Evidence level: C
上一篇担心等分假设一破、天花板就被压得更低。把它算成闭式之后,结论要反过来写:那个假设从来不是承重件,真正控制一切的是一个比值——主档差有多大,相对于主档外面还剩多少空间。
正文
上一篇把整条天花板推导压在一个参数化上:两位判读者把主档(比如「无变化」)之外的质量各自等分给另外两档。等分带来一个漂亮的巧合——两条边际的总变差距离恰好等于主档占比差 δ,于是上限 1−δ 干净得能背下来。上一篇担心:真实判读里可能一位偏乐观、一位偏保守,次档分配不同向,等分一破,TVD 就大于 δ,上限被压得更低。
这个担心的方向对,但幅度算错了——而且错在「等分假设很脆弱」这个前提本身。
先把闭式解摆出来。三档裁决下,两条边际分布 p、q 的总变差距离有一个恒等式:TVD = max{|p₁−q₁|, |p₂−q₂|, |p₃−q₃|}。因为三档概率差相加为零,必有一档的差恰等于另两档之差的和,所以半 L1 范数就等于最大的那条边际差。这与参数化无关,是纯代数,随机取三百万组边际验证误差在 10⁻¹⁶ 量级。
第一个推论就解掉了上一篇一半的担心:TVD ≥ |p主−q主| = δ,恒成立。所以 p₀ 的上限 1−δ 从来不会被违反,等分假设根本不是这个上界的承重件。
第二个推论更硬:决定「TVD 会不会超过 δ」的,是主档那一档的差是不是三条差里最大的那条。等分假设只影响另两档,而主档差 δ 已经把地板焊死了。
再看方向性。同向错位(两人都把剩余的更多分给「改善」)完全无害:两人的偏移互相抵消,TVD 恒等于 δ,一分不动——枚举扫过 ε 从 0 到 0.4,TVD 一直是 0.0500。只有反向错位(一位偏乐观、一位偏保守)才有代价,而且有死区。设两人把剩余质量分给第二档的比例分别是 0.5+ε 与 0.5−ε,两侧剩余质量 s_A、s_B,可解出临界点:
ε* = δ / (2(s_A+s_B))
当 ε ≤ ε*,TVD 精确等于 δ,错位被主档差完全盖住;越过后 TVD 线性上升,斜率就是 s_A+s_B。
对称情形 m_A = m̄+δ/2、m_B = m̄−δ/2 时,ε* = δ / (4(1−m̄))。这个式子很反直觉:主档越集中,死区越宽,等分假设越安全。m̄=0.90、δ=0.05 时 ε*=0.125——两位读者对剩余质量的分配要相差 25 个百分点(一人 62.5%/37.5%,另一人 37.5%/62.5%),TVD 才开始动。而 m̄=0.70、δ=0.05 时 ε*=0.042,死区只剩 8.3%。
唯一的控制器是比值 r = δ/(1−m̄):主档差相对于主档外面剩余空间有多大。死区边界就是 ε* = r/4,死区占「最大可能错位 0.5」的比例则是 r/2。r=0.1 时只覆盖 5%(ε*=0.025),r=0.2 覆盖 10%,r=0.5 覆盖 25%,r=1 覆盖 50%,r=2 覆盖 100%(此时 ε*=0.5,任何偏离都安全)。整条推导只剩一个无量纲数。
越过死区之后,TVD 的形状也是线性的:令 t = ε/ε*,则 TVD/δ = (1+t)/2(t>1)。也就是说 TVD 相对 δ 的超出量只由 t 决定,与 δ 本身、与 m̄ 本身都无关——m̄ 与 δ 的全部作用都被吸收进 ε* 这一个尺度里。t=2 时 TVD 是 δ 的 1.5 倍,t=4 时是 2.5 倍。相应地 κ_max 的跌落:m̄=0.90、δ=0.05 下 t=1 时 0.731,t=2 时 0.599(跌 18.1%),t=3 时 0.470(跌 35.8%),t 到 4 顶到 ε=0.5 的理论极限,只剩 0.344(跌 52.9%)。注意这个「跌 18%」正是上一篇举出的那个数字——但它对应的是 t=2、即 ε 恰好是死区边界的两倍,不是「ε 稍大一点」。
回到上一篇的具体担心:ε 的典型量级若真是 0.02–0.05,在 m̄=0.90、δ=0.05 的死区(ε*=0.125)里被整个包住,t 只有 0.16–0.4,TVD 仍精确等于 δ,映射照用不误。要造出上一篇举的那个例子(A=(0.925,0.05,0.025)、B=(0.875,0.025,0.10)),得让 A 把剩余 0.075 里的 0.05 全押给第二档、B 把剩余 0.125 里的 0.10 押给第三档——两人的偏离量分别是 0.167 与 0.30,且方向相反。折算到对称参数族,这个例子等价于 t=2(TVD/δ=1.5),即 ε 恰好是死区边界的两倍——那是两位读者把剩余质量几乎反向推向两个极端,不是「有点风格差异」。
所以上一篇留下的四条判据需要整体改写。原判据把「ε<0.01」当作安全线、把「ε>δ 就说明主档差作代理失效」当作失效线,两条线的量纲都不对:ε 与 δ 不可直接比,必须除上剩余空间。正确形式是:r = δ/(1−m̄) ≤ 0.5 时(死区已覆盖 25% 的错位空间)等分假设可作工程近似;r 升高、ε 越过 r/4 才需要改用 TVD 报「两边际总变差距离」;只有当 ε 越过死区且「两人次档偏向确实系统相反」时,才谈得上「判的方向不同」而非「判得不准」。
最后是报告口径。既然 κ_max = 1 − TVD/(1−Σp_k q_k),而 TVD 不是 δ 的单调函数,「偏倚」这个词就该换。实测固定 (m̄, TVD) 后扫遍不同边际形状,κ_max 的跨度只有 0.0015–0.0078,远小于 0.05 的带宽——说明 (m̄, TVD) 这个二元组几乎是充分的,可替代单一 δ 作报告项。规范上不妨直接要求同报 m̄ 与 TVD,δ 退为参考。
这里有个容易读反的地方:κ_max 对 TVD 是单调下降的(p_o^max = 1−TVD 是恒等式,与边际形状无关),所以真正决定天花板的从头到尾都是 TVD,δ 只是 TVD 的一个下界和一个常用代理。等分假设的全部作用,是让这个下界变成等号。换句话说,上一篇问「δ 能不能当 TVD 用」,答案是:δ 永远是 TVD 的下界,等号条件只需 ε ≤ ε*,而这个条件在高集中区(m̄≥0.9)宽到几乎必然满足。
一句话收口:等分假设不承重,δ→上限的方向永远成立;数值上只需要一个比值 r = δ/(1−m̄),r ≤ 0.5 时放心用 δ,超出后按 TVD/δ = (1+t)/2 换算,t = ε/(r/4)。
证据与局限
已验证:三档下 TVD = 三条边际差绝对值的最大值,这是纯代数恒等式(随机三百万组边际,误差 10⁻¹⁶);TVD ≥ δ 恒成立,故 1−δ 上界无条件有效;同向错位下 TVD 精确等于 δ(ε 扫 0–0.4 不变);反向错位的临界点 ε* = δ/(2(s_A+s_B)) 与二分解实测逐点吻合;闭式 TVD = max(δ, δ/2+ε(s_A+s_B)) 全参数域误差 10⁻¹⁶;κ_max = 1 − TVD/(1−p_e) 由恒等式 p_o^max = Σmin(p,q) = 1−TVD 直接得出。
部分验证:把 δ 换成 (m̄, TVD) 二元组的充分性只是数值上「近似」成立(固定 (m̄,TVD) 后 κ_max 跨度 0.0015–0.0078),没有解析证明残差上界;死区宽度依赖对称参数化 m_A = m̄±δ/2,真实研究里两位读者的主档占比未必对称,非对称情形 ε* 会移动(主档差越大、动作越晚,方向不变)。文献侧只印证了「边际是 kappa 的独立自由度」这一大类结论(PMID 8501467 把 κ 分解成偏倚、流行度、观察一致三项;PMID 8621993 指出 κ 的上下限完全由边际决定;PMID 26095449 指出 κ 类扩展系数「prone to prevalence and marginal distribution issues」,并改用有序 GLMM 模型化路径;PMID 41230191 同样指出加权 κ 类系数对边际分布与类别流行度敏感),没有任何文献直接给过本文这条三档 TVD 闭式与死区。
待验证:真实三档裁决数据里 ε 的典型量级到底是多少——本文只能给出「若 ε≤0.05 则安全」的判别,没有实测分布;δ/(1−m̄) 在已发表研究中的实际落点未知;反向错位是否与读者资历、判读顺序、是否有回看机制相关,完全没查。
相关阅读
L1 网络状态:CORRESPONDENCE_NETWORK.md 边表止于 E-050(纳音↔频率,
L1 | open,现代证据栏原文「尚无任何可测来源」);网络对齐指标行原文「L2 以上边数 | 0」。三域节点表无「判读 / 评分者 / 边际分布 / 一致率」任何登记位(全表检索「判读 / 裁决 / 一致率 / 边际 / 偏倚 / 带宽 / TVD」零命中),物理侧「相位 / 节律」语义为「两振荡锁相」而非边际分布比较。故本文证据等级封顶 C,且该主题尚未进入 L1 对应网络(已登记至 l1-gaps)。证据等级: C
参考文献: PMID 8501467, PMID 8621993, PMID 2348207, PMID 25168681, PMID 26095449, PMID 41230191, PMID 37127220
免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。
The assumption that \”two people each take half of the remaining mass\” does not bear weight: there is a closed-form solution for the total variation distance across three tiers,…
Evidence level: C
In the previous discussion, there was concern that breaking the equal-split assumption would depress the ceiling further. By deriving a closed-form expression, the conclusion must be inverted: that assumption never bears structural weight; what truly controls everything is a ratio—the magnitude of the primary class difference relative to the remaining space outside the primary class.
Body
The previous derivation compressed the entire ceiling argument into a single parameterization: two raters split the mass equally between the other two classes, excluding the primary class (e.g., “No Change”). This equal-split assumption yields a convenient coincidence—the sum of the total variation distances at the two margins exactly equals the primary class proportion difference δ, making the upper bound 1−δ clean enough to memorize. The previous concern was: in real raters, one might be optimistic and the other conservative, splitting the secondary classes differently; if equal-split fails, TVD exceeds δ, and the ceiling is depressed further.
This direction of concern is correct, but the magnitude calculation was wrong—and the error lies in the premise itself that “the equal-split assumption is fragile.”
Let us first present the closed-form solution. Under three-class adjudication, the total variation distance between two marginal distributions p and q has an identity: TVD = max{|p₁−q₁|, |p₂−q₂|, |p₃−q₃|}. Because the sum of probability differences across three classes is zero, exactly one class difference equals the sum of the other two; thus, half the L1 norm equals the largest marginal difference. This result is independent of parameterization and purely algebraic; random sampling of three million marginal pairs verified the error at 10⁻¹⁶ magnitude.
The first corollary resolves half of the previous concern: TVD ≥ |p_main−q_main| = δ, always holds. Therefore, the upper bound 1−δ for p₀ is never violated; the equal-split assumption is not the load-bearing component of this bound.
The second corollary is stronger: whether “TVD exceeds δ” depends on whether the primary class difference is the largest among the three differences. The equal-split assumption affects only the other two classes, while the primary class difference δ already welds the floor shut.
Now consider directionality. Co-directional misalignment (both raters assign more of the remainder to “Improvement”) is entirely harmless: their offsets cancel each other, TVD remains exactly δ, unchanged—enumerating ε from 0 to 0.4 yields TVD consistently at 0.0500. Only counter-directional misalignment (one optimistic, one conservative) incurs cost, and it has a dead zone. Suppose the two raters allocate the remainder mass to the second class in proportions 0.5+ε and 0.5−ε, with remaining masses s_A and s_B; solving yields the critical point:
ε* = δ / (2(s_A+s_B))
When ε ≤ ε*, TVD exactly equals δ, with misalignment completely masked by the primary class difference; beyond this, TVD rises linearly with slope s_A+s_B.
In the symmetric case m_A = m̄+δ/2, m_B = m̄−δ/2, we obtain ε* = δ / (4(1−m̄)). This formula is counterintuitive: the more concentrated the primary class, the wider the dead zone, and the safer the equal-split assumption. With m̄=0.90, δ=0.05, ε*=0.125—raters must differ by 25 percentage points in remainder allocation (one at 62.5%/37.5%, the other at 37.5%/62.5%) before TVD moves. Conversely, with m̄=0.70, δ=0.05, ε*=0.042, the dead zone shrinks to 8.3%.
The sole controller is the ratio r = δ/(1−m̄): how large the primary class difference is relative to the remaining space. The dead zone boundary is ε* = r/4, and the dead zone occupies r/2 of the “maximum possible misalignment 0.5”. When r=0.1, it covers only 5% (ε*=0.025); r=0.2 covers 10%; r=0.5 covers 25%; r=1 covers 50%; r=2 covers 100% (ε*=0.5, so any deviation is safe). The entire derivation reduces to this single dimensionless number.
Beyond the dead zone, TVD’s shape remains linear: let t = ε/ε*, then TVD/δ = (1+t)/2 (t>1). That is, the excess of TVD over δ depends only on t, not on δ itself or m̄ itself—all effects of m̄ and δ are absorbed into the single scale ε*. When t=2, TVD is 1.5 times δ; when t=4, it is 2.5 times δ. Correspondingly, the drop in κ_max: with m̄=0.90, δ=0.05, t=1 gives 0.731, t=2 gives 0.599 (drop 18.1%), t=3 gives 0.470 (drop 35.8%); at t=4 (ε=0.5, the theoretical limit), it drops to 0.344 (drop 52.9%). Note that this “18% drop” matches the number cited previously—but it corresponds to t=2, i.e., ε exactly twice the dead zone boundary, not “ε slightly larger.”
Returning to the specific concern: if ε’s typical scale is indeed 0.02–0.05, it lies entirely within the dead zone (ε*=0.125) for m̄=0.90, δ=0.05, so t ranges from 0.16 to 0.4, and TVD remains exactly δ; the mapping holds without error. To construct the example cited previously (A=(0.925,0.05,0.025), B=(0.875,0.025,0.10)), rater A must assign all 0.05 of the remainder 0.075 to the second class, while rater B assigns all 0.10 of the remainder 0.125 to the third class—their deviations are 0.167 and 0.30, respectively, in opposite directions. In terms of the symmetric parameter family, this example is equivalent to t=2 (TVD/δ=1.5), i.e., ε exactly twice the dead zone boundary—meaning the raters pushed the remainder mass almost oppositely toward two extremes, not “a slight stylistic difference.”
Thus, the four criteria left by the previous discussion require comprehensive revision. The original criteria treated “ε<0.01” as the safety line and “ε>δ” as the failure line, but both have mismatched dimensions: ε and δ cannot be directly compared; they must be divided by the remaining space. The correct form is: when r = δ/(1−m̄) ≤ 0.5 (the dead zone covers 25% of misalignment space), the equal-split assumption serves as an engineering approximation; when r increases and ε exceeds r/4, one must switch to TVD reporting “the sum of marginal total variation distances”; only when ε crosses the dead zone and “the two raters’ secondary class biases are systematically opposite” can we speak of “different adjudication directions” rather than “inaccurate adjudication.”
Finally, reporting format. Since κ_max = 1 − TVD/(1−Σp_k q_k) and TVD is not a monotonic function of δ, the term “bias” should be replaced. After fixing (m̄, TVD) and sweeping different marginal shapes, κ_max spans only 0.0015–0.0078, far narrower than the 0.05 bandwidth—indicating (m̄, TVD) is nearly sufficient as a pair and can replace single δ as the report item. Normatively, we may directly require reporting both m̄ and TVD, with δ relegated to reference.
A point easy to misread: κ_max monotonically decreases with TVD (p_o^max = 1−TVD is an identity, independent of marginal shape), so what truly determines the ceiling throughout is TVD; δ serves only as a lower bound and common proxy for TVD. The full role of the equal-split assumption is to make this lower bound an equality. In other words, the previous question “Can δ serve as TVD?” has the answer: δ is always the lower bound of TVD; the equality condition requires ε ≤ ε*, and this condition holds almost certainly in high-concentration regions (m̄≥0.9), where the dead zone is wide.
In one sentence: the equal-split assumption does not bear weight; the direction δ→upper bound always holds; numerically, only one ratio r = δ/(1−m̄) matters: use δ confidently when r ≤ 0.5; beyond that, convert via TVD/δ = (1+t)/2, where t = ε/(r/4).
Evidence and Limitations
Verified: Under three classes, TVD equals the maximum of the absolute differences of the three marginals, a pure algebraic identity (randomly sampled three million marginal pairs yielded error 10⁻¹⁶); TVD ≥ δ always holds, so the upper bound 1−δ is unconditionally valid; under co-directional misalignment, TVD exactly equals δ (ε swept 0–0.4 remained constant); the critical point for counter-directional misalignment ε* = δ/(2(s_A+s_B)) matches empirical pointwise verification; the closed-form TVD = max(δ, δ/2+ε(s_A+s_B)) has error 10⁻¹⁶ across the full parameter domain; κ_max = 1 − TVD/(1−p_e) follows directly from the identity p_o^max = Σmin(p,q) = 1−TVD.
Partially Verified: The sufficiency of replacing δ with the (m̄, TVD) pair holds numerically as an “approximation” (fixing (m̄, TVD) yields κ_max spanning 0.0015–0.0078), but lacks analytical proof of residual upper bounds; the dead zone width depends on symmetric parameterization m_A = m̄±δ/2, whereas in real studies the primary class proportions of two raters may not be symmetric, so in asymmetric cases ε* shifts (larger primary class difference delays the onset of TVD increase, without changing direction). Literature only corroborates the general conclusion that “marginal distribution is an independent degree of freedom for kappa” (PMID 8501467 decomposes κ into bias, prevalence, and marginal agreement; PMID 8621993 states κ’s bounds are entirely determined by marginals; PMID 26095449 notes κ-type extension coefficients are “prone to prevalence and marginal distribution issues” and adopts ordered GLMM modeling; PMID 41230191 similarly highlights sensitivity of weighted κ-type coefficients to marginal distribution and category prevalence), but no literature directly provides this paper’s three-class TVD closed-form and dead zone.
Pending Verification: What is the typical scale of ε in real three-class adjudication data—this paper can only state “if ε≤0.05, it is safe,” without empirical distribution; the actual placement of δ/(1−m̄) in published studies remains unknown; whether counter-directional misalignment correlates with rater experience, adjudication order, or presence of review mechanisms has not been examined.
Related Reading
- Two raters’ primary class proportion difference of five points widens not the bandwidth but the ceiling
- How large is the δ skipped entirely in the derivation
- Bandwidth under 0.05 does not imply comparability
- Resonance Consciousness Field Research Pool
L1 Network Status: The CORRESPONDENCE_NETWORK.md edge table ends at E-050 (Na Yin ↔ Frequency,
L1 | open); the modern evidence column reads “No measurable source yet“; the network alignment indicator row reads “Edges ≥ L2 | 0“. The three-domain node table lacks any registration for “adjudication / rater / marginal distribution / agreement rate” (full table search for “adjudication / decision / agreement rate / marginal / bias / bandwidth / TVD” yields zero hits); the physical side “phase / rhythm” semantics refers to “two oscillators phase-locked,” not marginal distribution comparison. Thus, this paper’s evidence level caps at C, and the topic has not yet entered the L1 corresponding network (registered to l1-gaps).Evidence Level: C
References: PMID 8501467, PMID 8621993, PMID 2348207, PMID 25162681, PMID 26095449, PMID 41230191, PMID 37127220
Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.
Leave a Reply