 多个类比或理论可同等吻合同一数据——数据不充分决定理论选择。
-> Multiple analogies or theories equally fit the same data — data underdetermines theory choice.
-
-**数学类比**：有限数据点→无数函数可拟合；多个归纳假说可能同时与经验一致；理论选择需额外准则：简洁性、解释力、预测力。
-
-**哲学根源**：Quine-Duhem 论题——任何假说可通过调整辅助假说免于证伪。
-
-## 机器学习作为归纳 / Machine Learning as Induction
-
-> 机器学习是归纳推理的工程化——从数据推断一般规律。
-> Machine learning is the engineering of inductive reasoning — inferring general rules from data.
-
-**统计学习理论**：PAC 学习（Valiant 1984）——以高概率输出低误差假说；泛化界限——样本复杂度与假说类复杂度定量关系；VC 维度——衡量"归纳能力"的数学指标。
-
-**与数学归纳对比**：数学归纳=无限域确定性推理；统计归纳=有限样本概率性推理；泛化=归纳跳跃
-
-## 类比思维的力量与局限 / Power and Limits of Analogy
-
-**力量**：跨领域迁移知识（最重要创新来源之一）；将未知转化为已知；产生新思路和假说
-
-**局限**：类比不是证明——结论需独立验证；表面相似 ≠ 结构相似；每个类比有失效点——两个领域不可能完全同构

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

@@ -1,134 +0,0 @@
----
-name: information-theory
-description: |
-  Trigger when a problem needs to quantify uncertainty, measure information value as "information = reduction of uncertainty"; or compute entropy / mutual information / KL divergence / channel capacity, or design information criteria for compression / KV-cache compression / quantization / routing.
----
-
-# 📡 Information Theory
-
-> "Information is reduction of uncertainty — knowing more means doubting less."
->
-> — Information Theory, Coding Theory, Statistical Inference
-
-## Core Principle
-
-**Information is reduction of uncertainty — uncertainty is precisely quantified by entropy; compression and communication have unbreakable limits set by the entropy bound and channel capacity; information gain guides which observation to prioritize under uncertainty.**
-
-> **Mathematical Formalization**
->
-> Shannon entropy $H(X) = -\sum_x p(x)\log p(x)$ quantifies the average "surprise" of a random variable — the lower the probability of an event, the greater the surprise when it occurs; entropy is the expected value of surprise.
->
-> Mutual information $I(X;Y) = H(X) - H(X|Y) = \sum_{x,y} p(x,y)\log\frac{p(x,y)}{p(x)p(y)}$ measures the reduction in uncertainty about $X$ after observing $Y$ — this is precisely the mathematical definition of "information."
->
-> KL divergence $D_{KL}(P\|Q) = \sum_x p(x)\log\frac{p(x)}{q(x)}$ measures the information loss when $Q$ is used in place of $P$; it is **asymmetric**: $D(P\|Q)\neq D(Q\|P)$.
->
-> Source coding theorem: The average code length of the optimal compression satisfies $\ge H(X)$ bits/symbol; compression below the entropy bound inevitably loses information.
->
-> Channel coding theorem: The upper bound on the rate of reliable communication is the channel capacity $C = \max_{p(x)} I(X;Y)$; when $R<C$, there exist codes that drive the error rate to zero; when $R>C$, reliable communication is impossible.
->
-> Rate-distortion function $R(D) = \min_{p(\hat{x}|x):\,\mathbb{E}[d(x,\hat{x})]\le D} I(X;\hat{X})$ gives the minimum information rate for a given distortion $D$ — the fundamental limit of lossy compression.
->
-> See `original-texts.md` for detailed mathematical foundations.
-
-## GPU-Friendliness (Cross-Cutting Check)
-
-When information-theoretic quantities are used for **compression/pruning/quantization/KV-cache compression/routing design**, they must pass the `../../references/gpu-friendly-math.md` eight-dimension gate. Core criterion: **local, fusible entropy/KL estimation = friendly; global exact information estimation = unfriendly**.
-
-- **Quantization calibration (per-block entropy/Hessian)**: Local, batchable via GEMM, feasible at low precision — friendly (dimensions 1/2/5).
-- **KV-Cache compression (information bottleneck/block summary)**: Low-rank block summary compresses memory — friendly (dimension 4); Plücker-style block summaries are discussed in `../../references/books/algebraic-geometry-rising-sea.md`.
-- **Mutual information pruning / feature selection**: Adaptable if $I(X;Y)$ is estimated using local or low-rank approximations; unfriendly if full-distribution exact estimation is required.
-- **Information gain routing**: Global exact $I(X;Y)$ requires $O(n^2)$ memory, high precision, and is unfusable — anti-pattern; replace with continuous differentiable approximations such as softmax/tropical gating (dimensions 3/6/8).
-- **Anti-pattern**: Computing exact entropy/mutual information over the full token distribution — memory explosion, requires fp64, serial dependencies — "beautiful but intractable."
-
-Eight-dimension minimum criteria (formal terms): **Tensorization** checks whether entropy/KL/MI can be estimated per token/block/batch; **GEMM-mappability** checks whether compression, projection, and calibration reduce to linear algebra; **complexity** avoids full-distribution exact estimation; **memory and KV-cache** quantifies KV/activation/codebook footprint; **low-precision stability** checks log/softmax/KL dynamic range; **parallelism and communication** checks whether block-wise statistics can be reduced across devices; **sparse structure** checks whether codebooks/routing are block-structured; **operator fusion** checks whether statistics, masking, and quantization can be fused.
-
-> Used together with `../../references/books/matrix-analysis.md` (low-rank compression), `../../references/books/abstract-algebra.md` (coding/finite fields).
-
-## When NOT to Use
-
-- **The problem has no probabilistic structure** (e.g., purely symbolic reasoning, logical deduction) — entropy and information gain require probability distributions; without probability there is no information theory.
-- **Purely deterministic scenarios with no uncertainty** (e.g., mathematical problems with known exact answers) — when entropy is zero, information theory degenerates to trivial conclusions.
-- **Qualitative judgments that need no quantification** (e.g., aesthetic evaluations, emotional judgments) — information theory quantifies uncertainty in the probabilistic sense, not semantic ambiguity.
-
-## When to Use
-
-- Measuring the magnitude of uncertainty (entropy $H(X)$ quantifies the "disorder" of a random variable).
-- Comparing the value of different information sources (mutual information $I(X;Y)$ measures which observation $Y$ best reduces uncertainty about $X$).
-- Achieving optimal data compression (the source coding theorem guarantees the optimal compression limit is $H(X)$ bits/symbol).
-- Communicating reliably under noise (the channel coding theorem guarantees reliable transmission is feasible when $R<C$).
-- Feature selection or model selection (mutual information screens features; AIC/BIC/MDL serve as information criteria for model selection).
-- Bayesian model comparison (KL divergence $D(P\|Q)$ measures the information distance between distributions; Bayes factors quantify the evidence ratio between models).
-- **Designing information criteria for compression / KV-cache compression / quantization / routing**, and evaluating their GPU feasibility.
-
-## Method
-
-### Step 1: Identify Source and Uncertainty
-- What is the **random variable $X$**? — Define the information source and clarify the object of uncertainty.
-- The **probability distribution $p(x)$** — use a probability table for discrete distributions, a density function for continuous distributions.
-- **Compute $H(X) = -\sum p(x)\log p(x)$** — quantify the current level of uncertainty.
-- **Identify the uncertainty to be reduced** — clarify "knowing what would reduce uncertainty?"
-
-### Step 2: Quantify Information Gain
-- **Compute conditional entropy $H(X|Y)$** — the residual uncertainty about $X$ after observing $Y$.
-- **Compute mutual information $I(X;Y) = H(X) - H(X|Y)$** — the amount of information $Y$ provides about $X$.
-- **Identify the optimal observation** — which $Y$ maximizes $I(X;Y)$? That observation is the most worth acquiring.
-- **Chain rule** — $H(X_1,\dots,X_n) = H(X_1) + H(X_2|X_1) + \dots + H(X_n|X_1,\dots,X_{n-1})$, decomposing joint uncertainty variable by variable.
