Evidence level: C
把两族人群的三质指标分布做一维 Wasserstein 三项分解后,形状项能不能被分位数映射压到零是解析问题(答案是”能”),真正的裁决在第二问:形状被抹平时,同批数据的三质判别力还剩多少。
正文
跨族裔比较三质(瓦塔/皮塔/卡法)生理指标分布时,一维 Wasserstein 距离可拆成三个成分:位置项(中位数整体偏移)、离散项(方差/分位距差异)、形状项(偏度、峰度、多峰结构差异)。前两轮已经裁决了方向编码粒度与跨族裔判别力衰减的权衡(见相关阅读),本轮处理最后一个方法论缺口:形状项的处置。
第一问其实有解析答案。分位数映射的构造是:把甲分布的每个分位数数值改写成参考分布在同一分位数位置的取值。做完之后,甲分布与参考分布在每一个分位数上取值相同——分布完全重合,位置项、离散项、形状项全部归零。这在数学上是恒等式,不是经验问题。分位数映射作为偏置校正工具在气候健康影响投影中已有多方案比较研究,且不同映射方案给出的健康影响投影差异显著(PMID 33670819);把一个分布对齐到另一个分布的思路,在单细胞组学的多组学整合中也已是成熟工具,最优传输距离被直接用作样本层面的对齐目标(PMID 35050714;量化加速版本见 PMID 39808114)。
所以裁决的全部分量落在第二问:抹平形状的同时,判别信息是否一起被抹掉。这不是杞人忧天,组学文献里已有完整的先例链。分位数归一化的原始设计就假设所有样本共享同一潜在分布(PMID 12925520),后续方法学直接指出:当组间分布本身不同时,全局分位数归一化不再适用,必须分组建模(平滑分位数归一化,PMID 29036413)。归一化方案的选择会直接改变显著性检验的结论(PMID 15468753)。更尖锐的是批次校正方向:ComBat 类方法在移除批次效应时,会把与校正变量纠缠的真实生物学差异一起带走(PMID 33015620;泛化框架见 PMID 22257669),单细胞领域为此专门开发了不假设分布同形的稳健归一化(SCnorm,PMID 28418000)。另一侧,秩检验(Wilcoxon-Mann-Whitney)对分布形状本身敏感,形状差异不是可以随手丢掉的装饰成分(PMID 19247980)。跨祖裔预测模型的研究也表明,跨人群迁移时的信息保留与再校准之间存在真实的权衡结构(PMID 38835075;PRS 跨祖裔判别力系统性衰减的量化见 PMID 34995502)。
由此得到三条可检验的裁决路径。同批多族裔三质队列,逐指标做三项分解,然后执行分位数映射,前后各测一次三质判别 AUC:情形一,形状项降到近零且判别 AUC 不显著下降——形状差异与体质无关,属纯族裔测量偏置,映射安全;情形二,形状项降到近零但 AUC 显著下降——形状携带体质判别信息,”校正=抹信号”,正确动作是只校位置与离散、保留形状;情形三,形状项压不到零——重尾或多峰结构使分位数映射在边界分位数外插失控,该指标进入”不可校正区”,须换分布族或直接报缺口。
裁决为什么必须用同批数据做,而不是拿两个独立队列前后对比:三质判别的基线 AUC 本身随队列构成波动,独立队列的降幅里混着抽样噪声与族裔构成差异;只有同一批人先测原始指标、再测映射后指标,配对差值才把形状信息的有无单独隔离出来。这也是组学归一化方法学评估的标准做法:同一数据矩阵跑两套归一化,比较下游检出率的变化(PMID 15468753)。
三条实施细节决定裁决的可信度。第一,参考分布的选择(全样本合并 vs 多数族裔)本身是信息泄漏源,映射与判别评估须交叉拟合、分位数切点只在训练折内锚定。第二,边界分位数(低于 5%、高于 95%)的外插策略须预注册,这是情形三最常见的假阴性来源。第三,映射前后 AUC 比较用同批数据的配对自举,族裔分层重复,防止判别力下降被少数族裔样本量税伪装成形状信息。
证据与局限
已验证:一维 Wasserstein 三项分解与分位数映射的构造性性质(数学恒等);分位数归一化/映射在组间分布不同时失效或抹掉真实差异,有直接方法学文献(PMID 29036413、12925520、15468753);批次校正移除纠缠生物信号(PMID 33015620、22257669);最优传输用于分布对齐的工具链成熟(PMID 35050714、39808114)。
部分验证:分位映射方案选择影响下游结论,气候健康领域有比较研究(PMID 33670819);跨祖裔迁移中信息保留与校准的权衡结构(PMID 38835075、34995502)。
待验证:三质指标分布的形状项×族裔×判别力三角关系从未被同批数据测过;”只校位置与离散、保留形状”的折中方案在三质判别上的净收益无实测。该主题尚未进入 L1 对应网络,证据薄弱,结论以裁决设计形式给出。
相关阅读
L1 网络状态:EDGE_TABLE 仅 ayurveda-dosha 系列声频/HRV 方向 A 级边(
ayurveda-three-doshas-frequency18565332 /assr-signature-vs-dosha-bowl-selection29269119 等),CORRESPONDENCE_NETWORK 命中的”映射”均为背俞穴→HRV、E-060/E-061 等无关边——本主题(Wasserstein 形状项 × 分位数映射 × 三质判别)尚无对应边,尚未进入 L1 对应网络。证据等级: C
参考文献: PMID 29036413, PMID 12925520, PMID 15468753, PMID 28418000, PMID 33015620, PMID 22257669, PMID 35050714, PMID 39808114, PMID 33670819, PMID 19247980, PMID 38835075, PMID 34995502
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。
Correctability and Separability of Shape Components: Can Quantile Mapping Reduce the Wasserstein Shape Term to Near Zero, and Does This Simultaneously Erase Trimodal Discriminat…
Evidence level: C
After performing a one-dimensional Wasserstein three-term decomposition on the distribution of three-quality indicators for two populations, the question of whether the shape term can be compressed to zero via quantile mapping is an analytical problem (the answer is “yes”); the true verdict lies in the second question: how much of the three-quality discriminatory power remains in the same batch of data once the shape component is flattened.
