三类风压异常谁在背锅:共线栅格上的贡献分解为何做不出来

Written by

in

Language

Evidence level: B

想给”极值区、加速区、脉动区”排名,先得接受一个事实:在建筑角部,这三样东西常常同时出现、同时消失,本来就分不开。

正文

上一篇把风水师的形煞点位与风压异常带叠在一张图上,问重合度是否高于随机。这一篇问下一步:既然重合了,是哪一类异常在起作用——Cp 极值区的强负压,加速比超过 1.5 的高速带,还是脉动峰值区的湍流冲击?很多研究到这里就顺手跑个回归,把三类异常当三个自变量,看谁系数大、谁 p 值小,把系数当贡献报出去。这个动作在多数情况下无效。

三类异常物理同源。 它们不是三种独立现象,而是同一股分离流在三个通道上的读数。方形建筑角部发生流动分离,剪切层卷起成涡,涡脱落在侧壁造成强负压——这是 Cp 极值;同一股被挤压的流动绕过角部进入通道,形成加速比高值区;涡脱落的间歇性让压力信号变非高斯、尾部变厚——这是脉动峰值。一篇 2026 年风洞研究说得很直白:130 个工况下高斯区集中在仰角 60 度以下的中心,外围因涡脱落呈非高斯(PMID 42297878)。即脉动区的判据(非高斯)与极值区的成因(涡脱落)指向同一过程。同源,空间上就必然共线,共线到无法用常规回归拆开。

共线让系数分解变成随机数。 相关预测变量下可用的相对重要性方法至少有四种:Shapley 值分解、Genizi 法、相对权重、优势分析(PMID 42375422)。2026 年那篇基于鲍鱼数据的比较研究给了关键细节:同一份数据、同一个模型,PMVD 显示各变量重要性重叠,Genizi 却给出清晰分离,区间宽度也不同。同一问题、同一份数据,结论随尺子变。这不是方法不成熟,而是变量相关时”独立贡献”本身没有唯一解。广义可加模型更明显:生态学把这种共线性叫 concurvity,通行做法是按”平均共享方差”把总 R² 摊给各变量(PMID 39280972)。”摊”字是关键——分配规则人为约定,换摊法就换排名。

随机森林也救不了。 2010 年那篇专研置换重要性的模拟研究结论分情况:变量既相关又真有信号时,非条件置换会把更大份额分给相关的那一组,条件重要性会压低相关变量得分,缩放重要性在有信号与无信号下都明显有偏(PMID 20187966)。更早一项模拟发现基尼重要性在相关下偏倚、偏倚取决于相关强度与数量,置换型更稳(PMID 19460890)。合起来:重要性指标不中立,自带倾向;选非条件置换,等于默认”相关的一组一起加分”。

空间数据还多一层。 上面讲的是变量之间相关,栅格数据另有位置之间相关。2018 年一篇方法学研究用模拟证明:存在空间自相关时,常规双置换检验会抬高一类错误率,本不显著的关联被误判为显著;作者用 Moran 谱随机化保留原有空间结构,才把错误率控住(PMID 30289571)。这条格外致命:形煞点位成片出现(一栋楼的角部、一条街的夹缝),异常带也成片,两者的”重合”很容易只是因为都成片。

换三个能答的问法。 原问题无解,可换成:增量(加第三类带来多少额外解释力)、分层(先筛仅含一类的格点再比命中率)、可辨识性(量化共线,报最小可分辨的贡献差)。

判决实验骨架。 ①取子集:筛出仅含一类异常带的格点,若某类剩余不足总量 15%,退出排名。②多方法并列:同份数据同跑 Shapley、Genizi、相对权重、优势分析、条件置换重要性,五种排名并列报告。③空间零模型:不用简单随机置换,改 Moran 谱随机化保留评级点位的聚集结构(PMID 30289571)。④增量检验:以两类为基线逐个加入第三类,报 ΔAUC 及其在空间零模型下的置信区间,用区间而非 p 值(PMID 29906269)。⑤透明化:报方差膨胀因子、两两空间相关、每种方法排名在 1000 次 bootstrap 下的翻转频率。

三条判据。 若五种排名一致且增量在空间零模型下仍显著——可报”某一类为主”,须注明结论对方法选择不敏感,这是最强情形,实际很少见。若排名随方法变化(Genizi 与 PMVD 分道扬镳,已有先例)——结论只能是”三类在此数据上不可分辨”,改报合并指标。若某类筛后不足 15%——不作独立因子,只作描述。三种里中间那种最可能也最有用:它把”哪个最凶”换成了”这个尺度上问不出答案”。

落到形煞上。 风水对形煞是分条目的:尖角煞、天斩煞、反弓煞,各有各的凶应,各有各的化解。这套分类隐含一个假设——不同形状制造不同性质的麻烦,所以要用不同方法对付。风工程给的图像不同:不同形状制造的是同一股分离流的不同强度与位置,Cp 极值、加速、脉动是这股流的三个读数,不是三种病。若判决落在”不可分辨”,形煞的条目化分类就更接近一套便于传承的形状索引,而非病因学分类——它告诉你去看这个角,不告诉你这个角以哪种方式伤人。这个区分不是贬低传统,是把传统里可靠的部分(哪里该看)与过度延伸的部分(会怎么伤人)分开。古人讲”峦头一诀,百形皆应”,说的是看形方法通用;后人读成”每一形各有各的应”,是从方法跨到了病因,物理上没有支撑。

证据与局限

已验证:相关预测变量下相对重要性有多种分解方法且结论不一致(PMID 42375422,PMVD 显示重叠、Genizi 显示清晰分离);GAM 中共曲性需按平均共享方差摊分(PMID 39280972);随机森林非条件置换在相关且真有信号时把份额分给相关组、条件重要性压低相关变量、缩放重要性有偏(PMID 20187966);基尼重要性在相关下偏倚、置换型更稳(PMID 19460890);空间自相关抬高双置换检验一类错误率、Moran 谱随机化可控制(PMID 30289571);风压脉动区因涡脱落呈非高斯、与极值区同源(PMID 42297878);建筑凹角涡系同时决定通风与风舒适(PMID 32287964);转角优化改变双塔干扰效应(PMID 39953100)。

部分验证:”三类异常带在真实建筑栅格上达到何种共线水平”无直接测量;”贡献分解在此共线水平下是否可辨识”是方法学外推的判断,未在形煞数据上验证。

待验证:①三类异常两两空间相关的实际量级(预期 r>0.6);②筛出单类子集后各类剩余样本量是否够支撑分层比较;③五种分解方法的排名在 bootstrap 下的翻转频率;④控制空间自相关后,评级-异常重合度是否仍显著。

相关阅读

证据等级: B(贡献分解方法学、空间自相关对检验的影响、风压异常带同源性均有 PubMed 直接支撑;”在形煞栅格上做贡献分解”本身无先例,属研究设计提案)
参考文献: PMID 42375422, PMID 39280972, PMID 20187966, PMID 19460890, PMID 30289571, PMID 42297878, PMID 32287964, PMID 39953100, PMID 29906269


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。

Who Is Taking the Blame for Three Types of Wind Pressure Anomalies: Why Contribution Decomposition on Co-linear Grids Cannot Be Done

Evidence level: B

To rank “extreme zones, acceleration zones, and pulsation zones,” one must first accept a fact: at building corners, these three phenomena frequently co-occur and vanish simultaneously; they are inherently inseparable.

