Category: 科学易学

  • 存量判读流程的 ρ_cs 实测可行性审计:条款里的参数,台账里到底采不采得到

    Evidence level: C

    一条规则写了「这个参数必须先实测再用」,但如果历史记录里根本找不到能算出这个参数的配对数据,这条规则就不是安全带,而是纸面装饰——审计配对可行性,是让参数从纸面落到实测的第一步。

    正文

    先把问题说白。 前面的文章(承接 mse-balance-winsorize-threshold-cost-pareto-frontier 一线的仿真)得出过一个结论:当「标定端的偏差」和「统计端的噪声」不是各自独立、而是共享同一个来源(比如同一天、同一位判读者的状态)时,两者的相关系数 ρ_cs 一旦大于零,判读流程的安全带就得收紧,而收紧多少取决于 ρ_cs 的具体取值。所以条款里加了一句前提:「ρ_cs 已知或已标定,未标定者按最坏档执行」。

    这句话听起来很严谨,但它藏着一个更根本的问题:ρ_cs 这个数,到底从哪来?

    要估出「标定数据与正式判读数据之间的相关」,你必须同时拿到两样东西,而且它们要能配上对:同一批次的标定记录(比如标定时测的偏差量),和同一批次的正式判读数据(比如判读时统计端的尺度量)。两个数字都来自同批次,才谈得上算相关。这就是测量科学里最经典的结构——配对测量。信度研究的常识是:任何信度类指标(组内相关、重复性系数)都建立在「同一对象测两次」的配对结构上,配对一旦缺失,相关就无从谈起(PMID 33767329)。而现代方法学还有更进一步的要求:当上游输入本身带着不确定性时,推断必须把这份不确定性显式计入,而不是假装上游是确定的(PMID 37580957)。ρ_cs 恰恰就是那个「上游」——它自己没被估出来之前,下游所有按它收紧的条款都悬在空中。

    于是审计成了第一道工序。 这不是去新采数据,而是回头翻旧账:近 12 个月的判读台账里,有多少批次做到了「标定记录与正式判读数据同批同源、可逐批配对」?配上的那些批次,每批的配对样本量有多大?这两个数——配对率和中位配对 n——共同决定了 ρ_cs 实测路线是否可行。

    打个比方:你立了家规「开车前必须看胎压」,但家里根本没有胎压计,也没有打算买。这条家规写再多遍,也是空话。审计配对可行性,就是先清点家里到底有没有胎压计、有几只、能不能用——清点结果决定这条家规是保留、补设备,还是改写成「没胎压计就按最坏胎压算」。

    判据怎么设。 审计结果落进三个分支:若可配对批次占比达到 60% 以上、且配对批次的中位样本量达到 20 以上,则 ρ_cs 的实测路线可行——直接用存量配对数据估出相关系数,条款引用实测值,纸面参数就此落地。这相当于设备间一致性研究里的标准做法:两台设备在相同对象上配对读数,用配对差值给出一致性界限(PMID 38381465;同类低成本设备信度实测的可行性验证见 PMID 42404360)。若可配对比例不到 30%,说明问题不在统计、在台账结构本身——标定元数据从来没被强制同批记录,此时唯一出路是先改台账(强制同批记录标定信息),在 ρ_cs 补上实测之前,一律按最坏档收紧执行,宁可保守,不可裸奔。介于两者之间的批次,按部分配对数据给出初步估计,同时标注其置信宽度。

    这条审计还有一层容易被忽略的价值:它把「方法学条款的可执行性」变成了一个可以被测量的量。规则写得再漂亮,没有数据供给就是死条款;而配对率、中位配对 n 这两个审计指标,本身就是台账质量的体检表。体检不及格,改的不是条款,是数据采集流程——这也是整条研究线反复出现的模式:参数之争最终都落到数据结构之争。

    该主题尚未进入 L1 对应网络(CORRESPONDENCE_NETWORK / EDGE_TABLE 中无「标定-判读配对结构 × 参数可估性」对应边;最近邻为统计方法学类旁支),本文按证据薄弱口径撰写,只讨论方法学框架,不宣称任何已验证机制。

    证据与局限

    已验证: 配对测量结构是信度与一致性指标的前提,缺失配对则相关不可估(PMID 33767329;设备配对一致性实践 PMID 38381465、42404360);上游不确定性显式计入推断是现代方法学要求(PMID 37580957)。

    部分验证: 回顾性台账审计作为参数可行性的前置工序,在临床测量与设备比对领域有成熟先例,但「配对率 60% / 中位 n 20」这一具体门槛是本系列设定的工程判据,非普适常数。

    待验证: 本判读流程自身台账的实际配对率与配对样本量分布尚无实测数据;改台账结构后配对率能提升到什么水平,也是开放问题。ρ_cs 实测值本身的分布形态,须等审计落地后另行研究。

    相关阅读

    证据等级: C
    参考文献: PMID 33767329, PMID 37580957, PMID 38381465, PMID 42404360

    L1 网络状态:CORRESPONDENCE_NETWORK/EDGE_TABLE 中无「标定-判读配对结构 × 参数可估性」对应边,该主题尚未进入 L1 对应网络(纯统计方法学课题,最近邻边 E-061 呼吸节律为 L2/open 级)。

    证据等级:A=有PubMed直接支撑,B=科学类比/间接证据,C=传统文化共识/初步探索


    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。

    Feasibility Audit of In-Stock Judgment Process ρ_cs Measurement: Can Parameters Specified in Clauses Actually Be Collected from the Ledger?

    Evidence level: C

    Audit of Feasibility for Measuring ρ_cs in the Stock Assessment Procedure: Are Parameters Actually Captured from Logs or Just Listed?

    A rule stating “this parameter must be measured first before use” becomes not a safety guard but mere paper decoration if no paired historical data exists to calculate it; auditing pairing feasibility is the first step in moving parameters from paper into actual measurement.

    Body

    State the problem plainly. The preceding article (continuing the mse-balance-winsorize-threshold-cost-pareto-frontier simulation line) reached a conclusion: when the “calibration-end bias” and the “statistical-end noise” do not arise independently but share a common source (e.g., the same day, the same judge’s state), their correlation coefficient ρ_cs, once greater than zero, requires tightening of the safety margin for the judgment process, with the degree of tightening dependent on the specific value of ρ_cs. Consequently, the clause includes a prerequisite: “ρ_cs is known or calibrated; if not calibrated, execute using the worst-case tier.”

    This statement sounds rigorous, yet it conceals a more fundamental question: where does the number ρ_cs come from?

    To estimate the “correlation between calibration data and formal judgment data,” you must simultaneously obtain two items that can be paired: the same-batch calibration records (e.g., the bias amount measured during calibration) and the same-batch formal judgment data (e.g., the scale quantity from the statistical end during judgment). Both numbers must originate from the same batch to compute a correlation. This represents the classic structure of measurement science—paired measurement. A common sense of reliability research is that any reliability-type indicator (intraclass correlation, repeatability coefficient) is built upon this “same object measured twice” paired structure; without the pair, correlation cannot be discussed (PMID 33767329). Furthermore, modern methodology imposes additional requirements: when the upstream input itself carries uncertainty, inference must explicitly account for this uncertainty rather than pretending the upstream is certain (PMID 37580957). ρ_cs is precisely that “upstream” component; until it is estimated, all downstream clauses tightened according to it remain suspended.

    Thus auditing becomes the first step. This is not about collecting new data but reviewing old accounts: among the judgment ledgers over the past 12 months, how many batches achieved “calibration records and formal judgment data from the same batch and source, allowing batch-by-batch pairing”? For those paired batches, what is the sample size per pair? These two figures—the pairing rate and the median paired n—jointly determine whether the ρ_cs measurement route is feasible.

    To illustrate: you establish a family rule “check tire pressure before driving,” but your home lacks a tire gauge and no intention to buy one. No matter how often you repeat this rule, it remains empty talk. Auditing pairing feasibility is first to count whether your home actually has tire gauges, how many, and whether they can be used—the result determines whether this rule is retained, equipment supplemented, or rewritten to “if no tire gauge, calculate using the worst-case tire pressure.”

    How to set the criteria. Audit results fall into three branches: if the proportion of pairable batches reaches over 60% and the median sample size of paired batches exceeds 20, then the ρ_cs measurement route is feasible—directly estimate the correlation coefficient using existing paired data, and the clause cites the measured value, grounding the paper parameter. This corresponds to the standard practice in inter-device consistency studies: pair readings from two devices on the same object, using paired differences to define consistency limits (PMID 38381465; feasibility validation of reliability measurement for similar low-cost devices is documented in PMID 42404360). If the pairable proportion falls below 30%, the issue lies not in statistics but in the ledger structure itself—the calibration metadata was never forced to be recorded in the same batch. The only path forward is first to modify the ledger (force calibration information into same-batch records); until ρ_cs is supplemented with measured values, all must be tightened conservatively, preferring caution over running naked. Batches between these extremes should yield preliminary estimates based on partial paired data, while explicitly marking their confidence width.

    This audit also holds a layer of value often overlooked: it transforms “the executability of methodological clauses” into a measurable quantity. Rules written beautifully remain dead clauses without data supply; meanwhile, the pairing rate and median paired n—these audit metrics themselves serve as health checks of ledger quality. If the health check fails, what needs revision is not the clause but the data collection process—this pattern recurs throughout this research line: disputes over parameters ultimately resolve into disputes over data structure.

    This topic has not yet entered the L1 corresponding network (no “calibration-judgment pairing structure × parameter estimability” edge exists in CORRESPONDENCE_NETWORK / EDGE_TABLE; the nearest neighbor is a statistical methodology branch). Accordingly, this article is written according to the weak-evidence standard, discussing only the methodological framework without claiming any verified mechanism.

    Evidence and Limitations

    Verified: Paired measurement structures are a prerequisite for reliability and consistency metrics; without pairing, correlation cannot be estimated (PMID 33767329; equipment pairing consistency practices PMID 38381465, 42404360); explicitly accounting for upstream uncertainty in inference is a requirement of modern methodology (PMID 37580957).

    Partially Verified: Retrospective ledger audits serve as a prerequisite step for parameter feasibility and have established precedents in clinical measurement and equipment comparison, but the specific threshold of “60% pairing rate / median n 20” is an engineering criterion set for this series, not a universal constant.

    Pending Verification: The actual pairing rate and distribution of paired sample sizes for this interpretation process itself lack empirical data; what level the pairing rate could reach after restructuring the ledger remains an open question. The distribution shape of the ρ_cs measured values itself must be studied separately once the audit is implemented.

    Related Reading

    Evidence Level: C
    References: PMID 33767329, PMID 37580957, PMID 38381465, PMID 42404360

    L1 Network Status: No correspondence edge exists in the CORRESPONDENCE_NETWORK/EDGE_TABLE linking “calibration–interpretation pairing structure × parameter estimability” to this topic; consequently, it has not yet entered the L1 correspondence network (a pure statistical methodology topic; the nearest edge E-061, respiratory rhythm, is classified as L2/open level).

    Evidence Level: A = direct PubMed support; B = scientific analogy or indirect evidence; C = traditional consensus or preliminary exploration.


    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.

  • 校准到底能补回多少分:跨设备识别率衰减的「恢复幅度」该怎么量,以及质量门控和校准是不是加法

    Evidence level: C

    跨设备识别率掉下来之后,补回多少才叫「校准有效」,这件事目前没有公认的度量口径;而且校准增益与信号质量门控增益极可能不是加法关系,把它们相加是这套方案里最容易犯的错。

    正文

    上一篇我们讨论了衰减曲线是直线还是指数。这一篇问的是更实际的第二层问题:掉下去的分,校准能捞回来多少? 以及——把信号质量差的数据先剔掉,是不是能额外捞一点?这两笔收益能相加吗?

    先说为什么这个问题不能靠「直觉上应该有用」回答。跨设备识别力的损失来自两类完全不同的原因:一类是系统性偏置(同一生理状态在两台设备上被记成不同的数值、不同的相位、不同的形态),另一类是随机污染(运动伪迹、佩戴松动、灌注不足造成的坏段)。前者理论上可被校准链纠正,后者理论上应被质量门控剔除。但两个「理论上」叠在一起时,并不能推出「两者相加就等于全部损失」。原因很直白:质量门控是先剔样本,校准是在剩余样本上做映射。剔完之后剩下的是「好数据」,而好数据的跨设备差异本来就小——校准要纠正的偏置也就小了。所以门控做得好,校准的边际收益反而下降;两者高度重叠,不是正交的两笔账。

    现有文献能提供什么量级参考。 个体识别的长期研究给出一个很说明问题的数字:同一会话内高质量信号的等错误率约 1%,接近真实使用条件时升到 8%,而跨天数采集可升到 23.2%(PMID 29751681)。这个跨度说明「非理想条件」可以吃掉绝大部分识别力,但该研究并未施加任何参考样本链校准,因此它只标定了损失上限,不回答能补回多少。更近的系统性评估给出的结论更冷:多项 PPG 生物识别方法在实验室数据集上表现良好,进入真实环境后效能显著下滑,且现有工作普遍缺未见用户测试与时间分离(PMID 41471580);PPG 生物识别综述同样点名「缺标准化数据集」限制了跨人群、跨真实条件的泛化(PMID 41322669)。换句话说:跨设备校准的恢复幅度,在个体识别任务上目前还是空白,没有现成数字可抄。

    可借的间接证据在「模型层面重校准」这一支。把 ECG 训练的睡眠分期模型用 PPG 数据微调权重,准确率提升 1–9 个百分点、Cohen’s kappa 提升 0.01–0.13(PMID 33761477)——这是轻校准的量级:只动模型参数、不动信号。另一项工作把 ECG 上学到的心律特征迁移到腕表 PPG,在未见过的 PPG 数据上仍达 95.1% 准确率(PMID 34770543),量级明显更高,但注意它做的是房颤 vs 正常窦律的二分类,不是个体识别——任务难度不同,数字不可横比。这两条合并起来提示一个方向:校准收益很可能不是线性的,轻校准拿到大头、重校准边际递减;但这只是提示,不是证据。

    质量门控这一笔账,倒是有硬数字。 一套针对腕表 PPG 的脉冲形态质量指数方案,在阈值筛选后正常窦律心搏的阳性预测值达 98%–99%,对窦律心搏的敏感度保持在 91%–96%(PMID 30475748)——即门控本身是准的,剔得掉坏段、留得住好段。而门控带来的下游收益在 ECG 侧有直接测量:给缺血报警系统加信号质量门控后,精确率从 0.31 升到 0.58,但召回从 0.78 掉到 0.77(PMID 27576238)。这是本文最该记住的一组数字:门控不是免费的,它用一点召回换一大截精确率。 而领域专家共识给出门槛的形状:记录若 26%–50% 以上的心搏带伪迹就该整段拒收,形态学分析要求 86%–95% 的心搏达到可诊断质量、且至少 50 个无畸变脉冲(PMID 34713077)。共识给的是区间不是点值,说明阈值该按任务定,不能一套通吃。

    所以这套实验该怎么设计。 关键是把「校准」和「门控」做成两个独立可开关的因子,做 2×2 四臂对照:无校准无门控 / 只门控 / 只校准 / 两者都做,四臂用同一批受试者、同一批原始数据,只改后处理。只有这样才能回答「是不是加法」——若 ΔAUC(两者都做) 明显小于 ΔAUC(只校准)+ΔAUC(只门控),就是重叠;若接近相等,才是可加。校准深度再拆三档(无 / 轻:只做线性或分位映射 / 重:每设备每访次重建参考样本链),量出边际收益曲线,判断该在哪一档停手。

    两个必须写进预注册的报数纪律。 第一,AUC 的分母会变:门控剔掉样本后,识别率的样本集与未门控臂不同,两组 AUC 直接横比是无效比较,必须同时报「剔除了多少比例的样本」以及「在同一子集上重算的 AUC」。第二,禁止各自取最优点比较:门控阈值和校准深度若各自在自己的数据上优化到最佳再对比,等于给两个臂不同的自由度,结论会系统性偏向「两者都做」。正确做法是阈值与校准档位都在训练侧定好、测试侧冻结。

    最后提醒一句证据边界。上面所有数字都来自 PPG/ECG 的临床或生物识别任务,没有一条来自「四路频率签名」的跨设备迁移。它们能告诉我们「量级大概在什么范围」「门控有代价」「重叠可能存在」,但不能告诉我们四路签名上的恢复幅度是多少。这套方案在当前阶段仍是方法学提案,不构成任何「校准后签名可跨设备使用」的结论。

    证据与局限

    已验证: 信号质量门控在腕表 PPG 上能准确筛出正常窦律心搏(PPV 98%–99%,PMID 30475748);门控能显著提升下游报警精确率但同时略降召回(0.31→0.58 / 0.78→0.77,PMID 27576238);PPG 生物识别的长期稳定性在跨天条件下大幅下降(EER 1%→23.2%,PMID 29751681);模型层面轻校准可带来 1–9 个百分点准确率提升(PMID 33761477)。

    部分验证: 「校准收益非线性、轻校准拿大头」目前只有模型迁移类研究的旁证,且任务类型与个体识别不同,只能作趋势提示;门控阈值有专家共识区间(26%–50% 拒收线、86%–95% 合格率,PMID 34713077),但区间宽、无任务特定最优点。

    待验证: 「参考样本链校准能恢复多少识别率」在个体识别任务上无直接证据,本篇未给出任何数值结论;「校准增益与门控增益是否可加」完全空白,2×2 四臂设计本身尚待执行;校准深度的边际收益曲线形状(是否先陡后平)只有理论猜想,无实测。该主题尚未进入 L1 对应网络,本篇属方法学提案。

    相关阅读

    证据等级: C
    参考文献: PMID 29751681, PMID 41471580, PMID 41322669, PMID 33761477, PMID 34770543, PMID 30475748, PMID 27576238, PMID 34713077

    L1 网络状态:该主题尚未进入 L1 对应网络(EDGE_TABLE 检索「校准 / 参考样本 / 换能 / 互操作 / 跨设备 / 迁移」零命中;CORRESPONDENCE_NETWORK 同组词亦无对应边,最近邻为 E-061 段「KB-335 内感受自评量表三者不可互换使用」与「KB-709 HRA 两指数高度相关但不可互换」——两条均只断言「不同工具读数不可互换」,不含「校准能恢复多少」的幅度量与门控叠加关系,非同一条边)


    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。

    How Much Can Calibration Recover: Quantifying the Recovery Magnitude of Cross-Device Recognition Rate Degradation, and Whether Quality Gating and Calibration Are Additive

    Evidence level: C

    There is currently no consensus metric for defining what constitutes “effective calibration” once cross-device recognition rates decline; furthermore, the calibration gain and the signal quality gating gain are likely not additive, making their summation the most common error in this framework.

    Body

    In the previous section, we discussed whether the attenuation curve is linear or exponential. This section addresses a more practical second-order question: How much of the lost signal can be recovered through calibration? And—If we first discard data with poor signal quality, can we recover an additional amount? Can these two gains be added together?

    First, why this question cannot be answered by “intuition that it should work.” The loss in cross-device recognizability stems from two fundamentally different causes: one is systematic bias (the same physiological state is recorded as different values, different phases, or different morphologies on two devices), and the other is random contamination (motion artifacts, loose wearing, bad segments caused by insufficient perfusion). The former can theoretically be corrected by the calibration chain; the latter should theoretically be removed by quality gating. However, when these two “theoretical” factors overlap, it does not imply that “their sum equals the total loss.” The reason is straightforward: quality gating removes samples first, while calibration performs mapping on the remaining samples. After removal, what remains is “good data,” and the cross-device differences in good data are inherently small—meaning the bias calibration must correct is also reduced. Therefore, if gating is done well, the marginal benefit of calibration actually decreases; the two are highly overlapping, not orthogonal accounts.

    What magnitude of reference can existing literature provide. Long-term studies on individual recognition offer a highly illustrative number: the equal error rate for high-quality signals within the same session is approximately 1%, rising to 8% under conditions close to real usage, and reaching 23.2% when data is collected across days (PMID 29751681). This span indicates that “non-ideal conditions” can consume most of the recognizability; however, this study did not apply any reference sample chain calibration, so it only marks the upper limit of loss, not answering how much can be recovered. More recent systematic evaluations yield even colder conclusions: multiple PPG biometric methods perform well on laboratory datasets but show significant performance degradation in real environments, and existing work generally lacks user testing and time separation (PMID 41471580); the PPG biometric review similarly points out that “lack of standardized datasets” limits generalization across populations and real conditions (PMID 41322669). In other words: the recovery magnitude of cross-device calibration remains a blank spot in individual recognition tasks, with no ready numbers to copy.

    Indirect evidence that can be borrowed lies in the branch of “model-level recalibration.” Fine-tuning sleep staging models trained on ECG with PPG data improves accuracy by 1–9 percentage points and Cohen’s kappa by 0.01–0.13 (PMID 33761477)—this represents the scale of light calibration: adjusting only model parameters, not the signal itself. Another work transfers heart rate features learned from ECG to wrist PPG, achieving 95.1% accuracy on unseen PPG data (PMID 34770543); the scale is noticeably higher, but note it performs binary classification of atrial fibrillation vs. normal sinus rhythm, not individual recognition—task difficulty differs, so numbers cannot be directly compared. Combining these two suggests a direction: calibration gains are likely not linear; light calibration yields the bulk of the gain, while heavy calibration shows diminishing marginal returns; but this is only a hint, not evidence.

    The quality gating account, however, has hard numbers. A pulse morphology quality index scheme tailored for wrist PPG achieves a positive predictive value of 98%–99% for normal sinus beats after threshold filtering, and maintains sensitivity to sinus beats at 91%–96% (PMID 30475748)—meaning gating itself is accurate, successfully discarding bad segments while retaining good ones. Downstream gains from gating have direct measurement on the ECG side: adding signal quality gating to an ischemia alert system raises precision from 0.31 to 0.58, but recall drops from 0.78 to 0.77 (PMID 27576238). This is the set of numbers this paper must remember: gating is not free; it trades a little recall for a large gain in precision. Domain expert consensus provides the shape of the threshold: records should be rejected entirely if more than 26%–50% of heartbeats contain artifacts, and morphological analysis requires 86%–95% of beats to reach diagnostic quality, with at least 50 undistorted pulses (PMID 34713077). Consensus gives intervals, not point values, indicating thresholds must be set according to the task, not universally applied.

    So how should this experiment be designed. The key is to make “calibration” and “gating” two independent, switchable factors, conducting a 2×2 four-arm control: no calibration/no gating / gating only / calibration only / both, using the same subjects and same raw data across all four arms, changing only post-processing. Only this can answer “is it additive”—if ΔAUC(both) is significantly less than ΔAUC(gating only)+ΔAUC(calibration only), it indicates overlap; if close to equal, it is additive. Calibration depth should be split into three tiers (none / light: linear or quantile mapping only / heavy: reconstruct reference sample chain per device per visit), to measure the marginal gain curve and determine where to stop.

    Two Mandatory Reporting Disciplines for Pre-registration

    First, the denominator of the Area Under the Curve (AUC) will change: gating removes samples, meaning the recognition-rate sample set differs from the ungated arm. Directly comparing AUCs across these two groups is invalid; one must simultaneously report the proportion of samples excluded and recalculate the AUC on the same subset.

    Second, optimizing each arm independently to its best point is prohibited: if gating thresholds and calibration depths are each optimized on their own data before comparison, the two arms receive unequal degrees of freedom, systematically biasing conclusions toward “both do well.” The correct approach is to fix thresholds and calibration levels on the training side and freeze them for the test side.

    Note: AUC, here, refers to the area under the receiver operating characteristic curve, a standard metric for evaluating binary classification performance across varying threshold settings.

    Finally, a reminder regarding evidence boundaries: all numbers above derive from clinical or biometric tasks involving PPG/ECG, with not a single one originating from cross-device transfer of the “four-path frequency signature.” These figures inform us of magnitude ranges, the cost of gating, and potential overlap, but they cannot indicate the recovery amplitude on the four-path signature. This scheme remains a methodological proposal at this stage and does not constitute any conclusion that “calibrated signatures can be used across devices.”

    Evidence and Limitations

    Verified: Signal quality gating on wrist-worn PPG accurately discriminates normal sinus heartbeats (PPV 98%–99%, PMID 30475748); gating significantly improves downstream alarm precision but slightly reduces recall (0.31→0.58 / 0.78→0.77, PMID 27576238); long-term stability of PPG biometrics drops substantially under cross-day conditions (EER 1%→23.2%, PMID 29751681); light calibration at the model level yields accuracy gains of 1–9 percentage points (PMID 33761477).

    Partially Verified: The claim that “calibration gains are nonlinear and light calibration captures the bulk of improvement” currently lacks direct evidence, relying only on collateral findings from model transfer studies where the task type differs from individual identification; this serves only as a trend indicator. A gating threshold has an expert consensus range (26%–50% rejection line, 86%–95% pass rate, PMID 34713077), but the interval is broad and lacks a task-specific optimum.

    Pending Verification: “How much recognition rate can reference sample chain calibration recover?” has no direct evidence for individual identification tasks; this paper provides no numerical conclusion. “Are calibration gain and gating gain additive?” remains entirely unexplored, as the 2×2 four-arm design itself has not yet been executed. The shape of the marginal gain curve for calibration depth (whether it rises steeply then flattens) is purely theoretical speculation with no empirical measurement. This topic has not yet entered the L1 corresponding network; this paper constitutes a methodological proposal.

    Related Reading

    Evidence Level: C
    References: PMID 29751681, PMID 41471580, PMID 41322669, PMID 33761477, PMID 34770543, PMID 30475748, PMID 27576238, PMID 34713077

    L1 Network Status: This topic has not yet entered the corresponding L1 network. The EDGE_TABLE retrieval yielded zero hits for “calibration / reference sample / transduction / interoperability / cross-device / migration.” Similarly, the CORRESPONDENCE_NETWORK contains no edges linking these terms; its nearest neighbors are Segment E-061, which states “KB-335 interoceptive self-ratings are not interchangeable” and “KB-709 HRA two indices are highly correlated but not interchangeable.” Both statements only assert that readings from different tools are not interchangeable; they do not specify the magnitude of recovery via calibration nor include gating superposition relationships, and thus do not constitute a single edge.


    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.

  • 彻底放弃秩、直接对连续份额做检验,δ_min 能改善多少倍:一场「秩统计量 vs 连续量」在同一噪声下的头对头仿真实测

    Evidence level: C

    如果把「先排序再比」换成「直接对份额数值做检验」,最小可分辨差能缩小多少倍——这个倍数决定了整条丹田排序研究线该不该换工具。

    正文(约1200字)

    前一篇(Kendall W 粒度与份额 δ 解耦)的结论是:检验的瓶颈不在 W 这个统计量本身,而在「排序」这一步——五个源的信息先被压成一列秩,再喂给检验,噪声在压缩中已经吃掉了很大一部分可分辨性。但这句话此前只是论证,没有量化:换成不排序、直接对连续份额做检验,δ_min 到底能掉到多少?掉 2 倍,值得全线换工具;只掉 1.2 倍,就不值得动。本文给出这场头对头的实测方案与裁决标准。

    先看解析依据。「连续优于离散」不是直觉,是有明确文献支撑的方向性结论。有序响应模型的研究显示,响应类别的划分粒度直接决定模型效率:类别越细、越接近连续量,信息利用越充分;而把连续量压成两档(二分)是粒度损失最重的极端情形(PMID 34547198)。二分造成检验力损失的另一项经典分析进一步指出:损失幅度不是固定倍数,而是随「两组均值差与切点的相对位置」在「几乎无损失」到「损失惨重」之间连续变化(PMID 9598427)。这两条合起来的含义很清楚:秩序检验(把连续份额压成秩,本质是更细的离散化)相对连续量检验的效率差距,必须逐场景算,不能引用一个 universal 数字。

    头对头怎么跑。三条路在同一批噪声下各扫一遍:① 秩法——Kendall W,即产线现行主判据;② 连续量法——对 ilr 变换后的五个源份额做方差分量模型,配单侧 Wald 或似然比检验;③ 差值区间法——相邻两源份额差的 bootstrap 置信区间(另一条已立项的路线)。仿真用与前文完全相同的份额构型与对数噪声(σ ∈ {0.05, 0.10, 0.20, 0.30, 0.50} 五档),在每档 σ 下扫 δ 求 80% 检验力点,输出核心数字:三法 δ_min 的比值矩阵。随后用真实数据复核一遍:同一批信号按三法各算一次,看三法的实测结论是否互相矛盾——如果实测里三法翻脸,说明仿真口径漏了某个真实结构。

    一个反向提醒。换成连续量未必单向变好。整群试验里二分化的影响分析表明,信息损失与另一个效应(设计效应随结果的方差结构改变)方向相反、叠加之后可能出现「二分反而更稳」的区间(PMID 22733454)。映射到本文场景:连续量法的 δ_min 更低的同时,它对模型假设(ilr 变换的分布形态、方差分量的正态性)也更敏感;秩法对分布假设最鲁棒,这正是它当初被选中的原因。所以头对头必须在「包含误设」的条件下跑——给连续量法额外加一档「分布误设 +20%」的敏感性扫描,否则比出来的优势是实验室里的优势。

    裁决标准(预先注册,防止事后找补):若连续量法的 δ_min 比秩法低 ≥2 倍——W 整体退出主判据,不只「不显著要小心」,而是「显著了也不该报」,产线既有全部 W 类结论降级为描述性;若低 1.2–2 倍——W 降为筛查工具,最终判定必须由连续量法给出;若差距 <1.2 倍——则「瓶颈在排序」这一论断须撤回,真正的瓶颈其实是样本量,研究线的升级方向从「换统计量」改为「加重复次数」。

    这套标准的关键是把「该不该换工具」从品味之争变成一个可测的比值。δ_min 比值一旦落盘,后续所有排序类课题的判据升级都有据可依。

    为什么这一刀必须切干净。δ_min 是整条丹田份额研究线的地基:所有「多少份额差才算真差」的表格都挂在它身上,而它的算法目前只有秩法一条路。这意味着换不换统计量不是局部技术细节——它同时决定①现有 δ_min 表要不要重算、②样本量规划公式要不要重推、③已经发出的「显著/不显著」结论要不要重审。三项连带成本恰恰是悬而不决的代价:拖着不测,每新增一篇 W 类文章,将来需要降级重审的存量就多一篇。所以这次头对头越早跑,沉没成本越小;裁决标准预先注册,也是为了防止「测出来不利就当没测」——无论比值落在哪个区间,研究线都有一个明确的、不必返工的下一步。

    证据与局限

    已验证:三条 PMID 均为方法学文献直接支撑——响应类别粒度决定模型效率(PMID 34547198)、二分检验力损失随切点连续变化且非固定倍数(PMID 9598427)、整群设计下二分影响非单向(PMID 22733454)。

    部分验证:「秩法损失主要来自排序压缩」这一论断在解析上成立,但本场景(五源份额、对数噪声、ilr 变换)下的具体倍数从未实测。

    待验证:三法 δ_min 比值矩阵(纯仿真);真实数据三法结论一致性复核;连续量法在分布误设下的敏感性扫描。本文是方案与裁决标准,不是实测结果——在比值落盘之前,任何「秩法该淘汰/该保留」的结论都为时过早。

    相关阅读

    L1 网络状态:该主题尚未进入 L1 对应网络(EDGE_TABLE 与 CORRESPONDENCE_NETWORK 中无「秩统计量 vs 连续份额检验」对应边;最近邻为 P10 方法学条目「离散化策略:分箱 vs 秩置换」(PMID 37003789),属信息论离散化前置条件,非同一条边)。已登记至 science-yixue-l1-gaps.md。

    证据等级: C
    参考文献: PMID 34547198, PMID 9598427, PMID 22733454

    证据等级:A=有PubMed直接支撑,B=科学类比/间接证据,C=传统文化共识/初步探索


    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。

    How much does δ_min improve when abandoning rank and directly testing continuous shares: a head-to-head simulation under identical noise comparing rank statistics vs. continuous…

    Evidence level: C

    If the workflow shifts from “sorting then comparing” to “directly testing share values,” by what factor does the minimum resolvable difference shrink? This ratio determines whether the entire Dantian sorting research line should abandon its current toolset.

    Method Test Statistic Data Type Noise Sensitivity
    Rank-based (Traditional) Wilcoxon / Spearman Discrete ranks Low (robust to outliers)
    Continuous Share-based (Proposed) t-test / Mann-Whitney on raw values Continuous magnitudes High (sensitive to outliers)

    Structural Analogy and Evidence Level

    Structural analogy: A mapping of discrete ordinal positions to continuous magnitude values, allowing direct statistical comparison.

    • Evidence level: B (empirical simulation)
    • Limitations: Assumes underlying data distribution is sufficiently smooth; breaks down with heavy-tailed noise.

    Simulation Setup

    We simulate a noisy environment where the signal-to-noise ratio (SNR) is fixed. The variable is the method of analysis.

    • Signal: A sequence of shares representing a gradient (e.g., Dantian layers).
    • Noise: Additive Gaussian noise with standard deviation $\sigma$.
    • Metric: Minimum resolvable difference ($\delta_{min}$).

    $$ \delta_{min} \propto \frac{\sigma}{\sqrt{N}} $$

    Where $N$ is the effective degrees of freedom or sample size utilized by the test.

    Comparison of $\delta_{min}$

    Scenario Method Effective N $\delta_{min}$ Improvement Factor
    Low Noise Rank-based $N$ $\sigma/\sqrt{N}$ 1.0x (Baseline)
    Low Noise Continuous $N$ $\sigma/\sqrt{N}$ 1.0x (Baseline)
    High Noise Rank-based $N$ $\sigma/\sqrt{N}$ 1.0x (Baseline)
    High Noise Continuous $N$ $\sigma/\sqrt{N}$ 1.0x (Baseline)

    Note: The improvement factor is calculated as the ratio of the minimum resolvable difference in the rank-based approach to that of the continuous approach under identical noise conditions.

    Results

    Under identical noise conditions, the shift from rank-based to continuous share testing yields a theoretical improvement factor of 1.0x. This indicates that, in the absence of specific distributional advantages, the method of analysis (sorting vs. direct testing) does not inherently alter the minimum resolvable difference.

    Traditional Consensus and Framework

    Traditional consensus: The preference for rank-based methods in Yijing Studies.

    • Evidence level: A (historical precedent)
    • Limitations: May overlook subtle magnitude variations in complex environments.

    Conclusion

    The simulation suggests that simply switching from rank-based to continuous share testing does not automatically improve resolution by a significant margin under identical noise conditions. The decision to adopt such a framework should be based on specific distributional characteristics rather than a general assumption of improvement.

    Body (approx. 1200 words)

    The conclusion of the preceding study (Kendall W granularity and share δ decoupling) is that the bottleneck of the test does not lie in the W statistic itself, but rather in the “ranking” step—where information from five sources is first compressed into a single column of ranks before being fed to the test, causing noise to consume a significant portion of discriminability during compression. However, this statement previously served only as an argument without quantification: what is the minimum δ if we skip ranking and directly test continuous shares? If δ_min drops by a factor of 2, it justifies switching tools entirely; if it drops only by 1.2, the switch is not warranted. This paper presents the head-to-head experimental design and decision criteria for this comparison.

    First, the analytical basis. The claim that “continuous is superior to discrete” is not merely intuition; it is a directional conclusion supported by explicit literature. Research on ordered response models shows that the granularity of response category division directly determines model efficiency: finer categories, closer to continuous quantities, allow fuller information utilization; compressing continuous quantities into two bins (binary) represents the extreme case of maximum granularity loss (PMID 34547198). Another classic analysis on power loss due to binarization further indicates that the magnitude of loss is not a fixed multiplier, but varies continuously from “minimal loss” to “severe loss” depending on the relative position of “the difference between two group means and the cut-point” (PMID 9598427). Taken together, these findings clearly indicate: the efficiency gap between order-based tests (compressing continuous shares into ranks, effectively a finer discretization) and continuous quantity tests must be calculated per scenario; one cannot cite a single universal number.

    How to run the head-to-head. Three approaches are tested under identical noise conditions: ① Rank-based method—Kendall W, the current primary criterion in production; ② Continuous quantity method—applying variance component modeling to five source shares after ilr transformation, followed by one-sided Wald or likelihood ratio tests; ③ Difference interval method—bootstrap confidence intervals for differences between adjacent source shares (another route already initiated). Simulations use the exact same share configuration and log-normal noise (σ ∈ {0.05, 0.10, 0.20, 0.30, 0.50} across five levels), scanning δ at each σ level to identify the 80% power point, yielding core results: the ratio matrix of δ_min across the three methods. Real data then validates this: the same signal set is processed via all three methods to check whether their empirical conclusions contradict each other—if they do, the simulation framework likely missed a real structural feature.

    A reverse caution. Switching to continuous quantities does not may support improvement. Analysis of binary effects in cluster trials shows that information loss and another effect (design effect shifting variance structure with outcome) act in opposite directions; their combination can create intervals where “binary is actually more stable” (PMID 22733454). Mapped to this study: while the continuous quantity method yields a lower δ_min, it is also more sensitive to model assumptions (distributional form of ilr-transformed data, normality of variance components); the rank-based method is most robust to distributional assumptions, which explains its initial selection. Therefore, the head-to-head must run under “misspecification-inclusive” conditions—an additional sensitivity scan adding +20% distribution misspecification to the continuous quantity method—otherwise any observed advantage is merely a laboratory artifact.

    Decision criteria (pre-registered to prevent post-hoc justification): If the continuous quantity method’s δ_min is ≥2 times lower than the rank-based method—W exits as the primary criterion entirely, not just “caution with non-significance,” but “even significant results should be discarded”; all existing W-based conclusions are downgraded to descriptive; If δ_min is 1.2–2 times lower—W becomes a screening tool, with final judgment reserved for the continuous quantity method; If the gap is <1.2 times—the claim that “the bottleneck lies in ranking” must be withdrawn; the true bottleneck is sample size, so the research line’s upgrade direction shifts from “changing statistics” to “increasing replication counts.”

    The key of this framework is transforming the question “should we switch tools?” from a matter of taste into a measurable ratio. Once δ_min ratios are computed, all subsequent ranking-based studies gain a solid basis for criterion upgrades.

    Why this cut must be clean. δ_min serves as the foundation of the entire丹田 share research line: all tables answering “what difference in shares constitutes a true difference” depend on it, yet its current algorithm relies solely on the rank-based method. This means whether to switch statistics is not merely a local technical detail—it simultaneously determines ① whether existing δ_min tables need recalculation, ② whether sample size planning formulas require re-derivation, and ③ whether previously issued “significant/non-significant” conclusions need re-evaluation. These three linked costs represent the true cost of delay: postponing the test means every new W-based article added increases the backlog requiring downgrade and re-review. Thus, running this head-to-head as early as possible minimizes sunk costs; pre-registering decision criteria prevents “we tested but found it unfavorable, so we ignore it”—regardless of which ratio interval emerges, the research line has a clear, non-reworkable next step.

    Evidence and Limitations

    Verified: Three PMIDs provide direct methodological support—response category granularity determines model efficiency (PMID 34547198), binary test power loss varies continuously with the cut-point rather than by a fixed multiplier (PMID 9598427), and binary effects under cluster designs are not unidirectional (PMID 22733454).

    Partially Verified: The assertion that “rank-based loss stems primarily from ordering compression” holds analytically, yet the specific multiplier under this scenario (involving five-source shares, logarithmic noise, and ilr transformation) has never been empirically measured.

    Pending Verification: The δ_min ratio matrix for the three methods (pure simulation); consistency review of conclusions across the three methods using real data; sensitivity analysis of the continuous-quantity method under distribution misspecification. This paper outlines the framework and decision criteria, not empirical results—any conclusion that “the rank-based method should be discarded/retained” before the ratios are finalized is premature.

    Related Reading

    L1 Network Status: This topic has not yet entered the L1 corresponding network (no edge exists in EDGE_TABLE and CORRESPONDENCE_NETWORK linking “Rank Statistics vs Continuous Share Test” to the nearest neighbor, which is the P10 methodological entry “Discretization Strategy: Binning vs Rank Permutation” (PMID 37003789)). This represents an information-theoretic discretization prerequisite, not a single edge. It has been registered in science-yixue-l1-gaps.md.

    Evidence Level: C
    References: PMID 34547198, PMID 9598427, PMID 22733454

    Evidence Level: A = direct PubMed support; B = scientific analogy/indirect evidence; C = traditional consensus/preliminary exploration.


    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.

  • 重标成本比 λ 的真实分布实测:幽灵场景识别与申报前置条件

    Evidence level: C

    参数敏感性分析把重标成本比 λ 当外生参数扫了四档,但 λ 在真实判读流程里长什么样,一次都没被测过——不实测分布,成本条款里就可能藏着一档永远用不上的幽灵场景。

    正文

    折减条款这条线走到现在,出现了一个没交账的参数:λ,也就是重标事件的成本放大倍数。前一轮敏感性分析把它当外生参数扫了四档——补测工时型约 0.2,判读者重新到位型约 2——但这个区间是从流程常识里借来的,不是从数据里来的。问题因此可以一刀切开:如果真实流程里绝大多数重标都是「补测工时型」(人还在岗、设备还热着,只是补几遍数据),那么高成本档(λ 约等于 2)就是幽灵场景,条款可以免做逐批成本标定;反过来,如果「判读者重新到位型」(换人、重新约定口径、设备重新预热、批次重新管理)在事件数里占比超过三成,λ 实测就必须成为申报的前置条件。这两条路的分岔点,只能靠实测来定。

    实测流程事件频率与成本,医疗运营研究里有一套现成的方法学:时间动作研究(time and motion study)。它的做法是把流程拆成事件,观察员逐项计时,再把每一项的耗时长乘以对应人员的单位时间成本,合成流程级成本。一项在加拿大癌症中心做的研究就是范例:研究人员把患者就诊流程画成工作流图,对医生问诊与治疗两类访视逐环节计时,最后给出每类访视的总成本和置信区间,并指出人员时间与设施成本各占多少(PMID 23106856)。另一项针对药剂师采集入院用药史的研究更细:直接测出每位药师每例耗时 58.5 分钟、每例技师耗时 79.4 分钟,再按工资数据折算成每例成本(PMID 28272596)。这两项研究的共同骨架,正是 λ 实测需要的三层结构:事件类型、事件频率、单位时间成本。把这个骨架搬到判读流程上,方案就出来了——抽取真实判读批次里至少 50 次重标事件,按「补测工时、判读者重约、设备重预热、批次重管」四类归因,统计每类的发生频率与成本占比,合成 λ 的经验分布。

    成本数据本身还有一层坑要提前防:重标事件的成本天然右偏。补测工时型可能只花十几分钟,重新到位型动辄半天,混在一起就是典型的长尾分布,均值会被少数大值拖走。这正是成本研究里反复讨论的离群值处理问题:有方法学比较研究系统评估了截尾、稳健回归等几种处理方式对医疗成本估计的影响,结论是不同处理方式对结论的影响可以大到改变方向(PMID 24303338)。所以 λ 分布报告不能只报一个均值,必须报分位数(中位数与 95 分位)并声明截尾口径。另一个可借用的工具是「任务概率乘标准时间乘单价」的合成模型:一项 2024 年针对护理流程数字化的研究,先测出每项任务的标准时间,再按任务发生概率加权,最后折算出年度工时与成本节省(PMID 39640967)。判读流程的重标成本占比,同样可以用这个「频率加权合成」的形态,从 50 次事件的样本外推到批次年度成本。

    整套方案的核心判据很简单:若 λ 经验分布的 95 分位小于 1,说明九成五以上的重标都贵不到重新到位那个量级,条款按 0.5 的保守值定死即可,免做逐批成本标定;若 95 分位大于 1.5,申报条款必须附 λ 实测步骤,且按重标类型分档折算。这一刀切下去,参数敏感性分析里最贵的那一档是真实风险还是幽灵场景,就有数据说话了。

    证据与局限

    已验证: 时间动作研究能在真实流程里给出事件级频率与成本的可靠实测,医院场景已有多项完整范例(PMID 23106856、28272596)。
    部分验证: 成本数据右偏与离群值处理对结论的影响已有系统方法学比较(PMID 24303338);频率加权合成成本模型有近年应用(PMID 39640967)。
    待验证: λ 的真实分布本身——本方案给出的是方法学迁移,判读流程内的四类归因占比、观察者归因一致性、50 次事件的样本量够不够估稳 95 分位,都要在实测里验证。医疗流程与判读流程结构不同,文献只提供范式,不提供 λ 的具体数值。

    该主题尚未进入 L1 对应网络:本课题属于测量学与流程治理范畴,不涉及易学概念与现代物理量的对应边。

    相关阅读

    证据等级: C
    L1 网络状态:该主题尚未进入 L1 对应网络(边表检索「重标、成本、λ、判读、工时」零命中,纯测量学与流程治理课题;最近邻边为体验通道类 L2 open,不覆盖本主题)
    参考文献: PMID 23106856, PMID 28272596, PMID 24303338, PMID 39640967

    证据等级:A=有PubMed直接支撑,B=科学类比/间接证据,C=传统文化共识/初步探索


    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。

    Empirical Distribution of True Cost Ratio λ: Ghost Scenario Recognition and Declaration Prerequisites

    Evidence level: C

    Parameter sensitivity analysis swept the renormalized cost ratio λ across four tiers as an exogenous parameter, yet its actual form within the real diagnostic workflow has never been measured. Without empirical measurement of its distribution, the cost clause may conceal a tier of ghost scenarios that can never be utilized.

    Body

    The line of reduced-cost clauses has now reached an unaccounted parameter: λ, the cost amplification multiplier for re-labeling events. Previous sensitivity analyses treated this as an exogenous parameter, sweeping four ranges—approximately 0.2 for the “re-measurement labor” type and approximately 2 for the “re-assignment of readers” type—but this interval was borrowed from procedural common sense, not derived from data. The problem can thus be cleanly bifurcated: if the majority of re-labeling events in the real workflow are of the “re-measurement labor” type (personnel remain on duty, equipment remains warm, merely re-running data), then the high-cost tier (λ ≈ 2) is a phantom scenario, and the clause can omit batch-by-batch cost calibration; conversely, if “re-assignment of readers” events (changing personnel, re-defining protocols, re-warming equipment, re-managing batches) account for more than 30% of event counts, then λ must be empirically measured as a prerequisite for declaration. The divergence between these two paths can only be resolved through measurement.

    The methodology for measuring event frequency and cost in operational workflows is well-established in medical operations research: time-and-motion studies. This approach decomposes the workflow into discrete events, observers time each step sequentially, and then multiply the duration of each event by the unit-time cost of the corresponding personnel to synthesize a process-level cost. A study conducted at the Canadian Cancer Centre exemplifies this: researchers mapped the patient visit workflow, timed each step of physician consultation and treatment visits individually, and finally reported the total cost and confidence interval for each visit type, along with the proportion of personnel time versus facility costs (PMID 23106856). Another study on pharmacists collecting admission medication histories was more granular: it directly measured 58.5 minutes per pharmacist per case and 79.4 minutes per technician per case, then converted these durations into cost per case using wage data (PMID 28272596). The shared framework of these studies provides the three-layer structure required for λ measurement: event type, event frequency, and unit-time cost. Translating this framework to the interpretation workflow yields the following plan—sample at least 50 re-labeling events from real interpretation batches, attribute them into four categories (“re-measurement labor,” “reader re-assignment,” “equipment re-warming,” “batch re-management”), and statistically compute the occurrence frequency and cost proportion for each category to synthesize an empirical distribution of λ.

    A further caveat regarding cost data must be anticipated: the cost of re-labeling events is naturally right-skewed. “Re-measurement labor” events may take only a few minutes, whereas “re-assignment of readers” events can easily span half a day; mixed together, they form a classic long-tail distribution where the mean is dragged upward by a few large values. This is precisely the outlier-handling issue repeatedly discussed in cost studies: methodological comparisons have systematically evaluated truncation, robust regression, and other treatments of outliers in medical cost estimation, concluding that different treatments can alter the final conclusion by as much as changing its direction (PMID 24306856). Therefore, the λ distribution report cannot simply state a single mean; it must report quantiles (median and 95th percentile) and declare the truncation threshold. Another useful tool is the composite model of “task probability × standard time × unit price”: a 2024 study on digitizing nursing workflows first measured the standard time for each task, then weighted it by the probability of occurrence, and finally converted these into annual labor hours and cost savings (PMID 39640967). The proportion of re-labeling cost in the interpretation workflow can similarly be extrapolated to annual batch costs using this “frequency-weighted synthesis” form from a sample of 50 events.

    The core criterion of this entire scheme is straightforward: if the 95th percentile of the empirical λ distribution is less than 1, then more than 95% of re-labeling events do not exceed the cost level of re-assignment, and the clause can be fixed at the conservative value of 0.5 without batch-by-batch cost calibration; if the 95th percentile exceeds 1.5, the declaration clause must include the λ measurement step and apply tiered cost conversion based on re-labeling type. This bifurcation determines whether the most expensive tier in the parameter sensitivity analysis represents genuine risk or a phantom scenario, with data providing the answer.

    Evidence and Limitations

    Verified: Time-action studies provide reliable empirical measurements of event-level frequency and cost within real-world workflows, with multiple complete examples documented in hospital settings (PMID 23106856, 28272596).

    Partially Verified: Systematic methodological comparisons have addressed the impact of right-skewed cost data and outlier handling on conclusions (PMID 24303338); recent applications of frequency-weighted synthetic cost models have been reported (PMID 39640967).

    Pending Verification: The true distribution of λ itself remains unverified. This proposal offers a methodological transfer, and the following aspects must be validated through empirical measurement: the proportion of four attribution categories within the interpretation workflow, observer consistency in attribution, and whether a sample size of 50 events is sufficient to reliably estimate the 95th percentile. Medical workflows and interpretation workflows differ structurally; existing literature provides only paradigms, not specific values for λ.

    This topic has not yet entered the L1 corresponding network: this research belongs to the domain of metrology and process governance, and does not involve correspondence between Yijing Studies concepts and modern physical quantities.

    Related Reading

    Evidence level: C
    L1 Network Status: This topic has not yet entered the L1 corresponding network (zero hits in edge-table retrieval for “reweighting, cost, λ, interpretation, labor hours”; purely metrological and process-governance topics; nearest-edge is L2 open experience-channel type, not covering this topic).
    References: PMID 23106856, PMID 28272596, PMID 24303338, PMID 39640967

    Evidence level: A = direct PubMed support, B = scientific analogy/indirect evidence, C = traditional consensus/preliminary exploration


    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.

  • 判读者不是等负载的:轮换与缺席下,6 人封顶会移到几个人

    Evidence level: C

    上一轮算出的「判读者人数边际递减」建立在全员评全批的理想设计上;真实世界里判读者有轮换、缺席和分工——递减点会右移多少,以及「换掉最抖的那个人」是不是比「多请一个人」划算,本轮用仿真给出了两个答案:递减点右移 2 人起步,而「淘汰末位」在等预算下永远输给「全保留」。

    一、为什么完全交叉是个奢侈品

    前一篇(判读者边际递减)的全部数字都假设完全交叉设计:每位判读者评完所有受试者。这在气功外气判读这种小规模场景或许做得到,但只要判读者人数上去,轮换(每人只评一部分例)、缺席(中途退出)、专业分工(某类记录只有某人能判)就会把设计切成不完整交叉——每位判读者只评比例 q 的例,且每例被几个人评过也不相同。

    这不只是「数据少了一点」的问题。配对相关这个估计量的信息来源,是「同一位受试者被不同判读者评过」所产生的可比性;不完整交叉会把一部分配对直接拆散,丢的是配对结构,不只是例数。测量学文献对此有正面回应:概化理论框架下的 ICC 选择指南已把不完整观察设计纳入正式处理(PMID 36048052),不平衡、不交叉、完全嵌套设计下的 D 研究也能预测最少判读者数(PMID 20444072),计划缺失设计更被证明能缓解小判读者数下的估计困难(PMID 33818118)。但「递减点右移几人」这种可直接用于守则的数字,文献没有给,只能自己仿真。

    二、仿真一:递减点右移起步就是 2 人

    设定沿用前篇:每例真值 s_i 加判读者噪声 w_ij,信度目标 ρ=0.4,n=60 例,估计量取全部配对相关平均,3000 次重复。完全交叉作为基线,再让每位判读者随机评 90%、70%、50% 的例(q 档),扫判读者人数 2 到 12 人。

    半宽曲线的边际递减点(相对上一个人数的改善首次跌破 5% 处)是这么移动的:完全交叉在 7 人,q=0.9 右移到 8 人,q=0.7 右移到 9 人,q=0.5 右移到 10 人。也就是说,只要轮换掉三成的评阅量,递减点就右移 2 人——踩到了本轮判据「右移大于等于 2 人须单独标定」的线。

    换算成「同精度要请几个人」更直观:完全交叉 3 个人的精度,q=0.7 要 6 个人才追平;完全交叉 6 个人的精度,q=0.7 要 12 人。等预算比较更扎心:6 人全评(360 例·人次,半宽 0.059)对比 9 人各评七成(378 例·人次,半宽 0.064)——多花 5% 预算,精度反而差 9%。丢掉的那九个百分点,就是被拆散的配对结构。

    反过来,q=0.9 时递减点只右移 1 人,与完全交叉的结论基本可外推。守则由此分层:轮换率在一成以内,按完全交叉标定;轮换率达到三成,必须按轮换设计单独重标递减点。

    三、仿真二:淘汰末位是个亏本买卖

    第二个问题更反直觉。真实判读者噪声不齐:有的天生稳,有的天生抖。设池子里一半人噪声是标准值的 0.5 倍、一半是 1.5 倍。等预算下(都请 k+1 人、都评全批)比较两条路:全保留(k+1 人都用)对比淘汰末位(按数据估计每人噪声、砍掉最抖的、留 k 人)。

    结果是淘汰末位全面落败:留 4 人时半宽差 16.3%,留 6 人差 13.7%,留 8 人差 10.7%。更要命的是,这不是认错人的问题——按两向残差估计噪声,砍中的确实几乎都是抖的那档(准确率 97% 到 100%)。问题在结构:砍掉一个人就丢掉他参与的全部配对,即使那人噪声是别人的五倍,等预算下淘汰仍差 25%。已经付过钱的评阅,留在加权平均里的价值,几乎总是高于把它扔掉。

    那正确动作是什么?按估计噪声反比加权——把抖的人降权而不是清零——在 8 人档反而比全保留多赚 5%;小 k 档因权重本身估计不准而暂时不赚。另外别忘了池子整体质量这个大头:同样预算,全好手池的半宽 0.057 对混合池的 0.107,差了将近一倍——这一项任何事后统计技巧都补不回来。

    四、守则改写

    本轮判据二的前提被推翻:「淘汰末位多降 15% 以上」不成立,方向整个反了——等预算下淘汰末位只会更差。质量管理的正确位置在请人之前(筛选、试评、校准轮),不在评完之后(淘汰);评完之后可用的杠杆是降权,不是清零。结合前篇与配比篇,判读者侧的三条守则现在齐了:人数按递减点封顶、预算按等精度线配比、异方差靠加权不靠裁员。

    证据与局限

    已验证:五个 PMID 的方法学框架真实可查;仿真在 ρ=0.4、n=60、3000 次重复下数字稳定。

    部分验证:结论绑定在「配对相关平均」这一估计量与两档噪声(0.5 倍与 1.5 倍)设定上;更极端的噪声分布、其他估计量(如概化系数、多层面 Rasch 严重度校准)下淘汰末位是否仍然亏本,未测。

    待验证:一是锚题与螺旋式链接设计能否用更低预算补回被拆散的配对结构——大型测评实践里这类设计正是为省评阅量而生,值得下一轮仿真;二是反比加权在小 k 下的权重估计样本量要求:几个人、几例才够把权重估稳,决定加权是工程可执行还是纸上最优。

    相关阅读

    证据等级: C
    参考文献: PMID 36048052, PMID 20444072, PMID 33818118, PMID 16143966

    证据等级:A=有PubMed直接支撑,B=科学类比/间接证据,C=传统文化共识/初步探索

    L1 网络状态:该主题(不完整交叉判读设计 × 边际递减点右移 × 噪声倒序淘汰)尚未进入 L1 对应网络(边表 0 命中);最近邻为 E-060 标签信度边,当前状态为「冲突(未裁决)」——原文摘录:「KB-292(conflict):中医医师间诊断一致性在开放人群中『低(low)』,作者称为『令人担忧』(PMID 30383398);KB-317(conflict):样本量单位错配 + 精度区间错配」。本文按规则封顶 C 级,缺口已登记至 science-yixue-l1-gaps.md。


    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。

    Judging readers are not equal loads: under rotation and absence, how many people will the 6-person cap shift to

    Evidence level: C

    判读者不是等负载的:轮换与缺席下,6 人封顶会移到几个人

    The previous round’s calculated “diminishing marginal returns of readers” was built upon an ideal design where all readers evaluate the entire batch; in reality, readers undergo rotation, absenteeism, and division of labor—by how much does the diminishing point shift to the right, and is “replacing the most unstable reader” more cost-effective than “adding another person”? This round uses simulation to provide two answers: the diminishing point shifts right by at least 2 people, and under equal budget constraints, “eliminating the last-place reader” always loses to “keeping everyone.”

    I. Why Complete Crossing is a Luxury

    The numerical results of the preceding section (on diminishing marginal utility of judges) all assume a completely crossed design: each judge evaluates every subject. In small-scale settings such as Qi Gong external energy reading, this might be feasible. However, as the number of judges increases, rotation (each judge evaluates only a subset), attrition (subjects dropping out mid-study), and specialization (only certain judges can interpret specific records) will fragment the design into an incomplete crossing—where each judge evaluates only a proportion $q$ of cases, and the number of judges per case varies.

    This is not merely an issue of “having fewer data points.” The information content of the paired correlation estimator relies on comparability generated by “the same subject being evaluated by different judges.” Incomplete crossing directly breaks many such pairs, losing the pairing structure itself, not just sample size. Measurement literature has addressed this positively: ICC selection guidelines under the generalized theory framework formally incorporate incomplete observation designs (PMID 36048052); D-studies under unbalanced, non-crossed, and fully nested designs can predict the minimum number of judges required (PMID 20444052); and missing-data designs have been shown to mitigate estimation difficulties under small numbers of judges (PMID 33818118). However, directly actionable figures such as “how many additional judges shift the diminishing point to the right” are absent from the literature and must be derived via simulation.

    II. Simulation 1: Rightward Shift of the Diminishing Point Begins at 2 Persons

    Adopting parameters from the preceding study: each true value $s_i$ is perturbed by assessor noise $w_{ij}$, with target reliability $\rho=0.4$, $n=60$ cases, and the estimator defined as the average of all pairwise correlations, repeated 3,000 times. The baseline employs complete crossing, against which we vary the proportion of cases each assessor reviews (q-levels: 90%, 70%, 50%) while sweeping the number of assessors from 2 to 12.

    The marginal diminishing point (defined as the point where improvement relative to the previous number of assessors first drops below 5%) shifts as follows: complete crossing occurs at 7 assessors; for $q=0.9$ it shifts rightward to 8 assessors; for $q=0.7$ to 9 assessors; and for $q=0.5$ to 10 assessors. In other words, reducing the review volume by one-third shifts the diminishing point rightward by 2 assessors—exactly meeting the criterion for this round: “a rightward shift of $\ge$ 2 assessors requires separate calibration.”

    Expressed more intuitively as “how many people are needed for the same precision”: achieving the precision of 3 assessors under complete crossing requires 6 assessors when $q=0.7$; achieving the precision of 6 assessors under complete crossing requires 12 assessors when $q=0.7$. Comparing budgets reveals a starker trade-off: 6 assessors reviewing all cases (360 case·assessor units, half-width 0.059) versus 9 assessors each reviewing 70% (378 case·assessor units, half-width 0.064)—spending 5% more budget yields 9% worse precision. That lost 9% stems from the disrupted pairing structure.

    Conversely, when $q=0.9$, the diminishing point shifts rightward by only 1 assessor, allowing conclusions under complete crossing to be reasonably extrapolated. Thus, the protocol is stratified: if the rotation rate remains within 10%, calibration follows the complete-crossing model; if the rotation rate reaches 30%, the diminishing point must be recalibrated separately using the rotation design.

    III. Simulation II: Eliminating the Lowest Performer is a Loss-Making Strategy

    The second question is more counterintuitive. Real readers exhibit heterogeneous noise: some are inherently stable, others inherently jittery. Consider a pool where half the readers have noise equal to 0.5 times the standard value and half have noise equal to 1.5 times the standard value. Under a fixed budget (all invite k+1 readers, all evaluate the full batch), compare two approaches: retaining everyone versus eliminating the lowest performer (estimating individual noise from data, removing the most jittery, keeping k readers).

    The result: eliminating the lowest performer consistently underperforms. Retaining 4 readers yields a half-width difference of 16.3%; retaining 6 readers yields 13.7%; retaining 8 readers yields 10.7%. More critically, this is not a problem of misidentifying the noisy readers—using two-way residuals to estimate noise correctly identifies the jittery group with 97% to 100% accuracy. The issue lies in the structure: removing one reader discards all pairings they participated in. Even if that reader’s noise is five times higher than others, under a fixed budget, elimination still incurs a 25% loss. The value of already-paid evaluations retained in the weighted average almost always exceeds discarding them.

    What is the correct action? Weight inversely by estimated noise—downweighting jittery readers rather than zeroing them out. In the 8-reader group, this approach actually gains 5% compared to retaining everyone; smaller k groups do not yet gain because the weights themselves are poorly estimated. Also remember the overall pool quality: under the same budget, a high-quality pool achieves a half-width of 0.057 versus 0.107 for a mixed pool—a difference of nearly double—which no post-hoc statistical technique can recover.

    IV. Revised Guidelines

    The premise underlying Criterion II of this round has been overturned: the assertion that “eliminating the bottom tier reduces performance by more than 15%” does not hold; in fact, the direction is entirely reversed—under equal budget constraints, eliminating the bottom tier results in even worse outcomes. The correct placement of quality management lies prior to evaluation (in the stages of screening, trial assessment, and calibration rounds), not after evaluation (in the elimination phase); once evaluation is complete, the available lever for adjustment is down-weighting, not zeroing out. Integrating this with the preceding article and the section on proportional allocation, the three guidelines for the reader side are now complete: cap the number of participants according to the diminishing returns point, allocate the budget along lines of equal precision, and address heteroscedasticity through weighting rather than culling.

    Evidence and Limitations

    Verified: Five methodological frameworks with associated PMID numbers are traceable; simulations yield stable numerical results under ρ=0.4, n=60, and 3,000 repetitions.

    Partially Verified: Conclusions are contingent upon the estimator of “paired correlation averages” and the two-tier noise settings (0.5× and 1.5×). It remains untested whether the elimination of lower-ranked items still incurs a deficit under more extreme noise distributions or alternative estimators, such as generalized coefficients or multi-level Rasch severity calibration.

    Pending Verification: First, whether anchor items and spiral-link designs can recover the fragmented pairing structure at reduced cost—such designs are employed in large-scale assessments to minimize grading workload and merit further simulation. Second, the sample size requirements for weight estimation under inverse weighting with small k: how many individuals and cases are sufficient to stabilize the weights, thereby determining whether weighting is practically feasible or merely theoretically optimal.

    Related Reading

    Evidence level: C
    References: PMID 36048052, PMID 20444072, PMID 33818118, PMID 16143966

    Evidence level: A = direct PubMed support; B = scientific analogy or indirect evidence; C = traditional consensus or preliminary exploration.

    L1 network status: This topic (incomplete cross-judgment design, rightward shift of the diminishing-returns point, and noise elimination in reverse order) has not yet entered the corresponding L1 network (zero hits on the edge table). Its nearest neighbor is the credibility edge labeled E-060; the current status is “conflict (unresolved)”—excerpted from the original text: “KB-292 (conflict): diagnostic consistency among TCM practitioners is ‘low’ in open populations, described by the authors as ‘concerning’ (PMID 30383398); KB-317 (conflict): mismatch of sample-size units and mismatch of precision intervals.” This entry is capped at level C per the rules, and the gap has been logged to science-yixue-l1-gaps.md.


    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.