Body
When comparing physiological indicator distributions across ethnic groups for the three constitutions (Vata, Pitta, Kapha), the one-dimensional Wasserstein distance decomposes into three components: location (median shift), dispersion (variance/quantile-distance differences), and shape (skewness, kurtosis, multimodality). Previous rounds have adjudicated the trade-off between encoding granularity and cross-ethnic discriminative power attenuation (see related reading). This round addresses the final methodological gap: handling the shape component.
The first question actually has an analytical answer. The construction of quantile mapping is: rewrite each quantile value of Distribution A as the corresponding quantile value in Reference Distribution. After this, both distributions take identical values at every quantile—the distributions completely overlap, and location, dispersion, and shape terms all become zero. This is a mathematical identity, not an empirical issue. Quantile mapping as a bias-correction tool has been compared across multiple schemes in climate-health impact projections, with significant differences observed in projected health impacts depending on the mapping scheme (PMID 33670819). Aligning one distribution to another is also a mature tool in multi-omics integration for single-cell genomics, where optimal transport distance is directly used as the alignment target at the sample level (PMID 35050714; quantified acceleration version see PMID 39808114).
Therefore, the full adjudication rests on the second question: when flattening shape, does discriminative information disappear along with it? This is not mere worry; omics literature already contains complete precedent chains. The original design of quantile normalization assumes all samples share the same underlying distribution (PMID 12925520); subsequent methodological work directly states that when inter-group distributions differ, global quantile normalization is no longer applicable and must be replaced by group-specific modeling (smoothed quantile normalization, PMID 29036413). The choice of normalization scheme directly alters the conclusion of significance testing (PMID 15468753). More critically, batch correction direction: ComBat-like methods, when removing batch effects, also remove true biological differences entangled with correction variables (PMID 33015620; generalized framework see PMID 22257669). For this reason, single-cell communities have specifically developed robust normalization methods that do not assume distribution homogeneity (SCnorm, PMID 28418000). On the other side, rank-based tests (Wilcoxon-Mann-Whitney) are sensitive to distribution shape itself; shape differences are not merely decorative components that can be discarded at will (PMID 19247980). Cross-ancestry prediction model research also demonstrates that cross-population transfer involves a real trade-off structure between information preservation and recalibration (PMID 38835075; quantified systematic attenuation of cross-ancestry discriminative power of PRS see PMID 34995502).
Thus, three testable adjudication paths emerge. For a same-batch multi-ethnic tri-constitution cohort, perform three-way decomposition per indicator, then execute quantile mapping, measuring tri-constitution discriminative AUC before and after: Case 1, shape term drops near zero and discriminative AUC does not significantly decrease—shape differences are unrelated to constitution, representing pure ethnic measurement bias; mapping is safe. Case 2, shape term drops near zero but AUC significantly decreases—shape carries constitution discriminative information; “correction = signal removal”; correct action is to calibrate only location and dispersion while preserving shape. Case 3, shape term cannot be reduced to zero—heavy tails or multimodal structure cause quantile mapping extrapolation failure at boundary quantiles; this indicator enters the “non-correctable zone,” requiring a change of distribution family or direct reporting of the gap.
Adjudication must use the same batch data rather than comparing two independent cohorts before and after mapping: baseline tri-constitution discriminative AUC itself fluctuates with cohort composition; the drop in independent cohorts mixes sampling noise and ethnic composition differences. Only by first measuring raw indicators and then mapped indicators on the same individuals can paired differences isolate the presence or absence of shape information. This is also the standard practice in omics normalization methodology evaluation: running two normalization schemes on the same data matrix and comparing downstream detection rate changes (PMID 15468753).
Three implementation details determine adjudication credibility. First, reference distribution selection (full sample merge vs. majority ethnicity) is itself an information leakage source; mapping and discriminative evaluation must use cross-validation, with quantile cutpoints anchored only within the training fold. Second, extrapolation strategy for boundary quantiles (below 5%, above 95%) must be pre-registered; this is the most common source of false negatives in Case 3. Third, AUC comparison before and after mapping uses paired bootstrapping on the same batch data with repeated ethnic stratification, preventing discriminative power decline from being masked as shape information by the tax burden of small ethnic sample sizes.
Evidence and Limitations
Verified: The constructive properties of the one-dimensional Wasserstein three-term decomposition and quantile mapping (mathematical identities); quantile normalization/mapping fails or erases true differences when inter-group distributions differ, with direct methodological literature (PMID 29036413, 12925520, 15468753); batch correction removes entangled biological signals (PMID 33015620, 22257669); optimal transport provides a mature toolchain for distribution alignment (PMID 35050714, 39808114).
Partially Verified: The choice of quantile mapping scheme affects downstream conclusions, with comparative studies in the climate-health domain (PMID 33670819); trade-off structures between information retention and calibration during cross-ancestry migration (PMID 38835075, 34995502).
Pending Verification: The shape term × ancestry × discriminability triangle of the three-quality metrics has never been measured on the same dataset; the compromise approach of “calibrating only location and dispersion while preserving shape” lacks empirical evidence regarding net gain in three-quality discrimination. This topic has not yet entered the L1 correspondence network, evidence is weak, and conclusions are presented in the form of adjudication design.
Related Reading
- Directional Encoding “Tier Granularity” Optimal Scanning
- “Intercept-Slope Intersection” Existence and Positional Arbitration
- Interaction Intensity Ranking “Third Variable Contamination” Recalculation
- Research Pool: Cross-Cultural Frequency Cognition
L1 Network Status: The EDGE_TABLE contains only A-level edges for the
ayurveda-doshaseries acoustic/HRV directional mappings (ayurveda-three-doshas-frequency18565332 /assr-signature-vs-dosha-bowl-selection29269119, etc.). The “mapped” edges hit by CORRESPONDENCE_NETWORK are irrelevant connections such as Back-Shu Acupoints→HRV, E-060/E-061, etc. — this topic (Wasserstein shape term × quantile mapping × three-quality discrimination) lacks corresponding edges and has not yet entered the L1 correspondence network.Evidence Level: C
References: PMID 29036413, PMID 12925520, PMID 15468753, PMID 28418000, PMID 33015620, PMID 22257669, PMID 35050714, PMID 39808114, PMID 33670819, PMID 19247980, PMID 38835075, PMID 34995502
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.
Leave a Reply