Evidence level: B
上一轮问”统一模型能不能跨五种楼形外推”,这一轮问反过来的问题:当统一模型失败时,”按截面类型分层”是不是答案?以及更扎心的——形煞的那些类,到底是不是物理上真实的分界?
正文
第一层:先把问题翻译对。 本课题的表面问题是”分层模型 vs 统一模型谁预测得更准”,但真正的问题是:风水师按形状分的那些类(方/圆/锥/切角/扭转),是否恰好落在 St 依赖方式的物理断点上? 分层建模只是手段,目的是检验”分类本身有没有物理依据”。若分类是真实的,分层该赢;若分类只是对连续形态空间的离散命名,分层赢就是假的——多出来的参数只是在拟合噪声。
第二层:分层”赢”常常是统计幻觉。 类数检出有个残酷的实证:潜在剖面分析里,当真实类间距小(Cohen’s d = 0.2-0.5)时,七种模型选择方法没有一个能可靠挑出正确的类数,只有类间距极大(d = 1.5)时 LMR、bootstrap LRT、BIC 才靠谱,而 AIC 与熵几乎总是选错(PMID 24489457)。更糟的是伪类:增长混合建模里,结果变量只是轻微偏离正态,就会提取出本不存在的类,并把实质意义硬安上去(PMID 29795894)。“分成几类效果更好”本身,可能是非正态误差造出来的。 所以裁决的第一步不是比模型,是确认这个类数在更严的零假设下还存在。
第三层:但”借力”本身有坚实的正面理由。 亚组分析里的经验贝叶斯收缩给出了经典结果:当存在 3 个以上亚组时,把每个亚组的估计向总体均值收缩,虽然单个估计可能被”拉偏”,但总预期误差下降——独立分析反而更差(PMID 20138396)。而且借用强度可量化:借用强度与借用指数用 Mallow 距离构造,能从 MCMC 样本算出每组之间实际借了多少信息(PMID 32341613)。分层不是”要不要借”,而是”借多少、该不该借”——这把一刀切的问题变成了一个可测的量。
第四层:问题是”该借多少”不是数据自己能定的。 basket trial 的方法学给出了最直接的教训:完全贝叶斯的分层模型在 10 个以下亚组时,没有足够信息确定借用程度,会显著膨胀一类错误;校准版把收缩参数定义成”亚组相似度”的函数、用模拟标定,才把一类错误压下来(PMID 29499621)。对应到我们的五类截面:样本少到一定程度,”该不该借”根本测不出来,分层模型的漂亮结果可能是借错了方向。
第五层:连裁判工具都不可靠。 用留一交叉验证做贝叶斯模型比较,有一个著名警示:PSIS-LOO 在多层回归+事后分层(MRP)的工作流里,恢复不了模型的正确排序——因为在聚合阶段,个体层面的预测误差互相抵消了(PMID 38146825)。本课题的外推误差正是”聚合到截面族层面”来比较的,恰好踩中这个抵消模式。所以裁决实验要先验证比较工具本身:用已知真结构的模拟数据测试 LOO 能否认出对的模型,再拿去判真实数据。
第六层:物理侧的真分界,少而粗。 有两个并行的证据:一方面,St-Re 的统一关系 St=1/(A+B/Re) 对多种形状成立,Re→∞ 时极限 St≈0.21 与形状无关,形状差异全被压进参数 B(PMID 26565336)——函数形式没变,类与类之间只是参数连续变化;另一方面,锐缘形状确实有质的不同:尖角让分离点位置被固定住,而圆缘的分离点随 Re 漂移(PMID 39851685)——这才是真正的物理断点。合起来指向一个判断:真分界可能只有”尖/圆”这一刀,其余(切角 30° 还是 45°、锥率 1:3 还是 1:5)都是连续谱上的采样点。
第七层:裁决设计与三种结局。 用五类截面 × 形态参数扫描做 CFD+风洞,并列三模型:统一共享参数、按截面族分层(组内共享)、全独立;先用模拟数据标定 LOO 工具的认回率,再用留一族交叉验证比外推误差,最后把组间参数的散度画进形态参数空间找断点。① 分层显著赢且断点与”尖/圆”重合:形煞的最大分类(尖圆)是真实物理分界,细分则不是——一堵粗墙,墙内是带;② 分层赢但断点落在风水分类完全不同的位置:形煞分类与物理分界错位,评级按形煞走会系统性误判;③ 统一模型不输:五种楼形之间根本没有墙,”形煞分类”只是命名习惯。
回到易学。 风水把形分成尖圆方劈诸类,本质是给一个连续的形状空间起名字。起名字有大用——它让师徒能指物而谈、让吉凶有个粗分的框架,这不丢人。但古人没有仪器,分不出”这一刀是真的断,那几刀只是刻度”。现代测量说得很清楚:名字不等于边界。 若最终裁决是”尖圆一刀、其余连续”,那么形煞口诀里最该保留的是”尖者带煞、圆者中和”这一类极粗判断,而”切角几度、扭转几分”这类细分,必须降格为经验描述、不得当公式用。给连续的东西命名,是人类的天性;把命名当成发现了边界,是学科的陷阱。
证据与局限
已验证: 类间距小时没有任何常用统计量能可靠检出真实类数,AIC 与熵尤其差(PMID 24489457);结果变量非正态时混合建模会提取伪类并赋予实质意义(PMID 29795894);3 个以上亚组时经验贝叶斯收缩降低总预期误差(PMID 20138396);借用强度/借用指数可量化分层模型的实际信息借用(PMID 32341613);小样本下完全贝叶斯分层模型定不了借用程度、一类错误膨胀,需校准(PMID 29499621);PSIS-LOO 在多层模型聚合层面对模型排序不可靠(PMID 38146825);St-Re 统一关系跨形状成立、形状只进参数 B(PMID 26565336);锐缘固定分离点、圆缘分离点随 Re 漂移(PMID 39851685)。
部分验证: “统一函数形式+参数连续变化”与”尖圆存在质的分界”两个主张各自有证据,但”风水形煞分类是否与 St 依赖方式的物理分界重合”从未被同一批算例直接检验过;形煞分类的类数与间距(切角/锥率/扭转如何取值)没有公开的标准化数据集可供裁决。
待验证: ① 分层模型相对统一模型的增益在留一族外推下是否显著;② 若显著,组间参数散度在形态参数空间是否形成可定位的断点,断点是否与”尖/圆”重合;③ 风水形煞分类(方/圆/锥/切角/扭转)与物理分界的重合度;④ 用模拟数据标定后,LOO 工具在本课题的聚合结构下能否正确认出真实模型。
相关阅读
- 一个特征量管五种楼形:尾流宽度能否统一预测涡脱频率的裁决
- 一条公式管所有楼形:Strouhal 数回归模型跨截面类型的泛化验证
- 建筑形态=风涡频率控制器:风水形制如何定量调控涡旋脱落
- 卡门涡街次声频谱与人体器官共振频率的匹配
- 研究池子:环境-生命频率接口
证据等级: B
参考文献: PMID 24489457, PMID 29795894, PMID 20138396, PMID 32341613, PMID 29499621, PMID 38146825, PMID 26565336, PMID 39851685
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。
Formal Classification: Is It a Wall or a Strip? Asking Where the Class Boundary Lies Before Layered Modeling
Evidence level: B
The previous round asked whether a unified model can be extrapolated across five building shapes. This round asks the reverse question: When the unified model fails, is “layered modeling by cross-section type” the answer? And more piercingly—do the categories of Form Sha represent physically real boundaries?
Main Text
Layer 1: Translate the problem correctly. The surface issue of this topic is “which predicts more accurately: layered models vs. unified models,” but the real question is: Do the categories used by Feng Shui practitioners (square/circular/conical/angular/twisted) align with the physical breakpoints of St-dependence? Layered modeling is merely a means; the goal is to test whether “the classification itself has physical basis.” If the classification is real, layered modeling should win; if it is merely a discrete naming of a continuous morphological space, winning with layered modeling is false—the extra parameters are just fitting noise.