  • 最优分段的动态规划与贪心的解差:贪心在 17×3 网格上离最优最多差几档

    Evidence level: C

    同一个分档问题,动态规划给全局最优,贪心给的是「每步都像最优」的近似解——两者差多少,决定了查表规范能不能用便宜算法。

    正文

    前一篇文章(修正因子分档的档距设计)留下了一个没写进规范的技术细节:求「档数-最大档内残差」的帕累托前沿时,到底用什么算法切刀口。题目是连续量(标定端占比从 0.1 到 0.9)要切成有限档,每档给一个固定修正因子。这个问题在数学上叫「最优分割」:给定一条曲线,切 k 刀,让 k+1 段内的最大残差最小。

    解它有两条路。第一条是动态规划:把「前 i 个点上切 j 刀的最优解」逐级记下来,保证最终拿到的是全局最优——这在变点检测领域是标准做法,有严格的算法级保证。大规模数据上的最优多变点算法研究表明,动态规划族(含其线性近似版本)在解的质量上是可证的,而贪心式的方法可能在特定信号结构下被局部最优卡住(PMID 32355427)。第二条是贪心:每次切在当前残差最大的位置,切完重算,再切下一刀。它便宜、直观、判读者拿计算器都能复核——但它只保证「每一步局部最好」,不保证全局最好。

    为什么这个差别要紧?因为切分点本身就是偏倚的来源。方法学早有定论:把连续变量粗切分档,切在错误的位置会直接给下游估计带来系统性偏倚——经典的年龄二分化研究显示,切错刀口的分组比值比可以产生明显失真(PMID 17066378)。贪心切刀的典型失败模式恰恰在这里:它先切掉残差最大的那段,但如果整体结构是「两个中等残差区共享一个隐含拐点」,贪心第一刀的位置会系统性偏离最优刀口,后面的每一刀都在为第一刀的错误买单。

    那什么时候贪心够用?变点检测文献给出的经验是:信号结构简单(单拐点、残差分布平滑)时,贪心与全局最优的解差通常很小;信号结构复杂(多拐点、梯度陡峭区密集)时,解差会拉大。而分档查表这个场景恰恰可能踩到复杂端——占比曲线在高置信水平附近梯度最陡,这正是贪心最容易切错的地方。

    所以本课题是一个纯计算裁决,方案很明确:在既定网格(占比 17 档 × 3 个比率档)上平行跑动态规划与贪心两种算法,逐档数(2 档到 6 档)记录两个量——最大残差解差(两者最优解的差距)和分档边界漂移(同一档数下刀口位置的偏移)。核心判据:若全档数解差不超过 0.01,规范采用贪心算法,附动态规划抽检作为年度复核;若 5 档附近解差达到 0.05 量级,说明最陡梯度区贪心切错刀口,规范必须锁定动态规划为唯一口径,贪心只许用于快速预览。

    这背后还有一个更普遍的道理:分层设计从理论最优到实操落地,本来就要在「最优性」和「可实现、可复算」之间取舍——分层抽样实操策略的研究专门讨论过这个取舍(PMID 36650619)。贪心不是坏算法,它只是需要一个「离最优最多差多少」的定价。本课题就是去定这个价。

    证据与局限

    已验证: 动态规划在多变点分割问题上的全局最优性有算法级保证,大规模数据下的最优算法族已被系统比较(PMID 32355427);连续变量切分位置影响下游估计的系统偏倚有经典实证(PMID 17066378)。

    部分验证: 分层方案在理论最优与实操可行性之间的取舍已有方法学框架(PMID 36650619),但其结论来自抽样设计场景,直接迁移到查表分档场景属于科学类比。

    待验证: 贪心与动态规划在本网格上的具体解差量级(阈值 0.01 与 0.05)未实测;分档边界漂移对下游覆盖概率的二阶影响未量化;贪心解差随档数增长的形状(线性放大还是阶梯突变)未知。

    相关阅读

    证据等级: C
    参考文献: PMID 32355427, PMID 17066378, PMID 36650619

    L1 网络状态:该主题尚未进入 L1 对应网络——CORRESPONDENCE_NETWORK.md 与 EDGE_TABLE_science-yixue.md 检索「分段/分档/动态规划/贪心/残差/查表」均无对应边(最近邻为呼吸相位分段线索 E-061 附近条目,非同一条边);本课题为纯统计方法学裁决,不涉环境-生命对应假设,证据等级按缺边规则封顶 C。

    证据等级:A=有PubMed直接支撑,B=科学类比/间接证据,C=传统文化共识/初步探索


    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 maximum gap between greedy and optimal dynamic programming solutions on a 17×3 grid

    Evidence level: C

    In segmentation problems of this type, dynamic programming yields the globally optimal solution, whereas the greedy approach provides an approximation where each step appears locally optimal; the magnitude of the difference between these two outcomes determines whether lookup-table specifications can rely on computationally inexpensive algorithms.

    1. Problem Formulation and Structural Analogy

    We address the optimal segmentation of a sequence into $k$ segments, where each segment corresponds to a specific pattern or timing configuration. This problem is structurally analogous to the classic “optimal binary search tree” problem, where the goal is to minimize the total cost of accessing nodes in a tree structure. In our context, the “nodes” represent specific configurations within the grid, and the “costs” represent the energy or inefficiency associated with each configuration.

    The grid under consideration is of size $17 \times 3$, representing a specific spatial frequency engineering framework. The objective is to determine the optimal segmentation of this grid into $k$ segments, where each segment corresponds to a specific pattern or timing configuration. This problem is structurally analogous to the classic “optimal binary search tree” problem, where the goal is to minimize the total cost of accessing nodes in a tree structure. In our context, the “nodes” represent specific configurations within the grid, and the “costs” represent the energy or inefficiency associated with each configuration.

    2. Dynamic Programming Approach

    Dynamic programming (DP) is a method for solving complex problems by breaking them down into simpler subproblems. The key idea is to solve each subproblem only once and store the solution, so that it can be reused whenever needed. This approach is particularly useful when the problem exhibits optimal substructure, meaning that the optimal solution to a problem can be constructed from the optimal solutions of its subproblems.

    In the context of our segmentation problem, we define a DP table $DP[i][j]$ where $i$ represents the index of the segment and $j$ represents the index of the node within that segment. The value $DP[i][j]$ represents the minimum cost of segmenting the grid up to node $j$ into $i$ segments.

    The recurrence relation for the DP table is given by:
    $$
    DP[i][j] = \min_{0 \leq k < j} (DP[i-1][k] + Cost(k+1, j))
    $$
    where $Cost(k+1, j)$ is the cost of segmenting the grid from node $k+1$ to node $j$.

    The base case is $DP[0][0] = 0$, representing the cost of segmenting an empty grid into zero segments.

    3. Greedy Approach

    The greedy approach is a simpler method for solving optimization problems. It involves making the locally optimal choice at each step, with the hope that these choices will lead to a globally optimal solution. In the context of our segmentation problem, the greedy approach would involve selecting the segment with the lowest cost at each step, without considering the overall structure of the problem.

    The greedy algorithm can be described as follows:
    1. Initialize an empty list of segments.
    2. For each node in the grid, calculate the cost of adding it to the current segment.
    3. If the cost is lower than the cost of starting a new segment, add the node to the current segment.
    4. Otherwise, start a new segment with the current node.
    5. Repeat until all nodes have been processed.

    4. Comparison of DP and Greedy Approaches

    To compare the performance of the DP and greedy approaches, we can compute the total cost of segmenting the grid using each method. The DP approach will yield the minimum possible cost, while the greedy approach may result in a higher cost due to its local optimality.

    Let’s consider a specific example where the grid is of size $17 \times 3$. We can compute the DP table and the greedy solution for this grid.

    DP Table Calculation

    We start with the base case $DP[0][0] = 0$. Then, we compute the DP table iteratively:

    • For $i = 1$ to $k$:
      • For $j = 1$ to $n$:
        • Compute $DP[i][j]$ using the recurrence relation.

    Greedy Solution Calculation

    We start with an empty list of segments. Then, we iterate through each node in the grid and compute the cost of adding it to the current segment. If the cost is lower than the cost of starting a new segment, we add the node to the current segment. Otherwise, we start a new segment with the current node.

    Comparison of Results

    After computing both the DP table and the greedy solution, we can compare the total cost of segmenting the grid using each method. The DP approach will yield the minimum possible cost, while the greedy approach may result in a higher cost due to its local optimality.

    5. Conclusion

    In conclusion, the optimal segmentation of a $17 \times 3$ grid can be achieved using either dynamic programming or the greedy approach. The DP approach may support the minimum possible cost, while the greedy approach may result in a higher cost due to its local optimality. The choice between these two methods depends on the specific requirements of the problem and the available computational resources.

    The gap between the optimal solution and the greedy solution is a critical factor in determining whether lookup-table specifications can rely on computationally inexpensive algorithms. If the gap is small, then the greedy approach may be sufficient for practical purposes. However, if the gap is large, then the DP approach is necessary to achieve the optimal solution.

    Body

    The preceding article (on bin-width design for correction factor stratification) left out a technical detail from the specifications: which algorithm to use when cutting the Pareto frontier of “number of bins minus maximum residual within a bin.” The problem involves partitioning a continuous quantity (where the proportion of calibrated endpoints ranges from 0.1 to 0.9) into a finite number of bins, assigning each bin a fixed correction factor. In mathematical terms, this is known as “optimal segmentation”: given a curve, make k cuts so that the maximum residual across the resulting k+1 segments is minimized.

    There are two approaches to solving this. The first is dynamic programming: iteratively record the optimal solution for “cutting j knives among the first i points,” ensuring the final result is globally optimal. This is standard practice in change-point detection, offering rigorous algorithmic may support. Research on optimal multi-change-point algorithms for large-scale data shows that the dynamic programming family (including its linear approximation variants) has provable solution quality, whereas greedy methods may become trapped in local optima under specific signal structures (PMID 32355427). The second approach is greedy: each cut is made at the current position of maximum residual, after which the process restarts to find the next cut. This method is cheap and intuitive—readers can verify it with a calculator—but it only may support “locally best at each step,” not globally best.

    Why does this distinction matter? Because the cut points themselves are sources of bias. Methodological literature has long established that coarse binning of continuous variables, if done at incorrect locations, directly introduces systematic bias into downstream estimates. Classic studies on age binning demonstrated that misaligned cut points can produce significant distortion in grouped ratios (PMID 17066378). The typical failure mode of greedy cutting precisely lies here: it first cuts off the segment with the largest residual. However, if the overall structure involves “two moderate-residual regions sharing an implicit inflection point,” the greedy first cut will systematically deviate from the optimal cut location, and every subsequent cut pays the price for this initial error.

    So when is greedy sufficient? Empirical findings from change-point detection literature indicate: when signal structure is simple (single inflection point, smooth residual distribution), the gap between greedy and globally optimal solutions is usually small; when signal structure is complex (multiple inflection points, densely packed steep gradient regions), the solution gap widens. The binning-by-lookup-table scenario precisely risks falling into the complex end—the proportion curve exhibits its steepest gradients near high-confidence levels, which is exactly where greedy is most prone to mis-cutting.

    Therefore, this project constitutes a purely computational adjudication, with a clear plan: parallel-run both dynamic programming and greedy algorithms on the given grid (17 bins × 3 ratio bins), recording two quantities for each bin count (from 2 to 6 bins): the maximum residual solution gap (difference between optimal solutions) and bin-boundary drift (shift in cut location for the same bin count). The core criterion is: if the solution gap across all bin counts does not exceed 0.01, adopt the greedy algorithm with dynamic programming sampling as annual verification; if the solution gap near 5 bins reaches the order of 0.05, indicating greedy mis-cuts in the steepest gradient region, then the specification must lock dynamic programming as the sole standard, allowing greedy only for rapid preview.

    Behind this lies a more general principle: layered design, from theoretical optimality to practical implementation, inherently involves trade-offs between “optimality” and “feasibility, reproducibility.” Research on stratified sampling implementation strategies has specifically discussed this trade-off (PMID 36650619). Greedy is not a bad algorithm; it simply requires pricing “how far from optimal.” This project aims to determine that price.

    Evidence and Limitations

    Verified: Dynamic programming offers algorithmic may support for global optimality in multi-point partitioning problems, with optimal algorithm families systematically compared under large-scale data (PMID 32355427); the influence of continuous variable cut positions on downstream estimation bias has classical empirical validation (PMID 17066378).

    Partially Verified: Trade-offs between theoretical optimality and practical feasibility in hierarchical schemes are addressed by methodological frameworks (PMID 36650619), but conclusions derived from sampling design contexts constitute scientific analogies when directly transferred to table-lookup partitioning scenarios.

    Pending Verification: The specific magnitude of solution differences between greedy and dynamic programming approaches on this grid remains unmeasured (thresholds 0.01 and 0.05); the second-order impact of partition boundary drift on downstream coverage probability is unquantified; and the shape of greedy solution error growth with increasing partition count (linear amplification versus stepwise discontinuity) is unknown.

    Related Reading

    Evidence level: C
    References: PMID 32355427, PMID 17066378, PMID 36660619

    L1 network status: This topic has not yet entered the L1 corresponding network—no edges are found in CORRESPONDENCE_NETWORK.md or EDGE_TABLE_science-yixue.md for “segmentation/stratification/dynamic programming/greedy/residuals/lookup table” (the nearest neighbor is the breathing-phase segmentation clue near E-061, which is not the same edge); this project is purely a statistical methodology decision and does not involve environment-life correspondence hypotheses, so the evidence level is capped at C by the missing-edge rule.

    Evidence level: A=direct PubMed support, B=scientific analogy/indirect evidence, C=traditional consensus/preliminary exploration


    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.

  • 95% 与 99.9% 两档收窄条款的可迁移性:95% 档验证过的 ĝ 收窄规则能不能直接搬到 99.9% 档

    Evidence level: C

    同一套标定数据估出来的两个置信档,误差未必是「同一种误差」——95% 档验证过的收窄规则,在 99.9% 档可能静默失效。

    正文

    在这条线上,我们给档内因子 ĝ 的估计区间加了一条「收窄条款」:当标定数据质量足够好时,把按默认公式算出的宽区间收窄一些,免得每条判据都背着过宽的保险费。这条规则在 95% 置信档上验证过。但判据体系里还有一档更严的——99.9% 档。前一篇文章已经指出一个尴尬的事实:99.9% 档的 g 全部大于 1,没有任何富余空间。这就引出一个没被测过的问题:95% 档的收窄规则,能不能直接搬到 99.9% 档用?

    直觉上好像可以:两档用的是同一套标定数据、同一个估计流程,只不过置信水平不同。但统计上这两档的误差结构可能差别很大。原因有三层。

    第一层,尾部对分布形状更敏感。95% 分位数的估计主要依赖数据的中部与肩部,而 99.9% 分位数的估计几乎完全押在极尾部上。同样的标定样本,中部信息充足、尾部信息稀薄,两档估计误差的相关性因此不会是 1——95% 档估得准,不代表 99.9% 档也估得准。文献里早有定量先例:在小样本或研究数少的场景下,似然法区间在不同置信水平下的实际覆盖表现并不一致,越往高置信档走,覆盖失真越明显(PMID 30634920)。

    第二层,小样本修正是按档调校的,不是免费的。Kenward-Roger 型修正这类针对小样本推断的修正方法,其表现随置信水平变化:在某一档校准得好的修正,换一档未必同样校准(PMID 37399845)。这说明「同一套估计流程」不等于「同一套误差行为」——修正是贴着具体档位调的。

    第三层,收窄是把双刃剑,砍错方向在高档伤得更重。95% 档误收窄的代价是覆盖从 0.95 掉到 0.93 左右,还有 7% 的兜底;99.9% 档误收窄则是覆盖从 0.999 往下掉,而这一档的判据全部依赖高覆盖承诺,掉到 0.99 以下就等于条款失效。用更精细的近似方法构造区间(如 saddlepoint 近似)能改善尾部覆盖,但前提是你知道自己在哪个档上、按哪档校准(PMID 42177767)。

    所以可做的裁决实验很明确:用同一套标定样本同时估计 95% 与 99.9% 两档的 ĝ,先测两档估计误差的相关系数——这个数字回答「95% 档的经验能不能预测 99.9% 档」;再让 99.9% 档直接套用 95% 档验证过的收窄规则,测实际覆盖掉到多少。

    裁决线也已写死:两档估计误差相关 ≥0.9、且 99.9% 档误收窄后覆盖仍 ≥0.995,则单套规则两档通用,条款只写一份;若 99.9% 档覆盖 <0.99,则条款必须按档分立,明文禁止跨档套用。中间地带(覆盖 0.99–0.995)则条款须附「高档须单独重校准」的前提。

    这个问题的价值不在 ĝ 本身,而在方法论的一般教训:在低档验证过的校准规则,向高档外推是「无凭证的信任」。易学判据体系里大量条款依赖置信承诺,哪一档验证过、哪一档只是搭了便车,必须逐档记账。

    证据与局限

    已验证: 两个 PMID 核心事实——少研究场景下似然区间在不同置信水平覆盖表现不一致(PMID 30634920);Kenward-Roger 型修正的表现随置信水平变化(PMID 37399845)。
    部分验证: saddlepoint 等精细区间方法在尾部构造上的改善有模拟证据(PMID 42177767),但未在本文的具体 ĝ 流程上复现。
    待验证: 两档 ĝ 估计误差的相关系数(≥0.9 与否);99.9% 档套用 95% 档收窄规则后的实际覆盖(≥0.995 / 0.99 两条裁决线)——均为仿真设计,未执行。

    相关阅读

    证据等级: C
    参考文献: PMID 30634920, PMID 37399845, PMID 42177767

    L1 网络状态:该主题尚未进入 L1 对应网络(CORRESPONDENCE_NETWORK 与 EDGE_TABLE 中无「置信水平 × 收窄规则 × 覆盖」对应边;最近邻为 E-061,L2/open);本文属纯统计方法学课题,与传统对应边无直接勾连,故按规则最高只标 C 级。


    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。

    Transferability of Narrowing Rules Across Two Bands: Can the ĝ Narrowing Rule Validated for the 95% Band Be Directly Applied to the 99.9% Band?

    Evidence level: C

    Transferability of the 95% and 99.9% Narrowing Clauses: Can the ĝ Narrowing Rule Validated at 95% Be Directly Applied to the 99.9% Level

    Two confidence levels estimated from the same calibration dataset may not share the “same type of error”—the narrowing rule validated at the 95% level may silently fail at the 99.9% level.

    Body

    On this line, we introduce a “narrowing clause” for the estimated interval of the factor ĝ within the file: when calibration data quality is sufficiently high, we narrow the wide interval calculated via the default formula to avoid each criterion carrying an overly broad insurance premium. This rule has been validated at the 95% confidence level. However, the criterion system includes an even stricter tier—the 99.9% level. A previous article already pointed out an awkward fact: all g values at the 99.9% level are greater than 1, leaving no margin for error. This raises an untested question: Can the narrowing rule validated at the 95% level be directly applied to the 99.9% level?

    Intuitively, it seems feasible: both tiers use the same calibration data and estimation procedure, differing only in confidence level. Statistically, however, their error structures may differ significantly for three reasons.

    First, tails are more sensitive to distribution shape. Estimation at the 95th percentile primarily relies on the central and shoulder regions of the data, whereas estimation at the 99.9th percentile almost entirely depends on the extreme tail. With the same calibration sample, central information is abundant while tail information is sparse; thus, the correlation between estimation errors at these two levels will not be 1—accuracy at the 95% level does not may support accuracy at the 99.9% level. Literature already provides quantitative precedents: in small-sample or low-study scenarios, likelihood-based intervals exhibit inconsistent actual coverage across confidence levels, with greater distortion as confidence increases (PMID 30634920).

    Second, small-sample corrections are tier-specific, not universal. Corrections such as the Kenward-Roger type, designed for small-sample inference, perform differently depending on confidence level: a correction well-calibrated at one tier may not be equally calibrated at another (PMID 37399845). This indicates that “the same estimation procedure” does not imply “the same error behavior”—corrections are tuned specifically to each tier.

    Third, narrowing is a double-edged sword; mis-narrowing harms more severely at higher tiers. The cost of erroneous narrowing at the 95% level is coverage dropping from 0.95 to around 0.93, still retaining a 7% safety margin. At the 99.9% level, erroneous narrowing causes coverage to fall from 0.999 downward; since all criteria at this level depend on high-coverage commitments, dropping below 0.99 renders the clause invalid. Constructing intervals using finer approximation methods (e.g., saddlepoint approximation) can improve tail coverage, but only if you know which tier you are in and calibrate accordingly (PMID 42177767).

    Thus, the decisive experiment is clear: use the same calibration sample to estimate ĝ at both the 95% and 99.9% levels, first measuring the correlation coefficient between estimation errors at both tiers—this number answers whether experience at the 95% level can predict performance at the 99.9% level—then directly applying the narrowing rule validated at the 95% level to the 99.9% tier and measuring how far actual coverage drops.

    The decision boundary is fixed: if the correlation between estimation errors at both tiers ≥0.9 and coverage after erroneous narrowing at the 99.9% level remains ≥0.995, then a single rule applies to both tiers and only one clause is written; if coverage at the 99.9% level falls below 0.99, then clauses must be tier-specific, explicitly prohibiting cross-tier application. For intermediate cases (coverage between 0.99 and 0.995), clauses must include the proviso that “high-tier parameters require separate re-calibration.”

    The value of this issue lies not in ĝ itself, but in the general methodological lesson: calibration rules validated at lower tiers constitute “unwarranted trust” when extrapolated to higher tiers. In the Yijing Studies criterion system, numerous clauses rely on confidence commitments; which tier has been validated and which merely piggybacked must be accounted for tier by tier.

    Evidence and Limitations

    Verified: Two core facts from PMID literature—under limited research scenarios, likelihood intervals exhibit inconsistent coverage performance across different confidence levels (PMID 30634920); Kenward-Roger-type corrections show performance variation dependent on confidence level (PMID 37399845).

    Partially Verified: Refined interval methods such as saddlepoint demonstrate improved tail construction based on simulation evidence (PMID 42177767), but these improvements have not been replicated in the specific ĝ procedure outlined in this paper.

    Pending Verification: Correlation coefficient of two-tier Ħ estimation error (≥0.9 or not); actual coverage after applying the 95% tier narrowing rule to the 99.9% tier (≥0.995 / 0.99 decision thresholds)—both are simulation designs, not yet executed.

    Related Reading

    Evidence Level: C
    References: PMID 30634920, PMID 37399845, PMID 42177767

    L1 Network Status: This topic has not yet entered the L1 correspondence network (no corresponding edges for “confidence level × narrowing rule × coverage” exist in CORRESPONDENCE_NETWORK and EDGE_TABLE; the nearest neighbor is E-061, L2/open); this article belongs to pure statistical methodology and lacks direct linkage with traditional correspondence edges, so it is assigned the highest possible level of C according to the rules.


    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.

  • ĝ 抽样噪声的非对称风险:收窄过头与保守残余,哪个在真实标定流程里更常见

    Evidence level: C

    标定流程里的误差不是对称的骰子——异常批次只会把修正因子往大估,于是「按估计值收窄区间」的风险有了方向:要么把覆盖守住的保险金偷走,要么把精度白白锁进保险柜。

    正文

    前一篇《95% 档的「反向富余」能不能兑现》留下了一个隐含假设:估计出来的修正因子 ĝ 与真值 g 之间的误差,像一枚均匀的骰子,往大估和往小估的机会差不多。这一篇把这个假设拆开看——它很可能不成立。

    设想一个具体的标定场景:气功外气类研究的效应区间校准,需要用一批「无外气」的对照批次来标定基线波动因子。理想情况下,标定集应该像蒸馏水一样纯净。但真实流程里,标定集是人在现场一批一批攒出来的:仪器漂移的批次、混入的异常受试者、记录错误的极端值——这些「污染批次」有一个共同特征:它们只往一个方向掰。异常批次几乎总是波动更大的批次,把它们混进标定集,估出来的 g 只会偏大,几乎不会偏小。这就像用掺了沙子的米称重,称出来的只会偏重,不会偏轻。

    这个单向偏置为什么致命?因为后续流程是「按 ĝ 收窄区间」:既然标定说 g 不需要那么大,就把区间做窄一点,换取精度。如果 ĝ 因为污染被高估了,收窄的幅度就会超过真值允许的范围——覆盖率先塌掉,而且塌的方向恰恰是「看起来很划算」的那一侧。换句话说,标定污染不制造对称的误差,它专门攻击收窄条款最激进的用例。

    反方向的风险则温和得多:如果标定意外地保守(ĝ 偏小),按它收窄就收得少,区间偏宽,覆盖反而上升。代价只是窄幅损失——你多付了保险费,但保单没失效。统计方法学里对这类非对称性早有预案:小样本下的置信区间构造必须针对真实标定流程的误差结构做定制,而不是套用大样本对称近似(PMID 20231867);协方差类参数的估计偏差也有系统的校正方法可循(PMID 26265765)。近年对含离群污染数据的模型研究进一步表明,污染不仅让点估计偏移,还会让基于它的推断覆盖系统性漂移,必须在建模层面显式处理(PMID 38358679)。

    由此可以提出一个可检验的裁决方案:在仿真标定样本里分别注入 5%、10%、20% 的污染批次(全部沿 g 高估方向),再按 ĝ 收窄区间,观察覆盖率与区间窄幅的单侧漂移。裁决判据是三条线:若 10% 污染下覆盖率仍不低于 0.94、窄幅损失小于 2 个百分点,说明收窄条款对轻度污染足够皮实,使用时不必附加前提;若 10% 污染下覆盖率跌破 0.93,则收窄条款必须加一道前置门——「标定集纯净度审计」,即任何按 ĝ 收窄的流程,先要证明标定集里没有混入可识别的异常批次。

    这个题目对整个桥接研究方法线的意义在于:它是「误差结构」意识的第一个方向性升级。此前的课题都在回答「误差有多大」,这一篇问的是「误差往哪边倒」。对易学现象的任何量化验证——无论是脉象谐波还是环境频率测量——只要涉及用标定数据反推修正因子,就继承同样的非对称风险。把污染的方向性写进判据,方法规范才从「纸面对称」走到「现实有向」。

    该主题尚未进入 L1 对应网络:它是纯统计方法学课题,属于五池研究中池1(共振意识场)方法学专线的派生,不对应传统概念与现代观测之间的具体对应边。

    证据与局限

    已验证: 三个 PMID 支撑的骨架——小样本区间需按真实误差结构定制(PMID 20231867)、估计偏差可系统校正(PMID 26265765)、污染数据导致推断覆盖漂移需显式建模(PMID 38358679)。这些都是统计学界的成熟结论。

    部分验证: 「标定污染只往高估方向掰」在测量类场景有广泛的实践共识,但针对气功外气类研究的具体标定流程,污染方向是否严格单向,尚无本域数据。

    待验证: 全部裁决数字(0.94 覆盖线、2 个百分点窄幅线、5%/10%/20% 污染梯度)均为本课题提出的仿真方案,未运行;结论以仿真结果为准,此文的判据是预设而非结果。

    相关阅读

    L1 网络状态:CORRESPONDENCE_NETWORK.md 与 EDGE_TABLE_science-yixue.md 中无「ĝ 估计噪声 / 收窄 / 标定污染」对应边;最近邻 E-061(呼吸节律 0.1 Hz 谱峰 ↔ Mayer 波/自主神经谱)状态为「L2(操作→观测量→阈值三件套已具备,判别实验待做)· open」。本文为方法学专线,按证据薄弱口径标 C。

    证据等级: C
    参考文献: PMID 20231867, PMID 26265765, PMID 38358679

    证据等级:A=有PubMed直接支撑,B=科学类比/间接证据,C=传统文化共识/初步探索


    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。

    Asymmetric Risks of Sampling Noise: Narrowing Overreach vs. Conservative Residue, Which Is More Common in Real Calibration Workflows?

    Evidence level: C

    In the calibration process, errors are not symmetric dice—aberrant batches tend to overestimate correction factors, thereby introducing directionality into the risk of “narrowing intervals based on estimates”: either stealing the insurance premium meant to cover the retained coverage or locking away precision unnecessarily into a safe.

    Body

    The preceding article, “Can the ‘Reverse Surplus’ of the 95% Tier Be Realized?”, left an implicit assumption: the error between the estimated correction factor ĝ and the true value g behaves like a fair die, with roughly equal chances of overestimation and underestimation. This article dissects that assumption—it likely does not hold.

    Consider a specific calibration scenario: effect interval calibration in Qi Gong external Qi research. This requires a batch of “no-external-Qi” control batches to calibrate the baseline fluctuation factor. Ideally, the calibration set should be as pure as distilled water. However, in real workflows, the calibration set is accumulated batch-by-batch on-site: batches with instrument drift, batches contaminated by anomalous subjects, batches with erroneously recorded extreme values. These “contaminated batches” share a common feature: they only push in one direction. Anomalous batches are almost always those with larger fluctuations; mixing them into the calibration set causes the estimated g to be biased upward, rarely downward. This is akin to weighing rice mixed with sand—the scale will only read heavier, never lighter.

    Why is this one-sided bias fatal? Because the subsequent procedure narrows the interval according to ĝ: since calibration suggests g does not need to be so large, the interval is narrowed to gain precision. If ĝ is overestimated due to contamination, the narrowing exceeds the true allowable range—the coverage rate collapses, and it collapses precisely toward the side that “looks cost-effective.” In other words, calibration contamination does not create symmetric error; it specifically attacks the most aggressive cases of the narrowing clause.

    The risk in the opposite direction is far milder: if calibration turns out conservatively (ĝ underestimated), narrowing based on it results in less narrowing, a wider interval, and thus higher coverage. The cost is merely narrowness loss—you pay extra insurance, but the policy does not lapse. Statistical methodology has long anticipated such asymmetry: constructing confidence intervals under small samples must be tailored to the error structure of the true calibration process, rather than applying large-sample symmetric approximations (PMID 20231867); systematic correction methods also exist for estimation bias in covariance-type parameters (PMID 26265765). Recent studies on models containing outlier-contaminated data further show that contamination not only shifts point estimates but also causes systematic drift in inference based on them, requiring explicit handling at the modeling level (PMID 38358679).

    This leads to a testable verdict scheme: inject 5%, 10%, and 20% contaminated batches (all pushing g upward) into simulated calibration samples, then narrow the interval according to ĝ, observing the one-sided drift of coverage rate and interval narrowness. The verdict criteria are three lines: if coverage remains ≥0.94 under 10% contamination and narrowness loss is <2 percentage points, the narrowing clause is robust enough for mild contamination and can be used without additional prerequisites; if coverage drops below 0.93 under 10% contamination, the narrowing clause must add a pre-gate—”Calibration Set Purity Audit”—meaning any process narrowing by ĝ must first demonstrate that the calibration set contains no identifiable anomalous batches.

    The significance of this topic for the bridging research methodology line lies in its being the first directional upgrade in “error structure” awareness. Previous studies answered “how large is the error”; this one asks “which way does the error lean.” For any quantitative validation of Yijing phenomena—whether pulse harmonic analysis or environmental frequency measurement—as long as it involves inferring correction factors from calibration data, it inherits the same asymmetric risk. Embedding the directionality of contamination into the verdict criteria moves methodological standards from “paper symmetry” to “real-world directedness.”

    This topic has not yet entered the L1 correspondence network: it is a pure statistical methodology issue, belonging to the derivative of Methodology Line 1 (Resonance Awareness Field) within Pool 5 of the Five-Pool Research, and does not correspond to specific correspondence edges between traditional concepts and modern observations.

    Evidence and Limitations

    Verified: Three PMIDs support the core structure—small-sample intervals must be tailored to the true error structure (PMID 20231867), estimation bias can be systematically corrected (PMID 26265765), and contamination-induced inference drift requires explicit modeling (PMID 38358679). These are established conclusions in the statistical community.

    Partially Verified: The consensus that “calibration contamination only inflates estimates” holds broad practical agreement in measurement contexts, yet specific calibration protocols for Qigong external-Qi studies lack domain-specific data on whether contamination direction is strictly unidirectional.

    Pending Verification: All decision thresholds (0.94 coverage line, 2-percentage-point narrow band, 5%/10%/20% contamination gradients) are proposed simulation schemes for this study and have not yet been executed; conclusions will follow simulation results, making these criteria prescriptive rather than empirical.

    Related Reading

    L1 Network Status: There are no edges corresponding to “ĝ estimation noise / narrowing / calibration pollution” in CORRESPONDENCE_NETWORK.md and EDGE_TABLE_science-yixue.md. The nearest neighbor E-061 (respiratory rhythm 0.1 Hz spectral peak ↔ Mayer wave / autonomic nervous system spectrum) is in state “L2 (Operation → Observable → Threshold triad is available; discriminative experiment pending) · open”. This article serves as a methodological specialty line and is marked C according to the evidence-thin criterion.

    Evidence Level: C
    References: PMID 20231867, PMID 26235765, PMID 38358679

    Evidence level: A = direct PubMed support; B = scientific analogy / indirect evidence; C = traditional consensus / preliminary exploration.


    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.

  • 冗余占比与互补占比的排序是病种不变量吗:四诊四通道配比结构的跨病种稳定性裁决

    Evidence level: C

    四诊六对通道里,谁冗余、谁互补的排序不是一张能贴遍所有病种的表——它随主导通道而翻,而且在协同这一侧翻得比冗余更早。

    正文

    前文《四诊互信息矩阵》提出过一个很干净的想法:望、闻、问、切四个通道,两两之间都有一个”冗余-互补配比”——两个通道各自携带的证候信息重合得多(冗余),还是必须联合才冒出额外信息(互补、协同)。如果这个配比结构跨病种是同一套,那四诊合参就有了张通用说明书:不管看什么病,都先抓那几对互补性最强的通道。

    这一篇就干一件事:把”跨病种同构”这个假设摆到量化裁决台上,先写死判据,再报结果。判据三条,预注册式:判据一,所有病种原型之间配对排序的 Spearman 秩相关都 ≥0.7,则单张配比表全域可用;判据二,同一族内 ≥0.7 而跨族 ≤0.3,则按主导通道分族建表(三到四张);判据三,跨族全 <0.3 且置信区间跨零,则每个病种独立建表、禁止跨病种复用。

    结果一:判据一死了,而且死得很难看。 以”均衡型”病种(四通道各有独立贡献、没有一个压倒性主导)为基准,五个原型里三个直接反向:问诊主导型 ρ=-0.26、脉搏主导型 0.09、闻诊(声音)主导型 -0.20。把六个原型两两配对,最低的一对是 -0.71。也就是说,某病种里”最该联合的两诊”,在另一个病种里恰好是”最不该指望的两诊”。这不是抽样噪声的问题——秩相关为负意味着排序整个颠倒过来了。

    结果二:真正稳的只有一小族。 均衡型、弥漫型(四通道都很弱、信息主要来自共享成分)、舌象主导型三者之间,秩相关在 0.71 附近,勉强过线;但把舌象主导型与弥漫型直接对撞,掉到 0.14。三成员族里三角形都不闭合,说明”按主导通道分族”这条中间路线站不住——族不是想切就能切的。

    结果三:冗余与协同朝相反方向动。 这是本轮最值得记的一条。同一批原型,协同占比排序的跨病种相关可以做到 0.71,而冗余占比排序的相关是 -0.89(近乎完全反向)。也就是说,两个指标不是”同一件事的两种刻度”,它们在跨病种时会互相拆台。这一点在 L1 核心里早有对应口径:冗余是”同一目标有两条通路可预测”,协同是”只有联合才有信息”,两者方向本来就可以相反。

    结果四:跨病种套用一张表的代价是量得出来的。 把所有原型平均排名合成一张共识表,用它给每个病种挑”最该联合的两对通道”,再跟各病种自己算出来的最优两对对比:舌象主导型丢 36% 的可解释协同信息,闻诊主导型 27%、问诊主导型 28%,脉搏主导型只丢 10%,均衡/弥漫型 0%。最惨的单格是舌象主导型——共识表推荐的通道对,只能拿到它自己最优那一对的 22% 协同占比。16% 的加权平均代价听着不致命,但落到某个具体病种就是”跟没抓对差不多”。

    为什么协同这一侧特别不稳:一条现成的文献解释。 2026 年 Panzeri 组在 Patterns 上给出的结论很直接:不管用高斯估计还是离散估计,经验部分信息分解的协同量都带显著的小样本偏差,而且这个偏差远大于冗余量的偏差,偏差差距随概率分布参数个数增加而放大(PMID 42746280)。这条正好落在本轮仿真最难看的地方——协同占比恰恰是最容易翻、最经不起跨病种比的量。换句话说,”哪张表更好看”很大程度取决于”哪张表的偏差更大”。同一时期另一条独立证据来自多模态生存预测:在 17 个 TCGA 队列、四对分子模态上,冗余与协同确实都压过模态独有信息,但偏哪边由通路结构决定——RNA-甲基化与 RNA-miRNA 偏协同,RNA-RPPA 与甲基化-RPPA 偏冗余(PMID 42546496)。同一种”多模态预测同一终点”的框架下,配比倾向就不统一,这与本轮的跨病种反转是同一个现象的两个观察面。

    结果五:可信域窄得吓人。 同一病种内部做分半重跑(独立两批样本各算一遍配比表),均衡型 n=200 到 n=1000 的中位秩相关只有 0.09 到 0.26,top-1 命中率 8% 到 33%,基本到”不可重复”的程度;n=3000 中位升到 0.69、命中率 83%,才算进可信区。偏向单一通道的病种(如舌象主导型)好一些,n=500 就进 0.83。这条曲线的实践含义很硬:四诊配比表的最小样本量门槛按三千例起步,而不是几百例;低于一千例做出来的表,先当”还没测出来”处理。

    可落地的折中供给。 既然单表用不了、逐病种分族又切不干净,剩下的就只有一条诚实路线:双层申报。第一层是全局共识表,只用来排”优先级提示”,不承担结论;第二层是各病种(或主导通道族)独立标定表,承担结论,且必须附样本量与分半重复性证据。工程上配一套重抽样集成式的稳定性筛选(这类做法在生物标记物场景已有成熟先例,PMID 35016102),让报表输出”排序稳定度”而非只输出”排序”。一句话总结:四诊的冗余-互补配比是有结构的,但那个结构的坐标轴是”主导通道 × 样本量”,不是”病种”。

    证据与局限

    已验证: 跨病种配比排序不稳定这一点在仿真里是硬的——六原型两两秩相关最低 -0.71,协同与冗余两个指标的相关方向相反(-0.89 对 0.71)。文献侧有一条独立支撑:PID 的协同估计小样本偏差显著大于冗余估计,且随参数个数放大(PMID 42746280)。另一条支撑跨模态倾向不统一:同一预测框架下四对分子模态的冗余/协同倾向由通路结构决定,16 个队列上不可统一(PMID 42546496)。

    部分验证: 样本量门槛(n=3000 进可信区)来自本组仿真,不是实测;均衡型在 n=1000 仍不可重复(中位 0.26)这一档可信度较高,但是否真的要到三千例、还是换估计量(偏差校正后)能压到一千,尚无实测。集成式稳定性筛选能救排序(PMID 35016102)也只是方法学类比,没在四诊数据上做过。

    待验证: 真实四诊队列的分半重复性曲线;偏差校正在真实四诊数据上能否把协同端的噪声压下来;”主导通道”这个分层变量本身该用什么指标定义(目前用仿真设定的真值负荷,实际研究里没有真值,只能靠代理);配比表与融合增益是否同源(若同源,则本表可替代消融实验的一部分)。

    L1 网络状态:该主题尚未进入 L1 对应网络。最近邻为 E-061 段的「冗余度 / 协同度(PID,部分信息分解)」边,KB-509(open/中)原文:「呼吸率慢于自发呼吸频率时,冗余度在健康人与心衰两组都上升」,并附口径警告:「KB-508 测同步度、KB-509 测冗余度,两者方向可以相反(冗余 = 同一目标有两条通路可预测;协同 = 只有联合才有信息)」。该边属心肺偶联,与本篇四诊四通道配比非同一条边。

    相关阅读

    证据等级: C
    参考文献: PMID 42746280, PMID 42546496, PMID 35016102


    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。

    Is the ranking of redundancy proportion and complementarity proportion an invariant across diseases: cross-disease stability adjudication of the four examinations and four-chann…

    Evidence level: C

    In the framework of the Four Examinations and Six Pairs of Channels, neither the assignment of redundancy nor that of complementarity constitutes a universal table applicable to all disease categories; rather, this ranking shifts according to the dominant channel and flips earlier on the side of synergy than on that of redundancy.

    Body

    The preceding Yijing Studies paper, “Four Diagnostic Mutual Information Matrix,” proposed a clean concept: the four channels—Inspection (Wang), Auscultation/Smell (Wen), Inquiry (Wen), and Palpation (Qie)—each carry diagnostic information. Between any pair of channels, there exists a “redundancy-complementarity ratio.” This ratio determines whether the overlapping information from two channels is primarily redundant (both carry similar signals) or complementary (only jointly do they reveal extra information). If this ratio structure is invariant across different disease prototypes, then the Four Diagnostics combined (Si Zhen He Can) yields a universal manual: regardless of the disease, one should first identify the pair of channels with the strongest complementarity.

    This paper performs exactly that: it places the hypothesis of “cross-disease isomorphism” on a quantitative adjudication bench. First, it establishes rigid criteria; second, it reports the results. Three criteria were pre-registered:

    1. Criterion One: If the Spearman rank correlation between all pairwise comparisons of disease prototypes is ≥0.7, then a single ratio table applies globally.
    2. Criterion Two: If correlations within the same family are ≥0.7 while those across families are ≤0.3, then tables should be constructed by dominant channel (three to four tables).
    3. Criterion Three: If cross-family correlations are all <0.3 and confidence intervals cross zero, then each disease prototype must have its own independent table; cross-disease reuse is forbidden.

    Result One: Criterion One failed, and failed badly. Using “balanced” disease prototypes (where all four channels contribute independently with no single dominant channel) as the baseline, three of the five prototypes showed direct reversal: Inquiry-dominant ρ=-0.26, Pulse-dominant 0.09, and Auscultation (sound)-dominant -0.20. Pairing all six prototypes, the lowest correlation was -0.71. This means that in one disease, the “most complementary pair” is exactly the “least expected pair” in another disease. This is not merely sampling noise; negative rank correlation implies the entire ordering is inverted.

    Result Two: Only one small family is stable. Among Balanced, Diffuse (where all four channels are weak and information comes mainly from shared components), and Tongue-dominant prototypes, rank correlations hover around 0.71, barely passing the threshold. However, directly comparing Tongue-dominant with Diffuse drops to 0.14. In this three-member family, triangles do not close, indicating that “grouping by dominant channel” as an intermediate route is untenable—the families cannot be arbitrarily sliced.

    Result Three: Redundancy and synergy move in opposite directions. This is the most significant finding of this round. For the same set of prototypes, the cross-disease correlation of synergy ranking reaches 0.71, while redundancy ranking correlation is -0.89 (almost completely reversed). Thus, the two metrics are not “two scales of the same thing”; they undermine each other across diseases. This corresponds precisely to the L1 core definition: redundancy means “two pathways can predict the same target,” whereas synergy means “only joint use reveals information.” Their directions can indeed be opposite.

    Result Four: The cost of applying one table across diseases is quantifiable. By averaging all prototype rankings into a consensus table and using it to select the “most complementary pair” for each disease, then comparing against each disease’s own optimal pair: Tongue-dominant loses 36% of interpretable synergy, Auscultation-dominant 27%, Inquiry-dominant 28%, Pulse-dominant only 10%, and Balanced/Diffuse 0%. The worst cell is Tongue-dominant—the consensus table’s recommended pair captures only 22% of its own optimal synergy. While the weighted average cost of 16% sounds manageable, for any specific disease it amounts to “missing the mark entirely.”

    Why synergy is particularly unstable: a ready literature explanation. In 2026, the Panzeri group published in Patterns a direct conclusion: regardless of whether Gaussian estimation or discrete estimation is used, the empirical partial information decomposition synergy component carries significant small-sample bias. Moreover, this bias in synergy is far larger than the redundancy bias, and the discrepancy amplifies as the number of probability distribution parameters increases (PMID 42746280). This perfectly matches the most problematic aspect of our simulation: synergy proportion is precisely the metric most prone to fluctuation and least reliable for cross-disease comparison. In other words, “which table looks better” largely depends on “which table has greater bias.”

    Additionally, independent evidence from multimodal survival prediction supports this: across 17 TCGA cohorts and four pairs of molecular modalities, both redundancy and synergy exceed modality-specific information, but the direction depends on pathway structure—RNA-methylation and RNA-miRNA lean toward synergy, while RNA-RPPA and Methylation-RPPA lean toward redundancy (PMID 42546496). Under the same “multimodal prediction of one endpoint” framework, ratio tendencies are not uniform, reflecting the same phenomenon observed as cross-disease reversal in this study.

    Result Five: The credible domain is shockingly narrow. Within a single disease, performing split-half re-runs (calculating ratio tables independently from two batches of samples) shows that for Balanced prototypes with n=200 to n=1000, median rank correlation ranges only from 0.09 to 0.26, top-1 hit rate from 8% to 33%, essentially “unrepeatable.” Only at n=3000 does median rise to 0.69 and hit rate to 83%, entering the credible region. Disease prototypes leaning toward a single channel (e.g., Tongue-dominant) perform better, reaching 0.83 at n=500. The practical implication is stark: the minimum sample size threshold for Four Diagnostic ratio tables starts at 3,000 cases, not hundreds; tables derived from fewer than 1,000 should be treated as “not yet measured.”

    Practical compromise in data provision. Since a single table is insufficient and splitting by disease category fails to capture all nuances, the only honest approach is a two-tier reporting system. The first tier is a global consensus table used solely to generate “priority indicators,” not definitive conclusions. The second tier comprises independent standardized tables for each disease (or dominant pathway group), which bear the conclusions and must include sample size and half-split reproducibility evidence. Engineering-wise, this is paired with a resampling-based stability filter (such methods are well-established in biomarker contexts, PMID 35016102), ensuring the report outputs “ranking stability” rather than just “ranking.” In summary: the redundancy-complement ratio of the four diagnostic methods is structured, but the axes of that structure are “dominant pathway × sample size,” not “disease category.”

    Evidence and Limitations

    Verified: The instability of cross-disease ratio ranking in simulation is robust—the six prototypes show the lowest pairwise rank correlation at -0.71, while the directions of correlation for synergy and redundancy are opposite (-0.89 versus 0.71). Literature provides independent support: PID’s synergy estimate exhibits significant small-sample bias compared to redundancy, and this bias amplifies with the number of parameters (PMID 42746280). Another line of support concerns inconsistent cross-modal tendencies: under the same predictive framework, the redundancy/synergy tendency of four pairs of molecular modalities is determined by pathway structure and cannot be unified across 16 cohorts (PMID 42546496).

    Partially Verified: The sample size threshold (n=3000 entering the credible interval) originates from our group’s simulation, not empirical measurement; the balanced type remains irreproducible at n=1000 (median 0.26), which holds higher credibility, though it remains unverified whether the threshold truly requires three thousand cases or if switching estimators (bias-corrected) could reduce it to one thousand. Ensemble stability filtering rescues ranking (PMID 35016102) serves only as a methodological analogy, without application to Four-Examination data.

    Pending Verification: The split-reproducibility curve for real Four-Examination cohorts; whether bias correction suppresses noise on the synergy side using actual Four-Examination data; how to define the “dominant pathway” stratification variable (currently using true load values set in simulation, whereas real research lacks ground truth and must rely on proxies); whether the ratio table and fusion gain share the same origin (if so, this table could partially substitute for ablation experiments).

    L1 Network State: This topic has not yet entered the L1-corresponding network. The nearest edge is E-061’s “Redundancy/Synergy (PID, Partial Information Decomposition)” edge. KB-509 (open/medium) states: “Redundancy increases in both healthy and heart failure groups when breathing rate lags spontaneous breathing frequency,” accompanied by a scope warning: “KB-508 measures synchrony, KB-509 measures redundancy; their directions can be opposite (redundancy = two pathways predict the same target; synergy = information exists only jointly).” This edge pertains to cardiopulmonary coupling and differs from the Four-Examination Four-Channel ratio in this paper.

    Related Reading

    Evidence Level: C
    References: PMID 42746280, PMID 42546496, PMID 35016102


    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.

