Skip to content

[信息几何/判别] rank-sector K-only 信息损失:MLR、Wasserstein、Bayes threshold 与渐近无力化 #789

Description

@LightChainr

Exact finite theorem

对固定 cardinality k,R_k 随插入单调,因此

q0(k)=P(R_k=0) nonincreasing,
q2(k)=P(R_k=2) nondecreasing.

故在共同支撑上

C[2,k]/C[0,k]=q2(k)/q0(k)

随 k 单调增加。于是任意 Bernoulli(p) 下

K | rank=2 >=_MLR K | rank=0.

立即有 stochastic dominance,且

W1(mu2,mu0)=E[K|2]-E[K|0]=g1=-2 db/dz.

在 balance root P0=P2,用 K 区分 rank0 vs rank2 的 Bayes-optimal classifier必是单 threshold;最佳错误率

err_K=(1-TV(mu0,mu2))/2.

任何更复杂的 K-only nonlinear score都不能更好。

条件 scaling conjecture

near critical 当前程序给 g1~L^(3/4),而 raw Bernoulli sd(K)~L。若两个条件 K law 有共同 leading Gaussian width,则 standardized separation

delta_K ~ L^-1/4,
TV ~ L^-1/4,
KL/Hellinger^2 ~ L^-1/2,
I(rank;K) ~ L^-1/2.

因此绝对均值差越变越大,但仅靠总 occupancy K 的归一化拓扑信息反而趋零

#578 的接口

这给 conditioning hierarchy 一个理论排序:K-only / rank+K 可以有有限尺寸收益,但若上述 scaling 成立,想保持非消失的 topological variance reduction 必须保留 cut/connectivity geometry。

交付

不新增 sampling。相关:#775, #578, draft #773

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    priority:P2Deferred research or on-demand support; no default new compute allocation.

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions