math-skill 1.0.0 → 2.1.0

package/skills/transformation/SKILL.en.md

@@ -0,0 +1,120 @@
+---
+name: transformation
+description: |
+  Trigger when a problem is intractable in its current representation and needs a mathematical transform into an equivalent simpler problem; or finding an equivalent but more GPU-friendly representation (e.g. convolution->GEMM, spectral transform) for algorithm/operator design.
+---
+
+# Transformation
+
+> "Complex problem -> equivalent simple problem; the key is finding the right transformation and its inverse."
+>
+> -- Fourier, Laplace, Coordinate Transform
+
+## Core Principle
+
+**The same problem can have different representations. Choosing a good representation (or transformation) can make a difficult problem simple.** The essence of a transformation is not to change the problem itself, but to change the perspective from which it is viewed.
+
+> **Mathematical Formalization**
+>
+> Let a transformation $T: D_1 \to D_2$. $T$ is useful if and only if three conditions hold simultaneously: (1) **Computability**: $T(x)$ can be explicitly computed; (2) **Simplification**: the problem is more tractable in $D_2$ (e.g., differentiation -> multiplication, convolution -> multiplication); (3) **$T^{-1}$ exists**: the solution in $D_2$ can be mapped back to the original solution in $D_1$.
+>
+> Core question: **Under what conditions does $T^{-1}$ exist?** Existence determines equivalence, and the domain of convergence determines validity. Properties: equivalence (no information loss, requiring $T^{-1}$ to exist), simplification, invertibility (when $T^{-1}$ exists and is computable, the transformation is exact rather than approximate).
+
+## GPU-Friendliness (Cross-Cutting Check)
+
+In algorithm design, the value of a transformation often manifests as "converting an unfriendly operation into a GPU-friendly one" -- pass the eight-dimensional gate in `../../references/gpu-friendly-math.md`:
+
+- **Convolution -> GEMM** (im2col / Winograd): Converts irregular convolution into matrix multiplication, fully utilizing Tensor Cores (friendly, but watch for memory expansion).
+- **Spectral transforms (FFT / DCT)**: $O(N^2)\to O(N\log N)$, batch-parallelizable; but butterfly communication patterns require attention to memory access.
+- **Duality transforms (Legendre / Fenchel)**: Converts constrained optimization into a dual problem -- friendly if the dual is more GEMM-amenable.
+- **Transforms as compression**: Frequency-domain sparsity means only significant coefficients need to be stored (KV / activation compression, friendly).
+- **Anti-patterns**: The transform itself requires $O(n^2)$ global rearrangement and cannot be fused -- "beautiful but incomputable"; the inverse transform is not numerically stable (ill-conditioned).
+
+Eight-dimensional minimum assessment (formal terms): **Tensorization** -- whether the transform / inverse transform can be batched; **GEMM-mappability** -- whether it can be rewritten as GEMM / convolution / FFT / batched solve; **Complexity** -- the transform must reduce, not increase, the computational order; **Memory & KV-Cache** -- whether im2col, frequency-domain caches, and block summaries save memory rather than explode it; **Low-precision stability** -- inverse transform, orthogonalization, normalization; **Parallelism & communication** -- butterfly / scan / block communication patterns; **Sparse structure** -- whether frequency-domain / block sparsity is structured; **Operator fusion** -- whether the transform-core operation-inverse transform chain can be fused or recomputed.
+
+> Cross-reference `../../references/books/matrix-analysis.md` (spectral / low-rank), `optimization-ml.md` (duality).
+
+## When NOT to Use
+
+- **The problem is already simple enough** -- no transformation is needed.
+- **The transformation loses critical information** (irreversible, and the lost information is precisely what matters) -- choose an information-preserving transformation.
+- **Only qualitative understanding is needed** -- transformations are typically quantitative tools.
+- **Convergence conditions are not satisfied** -- forcing a transformation produces meaningless results.
+
+## When to Use
+
+- The problem is difficult to analyze or solve in its current form; revealing hidden structure in data (periodic signals -> frequency spectrum).
+- Simplifying complex operations into simpler ones (convolution -> multiplication); linearizing nonlinear problems; decoupling coupled variables.
+- **Finding an equivalent but more GPU-friendly representation for an operator** (convolution -> GEMM, sparse -> structured, frequency-domain compression).
+
+## Method
+
+### Step 1: Analyze the Difficulty of the Current Representation
+Why is the current form hard to work with? Is the computation complex (e.g., ODEs are hard to solve directly)? Is the structure opaque (periodicity is invisible in the time domain)? Are variables coupled? Identifying "where the difficulty lies" is a prerequisite for choosing a transformation; selecting a transform without diagnosis is blind operation.
+
+### Step 2: Select the Transformation
+Choose based on the type of difficulty; for each candidate, verify the formula, domain mapping, convergence conditions, and simplification effect:
+
+| Transform | Formula | Domain Mapping | Convergence / Validity | Simplification Effect |
+|---|---|---|---|---|
+| Fourier | $F(\omega)=\int f(t)e^{-i\omega t}\,dt$ | $t\in\mathbb{R}\to\omega\in\mathbb{R}$ | Dirichlet: absolutely integrable, finitely many extrema/discontinuities | Differentiation -> multiplication, convolution -> multiplication |
+| Laplace | $F(s)=\int_0^\infty f(t)e^{-st}\,dt$ | $t\in[0,\infty)\to s\in\mathbb{C},\;\text{Re}(s)>\alpha$ | $\exists\alpha: \int|f(t)|e^{-\alpha t}\,dt<\infty$ | Constant-coefficient ODE -> algebraic equation, incorporates initial conditions |
+| Z-transform | $F(z)=\sum f[n]z^{-n}$ | $n\in\mathbb{N}\to z\in\mathbb{C},\;|z|>R$ | $\exists R: \sum|f[n]|R^{-n}<\infty$ | Difference equation -> algebraic equation |
+| Generating function | $G(x)=\sum a_n x^n$ | $n\in\mathbb{N}\to x,\;|x|<\rho$ | $\exists\rho: \sum|a_n|\rho^n$ converges | Recurrence -> differential equation |
+| Legendre | $f^*(p)=\sup_x(px-f(x))$ | $x\to p=f'(x)$ | When $f$ is convex and differentiable, $p\leftrightarrow x$ is a bijection | Convex optimization -> dual, Lagrangian -> Hamiltonian |
+| Wavelet | $W(a,b)=\int f(t)\psi_{a,b}(t)\,dt$ | $t\in\mathbb{R}\to(a,b)$ | $f\in L^2$, $\psi$ admissible | Time-frequency localization, multi-scale analysis |
+
+### Step 3: Execute the Transformation
+Transform the problem into the new representation space, strictly following the formula. A transformation is not an escape from the problem; it is a re-expression of the same problem in a more effective language.
+
+### Step 4: Verify Convergence and Domain Conditions
+**Before applying transformation results, one must verify convergence conditions**: For Fourier, check Dirichlet conditions ($\int|f|\,dt<\infty$); for Laplace, determine the region of convergence $\text{Re}(s)>\alpha$ (outside this region $F(s)$ is undefined); for the Z-transform, determine $|z|>R$ (the inverse transform depends on the choice of region of convergence); for generating functions, determine the radius of convergence $\rho$; for wavelets, verify the admissibility condition of the mother wavelet. Skipping this step is the largest source of errors.
+
+### Step 5: Solve in the Transformed Space
+In the new representation the problem is often simpler; all operations must be carried out within the valid domain of the transform. The value of the transformed space lies in making originally hidden structure visible.
+
+### Step 6: Apply the Inverse Transform to Return to the Original Space
+Map the solution back to the language of the original problem. The existence of the inverse transform requires conditions: Fourier inverse $f(t)=(1/2\pi)\int F(\omega)e^{i\omega t}\,d\omega$ (requires $F$ to be absolutely integrable); Laplace inverse via the Bromwich integral $f(t)=(1/2\pi i)\int F(s)e^{st}\,ds$ ($\gamma>\alpha$, to the right of the region of convergence); Z-inverse $f[n]=(1/2\pi i)\oint F(z)z^{n-1}\,dz$ (contour within the region of convergence); generating function $a_n=G^{(n)}(0)/n!$ or $[x^n]G(x)$. The transformation is merely a means; the final answer must be in the original space.
+
+### Step 7: Verify Equivalence
+Confirm that the transformation has not lost critical information and that the inverse-transformed result is indeed a solution of the original problem: Was the inverse transform executed within the correct region of convergence? Does the original function satisfy the prerequisite conditions? Were boundary / initial conditions correctly encoded? Equivalence is the bottom line of transformation theory -- a transformation must be a verifiably equivalent operation, not an approximation or an evasion.
+
+## Common Errors
+
+| Error | Critique | Correct Approach |
+|---|---|---|
+| Choosing an inappropriate transformation | Fails to simplify and instead increases complexity | Select the transformation based on problem characteristics |
+| Using an irreversible transformation that loses information | Cannot return to the original problem after transformation | Verify invertibility or confirm that the lost information is not critical |
+| Forgetting the inverse transform | After obtaining a solution in the new space, forgetting to transform back | The final answer must be in the original space |
+| Ignoring domain changes after transformation | The transformation may alter the domain or introduce singularities | Check the post-transformation domain and boundary conditions |
+| Treating the transformation as magic | A transformation only changes the representation, not the essence of the problem | A transformation is a tool; understanding is the core competency |
+| Applying Fourier to non-integrable functions | When $\int|f|\,dt=\infty$, $F(\omega)$ may be meaningless | First check Dirichlet conditions; if not satisfied, use generalized functions or Laplace |
+| Ignoring the Laplace region of convergence | $F(s)$ is defined only for $\text{Re}(s)>\alpha$ | Restrict all operations to the region of convergence |
+| Assuming all transformations are invertible | Some transformations are not invertible under certain conditions (e.g., non-convex Legendre) | Verify invertibility conditions before applying the inverse transform |
+| Overlooking discrete transforms (DFT/FFT) | Continuous transforms are not applicable to discrete data | Use the DFT for discrete data: $X[k]=\sum x[n]e^{-i2\pi kn/N}$ |
+| Post-transformation incomputability / non-fusibility | The transform requires $O(n^2)$ global rearrangement or an ill-conditioned inverse | Evaluate the GPU eight-dimensional gate; if necessary, choose a fusible, numerically stable equivalent transformation |
+
+## Operating Procedure
+
+The output must include:
+
+1. **Difficulty of the current representation**: `[Difficulty]: [Description]`
+2. **Transformation selection**: `[Transform]: [Choice] because [Reason], formula [Formula], convergence conditions [Conditions], expected simplification [Effect]`
+3. **Transformation execution**: The problem form after transformation
+4. **Convergence and domain verification**: `[Verification]: [Whether conditions are satisfied, region of convergence]`
+5. **Solution in transformed space**: The solution method under the new representation
+6. **Inverse transform**: The solution translated back to the original space, noting inverse-transform conditions
+7. **Equivalence verification**: Is it invertible? Is information lost? Is the region of convergence correct?
+8. **[GPU viability]** (if used for operator design) -- whether the transformation converts the operation into GEMM / fusible / numerically stable form; pass the eight-dimensional gate
+
+**Output must not consist of analysis alone without conclusions.**
+
+## Relations to Other Skills
+
+- **Abstraction**: A transformation is also an act of abstraction -- representing the same object in a new structure.
+- **Symmetry and invariance**: Properties that remain unchanged under a transformation are precisely the symmetries of that transformation.