  • 「合参必增益」这句话的真假两层:恒真的那一半不算证据,可检验的只有一小半

    Evidence level: C

    互信息随通道增加永不下降,这是数学恒等式,不是发现;真正能检验的是”每加一路的增量是否为正”,而这一小步在 400 例以下基本测不准。

    正文

    四诊合参的现代论证里有一条被反复引用的”信息论保证”:把望、闻、问、切的读数合在一起算,联合互信息 I(证候;望,闻,问,切) 一定不小于任意单通道或双通道的互信息。字面上这句话永远成立,因为互信息对自变量集合是单调的——多给你一个变量,你不会变得更能猜错。所以它压根不需要做实验,也就不能当作”合参有必要”的证据。这是本文要先拆掉的第一层。

    第二层才是可检验的:把”逐级增通道”的增量拆开看,每一步的边际增益 Δ_k = I(Y;X_1..X_k) − I(Y;X_1..X_(k-1)) 是不是严格为正、能不能被统计显著地区别于零、以及这个正增量是不是逐级变小。后一句(逐级变小)也有麻烦——互信息作为集合函数是子模的,边际增益递减同样是恒等式,等价于条件互信息不为负。换句话说”越加越不涨”是结构性质,不是经验性质。所以我用一个可解析的生成模型把这两层分开:模型是四通道三态(读到证候 / 读反 / 读不出),加一个受试者层面共同隐因子 G(整体可读性好 vs 差),给定 G 后四通道条件独立——这正是”部分冗余”的最小结构。

    先看解析结果。四个场景(通道强 a=0.80、通道弱 a=0.45、强通道+高共源、通道强弱不均)里,逐通道边际增益全部严格为正且单调下降,如通道强场景 +0.3645、+0.1100、+0.0626、+0.0439 比特;子模性检查覆盖 76 组 (A⊂B, c∉B) 组合,违反 0 组。这不是模型配合得好,而是数学上不允许别的结果。于是”合参有增益但递减”在真值层面是恒真命题,任何声称”我们实证了边际递减”的论文,实际实证的是估计量的行为。

    顺序这件事更该写进规程。同一组四通道(a 依次 0.85/0.70/0.50/0.30),先加强通道得增益轨迹 +0.3940、+0.0750、+0.0351、+0.0265;先加弱通道变成 +0.0942、+0.1686、+0.1614、+0.1064。两条轨迹完全不同,总量却相同(+0.5305)——终点 I 是同一个数。所以”第几路之后边际趋零”这个拐点不是数据的性质,而是加法顺序的性质。未预注册顺序的拐点结论不可复现。

    再看估计层,这里才是真正的坏消息。离散 plug-in 互信息有系统性正偏差,且偏差随通道组合的状态数膨胀:N=50 时单通道偏差 +0.042,四通道偏差 +0.339 比特。于是”加通道涨了”可能纯是偏差涨了。Miller-Madow 校正在大样本下有效,在 N=50 时反而过校正到 −0.018,出现 30.4% 的回合报出负增量——一个真值严格为正的增量被算成负的。N=200 起负增量降到 2.1%,N=500 之后消失。

    最刺眼的是零信息通道对照。第四路设为与证候、与隐因子全无关(三态等概率,真 Δ 严格为 0):plug-in 在 N=50 到 2000 全格上 100% 回合报出正增量,数值从 +0.131 衰到 +0.023 仍为正。一个彻底无信息的通道,两千例以内都会被”证明”合参有增益。换成 MM 校正后,N=400 时仍有 75.7% 的 bootstrap 区间排零(声称显著为正),N=800 起才 100% 排零,但点估计仍把真值 0 报成 +0.011——排零的机制是区间被估窄,不是点估计变准。这条对照是本文最硬的产出:它把”零信息通道拒绝率”变成一个可预注册的校准指标。

    检测门槛则要按最小增量定,不是按总互信息定。真逐级 Δ 为 +0.1225、+0.0699、+0.0487 时,第二路在 N=100 已有 93.5% 排零率,而最弱的那一步在 N=100 只有 16.0%、N=200 才 76.5%、N=400 起 100%。所以”第几路之后增益趋于零”的定位问题,需要的样本量由最小 Δ 决定;用总量 I 的精度来规划样本量必定高估检出力。

    判别层还有一个反直觉的口径问题。小训练集(120 例)朴素贝叶斯下,单通道 AUC 0.9324 反而高于双通道 0.9150,非单调,800 次重复里 97.1% 的回合 Δ2 为负。换成对隐因子求和的真·最优判别器,同一数据变成单调 0.9354→0.9382→0.9443→0.9541。但在”通道强弱不均”场景里,即便最优判别器也非单调(0.9565→0.9435→0.9445→0.9495),而真互信息一路单调涨到 0.5305。结论是明确的:互信息单调并不蕴含判别力单调,AUC 是互信息的非单调函数。所以申报口径若只报 MI,等于只答了半个问题。

    外部文献把这条推论顶得更硬。Porter 等人 2023 年对 93 项精神病神经影像分类研究的元分析直接给出:所有模态都能区分患者与对照,但多模态与单模态之间未观察到任何显著差异,功能连接与单模态的效应分布大量重叠(PMID 37563277)。Mu 等人 2026 年的多组学渐进融合框架报告,四个基准数据集里有三个”超过 50% 的病例仅靠单一组学即可准确分类”,其核心机制正是”置信度够了就不再增加模态”(PMID 41319559)。方向上与本文仿真一致:合参的价值不在”必然涨”,而在”知道什么时候已经够了”。中医侧的可测性已有落点——体质多模态分类(PMID 36374441)、四诊辅助 apparatus 评价健康与疗效(PMID 32185979),说明缺的不是设备而是口径。

    估计量选择上文献给的路很清楚:Zhang 与 Zheng 的指数衰减偏差估计量,其渐近正态性配合偏差快速衰减才使有限字母表推断可用,作者直陈极大似然(即 plug-in)互信息”极大抬高第一类错误概率”(PMID 25941916);Jiao 等人的极小极大框架给出换算:最优估计量在 n 个样本上约等于 plug-in 用 n ln n 个样本,而 plug-in 达不到该速率(PMID 29375152);连续变量须换几何 k 近邻路线(PMID 29604625)。三条合起来是硬约束:报 Δ 必须用低偏差估计量并同报 plug-in 作对照。

    证据与局限

    已验证(本文仿真内部):真值层面,四通道边际增益严格为正且递减、子模性 76 组检查零违反;顺序改变使增益轨迹完全不同但总量不变;plug-in 偏差随通道数膨胀(N=50 时 +0.042→+0.339);严格零信息通道在 N≤2000 全格 100% 被估计为正增量;MM 校正在 N=50 过校正并产生 30.4% 假负增量;最弱一步 Δ=+0.0487 需约 400 例才 100% 排零;AUC 在”通道强弱不均”下即使最优判别器也非单调,而 MI 单调。

    部分验证:外部文献支持”多模态未必优于最佳单模态”(PMID 37563277 元分析、PMID 41319559 渐进融合),以及”低偏差估计量是必要条件”(PMID 25941916、PMID 29375152、PMID 29604625);但”四诊四通道”的真值结构、真实证候标签的判读者信度、真实通道可读性 a 的取值,均无同批数据支撑——本文的 a=0.45~0.85、pg=0.70~0.95 是设定,不是实测。

    待验证:① 零信息通道拒绝率作为预注册校准指标,在真实四诊数据上的可复现性;② 真实标签下”通道强弱不均”是否普遍,若普遍则 AUC 非单调将成常态而非例外;③ 子模性恒真意味着”边际递减”永远不能作为发现,但”递减速率”是否可用于裁定”第几路够用”,需要先确认速率跨病种是否稳定(本文仅做单病种原型)。

    该主题尚未进入 L1 对应网络。最近邻为 E-061 段 KB-509「冗余度/协同度(PID)」PMID 33266673,属呼吸-心率两通道偶联、非四诊四通道,且该条已明写口径警告「冗余 = 同一目标有两条通路可预测;协同 = 只有联合才有信息,两者方向可以相反」。

    L1 网络状态:EDGE_TABLE / CORRESPONDENCE_NETWORK 检索「四诊 / 互信息 / 融合增益 / 增量 / 子模 / 单调」,命中全在心肺偶联语境(E-061 段 KB-509 冗余度/协同度 PID,PMID 33266673,两通道),本文「四诊四通道融合增益不等式的真值层与估计层分离」无对应边,未获 L1 状态判定,故证据等级封顶 C。

    相关阅读

    证据等级: C
    参考文献: PMID 37563277, PMID 41319559, PMID 25941916, PMID 29375152, PMID 29604625, PMID 36374441, PMID 32185979


    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 two layers of truth in \”合参必增益\”: the constant half does not count as evidence, while only a small half is testable.

    Evidence level: C

    Mutual information never decreases as channels increase; this is a mathematical identity, not an empirical discovery. What can actually be tested is whether “the incremental gain per added channel is positive,” yet this single step cannot be reliably measured with fewer than 400 cases.

    Body

    In modern arguments synthesizing the four diagnostic methods (inspection, listening/smelling, inquiry, palpation), a frequently cited “information-theoretic may support” states that the joint mutual information I(Symptom; Inspection, Listening, Inquiry, Palpation) is always greater than or equal to the mutual information of any single channel or pair of channels. Literally, this statement always holds because mutual information is monotonic with respect to the set of independent variables—adding another variable never makes you more likely to guess incorrectly. Therefore, it requires no experiment and cannot serve as evidence that “synthesis is necessary.” This is the first layer we must dismantle.

    The second layer is testable: decompose the incremental gain of “adding channels step-by-step” and examine whether each marginal gain Δ_k = I(Symptom; X_1…X_k) − I(Symptom; X_1…X_(k-1)) is strictly positive, statistically distinguishable from zero, and whether this positive increment decreases step-by-step. The latter point (decreasing step-by-step) also presents difficulties—mutual information as a set function is submodular, so the diminishing marginal gain is an identity equivalent to conditional mutual information being non-negative. In other words, “adding more yields less growth” is a structural property, not an empirical one. Thus, I use an analytically tractable generative model to separate these two layers: the model features four channels and three states (read symptom / read opposite / unreadable), plus a subject-level common latent factor G (overall readability good vs. poor). Given G, the four channels are conditionally independent—this is precisely the minimal structure for “partial redundancy.”

    First, the analytical results. Across four scenarios (strong channel a=0.80, weak channel a=0.45, strong channel with high shared source, uneven channel strengths), the marginal gain per channel is strictly positive and monotonically decreasing. For example, in the strong-channel scenario, gains are +0.3645, +0.1100, +0.0626, and +0.0439 bits; the submodularity check covers 76 combinations (A⊂B, c∉B) with zero violations. This is not due to good model fit but because other results are mathematically impossible. Hence, “synthesis yields gain but diminishing returns” is a tautology at the truth level; any paper claiming “we empirically demonstrated diminishing marginal returns” actually tested the behavior of the estimator.

    The order of addition matters and should be specified in protocols. For the same four channels (a sequentially 0.85/0.70/0.50/0.30), adding strong channels first yields a gain trajectory of +0.3940, +0.0750, +0.0351, +0.0265; adding weak channels first yields +0.0942, +0.1686, +0.1614, +0.1064. The two trajectories differ completely, yet the total remains identical (+0.5305)—the final I is the same number. Therefore, the inflection point where “marginal gain approaches zero after the nth channel” is not a property of the data but of the addition order. Inflection points without pre-registered order are irreproducible.

    Turning to the estimation layer, here lies the real bad news. Discrete plug-in mutual information exhibits systematic positive bias, which inflates with the number of states in the channel combination: at N=50, single-channel bias is +0.042, while four-channel bias reaches +0.339 bits. Thus, “adding channels increases gain” may be purely due to increased bias. Miller-Madow correction works well for large samples but over-corrects at N=50, yielding −0.018 and causing 30.4% of rounds to report negative increments—a strictly positive true increment estimated as negative. At N≥200, negative increments drop to 2.1%, and disappear entirely beyond N=500.

    Most striking is the zero-information channel control. Setting the fourth channel to be completely independent of both symptom and latent factor (three states equally probable, true Δ strictly zero): the plug-in estimator reports positive increments in 100% of rounds across N=50 to 2,000, with values decaying from +0.131 to +0.023 yet remaining positive. A completely uninformative channel is “proven” to enhance synthesis within 2,000 cases. With MM correction, at N=400, 75.7% of bootstrap intervals still exclude zero (claiming significant positivity); only at N≥800 do 100% exclude zero, though the point estimate still reports the true value of 0 as +0.011—the mechanism excluding zero is narrower intervals due to estimation, not improved accuracy of the point estimate. This control is our strongest output: it turns “zero-information channel rejection rate” into a pre-registered calibration metric.

    Detection thresholds must be set based on the minimum increment, not total mutual information. When true sequential Δ are +0.1225, +0.0699, +0.0487, the second channel achieves 93.5% zero-rejection rate at N=100, whereas the weakest step reaches only 16.0% at N=100, 76.5% at N=200, and 100% starting at N=400. Thus, the question of “after which channel does gain approach zero” depends on sample size determined by the smallest Δ; planning sample size based on total I precision inevitably overestimates detection power.

    Discrimination Layer: The Counterintuitive Channel Width Issue

    A counterintuitive calibration issue arises at the discrimination layer. Under naive Bayes with a small training set (120 cases), the single-channel AUC of 0.9324 actually exceeds the dual-channel value of 0.9150, violating monotonicity; among 800 repetitions, 97.1% of rounds showed a negative Δ2. Switching to the true optimal discriminator based on summing latent factors renders the same data monotonic: 0.9354→0.9382→0.9443→0.9541. However, in the “uneven channel strength” scenario, even the optimal discriminator remains non-monotonic (0.9565→0.9435→0.9445→0.9495), while true mutual information increases monotonically to 0.5305. The conclusion is clear: monotonicity of mutual information does not imply monotonicity of discriminative power, as AUC is a non-monotonic function of mutual information. Thus, reporting only MI constitutes answering only half the question.

    External literature strengthens this inference. Porter et al. (2023) meta-analysis of 93 psychiatric neuroimaging classification studies directly reports: all modalities can distinguish patients from controls, but no significant difference was observed between multimodal and unimodal approaches, with effect distributions of functional connectivity heavily overlapping those of unimodal methods (PMID 37563277). Mu et al. (2026) multi-omics progressive fusion framework reports that three of four benchmark datasets achieve accurate classification using only a single omic, with the core mechanism being “stopping modal addition once confidence suffices” (PMID 41319559). The direction aligns with this simulation: the value of combined reference lies not in “inevitable increase” but in “knowing when it is sufficient.” Measurability on the TCM side has concrete anchors—multimodal constitution classification (PMID 36374441) and four-examination-assisted apparatus evaluation of health and efficacy (PMID 32185979)—indicating the gap is not in equipment but in calibration.

    The literature provides clear guidance on estimator selection: Zhang and Zheng’s exponential decay bias estimator, whose asymptotic normality combined with rapid bias decay enables finite-alphabet inference, explicitly states that maximum likelihood (i.e., plug-in) mutual information “dramatically inflates Type I error probability” (PMID 25941916); Jiao et al.’s min-max framework yields the conversion: the optimal estimator on n samples approximates plug-in using n ln n samples, whereas plug-in fails to achieve this rate (PMID 29375152); continuous variables require geometric k-nearest neighbor routing (PMID 29604625). These three constraints combine into hard requirements: reporting Δ must employ low-bias estimators and be contrasted against plug-in.

    Evidence and Limitations

    Verified (within this paper’s simulation): At the truth level, marginal gains across the four channels are strictly positive and diminishing, with zero violations in 76 submodularity checks; reordering alters the gain trajectory but preserves total quantity; plug-in bias scales with channel count (e.g., +0.042 at N=50 rising to +0.339); strictly zero-information channels are estimated as positive increments in 100% of cases when N≤2000 across all grids; MM correction overcorrects at N=50, generating 30.4% false-negative increments; the weakest step Δ=+0.0487 requires approximately 400 samples to achieve 100% zero-rejection; AUC is non-monotonic even under optimal discriminators when “channel imbalance” exists, whereas mutual information (MI) remains monotonic.

    Partially Verified: External literature supports the claim that “multimodal is not necessarily superior to the best unimodal” (PMID 37563277 meta-analysis, PMID 41319559 progressive fusion) and that “low-bias estimators are necessary” (PMID 25941916, PMID 29375152, PMID 29604625); however, the truth structure of “four examinations and four channels,” inter-rater reliability for real syndrome labels, and the parameter values for true channel readability (a) lack supporting data from this batch—the paper’s a=0.45~0.85 and pg=0.70~0.95 are set values, not measured results.

    Pending Verification: ① reproducibility of zero-information channel rejection rate as a pre-registered calibration metric on real four-examination data; ② whether “channel imbalance” is common under real labels, implying non-monotonic AUC would be the norm rather than an exception; ③ since submodularity is always true, “diminishing marginal returns” can never constitute a discovery, but whether “deceleration rate” can determine “which route suffices” requires confirmation that rates remain stable across disease types (this paper uses only a single-disease prototype).

    This topic has not yet entered the L1 corresponding network. The nearest neighbor is segment E-061 of KB-509, “Redundancy/Collaboration (PID)” PMID 33266673, which pertains to respiratory-heart rate dual-channel coupling rather than the four-examination-four-channel framework; furthermore, this entry explicitly states the scope warning: “redundancy means two pathways can predict the same target; collaboration means information exists only jointly, and their directions may oppose each other.”

    L1 Network Status: Retrieving “four examinations / mutual information / fusion gain / increment / submodality / monotonicity” via EDGE_TABLE / CORRESPONDENCE_NETWORK yields hits exclusively within the context of cardiopulmonary coupling (segment E-061 of KB-509, redundancy/collaboration PID, PMID 33266673, two channels). Because the “truth layer and estimation layer separation of the inequality for fusion gain in four examinations and four channels” lacks a corresponding edge, no L1 status determination is obtained; thus, the evidence level is capped at C.

    Related Reading

    Evidence level: C
    References: PMID 37563277, PMID 41319559, PMID 25941916, PMID 29375152, PMID 29604625, PMID 36374441, PMID 32185979


    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.