-
-### Step 3: Choose Coding Strategy
-- **Source coding (compression)**: Huffman coding (greedy optimal prefix code, average length approaching $H(X)$), arithmetic coding (closer to the entropy bound), universal coding (LZ77/LZ78/LZW, no prior distribution needed).
-- **Channel coding (error correction)**: Hamming codes (minimum distance 3, corrects 1-bit errors), Reed–Solomon codes (burst error correction), LDPC/Turbo codes (approaching the Shannon limit).
-- **Coding selection principle**: Compression needs → source coding → approach $H(X)$; noise protection needs → channel coding → approach the Shannon limit ($R\to C$).
-
-### Step 4: Evaluate Channel Capacity
-- **Compute channel capacity $C = \max_{p(x)} I(X;Y)$** — maximize mutual information over all input distributions.
-- **Compare transmission rate $R$ with capacity $C$**: $R<C$ → reliable communication is feasible; $R>C$ → errors are inevitable.
-- **Noise models**: BSC (binary symmetric channel, flip probability $p$), BEC (binary erasure channel, erasure probability $\varepsilon$), AWGN (additive white Gaussian noise channel).
-- **Capacity formula examples**: BSC capacity $C = 1 - H(p)$; AWGN capacity $C = \frac{1}{2}\log(1 + S/N)$.
-
-### Step 5: Apply Information Criteria
-- **AIC (Akaike Information Criterion)**: $\text{AIC} = -2\ln L + 2k$ — favors goodness of fit, suited for prediction objectives.
-- **BIC (Bayesian Information Criterion)**: $\text{BIC} = -2\ln L + k\ln n$ — favors parsimony, suited for explanation objectives.
-- **KL divergence $D_{KL}(P\|Q) = \sum p(x)\log\frac{p(x)}{q(x)}$** — measures the information loss when $Q$ is used in place of $P$; note the asymmetry.
-- **MDL principle (Minimum Description Length)**: Choose the model that minimizes "data description length + model description length" — the information-theoretic version of Occam's razor.
-
-### Step 6: Make Information-Optimal Decision
-- **Bayesian experimental design**: Choose the experiment that maximizes expected information gain $\max\,\mathbb{E}[I(\theta;Y)]$ — prioritize acquiring the data that most reduces uncertainty.
-- **Minimize KL divergence**: The output distribution $Q$ of a decision should be as close as possible to the target distribution $P$, i.e., $\min D(P\|Q)$.
-- **Maximum entropy principle**: Under known constraints, choose the distribution that maximizes $H(X)$ — fewest assumptions, most conservative inference.
-- **Information bottleneck**: $\min I(X;T) - \beta I(T;Y)$ — when compressing $X$ into $T$, retain the maximum relevant information about $Y$.
-
-## Common Errors
-
-| Error | Critique | Correct Approach |
-|-------|----------|-----------------|
-| Equating information with bits rather than probabilistic reduction | Entropy $H(X)$ is a function of the probability distribution; bits are merely the unit of measurement | Understand information as reduction of uncertainty: $I(X;Y)=H(X)-H(X\|Y)$ |
-| Ignoring channel capacity limits | When $R>C$, no coding scheme achieves reliable communication | Compute capacity $C=\max I(X;Y)$ and ensure $R<C$ |
-| Confusing entropy with variance | Entropy measures the "spread" of the probability structure; variance measures the "spread" of numerical values; they are not equivalent | Use entropy for probabilistic uncertainty; use variance for numerical deviation; continuous entropy can be negative |
-| Over-compressing below the entropy bound | Optimal compression satisfies $\ge H(X)$ bits/symbol; compression below this bound inevitably loses information | Accept the entropy bound and design codes accordingly |
-| Ignoring the asymmetry of KL divergence | $D(P\|Q)\ne D(Q\|P)$; the direction determines the meaning | Be explicit about direction: $D(P\|Q)$ means "the extra cost of encoding $P$ using $Q$" |
-| Equating correlation with information | Correlation $\rho$ measures only linear association; $I(X;Y)=0 \Leftrightarrow$ independence, but $\rho=0 \not\Rightarrow$ independence | Use mutual information $I(X;Y)$ to assess dependence |
-| Forcing quantification on qualitative judgments | Information theory quantifies probabilistic uncertainty, not semantic ambiguity or subjective experience | Distinguish probabilistic uncertainty from semantic ambiguity; do not apply entropy formulas to qualitative problems |
-| Global exact entropy / mutual information is intractable | Exact estimation of $I(X;Y)$ over the full distribution requires $O(n^2)$ memory and high precision; infeasible on GPU | Use local/sampled/low-rank estimation; pass the GPU eight-dimension gate |
-
-## Operating Procedure
-
-When this skill is triggered, the output must include:
-
-1. **[Information source]: [description]** $H(X)=$ [value] — define the random variable $X$, compute its entropy, and quantify the current uncertainty.
-2. **[Information gain]: [description]** $I(X;Y)=$ [value] — compute mutual information and identify the most valuable observation $Y$.
-3. **[Coding strategy]: [choice]** — source coding (compression) or channel coding (error correction), stating which limit is being approached.
-4. **[Channel capacity]: [description]** $C=$ [value] — compute channel capacity and compare transmission rate $R$ with $C$.
-5. **[Information criterion]: [AIC/BIC/KL/MDL]** — state the chosen information criterion and the rationale.
-6. **[Optimal decision]: [explanation]** — decision recommendation based on information gain maximization or KL divergence minimization.
-7. **[GPU feasibility]** (if used for compression/KV-cache compression/quantization/routing) — is entropy/KL/mutual information estimation local and fusible, or global and exact? Pass the eight-dimension gate; annotate as friendly / retrofittable / unfriendly + adaptation suggestions.
-
-**The output must not present analysis alone without a conclusion.**
-
-## Relations to Other Skills
-
-- **Probability and statistics**: Entropy and information gain complement probabilistic reasoning — $H(X)$ measures distributional uncertainty, and $I(X;Y)$ is the information-theoretic expression of Bayesian updating.
-- **Optimization thinking**: Channel capacity maximization is an optimization problem — $C=\max_{p(x)} I(X;Y)$ is optimization over input distributions.
-- **Transformation thinking**: Coding is a transformation of the information space — source coding transforms into efficient representations; channel coding transforms into noise-resistant representations.
-- **Modeling thinking**: Information criteria guide model selection — AIC/BIC/MDL quantify the fit-complexity trade-off from an information-theoretic perspective.
-- **Algorithmic thinking**: Compression algorithms are computational implementations — Huffman/LZ/LDPC and related algorithms are approximation implementations of information-theoretic limits.
-- **Modern mathematics activation**: `../../references/books/matrix-analysis.md` (low-rank compression), `../../references/books/abstract-algebra.md` (coding/finite fields), `../../references/books/algebraic-geometry-rising-sea.md` (Plücker KV compression).

package/skills/information-theory/SKILL.md

@@ -1,140 +0,0 @@
----
-name: information-theory
-description: |
-  触发：问题需量化不确定性、以"信息=不确定性的减少"度量信息价值；或需计算熵/互信息/KL 散度/信道容量，为压缩/KV 压缩/量化/路由设计信息准则时调用。
-  English: Trigger when a problem needs to quantify uncertainty, measure information value as "information = reduction of uncertainty"; or compute entropy / mutual information / KL divergence / channel capacity, or design information criteria for compression / KV-cache compression / quantization / routing.
----
-
-> **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。
-
-# 📡 信息论思想 / Information Theory
-
-> "信息是不确定性的减少——知道更多意味着怀疑更少。"
-> "Information is reduction of uncertainty — knowing more means doubting less."
->
-> —— 信息论、编码理论、统计推断 / Information Theory, Coding Theory, Statistical Inference
-
-## 核心原则 / Core Principle
-
-**信息是不确定性的减少——不确定性可由熵精确量化，压缩与通信存在由熵限与信道容量刻画的不可逾越极限，信息增益则指导在不确定性下应优先获取何种观察。**
-
-**Information is reduction of uncertainty — uncertainty is precisely quantified by entropy; compression and communication have unbreakable limits set by the entropy bound and channel capacity; information gain guides which observation to prioritize under uncertainty.**
-
-> **数学形式化 / Mathematical Formalization**
->
-> Shannon 熵 $H(X) = -\sum_x p(x)\log p(x)$ 量化随机变量的平均"惊奇度"——概率越低的事件发生时惊奇越大，熵是惊奇度的期望。
->
-> 互信息 $I(X;Y) = H(X) - H(X|Y) = \sum_{x,y} p(x,y)\log\frac{p(x,y)}{p(x)p(y)}$，即观察 $Y$ 后对 $X$ 不确定性的减少——这正是"信息"的数学定义。
->
-> KL 散度 $D_{KL}(P\|Q) = \sum_x p(x)\log\frac{p(x)}{q(x)}$ 衡量用 $Q$ 替代 $P$ 的信息损失，**不对称** $D(P\|Q)\neq D(Q\|P)$。
->
-> 信源编码定理：最优压缩的平均编码长度 $\ge H(X)$ bits/symbol，低于熵限则必然丢失信息。
->
-> 信道编码定理：可靠通信的速率上限为信道容量 $C = \max_{p(x)} I(X;Y)$；$R<C$ 时存在编码使误码率趋零，$R>C$ 时可靠通信不可能。
->
-> 率失真函数 $R(D) = \min_{p(\hat{x}|x):\,\mathbb{E}[d(x,\hat{x})]\le D} I(X;\hat{X})$ 给定失真 $D$ 下的最小信息率——有损压缩的极限。
->
-> 详细数学依据见 `original-texts.md`。
-
-## GPU 友好性 / GPU-Friendliness（横切检查）
-
-信息论量用于**压缩/剪枝/量化/KV 压缩/路由设计**时，须过 `../../references/gpu-friendly-math.md` 八维门。核心判据：**局部、可融合的熵/KL 估计 = 友好；全局精确信息估计 = 不友好**。
-
-- **量化校准（per-block 熵/Hessian）**：局部、可批量 GEMM 化、低精度可行——友好（维度 1/2/5）。
-- **KV-Cache 压缩（信息瓶颈/块摘要）**：低秩块摘要压缩显存——友好（维度 4）；Plücker 式块摘要见 `../../references/books/algebraic-geometry-rising-sea.md`。
-- **互信息剪枝/特征选择**：用局部或低秩近似估计 $I(X;Y)$ 则可改造；若需全分布精确估计则不友好。
-- **信息增益路由**：全局精确 $I(X;Y)$ 需 $O(n^2)$ 显存、高精度、不可融合——反模式；改用 softmax/热带门控等连续可微近似（维度 3/6/8）。
-- **反模式**：在完整 token 分布上精确计算熵/互信息——显存爆炸、需 fp64、串行依赖，"美但不可算"。
-
-八维最低判定（正式术语）：**张量化**看熵/KL/MI 是否可按 token/block/batch 估计；**GEMM 可映射**看压缩、投影、校准是否落线性代数；**复杂度**避免全分布精确估计；**显存与 KV-Cache**量化 KV/激活/码本占用；**低精度稳定**检查 log/softmax/KL 动态范围；**并行与通信**看分块统计是否可跨设备归约；**稀疏结构**看码本/路由是否块结构化；**算子融合**看统计、mask、量化能否融合。
-
-> 配合 `../../references/books/matrix-analysis.md`（低秩压缩）、`../../references/books/abstract-algebra.md`（编码/有限域）。
-
-## 不适用场景 / When NOT to Use