Main Text

The previous article overlaid Feng Shui practitioners’ Form Sha locations with wind pressure anomaly bands and asked whether the overlap exceeded random chance. This article asks the next step: given the overlap, which specific anomaly is driving the effect—the strong negative pressure in the Cp extreme zone, the high-speed band with an acceleration ratio exceeding 1.5, or the turbulent impact in the pulsation peak zone? Many studies stop here and casually run a regression, treating the three anomalies as independent predictors to see which has the largest coefficient or smallest p-value, then report the coefficients as “contributions.” This action is largely ineffective in most cases.

Physical Homogeneity of the Three Anomalies. They are not three independent phenomena but rather readings from a single separated flow across three channels. At the corner of a square building, flow separation occurs; the shear layer rolls into vortices, and vortex shedding creates strong negative pressure against the side wall—this is the Cp extreme. The same compressed flow bypasses the corner and enters a channel, forming a high acceleration ratio zone; the intermittent nature of vortex shedding makes the pressure signal non-Gaussian with fat tails—this is the pulsation peak. A 2026 wind tunnel study states plainly: under 130 test cases, Gaussian regions concentrated below a 60-degree elevation angle at the center, while peripheries became non-Gaussian due to vortex shedding (PMID 42297878). Thus, the criterion for the pulsation zone (non-Gaussian) and the cause of the extreme zone (vortex shedding) point to the same process. Being homogenous in origin means they are inevitably co-linear in space, co-linear to the point where conventional regression cannot separate them.

Co-linearity Turns Coefficient Decomposition into Random Numbers. For relative importance methods under correlated predictors, there are at least four approaches: Shapley value decomposition, Genizi method, relative weights, and advantage analysis (PMID 42375422). A 2026 comparative study based on abalone data provided key details: with the same dataset and model, PMVD showed overlapping variable importance, while Genizi yielded clear separation with different interval widths. The same problem, same dataset, yet conclusions vary based on the metric used. This is not a lack of method maturity; when variables are correlated, “independent contribution” itself has no unique solution. Generalized Additive Models (GAMs) are even more explicit: ecology refers to this co-linearity as concurvity, and the standard practice is to allocate total $R^2$ among variables based on “average shared variance” (PMID 39280972). The word “allocate” is critical—the distribution rule is a human convention; changing the allocation method changes the ranking.

Random Forests Cannot Save It. A 2010 simulation study on permutation importance concluded differently based on conditions: when variables are correlated yet truly signal-bearing, unconditional permutation assigns a larger share to the correlated group, while conditional importance suppresses scores for correlated variables, and scaled importance is significantly biased whether signal exists or not (PMID 20187966). An earlier simulation found that Gini importance is biased under correlation, with bias depending on correlation strength and count, whereas permutation-based methods are more stable (PMID 19460890). Combined: importance metrics are not neutral and carry inherent biases; selecting unconditional permutation defaults to “adding points to the correlated group together.”

Spatial Data Adds Another Layer. The above discusses correlation between variables; grid data also involves correlation between locations. A 2018 methodological study using simulations proved that when spatial autocorrelation exists, conventional double permutation tests inflate the Type I error rate, falsely flagging non-significant associations as significant; authors controlled the error rate only by using Moran spectral randomization to preserve the original spatial structure (PMID 30289571). This is particularly fatal: Form Sha points appear in clusters (corners of buildings, gaps between streets), and anomaly bands also appear in clusters; their “overlap” is easily just because both are clustered.

Three Alternative Questions. The original question has no solution; it can be reframed as: incremental gain (how much extra explanatory power does adding the third type bring?), stratification (screen for grid cells containing only one type first, then compare hit rates), or identifiability (quantify co-linearity and report the minimum resolvable contribution difference).

Judgment Experiment Skeleton. ① Subset selection: screen for grid cells containing only one type of anomaly band; if a category remains below 15% of the total, exit the ranking. ② Parallel methods: run Shapley, Genizi, relative weights, advantage analysis, and conditional permutation importance on the same dataset simultaneously, reporting all five rankings side-by-side. ③ Spatial null model: do not use simple random permutation; instead, use Moran spectral randomization to preserve the clustering structure of rating points (PMID 30289571). ④ Incremental testing: add the third category one by one using two categories as baselines, report $\Delta$AUC and its confidence interval under the spatial null model, using intervals rather than p-values (PMID 29906269). ⑤ Transparency: report variance inflation factors, pairwise spatial correlations, and the flip frequency of rankings across 1000 bootstrap iterations.

Three Criteria. If all five rankings are consistent and the increment remains significant under the spatial null model—report “one category dominates,” noting that the conclusion is insensitive to method choice. This is the strongest scenario but rare in practice. If rankings vary by method (e.g., Genizi and PMVD diverge, as seen in prior cases)—the conclusion must be “these three categories are indistinguishable on this dataset,” and report a combined metric instead. If a category falls below 15% after screening—do not treat it as an independent factor; only describe it. Among the three, the middle case is likely and most useful: it transforms “which is the worst” into “the question cannot be answered at this scale.”

Applying to Form Sha. Feng Shui categorizes Form Sha into specific items: sharp corner Sha, sky-cut Sha, reverse bow Sha, each with its own凶应 (adverse response) and remediation. This classification implies an assumption that different shapes create different natures of trouble, requiring different countermeasures. Wind engineering provides a different image: different shapes produce different intensities and positions of the same separated flow. Cp extremes, acceleration, and pulsation are merely three readings of this single flow, not three distinct diseases. If the judgment falls on “indistinguishable,” the itemized classification of Form Sha approaches a shape index for transmission rather than an etiological classification—it tells you to look at this corner, not how that corner injures. This distinction does not demean tradition but separates the reliable parts of tradition (where to look) from over-extended parts (how it injures). Ancient texts say “One principle of landscape, all forms respond,” referring to universal viewing methods. Later readers interpret this as “each form has its own response,” crossing from method to etiology, which lacks physical support.

Evidence and Limitations

Verified: Multiple decomposition methods exist for relative importance under correlated predictors with inconsistent conclusions (PMID 42375422; PMVD shows overlap, Genizi shows clear separation); concurvity in GAMs requires allocation by average shared variance (PMID 39280972); in Random Forests, unconditional permutation assigns shares to correlated groups when signal exists, conditional importance suppresses correlated variables, and scaled importance is biased (PMID 20187966); Gini importance is biased under correlation with bias dependent on strength and count, while permutation is more stable (PMID 19460890); spatial autocorrelation inflates Type I error in double permutation tests, controllable via Moran spectral randomization (PMID 30289571); wind pressure pulsation zones become non-Gaussian due to vortex shedding and share origin with extreme zones (PMID 42297878); building re-entrant corners vortex systems simultaneously determine ventilation and wind comfort (PMID 32287964); corner optimization alters double-tower interference effects (PMID 39953100).

Partially Verified: The degree of co-linearity reached by “three anomaly bands on real building grids” lacks direct measurement; whether “contribution decomposition is identifiable at this level of co-linearity” is a methodological extrapolation not verified on Form Sha data.

To Be Verified: ① Actual magnitude of pairwise spatial correlation between the three anomaly types (expected $r > 0.6$); ② Whether sample sizes per category after screening single-type subsets are sufficient to support stratified comparisons; ③ Flip frequency of rankings across the five decomposition methods under bootstrap; ④ Whether rating-anomaly overlap remains significant after controlling for spatial autocorrelation.

Related Reading

Evidence Level: B (Methodological support for contribution decomposition, impact of spatial autocorrelation on testing, and homogeneity of wind pressure anomaly bands all have direct PubMed backing; “performing contribution decomposition on Form Sha grids” itself has no precedent and constitutes a research design proposal)
References: PMID 42375422, PMID 39280972, PMID 20187966, PMID 19460890, PMID 30289571, PMID 42297878, PMID 32287964, PMID 39953100, PMID 29906269


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.

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *