20260625:很多新内容
This commit is contained in:
49
concepts/gaussian-width.md
Normal file
49
concepts/gaussian-width.md
Normal file
@@ -0,0 +1,49 @@
|
||||
---
|
||||
title: "Gaussian Width (高斯宽度)"
|
||||
created: 2026-06-23
|
||||
updated: 2026-06-23
|
||||
type: concept
|
||||
tags: ["high-dimensional-probability", "convex-geometry", "complexity-measure", "learning-theory"]
|
||||
sources: ["https://arxiv.org/abs/2606.18306"]
|
||||
---
|
||||
|
||||
# Gaussian Width (高斯宽度)
|
||||
|
||||
**Gaussian width** 是高维概率论和凸几何中的核心复杂度度量。对于集合 T ⊂ ℝᵈ,定义为:
|
||||
|
||||
```
|
||||
w(T) = E_{g∼N(0,I_d)} [sup_{v∈T} ⟨g, v⟩]
|
||||
```
|
||||
|
||||
## 直觉
|
||||
|
||||
- 以**随机高斯方向**探测集合 T,取其最大投影,再对随机方向取期望
|
||||
- 大宽度 → 集合在高维空间中"覆盖广" → 复杂度高
|
||||
- 小宽度 → 集合集中在小范围 → 复杂度低
|
||||
|
||||
## 关键性质
|
||||
|
||||
1. **单调性**:T₁ ⊆ T₂ ⇒ w(T₁) ≤ w(T₂)
|
||||
2. **齐次性**:w(aT) = |a|·w(T)
|
||||
3. **凸包不变**:w(conv(T)) = w(T)
|
||||
4. **次可加性**:w(T₁+T₂) ≤ w(T₁)+w(T₂)
|
||||
|
||||
## 在机器学习中的角色
|
||||
|
||||
Gaussian width 与 [[rademacher-complexity|Rademacher 复杂度]]等价(常数级),是假设类泛化能力的核心度量:
|
||||
|
||||
- **压缩感知** (Chandrasekaran et al., 2012):描述恢复相变
|
||||
- **凸优化** (Amelunxen et al., 2014):统计维度的几何刻画
|
||||
- **经验过程** (Bartlett & Mendelson, 2002):控制一致偏差
|
||||
|
||||
## 局限性
|
||||
|
||||
Gaussian width 本质上是**欧几里得**的——所有方向等权看待。当参数空间携带非平凡黎曼度量时(如统计模型中的 Fisher 信息度量),欧几里得宽度无法捕捉方向的统计敏感性差异。
|
||||
|
||||
[[fisher-width|Fisher Width]] 将 Gaussian width 推广到[[statistical-manifold|统计流形]]上。
|
||||
|
||||
## 参考
|
||||
|
||||
- [[statistical-manifold|Statistical Manifold]]
|
||||
- [[fisher-width|Fisher Width]]
|
||||
- [[generalization-bounds|Generalization Bounds]]
|
||||
Reference in New Issue
Block a user