-
-- **问题无概率结构**（如纯粹符号推理、逻辑演绎）——熵与信息增益需要概率分布，无概率则无信息论。
-- **纯确定性场景无不确定性**（如已知精确答案的数学问题）——熵为零时信息论退化为平凡结论。
-- **定性判断无需量化**（如美学评价、情感判断）——信息论量化的是概率意义上的不确定性，非语义歧义。
-
-## 何时使用 / When to Use
-
-- 需要度量不确定性的大小（熵 $H(X)$ 量化随机变量的"混乱程度"）。
-- 需要比较不同信息源的价值（互信息 $I(X;Y)$ 衡量哪个观察 $Y$ 最能减少关于 $X$ 的不确定性）。
-- 需要最优数据压缩（信源编码定理保证最优压缩极限为 $H(X)$ bits/symbol）。
-- 需要在噪声环境下可靠通信（信道编码定理保证 $R<C$ 时可靠传输可行）。
-- 需要特征选择或模型选择（互信息筛选特征，AIC/BIC/MDL 作为信息准则选择模型）。
-- 需要贝叶斯模型比较（KL 散度 $D(P\|Q)$ 衡量分布间信息距离，贝叶斯因子量化模型间证据比）。
-- **为压缩 / KV-Cache 压缩 / 量化 / 路由设计信息准则**，并评估其 GPU 可行性。
-
-## 方法流程 / Method
-
-### 第一步：识别信息源与不确定度 / Identify Source and Uncertainty
-- **随机变量 $X$** 是什么？——定义信息源，明确要研究的不确定性对象。
-- **概率分布 $p(x)$**——离散分布用概率表，连续分布用密度函数。
-- **计算 $H(X) = -\sum p(x)\log p(x)$**——量化当前不确定性水平。
-- **识别要减少的不确定性**——明确"知道什么之后不确定性会降低？"
-
-### 第二步：量化信息增益 / Quantify Information Gain
-- **计算条件熵 $H(X|Y)$**——观察 $Y$ 之后 $X$ 的剩余不确定性。
-- **计算互信息 $I(X;Y) = H(X) - H(X|Y)$**——$Y$ 对 $X$ 提供的信息量。
-- **识别最优观察**——哪个 $Y$ 使 $I(X;Y)$ 最大？该观察最值得获取。
-- **链式规则**——$H(X_1,\dots,X_n) = H(X_1) + H(X_2|X_1) + \dots + H(X_n|X_1,\dots,X_{n-1})$，逐变量拆解联合不确定性。
-
-### 第三步：选择编码策略 / Choose Coding Strategy
-- **信源编码（压缩）**：Huffman 编码（贪心最优前缀码，平均长度接近 $H(X)$）、算术编码（更接近熵限）、通用编码（LZ77/LZ78/LZW，无需已知分布）。
-- **信道编码（纠错）**：Hamming 码（最小距离 3，纠正 1 位错）、Reed-Solomon 码（突发纠错）、LDPC/Turbo 码（逼近 Shannon 极限）。
-- **编码选择原则**：压缩需求用信源编码→逼近 $H(X)$；噪声防护用信道编码→逼近 Shannon 极限 ($R\to C$)。
-
-### 第四步：评估信道容量 / Evaluate Channel Capacity
-- **计算信道容量 $C = \max_{p(x)} I(X;Y)$**——在所有输入分布上最大化互信息。
-- **比较传输速率 $R$ 与容量 $C$**：$R<C$ → 可靠通信可行；$R>C$ → 必定出错。
-- **噪声模型**：BSC（二元对称信道，翻转概率 $p$）、BEC（二元擦除信道，擦除概率 $\varepsilon$）、AWGN（加性白高斯噪声信道）。
-- **容量公式示例**：BSC 容量 $C = 1 - H(p)$；AWGN 容量 $C = \frac{1}{2}\log(1 + S/N)$。
-
-### 第五步：应用信息准则 / Apply Information Criteria
-- **AIC（赤池信息准则）**：$\text{AIC} = -2\ln L + 2k$——偏重拟合，适合预测目标。
-- **BIC（贝叶斯信息准则）**：$\text{BIC} = -2\ln L + k\ln n$——偏重简约，适合解释目标。
-- **KL 散度 $D_{KL}(P\|Q) = \sum p(x)\log\frac{p(x)}{q(x)}$**——衡量用 $Q$ 替代 $P$ 的信息损失，注意不对称性。
-- **MDL 原则（最小描述长度）**：选择使"数据描述长度 + 模型描述长度"最小的模型——信息论版本的奥卡姆剃刀。
-
-### 第六步：做出信息最优决策 / Make Information-Optimal Decision
-- **贝叶斯实验设计**：选择使期望信息增益 $\max\,\mathbb{E}[I(\theta;Y)]$ 最大的实验——优先获取最能减少不确定性的数据。
-- **最小化 KL 散度**：决策输出分布 $Q$ 应尽量接近目标分布 $P$，即 $\min D(P\|Q)$。
-- **最大熵原则**：在已知约束下选择使 $H(X)$ 最大的分布——最少假设，最保守推断。
-- **信息瓶颈**：$\min I(X;T) - \beta I(T;Y)$——在压缩 $X$ 为 $T$ 时保留关于 $Y$ 的最大相关信息。
-
-## 常见错误 / Common Errors
-
-| 错误 / Error | 批评 / Critique | 正确做法 / Correct Approach |
-|---|---|---|
-| 把信息等同于比特而非概率减少 | 熵 $H(X)$ 是概率分布的函数，比特只是量化单位 | 理解信息为不确定性的减少 $I(X;Y)=H(X)-H(X\|Y)$ |
-| 忽视信道容量极限 | $R>C$ 时无论何种编码都不能可靠通信 | 计算容量 $C=\max I(X;Y)$，确保 $R<C$ |
-| 混淆熵与方差 | 熵度量概率结构的"扩散度"，方差度量数值"展开度"，两者不等价 | 概率不确定性用熵，数值偏差用方差；连续熵可为负 |
-| 过度压缩低于熵限 | 最优压缩 $\ge H(X)$ bits/symbol，低于此限必然丢信息 | 接受熵限并据此设计编码 |
-| 忽视 KL 散度不对称性 | $D(P\|Q)\ne D(Q\|P)$，方向不同含义不同 | 明确方向：$D(P\|Q)$ 用于"用 $Q$ 编码 $P$ 的额外代价" |
-| 把相关性等同于信息 | 相关 $\rho$ 仅衡量线性关联；$I(X;Y)=0 \Leftrightarrow$ 独立，但 $\rho=0 \not\Rightarrow$ 独立 | 用互信息 $I(X;Y)$ 评估依赖关系 |
-| 把定性判断强行量化 | 信息论量化概率不确定性，非语义歧义或主观感受 | 区分概率不确定性与语义歧义，不对定性问题套熵公式 |
-| 全局精确熵/互信息不可算 | 在完整分布上精确估计 $I(X;Y)$ 需 $O(n^2)$ 显存与高精度，GPU 不可行 | 用局部/采样/低秩估计，过 GPU 八维门 |
-
-## 操作规程 / Operating Procedure
-
-当本 skill 被触发时，输出必须包含：
-
-1. **[信息源]:[描述]** $H(X)=$ [值]——定义随机变量 $X$，计算其熵，量化当前不确定性。
-2. **[信息增益]:[描述]** $I(X;Y)=$ [值]——计算互信息，识别最有价值的观察 $Y$。
-3. **[编码策略]:[选择]**——信源编码（压缩）或信道编码（纠错），说明逼近哪种极限。
-4. **[信道容量]:[描述]** $C=$ [值]——计算信道容量，比较传输速率 $R$ 与 $C$。
-5. **[信息准则]:[AIC/BIC/KL/MDL]**——说明选择的信息准则及理由。
-6. **[最优决策]:[说明]**——基于信息增益最大化或 KL 散度最小化的决策建议。
-7. **[GPU 可行性]**（若用于压缩/KV 压缩/量化/路由）——熵/KL/互信息估计是局部可融合还是全局精确？过八维门，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。**
-
-## 与其他 skill 的关系 / Relations to Other Skills
-
-- **概率与统计**：熵与信息增益补充概率推理——$H(X)$ 度量分布不确定性，$I(X;Y)$ 是贝叶斯更新的信息论表达。
-- **优化思想**：信道容量最大化是优化问题——$C=\max_{p(x)} I(X;Y)$ 是输入分布上的最优化。
-- **变换思想**：编码是信息空间的变换——信源编码变换为高效表示，信道编码变换为抗干扰表示。
-- **建模思想**：信息准则指导模型选择——AIC/BIC/MDL 从信息论角度量化拟合与复杂度权衡。
-- **算法思想**：压缩算法是计算实现——Huffman/LZ/LDPC 等算法是信息论极限的逼近实现。
-- **现代数学激活**：`../../references/books/matrix-analysis.md`（低秩压缩）、`../../references/books/abstract-algebra.md`（编码/有限域）、`../../references/books/algebraic-geometry-rising-sea.md`（Plücker KV 压缩）。