一层还是两层?用高斯混合模型给历代星表坐标”分堆

Written by

in

Language

Evidence level: B

一份星表的坐标残差,到底是”一个层”还是”两个层”?这看似直观,其实是一道要过三道统计关的模型选择题——选错了,整个”新鲜度”判断就崩了。

正文

上一篇把”抄的还是重测的”拆成三种误差指纹。这一篇把其中第三层——”内部一致性指纹”——单独拎出来,做成一道标准统计选择题:一份星表的坐标残差分布,到底是”一个层”还是”两个层”?

为什么”几层”这么关键?因为一层意味着单历元(纯抄录,或一次性实测),两层意味着”旧底子+新刷新”(抄录+重测混合)。这不是文风问题,而是”这份星表有多少是当朝实拍”这一核心判据的分水岭。

星表残差的分布,在数学上就是一堆数字堆出的直方图。判断它是单峰还是双峰,最现成的工具是高斯混合模型(GMM)——假设数据由若干个高斯”堆”叠加而成,再用信息准则(最常用 BIC)选”几个堆最合适”。听上去很干净,但这里藏着三个坑。

第一个坑:BIC 数的不是”层”,而是”几个高斯够拟合密度”。 一个非高斯的单峰分布(比如厚尾、偏斜),完全可能被拆成两三个高斯,BIC 反而更偏爱这个多高斯解——于是”两层”的结论可能纯粹是拟合出来的假象(PMID 20953302)。更麻烦的是,BIC 这套自动判决在已知真类数的模拟里极少是选出正确答案的那个(PMID 34548917)。换句话说,”BIC 说两层”离”真有两层”还有一大段距离。

第二个坑:层的算法不止一种,结论常常打架。 除 BIC 外,还有 gap 统计量(用零分布对照判簇数,PMID 17425640、PMID 18397779)、贝叶斯有限混合的数类准则(PMID 28081166)、”过拟合时多余成分会渐近变空”的 Zmix 思路(PMID 26177375),甚至用数值代数几何直接算出成分数(PMID 33867617)。这些方法各有前提,结果不一致是常态,单凭一个准则定层不可靠。

第三个坑:想加一层,得先过检验。 从一层加到两层常用似然比检验,但检验”多一个成分”时参数落在边界上,检验统计量不再服从标准卡方分布,直接用会严重高估显著性(PMID 16474914)。而且”该选哪个模型”本身有哲学分歧:AIC 偏预测、BIC 偏解释,二者常给出不同答案(PMID 22309957)。

那么有没有办法绕开这些坑?有——别只问”几层”,先问”这两个峰能不能被一个整体平移消掉”。

这是本篇最要紧的一步。古代仪器若有约 1° 的系统刻度误差,它给全表带来的是整体平移,不是双峰;而真实”两次观测”带来的是”一次平移吸收不掉”的层结构。于是判别可以归约成一个可操作的问句:把残差减去某个统一偏移后,双峰还在不在? 还在,才是真·两层;一平移就并成一个峰,那多半是系统误差假扮的。这正好接上《石氏星经》的双观测层之争——该研究认为这份星表含公元前 4 世纪与公元 2 世纪两个观测层(PMID 40273251)。

另外两类因素要一起考虑。一是厚尾:抄写/计算错误会产生少量大离群,使残差偏离高斯,此时普通 GMM 不合适,得用能容纳厚尾的指数幂混合(PMID 38785671)。二是外部锚:残差要换算成”年代”,得靠 ΔT 这条独立时间尺子(PMID 28119545、PMID 10800374)——锚不准,”两层”的年代也会整体漂。现代对照可以看 Gaia 卫星的天体测量误差处理(PMID 23630374):当代星表尚且存在系统差与不确定度传播,古人星表更需谨慎。

落地产出,就是上一篇预告的”坐标新鲜度指数“:每份星表给出一个”当朝实测占比”分数,并附上”层数判决的不确定度”。三种结局——①单成分胜出:这份星表是一次观测,指数直接给高或低;②双成分且平移吸收不掉:确认”抄录层+重测层”,指数按两层比例算;③样本量太小、两峰分不开:不强行定层,只给上界,等更多数据。

一句话:“几层”不是看直方图顺眼不顺眼,而是一道要过三道统计关的模型选择题。 把系统误差的”平移”和真实观测的”分层”分开,星表的”新鲜度”才不是自说自话。

证据与局限

已验证: 《石氏星经》坐标含双观测层(公元前4世纪+公元2世纪),证明”单层 vs 双层”可用混合模型分离(PMID 40273251);BIC 成分数不等于真实簇数、非高斯单峰易被高估(PMID 20953302);BIC 自动判决在已知真类模拟中极少最优(PMID 34548917);gap 统计量可定簇数(PMID 17425640、PMID 18397779);贝叶斯有限混合数类准则与”多余成分渐近变空”(Zmix)思路可判层(PMID 28081166、PMID 26177375);数值代数几何可直接求成分数(PMID 33867617);边界条件下似然比检验不服从卡方(PMID 16474914);AIC 与 BIC 目标不同(PMID 22309957);指数幂混合可容纳厚尾(PMID 38785671);ΔT 与历史交食提供独立时间锚(PMID 28119545、PMID 10800374);Gaia 天体测量误差处理为现代对照(PMID 23630374)。

部分验证: “先扣除整体平移再看双峰”这一判别步骤逻辑完整,但目前只在《石氏星经》等个案上被间接印证;”坐标新鲜度指数”的构建框架已成型,尚未有覆盖全部中国历代星表的系统结果发布。

待验证: 转抄过程中”部分重测”的混合比例对层数判别置信度的影响(需合成注入标定);”坐标新鲜度指数”与史载仪器精度(浑仪/简仪刻度)的定量对接;多准则(BIC/gap/贝叶斯/代数几何)层数判决的一致性检验;文本证据与数据指纹互相矛盾个案的系统梳理。

相关阅读

证据等级: B
参考文献: PMID 40273251, PMID 20953302, PMID 34548917, PMID 17425640, PMID 18397779, PMID 28081166, PMID 26177375, PMID 33867617, PMID 16474914, PMID 22309957, PMID 38785671, PMID 28119545, PMID 10800374, PMID 23630374


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。

One Layer or Two Layers? Using Gaussian Mixture Models to Cluster Coordinates from Historical Star Catalogs

Evidence level: B

Does a coordinate residual distribution from a star catalog represent “one layer” or “two layers”? This seems intuitive, yet it is a model selection problem requiring three statistical checkpoints. Choosing incorrectly causes the entire “freshness” assessment to collapse.

Main Text

The previous article decomposed “copied versus re-measured” into three error fingerprints. This article isolates the third layer—the “internal consistency fingerprint”—and frames it as a standard statistical multiple-choice question: Does the coordinate residual distribution of a star catalog represent “one layer” or “two layers”?

Why is the number of layers critical? One layer implies a single epoch (pure transcription or one-time measurement), while two layers imply “old foundation + new refresh” (transcription mixed with re-measurement). This is not a stylistic issue but the watershed determining whether “what proportion of this catalog reflects contemporary actual observation.”

Mathematically, the distribution of star catalog residuals is simply a histogram constructed from numerical data. The most direct tool for determining whether it is unimodal or bimodal is the Gaussian Mixture Model (GMM)—assuming data consists of stacked Gaussian “clusters,” then selecting the optimal number of clusters using an information criterion (most commonly BIC). It sounds clean, but three pitfalls lurk here.

First Pitfall: BIC counts not “layers,” but “how many Gaussians suffice to fit the density.” A non-Gaussian unimodal distribution (e.g., heavy-tailed or skewed) can be completely decomposed into two or three Gaussians, and BIC may still prefer this multi-Gaussian solution. Thus, a “two-layer” conclusion might be purely an artifact of overfitting (PMID 20953302). More troublingly, BIC’s automatic decision rarely selects the correct answer in simulations with known true cluster counts (PMID 34548917). In other words, “BIC says two layers” is far from “there truly are two layers.”

Second Pitfall: There are multiple algorithms for determining layers, and conclusions often conflict. Besides BIC, there is the gap statistic (judging cluster count against a null distribution, PMID 17425640, PMID 18397779), the number of classes criterion for Bayesian finite mixtures (PMID 28081166), the Zmix approach where “extra components asymptotically vanish during overfitting” (PMID 26177375), and using numerical algebraic geometry to directly compute the number of components (PMID 33867617). Each method has specific premises; inconsistency is the norm. Relying on a single criterion to determine layers is unreliable.

Third Pitfall: Adding a layer requires passing a specific test. Moving from one to two layers typically uses the likelihood ratio test. However, when testing for “one additional component,” parameters often fall on the boundary, causing the test statistic to no longer follow the standard chi-square distribution; direct application severely overestimates significance (PMID 16474914). Furthermore, there is a philosophical divergence regarding “which model to select”: AIC favors prediction while BIC favors explanation, often yielding different answers (PMID 22309957).