+- **Optimization**: Transforming to the dual space sometimes makes optimization easier.
+- **Modeling**: A transformation is often a key step in solving a model.
+- **Algorithmic thinking**: The FFT reduces the DFT from $O(N^2)$ to $O(N\log N)$, exemplifying how transformation ideas serve computational efficiency.
+- **Information theory**: Coding is transformation -- source coding transforms data into an efficient representation; channel coding transforms it into a noise-resistant representation.
+- **Modern mathematics activation**: `../../references/books/matrix-analysis.md` (spectral / low-rank transforms), `algebraic-geometry-rising-sea.md` (Plücker coordinates for KV compression, tropical gating), `optimization-ml.md` (duality transforms).

package/skills/transformation/SKILL.md

@@ -1,264 +1,124 @@
----
-name: transformation
-description: |
-  触发条件：
-  科研模式：数学推导、方程求解、信号处理等问题在当前表示下难以处理，需要通过数学变换将其转化为更简单的等价问题。
-  生活模式：日常问题解决中换个角度思考，将看似棘手的问题从不同视角重新审视，使困难消失或变得可解。
-  English:
-  Research mode: Trigger for mathematical derivation, equation solving, signal processing — when a problem is intractable in its current representation and needs mathematical transformation into an equivalent simpler problem.
-  Life mode: Trigger for daily problem-solving, changing perspective, reframing problems — viewing a seemingly thorny problem from a different angle so the difficulty vanishes or becomes tractable.
----
-
-# 🔄 变换思想 / Transformation Thinking
-
-> "复杂问题→等价简单问题，关键是找到合适的变换和逆变换。"
-> "Complex problem → equivalent simple problem; the key is finding the right transformation and its inverse."
->
-> "变换的精髓不是改变问题本身，而是改变我们看问题的视角。"
-> "The essence of transformation is not changing the problem itself, but changing our perspective on the problem."
->
-> —— 傅里叶变换、拉普拉斯变换、坐标变换 / Fourier Transform, Laplace Transform, Coordinate Transformation
-
-## 核心原则 / Core Principle
-
-**同一个问题可以有不同的表示方式。选择一个好的表示（或变换），可以让困难的问题变得简单。变换的精髓不是改变问题本身，而是改变我们看问题的视角。**
-
-**The same problem can be represented in different ways. Choosing a good representation (or transformation) can make a difficult problem simple. The essence of transformation is not changing the problem itself, but changing our perspective on the problem.**
-
-> **数学形式化 / Mathematical Formalization**（科研模式参考 / Research Mode Reference）
->
-> 设变换 T: D₁ → D₂，T 有用当且仅当三个条件同时成立：
-> 1. T 可计算：对 D₁ 中元素 x，T(x) 可以显式求出
-> 2. 简化性：问题在 D₂ 中比在 D₁ 中更容易处理（如微分→乘法、卷积→乘法）
-> 3. T⁻¹ 存在：可以从 D₂ 的解恢复 D₁ 的原解
->
-> Let T: D₁ → D₂. T is useful iff three conditions hold simultaneously:
-> 1. T is computable: for x ∈ D₁, T(x) can be explicitly obtained
-> 2. Simplification: the problem is easier in D₂ (e.g., differentiation → multiplication, convolution → multiplication)
-> 3. T⁻¹ exists: the original solution in D₁ can be recovered from D₂
->
-> 核心追问：**T⁻¹ 在什么条件下存在？** 这是变换思想最本质的数学问题——存在性决定等价性，收敛域决定有效性。
->
-> Key question: **Under what conditions does T⁻¹ exist?** This is the most essential mathematical question of transformation — existence determines equivalence, convergence domain determines validity.
->
-> 变换的核心性质：
-> - **等价性**：变换前后的问题是等价的（信息不丢失），要求 T⁻¹ 存在 / **Equivalence**: the problem before and after transform is equivalent (no information loss), requiring T⁻¹ to exist
-> - **简化性**：变换后的问题更容易处理 / **Simplification**: the problem is easier to handle after transform
-> - **可逆性**：T⁻¹ 存在且可计算时，变换是等价的而非近似的 / **Invertibility**: when T⁻¹ exists and is computable, the transform is equivalent rather than approximate
-
-## 不适用场景 / When NOT to Use
-
-- **问题已经足够简单**——不需要变换 `[通用]` / The problem is already simple enough — no transformation needed `[General]`
-- **变换会丢失关键信息**（如不可逆变换且丢失的信息正是关心的）——需要选择保信息的变换 `[科研]` / Transformation loses critical information — choose an information-preserving transform `[Research]`
-- **只需要定性理解**——变换通常是定量工具 `[科研]` / Only qualitative understanding is needed — transformation is typically a quantitative tool `[Research]`
-- **收敛条件不满足**——强行变换会产生无意义结果 `[科研]` / Convergence conditions are not met — forcing a transform yields meaningless results `[Research]`
-- **换视角变成逃避现实**——不是所有问题换个角度就能解决，切换视角的前提是新视角确实能简化问题 `[生活]` / Perspective-switching becomes escapism — not all problems become easier from a different angle; the premise is that the new perspective genuinely simplifies `[Life]`
-- **新视角的结论回不到原问题**——视角切换必须可逆，否则你只是在另一个世界里自嗨 `[生活]` / Conclusions from the new perspective can't translate back — perspective switching must be reversible, otherwise you're just wandering in another world `[Life]`
-
-## 何时使用 / When to Use
-
-### 科研触发条件 / Research Triggers
-
-- 一个问题在当前形式下难以分析或求解 / A problem is hard to analyze or solve in its current form
-- 想要揭示数据的隐藏结构（如周期信号→频谱） / Revealing hidden structure in data (e.g., periodic signal → frequency spectrum)
-- 需要将复杂运算简化为简单运算（如卷积→乘法） / Simplifying complex operations (e.g., convolution → multiplication)
-- 需要将非线性问题线性化 / Linearizing nonlinear problems
-- 需要解耦耦合变量 / Decoupling coupled variables
-
-### 生活触发条件 / Life Triggers
-
-- 当前视角下问题显得无解或过于复杂，直觉告诉你换个角度看会不同 / The problem seems unsolvable or overly complex from the current angle, and intuition suggests a different perspective could help
-- 反复尝试同一方法失败，需要跳出思维惯性 / Repeated failures with the same approach — need to break out of thinking inertia
-- 问题包含了多个相互纠缠的要素，换个视角可能让纠缠变清晰 / Multiple intertwined factors in the problem — a different perspective may untangle them
-- 需要把短期困扰放到长期视野中理解 / Need to understand a short-term difficulty in a long-term perspective
-
-## 方法流程 / Method
-
-### 第一步：分析当前表示的困难 / Step 1: Analyze Difficulty in Current Representation
-
-#### 科研模式 / Research Mode
-
-为什么当前形式难以处理？/ Why is the current form hard to handle?
-- 运算复杂？（如微分方程直接求解困难）/ Complex operations? (e.g., solving differential equations directly)
-- 结构不清晰？（如时域看不出周期性）/ Unclear structure? (e.g., periodicity invisible in time domain)
-- 变量耦合？（如多变量相互依赖）/ Coupled variables? (e.g., mutual dependencies among variables)
-
-#### 生活模式 / Life Mode
-
-当前视角下问题的困难在哪？是什么让你觉得棘手——是太复杂、太模糊、还是方向不对？/ What is the difficulty in the current perspective? What makes it feel tricky — is it too complex, too vague, or is the direction wrong?
-
-#### 共通要点 / Common Insight
-
-明确"困难在哪里"是选择变换的前提。不诊断困难就选变换，是盲目操作。/ Identifying "where the difficulty lies" is the prerequisite for choosing a transformation. Choosing a transform without diagnosing the difficulty is blind operation.
-
-### 第二步：选择变换 / Step 2: Choose the Transformation
-
-#### 科研模式 / Research Mode
-
-根据困难类型选择变换，每种变换必须检查其公式、域映射、收敛条件与简化效果。/ Choose based on difficulty type; for each transform, check formula, domain mapping, convergence conditions, and simplification effect.
-
-| 变换 / Transform | 公式 / Formula | 域映射 / Domain Mapping | 收敛/有效性条件 / Convergence/Validity | 简化效果 / Simplification |
-|---|---|---|---|---|
-| 傅里叶变换 Fourier | F(ω) = ∫₋∞⁺∞ f(t)e⁻ⁱωᵗdt | t ∈ ℝ → ω ∈ ℝ | Dirichlet条件：f绝对可积、有限极值点数、有限间断点数 | 微分→乘法，卷积→乘法 |
-| 拉普拉斯变换 Laplace | F(s) = ∫₀⁺∞ f(t)e⁻ˢᵗdt | t ∈ [0,∞) → s ∈ ℂ, Re(s)>α | ∃α: ∫₀⁺∞ |f(t)|e⁻αᵗdt < ∞ (Re(s)>α 时收敛) | 常系数ODE→代数方程，含初值 |
-| Z变换 Z-transform | F(z) = Σₙ₌₀⁺∞ f[n]z⁻ⁿ | n ∈ ℕ → z ∈ ℂ, |z|>R | ∃R: Σ|f[n]|R⁻ⁿ < ∞ (|z|>R 时收敛) | 差分方程→代数方程 |
-| 生成函数 Generating | G(x) = Σₙ₌₀⁺∞ aₙxⁿ | n ∈ ℕ → x ∈ ℝ, |x|<ρ | ∃ρ: Σ|aₙ|ρⁿ收敛 (|x|<ρ时) | 递推关系→微分方程 |
-| Legendre变换 Legendre | f*(p) = supₓ(px − f(x)) | x → p = f'(x) | f凸且可微时 p↔x 一一对应 | 凸优化→对偶问题，Lagrange力学→Hamilton力学 |
-| 小波变换 Wavelet | W(a,b) = ∫ f(t)ψₐᵇ(t)dt, ψₐᵇ=(1/√a)ψ((t−b)/a) | t ∈ ℝ → (a,b) ∈ ℝ⁺×ℝ | f ∈ L²(ℝ), ψ容许：∫|Ψ(ω)|²/|ω|dω<∞ | 时频局部化，多尺度分析 |
-
-#### 生活模式 / Life Mode
-
-有没有一个完全不同的视角让问题变得简单？常见的视角切换：/ Is there a completely different perspective that makes the problem simple? Common perspective switches:
-
-- **时间→频率**（看趋势而非细节）/ **Time → Frequency**: see trends rather than details
-- **局部→整体**（看大图而非细节）/ **Local → Global**: see the big picture rather than details
-- **过程→结果**（看终点而非路径）/ **Process → Outcome**: see the destination rather than the path
-- **表面→本质**（看结构而非外观）/ **Surface → Essence**: see structure rather than appearance
-
-#### 共通要点 / Common Insight
-
-变换的选择必须针对第一步诊断的困难。不同困难对应不同变换。/ The transformation must target the difficulty diagnosed in Step 1. Different difficulties correspond to different transformations.
-
-### 第三步：执行变换 / Step 3: Apply the Transformation
-
-#### 科研模式 / Research Mode
-
-将问题变换到新的表示空间，严格按公式执行。/ Transform the problem to the new representation space, executing strictly by formula.
-
-#### 生活模式 / Life Mode
-
-在新视角下重新描述你的问题——同一个问题用不同语言讲述时，困难可能消失了。/ Redescribe your problem under the new perspective — when the same problem is told in a different language, the difficulty may vanish.
-
-#### 共通要点 / Common Insight
-
-变换/视角切换不是逃避问题，而是用一种更有效的语言重新表达同一个问题。/ Transformation/perspective-switching is not escaping the problem but re-expressing the same problem in a more effective language.
-
-### 第四步：验证收敛与域条件 / Step 4: Verify Convergence and Domain Conditions
-
-#### 科研模式 / Research Mode
-
-**在应用变换结果之前，必须验证收敛条件是否满足。** / **Before applying transform results, convergence conditions must be verified.**
-
-- 傅里叶变换：检查 f(t) 是否满足 Dirichlet 条件（绝对可积 ∫|f(t)|dt < ∞，有限极值点，有限间断点）；不满足则 F(ω) 无意义
-- 拉普拉斯变换：确定收敛域 Re(s) > α；所有运算必须在收敛域内进行，F(s) 在收敛域外无定义
-- Z变换：确定收敛域 |z| > R；逆变换依赖收敛域的选择
-- 生成函数：确定收敛半径 ρ；|x| < ρ 时级数才有意义
-- 小波变换：验证母小波 ψ 的容许条件 ∫|Ψ(ω)|²/|ω|dω < ∞
-
-#### 生活模式 / Life Mode
-
-这个新视角在什么条件下才有效？不是所有问题换个角度就能解决——确认你的问题满足视角切换的前提。/ Under what conditions is this new perspective valid? Not all problems become solvable by changing angle — confirm your problem satisfies the prerequisites for perspective switching.
-
-#### 共通要点 / Common Insight
-
-任何变换/视角切换都有适用条件。跳过这一步是最大的错误来源。/ Any transformation/perspective switch has applicability conditions. Skipping this step is the biggest source of error.
-
-### 第五步：在变换空间求解 / Step 5: Solve in the Transformed Space
-
-#### 科研模式 / Research Mode
-
-在新表示下，问题往往变得更简单。所有运算必须在变换的有效域内进行。/ In the new representation, the problem is often simpler. All operations must stay within the transform's valid domain.
-
-#### 生活模式 / Life Mode
-
-在新视角下分析——原来难以看到的模式、结构、解决方案可能变得一目了然。/ Analyze under the new perspective — patterns, structures, and solutions that were hard to see may become immediately clear.
-
-#### 共通要点 / Common Insight
-
-变换空间/新视角的价值在于让原本隐蔽的结构变得可见。/ The value of the transformed space / new perspective is making previously hidden structures visible.
-
-### 第六步：逆变换回原空间 / Step 6: Inverse Transform Back
-
-#### 科研模式 / Research Mode
-
-将解逆变换回原问题的语言。逆变换的存在性需要条件：/ Translate the solution back via inverse transform. Inverse existence requires conditions:
-
-- 傅里叶逆变换：f(t) = (1/2π)∫₋∞⁺∞ F(ω)eⁱωᵗdω，需要 F(ω) 绝对可积
-- 拉普拉斯逆变换：f(t) = (1/2πi)∫ᵧ₋ⁱ∞ᵧ₊ⁱ∞ F(s)eˢᵗds (Bromwich积分)，γ > α 确保积分路径在收敛域右侧
-- Z逆变换：f[n] = (1/2πi)∮ F(z)zⁿ⁻¹dz，围道必须在收敛域内
-- 生成函数逆映射：aₙ = G⁽ⁿ⁾(0)/n!（泰勒展开系数），或 [xⁿ]G(x)
-
-#### 生活模式 / Life Mode
-
-把新视角的结论翻译回原来的问题语言——确保换角度看问题不是逃避，而是真的找到了解决路径。/ Translate conclusions from the new perspective back into the original problem's language — ensure that perspective-switching is not escapism but a genuine solution path.
-
-#### 共通要点 / Common Insight
-
-变换只是手段，最终答案必须在原空间。逆变换/翻译回原问题这一步不可省略。/ Transformation is only a tool; the final answer must be in the original space. The inverse transform / translation-back step cannot be omitted.
-
-### 第七步：验证等价性 / Step 7: Verify Equivalence
-
-#### 科研模式 / Research Mode
-
-确认变换过程没有丢失关键信息，逆变换的结果确实是原问题的解。检查：/ Confirm no critical information was lost, and the inverse result is indeed the solution. Check:
-- 逆变换是否在收敛域内执行 / Was inverse executed within convergence domain?
-- 原函数是否满足变换的前置条件 / Does the original function satisfy transform prerequisites?
-- 边界条件/初值条件是否被正确编码 / Were boundary/initial conditions correctly encoded?
-
-#### 生活模式 / Life Mode
-
-验证切换视角没有丢失关键信息——新视角的结论回到原问题时，是否真的解决了原问题？/ Verify that perspective-switching did not lose critical information — when the conclusion from the new perspective returns to the original problem, does it genuinely solve it?
-
-#### 共通要点 / Common Insight
-
-等价性是变换思想的底线。变换/视角切换必须是可验证的等价操作，而非近似或逃避。/ Equivalence is the baseline of transformation thinking. A transformation / perspective switch must be a verifiable equivalent operation, not an approximation or escapism.
