提出新型几何不变量,可精准区分三值阈值函数的复杂性。
The Banach-Butterfly Invariant: Influence-Adaptive Walsh Geometry for Ternary Polynomial Threshold Functions
- 基于沃尔什-哈达玛蝴蝶分解,引入影响自适应的巴拿赫几何结构。
- 该不变量在n=4时对65,536个布尔函数给出精确最小支持度证明,均值6.42。
- 其与最小支持度相关性随维度变化,适合研究函数复杂性与结构特征。
我们提出巴拿赫-蝴蝶不变量(BBT),一种在沃尔什-哈达玛蝴蝶分解上的影响自适应巴拿赫几何。对于布尔函数 $f:\{-1,+1\}^n\to\{-1,+1\}$,其坐标影响为 $\mathrm{Inf}_\ell(f)$,BBT 在蝴蝶层 $\ell$ 分配指数 $p_\ell = 1+\mathrm{Inf}_\ell(f)$,生成收缩不变量 $μ(f)=\prod_\ell 2^{-\mathrm{Inf}_\ell/(1+\mathrm{Inf}_\ell)}$。我们证明了 $\log_2μ(f) \ge -I(f)/(1+I(f)/n)$,且 $μ$ 在影响向量上严格舒尔凸(模置换),给出不同尺度类:$μ\sim 2^{-n/2}$(奇偶性)、$2^{-Θ(\sqrt{n})}$(多数函数)、$2^{-1/2}$(字典函数)。$\log_2μ$ 为有理数但非傅里叶系数的多项式,而 $μ$ 为代数数;它能区分总影响相同函数($n=3$ 时有122对)。利用伴侣合成论文中 $n \le 4$ 的三值沃尔什阈值全集作为有限测试平台,我们计算出所有 65,536 个 $n=4$ 布尔函数的精确 MILP 最小支持度证书(平均 6.42,最大 9,全为奇数,由奇偶性论证)。在 $n=5$ 的 10,000 个共轭等价代表元中也完成枚举(匹配 OEIS A000370)。在 $n=4$ 的最大层中,条件斯皮尔曼相关性 $ρ(μ,|\mathrm{supp}|)$ 固定总影响下为 +0.571,但在 $n=5$ 两种采样方式下均反转为 -0.38:表明 $μ$ 是有效的舒尔凸集中不变量,而非跨维度的最小支持度通用预测器。伴侣应用论文验证了受此理论启发的实值 WHT 激活能代理,在五个预训练大模型(W2A16)上降低维基文本-2 熵 15%-58%,优于原始自动量化;该迁移为定性而非形式化。
原文摘要 · Abstract (English)
We introduce the Banach-Butterfly Invariant (BBT), an influence-adaptive Banach geometry on the Walsh-Hadamard butterfly factorization. For a Boolean function $f:\{-1,+1\}^n\to\{-1,+1\}$ with coordinate influences $\mathrm{Inf}_\ell(f)$, BBT assigns exponent $p_\ell = 1+\mathrm{Inf}_\ell(f)$ to butterfly layer $\ell$, yielding the contraction invariant $μ(f)=\prod_\ell 2^{-\mathrm{Inf}_\ell/(1+\mathrm{Inf}_\ell)}$. We prove a Jensen lower bound $\log_2μ(f) \ge -I(f)/(1+I(f)/n)$ and that $μ$ is strictly Schur-convex in the influence vector (modulo permutation), giving scaling classes $μ\sim 2^{-n/2}$ (parity), $2^{-Θ(\sqrt{n})}$ (majority), $2^{-1/2}$ (dictators). $\log_2μ$ is rational but not polynomial in the Fourier coefficients while $μ$ is algebraic, and $μ$ separates functions with identical total influence (122 pairs at $n=3$). Using the certified $n \le 4$ ternary Walsh-threshold universe from a companion synthesis manuscript as a finite testbed, we compute exact MILP minimum-support certificates for all 65,536 Boolean functions at $n=4$ (mean 6.42, max 9, all-odd by a parity argument) and on 10,000 of the 616,126 NPN-canonical representatives we enumerate at $n=5$ (matching OEIS A000370). Conditional Spearman $ρ(μ,|\mathrm{supp}|)$ at fixed total influence is $+0.571$ in the largest stratum at $n=4$ but reverses to $-0.38$ at $n=5$ under both function-uniform and NPN-canonical sampling: $μ$ is a valid Schur-convex concentration invariant, not a universal monotone predictor of minimum support across $n$. A companion application paper validates a real-valued WHT activation-energy proxy inspired by this theory on five pretrained LLMs at W2A16, cutting wikitext-2 perplexity by 15-58% versus vanilla auto-round; the transfer from Boolean theory to the real-valued proxy is qualitative, not formal.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品;中文卡片由大模型生成,请以原文为准。