Layer 2: Layered “wins” are often statistical illusions. There is a harsh empirical fact regarding class detection: in latent profile analysis, when the true inter-class distance is small (Cohen’s d = 0.2-0.5), none of the seven model selection methods can reliably identify the correct number of classes. Only when the inter-class distance is very large (d = 1.5) do LMR, bootstrap LRT, and BIC perform reliably, while AIC and entropy almost always select incorrectly (PMID 24489457). Worse still are pseudo-classes: in growth mixture modeling, if the outcome variable only slightly deviates from normality, the model extracts non-existent classes and imposes substantive meaning on them (PMID 29795894). “Dividing into more classes yields better results” itself may be generated by non-normal errors. Therefore, the first step in adjudication is not comparing models but confirming whether this class number exists under a stricter null hypothesis.
Layer 3: But “borrowing strength” itself has solid positive justification. Empirical Bayesian shrinkage in subgroup analysis yields a classic result: when there are three or more subgroups, shrinking each subgroup’s estimate toward the overall mean reduces the total expected error, even if individual estimates may be biased—analysis without borrowing performs worse (PMID 20138396). Moreover, borrowing strength is quantifiable: borrowing strength and the borrowing index are constructed via Mallow’s distance, allowing calculation of actual information borrowed between groups from MCMC samples (PMID 32341613). Layered modeling is not a “yes/no” question of whether to borrow, but “how much to borrow and whether to borrow”—turning a blunt-cut problem into a measurable quantity.
Layer 4: The question of “how much to borrow” cannot be determined by data alone. Methodological lessons from basket trials provide the most direct insight: fully Bayesian layered models with fewer than 10 subgroups lack sufficient information to determine the degree of borrowing, leading to significant inflation of Type I error; calibrated versions define the shrinkage parameter as a function of “subgroup similarity” and use simulation calibration to suppress Type I error (PMID 29499621). Corresponding to our five cross-section types: when sample sizes are too small, whether “to borrow” cannot be measured at all; the pretty results of layered models may stem from borrowing in the wrong direction.
Layer 5: Even the adjudication tools are unreliable. Using leave-one-out cross-validation for Bayesian model comparison carries a famous warning: PSIS-LOO fails to recover the correct model ranking in workflows involving multivariate regression plus post-hoc stratification (MRP)—because at the aggregation stage, individual-level prediction errors cancel each other out (PMID 38146825). Our topic’s extrapolation error is compared exactly at the “aggregated cross-section family level,” coincidentally stepping on this cancellation mode. Therefore, adjudication experiments must first validate the comparison tool itself: test whether LOO can identify the correct model using simulated data with known true structures, then apply it to real data.
Layer 6: True physical boundaries are few and coarse. There are two parallel lines of evidence: On one hand, the unified St-Re relationship St=1/(A+B/Re) holds for multiple shapes; as Re→∞, the limit St≈0.21 is shape-independent, with all shape differences compressed into parameter B (PMID 26565336)—the functional form does not change; classes differ only by continuous parameter variation; On the other hand, sharp-edged shapes indeed have qualitative differences: sharp corners fix the separation point location, while rounded edges cause the separation point to drift with Re (PMID 39851685)—this is the true physical breakpoint. Together, they point to a judgment: The true boundary may only be “sharp vs. rounded,” while the rest (30° vs. 45° cut angles, 1:3 vs. 1:5 taper ratios) are just sampling points on a continuous spectrum.
Layer 7: Adjudication design and three outcomes. Using CFD and wind tunnel testing with five cross-section types and morphological parameter scanning, compare three models: unified shared parameters, layered by cross-section family (shared within groups), and fully independent. First, calibrate LOO tool recognition rates using simulated data; second, compare extrapolation errors using leave-one-group-out cross-validation; finally, plot the dispersion of inter-group parameters into morphological parameter space to find breakpoints. ① If layered modeling wins significantly and the breakpoint coincides with “sharp/rounded”: Form Sha’s maximum classification (sharp/rounded) represents a true physical boundary, while subdivision does not—a thick wall with a band inside; ② If layered modeling wins but the breakpoint falls at a location completely different from Feng Shui classifications: Form Sha classification misaligns with physical boundaries, leading to systematic misjudgment when rating follows Form Sha; ③ If the unified model does not lose: there is no wall between the five building shapes; “Form Sha classification” is merely a naming convention.
Returning to Yijing Studies. Feng Shui divides forms into sharp, round, square, split, and other categories, essentially naming a continuous shape space. Naming is highly useful—it allows masters and apprentices to refer to objects and discuss them, providing a coarse framework for auspiciousness and inauspiciousness; this is not shameful. However, ancients lacked instruments to distinguish “whether this cut is a true break or those few cuts are just scales.” Modern measurement is very clear: names do not equal boundaries. If the final adjudication is “one cut for sharp/rounded, continuous elsewhere,” then the most valuable retention in Form Sha aphorisms is the extremely coarse judgment “sharp carries sha, round is neutral,” while subdivisions like “how many degrees of cut angle or how much twist” must be downgraded to empirical descriptions and should not be used in formulas. Naming continuous things is human nature; treating names as discovered boundaries is a disciplinary trap.
Evidence and Limitations
Verified: When inter-class distance is small, no common statistical metric can reliably detect the true number of classes; AIC and entropy perform particularly poorly (PMID 24489457); when the outcome variable is non-normal, mixture modeling extracts pseudo-classes and assigns them substantive meaning (PMID 29795894); empirical Bayesian shrinkage reduces total expected error when there are three or more subgroups (PMID 20138396); borrowing strength and borrowing index quantify actual information borrowing in layered models (PMID 32341613); fully Bayesian layered models cannot determine borrowing degrees with small samples, causing Type I error inflation and requiring calibration (PMID 29499621); PSIS-LOO is unreliable for model ranking at the aggregation level of multivariate models (PMID 38146825); the unified St-Re relationship holds across shapes with shape differences only entering parameter B (PMID 26565336); sharp edges fix separation points while rounded edges cause drift with Re (PMID 39851685).
Partially Verified: Both claims—”unified functional form with continuous parameter variation” and “qualitative boundary exists between sharp and round”—have supporting evidence, but whether “Feng Shui Form Sha classifications align with physical breakpoints of St-dependence” has never been directly tested using the same set of cases; there is no publicly available standardized dataset for adjudicating the number and spacing of Form Sha classes (how cut angles/taper ratios/twist are defined).
To Be Verified: ① Whether the gain of layered models over unified models is significant under leave-one-group-out extrapolation; ② If significant, whether the dispersion of inter-group parameters forms a locatable breakpoint in morphological parameter space and if it coincides with “sharp/rounded”; ③ The degree of overlap between Feng Shui Form Sha classifications (square/circular/conical/angular/twisted) and physical boundaries; ④ Whether, after calibration with simulated data, the LOO tool can correctly identify the true model under the aggregation structure of this topic.
Related Reading
- One Feature Quantity Governs Five Building Types: Adjudication on Whether Tail Width Can Unified Predict Vortex Shedding Frequency
- One Formula Governs All Building Types: Generalization Verification of Strouhal Number Regression Model Across Cross-Section Types
- Building Morphology = Wind Vortex Frequency Controller: How Feng Shui Forms Quantitatively Regulate Vortex Shedding
- Karman Vortex Street Infrasound Spectrum Matching with Human Organ Resonance Frequencies
- Research Pool: Environment-Life Frequency Interface
Evidence Level: B
References: PMID 24489457, PMID 29795894, PMID 20138396, PMID 32341613, PMID 29499621, PMID 38146825, PMID 26565336, PMID 39851685
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