-
-## 常见错误 / Common Errors
-
-| 错误 / Error | 批评 / Critique | 正确做法 / Correct Approach |
-|---|---|---|
-| 选择不合适的变换 `[科研]` / Choosing an unsuitable transform `[Research]` | 变换没有简化问题，反而增加复杂度 / Transform adds complexity instead of simplifying | 根据问题特征选择变换 / Choose transform based on problem characteristics |
-| 使用不可逆变换且丢失信息 `[科研]` / Using irreversible transform that loses information `[Research]` | 变换后无法回到原问题 / Cannot return to original problem after transform | 检查变换是否可逆，或确认丢失的信息不重要 / Verify invertibility or confirm lost info is irrelevant |
-| 忘记逆变换 `[通用]` / Forgetting the inverse transform `[General]` | 在新空间得到解后忘记变回来 / Solution in new space not translated back | 变换只是手段，最终答案必须在原空间 / Transform is a tool; final answer must be in original space |
-| 变换后忽略定义域变化 `[科研]` / Ignoring domain changes after transform `[Research]` | 变换可能改变定义域或引入奇异点 / Transform may change domain or introduce singularities | 检查变换后的定义域和边界条件 / Check transformed domain and boundary conditions |
-| 把变换当魔法 `[通用]` / Treating transform as magic `[General]` | 变换只是换了表示，没有改变问题本质 / Transform only changes representation, not the problem itself | 变换是工具，理解力才是核心 / Transform is a tool; understanding is fundamental |
-| 对不可积函数用傅里叶变换 `[科研]` / Applying Fourier transform to non-integrable functions `[Research]` | ∫|f(t)|dt = ∞ 时 F(ω) 可能不存在或无意义 / F(ω) may not exist or be meaningless when ∫|f(t)|dt = ∞ | 先检查 Dirichlet 条件，不满足时考虑广义函数或Laplace / Check Dirichlet conditions first; if unsatisfied, use generalized functions or Laplace |
-| 忽略Laplace收敛域 `[科研]` / Ignoring Laplace convergence domain `[Research]` | F(s)仅在Re(s)>α有定义，域外运算无意义 / F(s) is defined only for Re(s)>α; operations outside are meaningless | 明确收敛域，所有运算限制在Re(s)>α内 / Specify convergence domain; restrict all operations to Re(s)>α |
-| 假设所有变换可逆 `[科研]` / Assuming all transforms are invertible `[Research]` | 部分变换在特定条件下不可逆（如非凸函数的Legendre变换） / Some transforms are non-invertible under certain conditions | 验证可逆条件后再执行逆变换 / Verify invertibility conditions before applying inverse |
-| 忽略离散变换(DFT/FFT) `[科研]` / Ignoring discrete transforms (DFT/FFT) `[Research]` | 连续变换不适用于离散数据，强行使用引入误差 / Continuous transforms unsuitable for discrete data; forced use introduces error | 对离散数据使用DFT: X[k]=Σx[n]e⁻ⁱ2πkn/N / Use DFT for discrete data |
-| 换视角变成逃避现实 `[生活]` / Perspective-switching becomes escapism `[Life]` | 不是换个角度看就能解决所有问题，逃避不等于变换 / Not all problems are solved by changing perspective; escapism ≠ transformation | 确认新视角确实简化了问题，而非回避了问题 / Confirm the new perspective genuinely simplifies, rather than avoids, the problem |
-| 新视角的结论回不到原问题 `[生活]` / Conclusions from new perspective can't translate back `[Life]` | 在新视角里找到了"答案"但无法落实到原问题的行动中 / Found an "answer" in the new perspective but can't translate it into actions for the original problem | 确保视角切换是可逆的——必须能翻译回原问题的语言 / Ensure perspective switch is reversible — must be able to translate back into the original problem's language |
-| 强行套用不适用的视角 `[生活]` / Forcing an inapplicable perspective `[Life]` | 不是所有视角切换都有效，前提条件不满足时换视角毫无意义 / Not all perspective switches are effective; when prerequisites fail, switching is meaningless | 先验证新视角的适用条件（第四步不可跳过） / Verify applicability conditions of the new perspective first (Step 4 cannot be skipped) |
-
-## 操作规程 / Operating Procedure
-
-当本 skill 被触发时，首先判断模式：/ When this skill is triggered, first determine the mode:
-
-- **科研模式**：涉及数学推导、方程求解、信号处理等定量问题 → 使用科研模式输出格式 / **Research Mode**: involves mathematical derivation, equation solving, signal processing, etc. → use Research output format
-- **生活模式**：涉及日常问题解决、视角切换、困难重构等定性问题 → 使用生活模式输出格式 / **Life Mode**: involves daily problem-solving, perspective switching, difficulty reframing, etc. → use Life output format
-
----
-
-### 科研模式输出格式 / Research Mode Output Format
-
-1. **当前表示的困难**：描述为什么当前形式难以处理 `[困难]: [描述]`
-2. **变换选择**：`[变换]: [选择] 因为 [理由]，公式为 [公式]，收敛条件为 [条件]，预期简化效果是 [效果]`
-3. **变换执行**：描述变换后的问题形式
-4. **收敛与域条件验证**：`[验证]: [条件是否满足，收敛域是什么]`
-5. **变换空间中的求解**：在新表示下的解法
-6. **逆变换**：将解翻译回原空间，注明逆变换条件
-7. **等价性验证**：验证变换是否可逆、是否丢失信息、收敛域是否正确
-
-**输出必须包含以上 7 项，不得只输出分析性文字而不给出结论。每项必须含具体数学内容，不得仅写概述。**
-
----
-
-### 生活模式输出格式 / Life Mode Output Format
-
-1. **[当前困难]: [描述]** — 当前视角下问题的核心困难是什么 / What is the core difficulty under the current perspective
-2. **[新视角]: [切换方向]** — 选择一个不同的视角（时间→频率、局部→整体、表面→本质等） / Choose a different perspective (time→frequency, local→global, surface→essence, etc.)
-3. **[新视角下的分析]: [洞察]** — 在新视角下问题呈现什么新特征 / What new features does the problem reveal under the new perspective
-4. **[视角有效性]: [条件]** — 这个新视角在什么条件下才有效 / Under what conditions is this new perspective valid
-5. **[回到原问题]: [翻译]** — 把新视角的结论翻译回原问题语言 / Translate conclusions from the new perspective back into the original problem's language
-6. **[行动建议]: [具体步骤]** — 明确写出"接下来我将……" / Explicitly write out "Next I will..."
-
-**输出必须包含以上 6 项，不得只输出感悟性文字而不给出行动建议。每项必须含具体内容，不得仅写抽象原则。**
-
-## 与其他 skill 的关系 / Relations to Other Skills
-
-- **抽象化思想**：变换也是一种抽象——在新的结构中表示同一个对象
-- **对称与不变性**：变换下保持不变的性质就是对变换的对称性
-- **优化思想**：变换到对偶空间有时使优化更容易
-- **建模思想**：变换常常是模型求解的关键步骤
-- **算法思想**：FFT将DFT的O(N²)复杂度降为O(N log N)，是变换思想在计算效率上的体现
-- **信息论思想**：编码即变换——信源编码将信息变换为高效表示，信道编码变换为抗干扰表示
+---
+name: transformation
+description: |
+  触发：数学推导、方程求解、信号处理等问题在当前表示下难处理，需通过数学变换转为更简单的等价问题；或为算法/算子设计寻找等价但更 GPU 友好的表示（如卷积→GEMM、谱变换）时调用。
+  English: Trigger when a problem is intractable in its current representation and needs a mathematical transform into an equivalent simpler problem; or finding an equivalent but more GPU-friendly representation (e.g. convolution→GEMM, spectral transform) 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。
+
+# 🔄 变换思想 / Transformation
+
+> "复杂问题→等价简单问题，关键是找到合适的变换和逆变换。"
+> "Complex problem → equivalent simple problem; the key is finding the right transformation and its inverse."
+>
+> —— 傅里叶变换、拉普拉斯变换、坐标变换 / Fourier, Laplace, Coordinate Transform
+
+## 核心原则 / Core Principle
+
+**同一个问题可以有不同的表示方式。选择一个好的表示（或变换），可以让困难的问题变得简单。** 变换的精髓不是改变问题本身，而是改变看问题的视角。
+
+> **数学形式化 / Mathematical Formalization**
+>
+> 设变换 T: D₁ → D₂，T 有用当且仅当三条件同时成立：① **可计算**：T(x) 可显式求出；② **简化性**：问题在 D₂ 中更易处理（如微分→乘法、卷积→乘法）；③ **T⁻¹ 存在**：可从 D₂ 的解恢复 D₁ 原解。
+>
+> 核心追问：**T⁻¹ 在什么条件下存在？** 存在性决定等价性，收敛域决定有效性。性质：等价性（信息不丢失，需 T⁻¹ 存在）、简化性、可逆性（T⁻¹ 存在且可计算时为等价而非近似）。
+
+## GPU 友好性 / GPU-Friendliness（横切检查）
+
+变换的价值在算法设计中常体现为「把不友好的运算变成 GPU 友好的运算」——过 `../../references/gpu-friendly-math.md` 八维：
+
+- **卷积→GEMM**（im2col/Winograd）：把不规则卷积变成矩阵乘，吃满 Tensor Core（友好，但注意显存扩张）。
+- **谱变换（FFT/DCT）**：O(N²)→O(N log N)，可批量并行；但 butterfly 通信模式需注意访存。
+- **对偶变换（Legendre/Fenchel）**：把约束优化变成对偶问题——若对偶更易 GEMM 化则友好。
+- **变换作为压缩**：频域稀疏→只存显著系数（KV/激活压缩，友好）。
+- **反模式**：变换本身需 O(n²) 全局重排且不可融合→"美但不可算"；逆变换不可数值稳定（病态）。
+
+八维最低判定（正式术语）：**张量化**看变换/逆变换是否可批量；**GEMM 可映射**看能否改写为 GEMM/conv/FFT/batched solve；**复杂度**要求变换后降阶而非升阶；**显存与 KV-Cache**检查 im2col、频域缓存、块摘要是否省而不是爆；**低精度稳定**检查逆变换、正交化、归一化；**并行与通信**看 butterfly/scan/分块通信；**稀疏结构**看频域/块稀疏是否结构化；**算子融合**看变换-核心算子-逆变换能否融合或重计算。
+
+> 配合 `../../references/books/matrix-analysis.md`（谱/低秩）、`optimization-ml.md`（对偶）。
+
+## 不适用场景 / When NOT to Use
+
+- **问题已足够简单**——不需要变换。
+- **变换会丢失关键信息**（不可逆且丢失的正是关心的）——选保信息的变换。
+- **只需定性理解**——变换通常是定量工具。
+- **收敛条件不满足**——强行变换产生无意义结果。
+
+## 何时使用 / When to Use
+
+- 问题在当前形式下难分析或求解；揭示数据隐藏结构（周期信号→频谱）。
+- 将复杂运算简化为简单运算（卷积→乘法）；非线性问题线性化；解耦耦合变量。
+- **为算子寻找等价但更 GPU 友好的表示**（卷积→GEMM、稀疏→结构化、频域压缩）。
+
+## 方法流程 / Method
+
+### 第一步：分析当前表示的困难
+为什么当前形式难处理？运算复杂（如 ODE 直接求解难）？结构不清晰（时域看不出周期）？变量耦合？明确"困难在哪"是选择变换的前提，不诊断就选变换是盲目操作。
+
+### 第二步：选择变换
+根据困难类型选择，每种检查公式、域映射、收敛条件、简化效果：
+
+| 变换 | 公式 | 域映射 | 收敛/有效性 | 简化效果 |
+|---|---|---|---|---|
+| 傅里叶 Fourier | F(ω)=∫f(t)e⁻ⁱωᵗdt | t∈ℝ→ω∈ℝ | Dirichlet：绝对可积、有限极值/间断点 | 微分→乘法，卷积→乘法 |
+| 拉普拉斯 Laplace | F(s)=∫₀∞f(t)e⁻ˢᵗdt | t∈[0,∞)→s∈ℂ, Re(s)>α | ∃α: ∫|f(t)|e⁻αᵗdt<∞ | 常系数ODE→代数方程，含初值 |
+| Z 变换 | F(z)=Σf[n]z⁻ⁿ | n∈ℕ→z∈ℂ, |z|>R | ∃R: Σ|f[n]|R⁻ⁿ<∞ | 差分方程→代数方程 |
+| 生成函数 | G(x)=Σaₙxⁿ | n∈ℕ→x, |x|<ρ | ∃ρ: Σ|aₙ|ρⁿ收敛 | 递推→微分方程 |
+| Legendre | f*(p)=supₓ(px−f(x)) | x→p=f'(x) | f 凸可微时 p↔x 一一对应 | 凸优化→对偶，Lagrange→Hamilton |
+| 小波 Wavelet | W(a,b)=∫f(t)ψₐᵇ(t)dt | t∈ℝ→(a,b) | f∈L², ψ 容许 | 时频局部化，多尺度 |
+
+### 第三步：执行变换
+将问题变换到新表示空间，严格按公式执行。变换不是逃避问题，是用更有效的语言重新表达同一问题。
+
+### 第四步：验证收敛与域条件
+**应用变换结果前必须验证收敛条件**：傅里叶检查 Dirichlet（∫|f|dt<∞）；拉普拉斯定收敛域 Re(s)>α，域外 F(s) 无定义；Z 变换定 |z|>R（逆变换依赖收敛域选择）；生成函数定收敛半径 ρ；小波验证母小波容许条件。跳过此步是最大错误来源。
+
+### 第五步：在变换空间求解
+在新表示下问题往往更简单，所有运算须在变换有效域内进行。变换空间的价值在于让原本隐蔽的结构可见。
+
+### 第六步：逆变换回原空间
+将解逆变换回原问题语言，逆变换存在性需条件：傅里叶逆 f(t)=(1/2π)∫F(ω)eⁱωᵗdω（需 F 绝对可积）；拉普拉斯逆 Bromwich 积分 f(t)=(1/2πi)∫F(s)eˢᵗds（γ>α 在收敛域右侧）；Z 逆 f[n]=(1/2πi)∮F(z)zⁿ⁻¹dz（围道在收敛域内）；生成函数 aₙ=G⁽ⁿ⁾(0)/n! 或 [xⁿ]G(x)。变换只是手段，最终答案必须在原空间。
+
+### 第七步：验证等价性
+确认变换未丢关键信息、逆变换结果确是原问题解：逆变换是否在收敛域内执行？原函数是否满足前置条件？边界/初值条件是否正确编码？等价性是变换思想的底线——变换必须是可验证的等价操作，而非近似或逃避。
+
+## 常见错误 / Common Errors
+
+| 错误 / Error | 批评 / Critique | 正确做法 / Correct Approach |
+|---|---|---|
+| 选择不合适的变换 | 没简化反而增加复杂度 | 根据问题特征选择变换 |
+| 用不可逆变换且丢信息 | 变换后无法回到原问题 | 检查可逆性或确认丢失信息不重要 |
+| 忘记逆变换 | 新空间得解后忘记变回来 | 最终答案必须在原空间 |
+| 变换后忽略定义域变化 | 变换可能改定义域或引入奇点 | 检查变换后定义域与边界条件 |
+| 把变换当魔法 | 变换只换表示，不改问题本质 | 变换是工具，理解力才是核心 |
+| 对不可积函数用傅里叶 | ∫|f|dt=∞ 时 F(ω) 可能无意义 | 先检查 Dirichlet，不满足用广义函数或 Laplace |
+| 忽略 Laplace 收敛域 | F(s) 仅 Re(s)>α 有定义 | 所有运算限制在收敛域内 |
+| 假设所有变换可逆 | 部分变换特定条件下不可逆（如非凸 Legendre） | 验证可逆条件后再逆变换 |
+| 忽略离散变换 DFT/FFT | 连续变换不适用离散数据 | 离散数据用 DFT: X[k]=Σx[n]e⁻ⁱ2πkn/N |
+| 变换后不可算/不可融合 | 变换需 O(n²) 全局重排或病态逆 | 评估 GPU 八维，必要时选可融合/数值稳定的等价变换 |
+
+## 操作规程 / Operating Procedure
+
+输出必须包含：
+
+1. **当前表示的困难**：`[困难]: [描述]`
+2. **变换选择**：`[变换]: [选择] 因 [理由]，公式 [公式]，收敛条件 [条件]，预期简化 [效果]`
+3. **变换执行**：变换后的问题形式
+4. **收敛与域条件验证**：`[验证]: [条件是否满足，收敛域]`
+5. **变换空间求解**：新表示下的解法
+6. **逆变换**：解翻译回原空间，注明逆变换条件
+7. **等价性验证**：是否可逆、是否丢信息、收敛域是否正确
+8. **[GPU 可行性]**（若用于算子设计）— 变换是否把运算变成 GEMM/可融合/数值稳定，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
+
+- **抽象化思想**：变换也是一种抽象——在新结构中表示同一对象。
+- **对称与不变性**：变换下保持不变的性质就是对变换的对称性。
+- **优化思想**：变换到对偶空间有时使优化更容易。
+- **建模思想**：变换常是模型求解的关键步骤。
+- **算法思想**：FFT 将 DFT 的 O(N²) 降为 O(N log N)，是变换思想在计算效率上的体现。
+- **信息论思想**：编码即变换——信源编码变换为高效表示，信道编码变换为抗干扰表示。
+- **现代数学激活**：`../../references/books/matrix-analysis.md`（谱/低秩变换）、`algebraic-geometry-rising-sea.md`（Plücker 坐标压缩 KV、热带门控）、`optimization-ml.md`（对偶变换）。