Is there a way to bypass these pitfalls? Yes—do not merely ask “how many layers,” but first ask “can these two peaks be eliminated by a single global translation?”

This is the most crucial step in this article. If ancient instruments had a systematic scale error of approximately 1°, it would cause a global translation across the entire catalog, not a bimodal distribution. In contrast, genuine “two observations” produce a layered structure that cannot be fully absorbed by a single translation. Thus, the discrimination reduces to an actionable question: After subtracting a uniform offset from the residuals, do the two peaks persist? If they do, it is truly “two layers”; if they merge into one peak upon translation, it is likely a system error masquerading as layers. This directly connects to the debate in Shi Ji Xing Jing regarding the double-observation layer—this study posits that the catalog contains two observation layers from the 4th century BCE and the 2nd century CE (PMID 40273251).

Two other factors must be considered simultaneously. First is heavy tails: Transcription or calculation errors generate small numbers of large outliers, causing residuals to deviate from Gaussianity; under such conditions, standard GMM is inappropriate, and one must use power-law mixtures capable of accommodating heavy tails (PMID 38785671). Second is external anchors: Converting residuals into “dates” requires the independent time scale of ΔT (PMID 28119545, PMID 10800374). Without accurate anchoring, the “two layers” will drift chronologically overall. Modern comparisons can reference Gaia satellite astrometric error handling (PMID 23630374): contemporary catalogs already face systematic differences and uncertainty propagation; ancient star catalogs require even greater caution.

The output is the previously announced “Coordinate Freshness Index”: each catalog receives a score representing the “proportion of contemporary actual measurement,” accompanied by the “uncertainty of the layer count decision.” Three outcomes: ① Single component wins: the catalog represents a single observation; the index is directly assigned high or low; ② Double component with persistent peaks after translation: confirms “transcription layer + re-measurement layer”; the index is calculated based on the proportion of layers; ③ Sample size too small or peaks indistinguishable: do not force a layer determination, only provide an upper bound, awaiting more data.

In summary: “How many layers” is not about whether the histogram looks pleasing to the eye, but a model selection problem requiring three statistical checkpoints. Separating the “translation” of systematic error from the “layering” of genuine observation ensures the catalog’s “freshness” is not self-serving.

Evidence and Limitations

Verified: Shi Ji Xing Jing coordinates contain a double observation layer (4th century BCE + 2nd century CE), proving “single vs. double layer” can be separated using mixture models (PMID 40273251); BIC component count does not equal true cluster count, and non-Gaussian unimodal distributions are prone to overestimation (PMID 20953302); BIC’s automatic decision is rarely optimal in simulations with known true cluster counts (PMID 34548917); the gap statistic can determine cluster count (PMID 17425640, PMID 18397779); Bayesian finite mixture number-of-classes criteria and the “extra components asymptotically vanish” (Zmix) approach can determine layers (PMID 28081166, PMID 26177375); numerical algebraic geometry can directly compute component counts (PMID 33867617); likelihood ratio tests do not follow chi-square under boundary conditions (PMID 16474914); AIC and BIC have different objectives (PMID 22309957); power-law mixtures can accommodate heavy tails (PMID 38785671); ΔT and historical eclipses provide independent temporal anchors (PMID 28119545, PMID 10800374); Gaia astrometric error handling serves as a modern reference (PMID 23630374).

Partially Verified: The logic of the “subtract global translation then examine bimodality” discrimination step is complete, but has only been indirectly corroborated in cases like Shi Ji Xing Jing; the framework for the “Coordinate Freshness Index” is established, but no systematic results covering all Chinese historical star catalogs have been published yet.

To Be Verified: The impact of the mixed proportion of “partial re-measurement” during transcription on the confidence of layer determination (requires synthetic injection calibration); quantitative docking between the “Coordinate Freshness Index” and historical instrument precision (armillary sphere/simplifier scales); consistency testing of layer determination across multiple criteria (BIC/gap/Bayesian/algebraic geometry); systematic review of cases where textual evidence contradicts data fingerprints.

Related Reading

Evidence Level: B
References: PMID 40273251, PMID 20953302, PMID 34548917, PMID 17425640, PMID 18397779, PMID 28081166, PMID 26177375, PMID 33867617, PMID 16474914, PMID 22309957, PMID 38785671, PMID 28119545, PMID 10800374, PMID 23630374


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 *