Evidence level: B
把 28 颗距星的观测年代都反演出来之后,”它们是不是同一年的”这道题,不能靠肉眼看散点图——它是一道标准的统计模型选择题:数据该用一个总体(单峰)解释,还是用两个、三个总体(多峰)解释?而这正是现代混合模型领域最容易被”看错”的一步。
正文
上一篇《全二十八宿距星统一岁差反演》把问题定成:用同一岁差模型对全部 28 颗距星逐一反演观测年代,看答案收不收敛。收敛=一次观测,散成几簇=多次观测与修订。但“散成几簇”这句判断本身是一个独立的统计问题,而且最容易翻车。
坑在哪?最顺手的做法是拟合高斯混合模型,用 BIC 选成分数——BIC 最低点对应的就是”簇数”。问题是:BIC 选的是”几个高斯能把密度拟合好”,不是”数据里真有几个子群”。一个非高斯的单一总体(如偏态分布)常被拆成两三个高斯逼近,BIC 便大方地说”3 个成分”,你却以为发现了 3 次观测。Baudry 等人专门解决了这个”成分数 ≠ 簇数”的高估问题(PMID 20953302):先用 BIC 定成分总数 K,再按熵准则做层级合并,合并后的簇数才是能和史料对话的数。Gergely 与 Vargha 的实测更不客气——在”已知真类型数”的模拟里,BIC 的自动判决极少给出最优解,还需用质量系数在两候选间取舍(PMID 34548917)。
第二个坑更深:这类问题上”究竟有几个类”是出了名的统计难题。贝叶斯有限混合领域早已发现,成分数设得偏大(过拟合混合)时,多余的类在比例参数较小的 Dirichlet 先验下会渐近变成空类——Nasserinejad 等人据此提出”数非空类”的判据,并与 DIC、边际似然、贝叶斯因子等主流办法系统比较(PMID 28081166);van Havre 等人用 Zmix 把它做成现成工具(PMID 26177375);Malsiner-Walli 等人把稀疏先验加在成分均值上,从 MCMC 输出直接读非空成分数的后验(PMID 26900266)。所以“1 次 vs 2 次观测”不能只看一条 BIC 曲线的最低点,而应三法并行:BIC/ICL 合并解、贝叶斯”数存活成分”、引入外部分组信息的 ICAL 准则(PMID 26461845)。
第三步:在问”几簇”之前,先该问”到底有没有簇”。28 个点是真带群结构,还是本就该由一个(加误差的)单一总体解释?集簇可检验性(clusterability)干的就是这个——在降维后的一维投影上跑 Hartigan Dip 检验与 Silverman 临界带宽检验,先判”单峰 vs 多峰”(PMID 37003995)。这是”单次观测”假说的第一道闸门:连多峰性都检不出,其后所有”几次观测”的叙述都是空中楼阁。而”要不要加一个成分”也不能凭残差看着像疙瘩就加——Tuttle 等人在多模态质谱分析里立了规矩:用 F 检验(似然比)判定”加第二个子群是否被数据正当支持”,且单看任一指标都会带来假阳性(PMID 39837577)。搬到星表上就是:加”第二次观测”这个成分,必须过似然比这一关。
最有意思的一点浮出来:系统误差和真实多期留下的痕迹形状完全不同。古人仪器若有约 1° 的系统刻度误差,不会把 28 颗星随机打散,而是把它们整体平移——这恰好能被”单一观测 + 一个全局偏移量”吸收掉。真实存在两次观测,才会留下一次全局平移吸收不掉的双峰。判别于是归约成一句可操作的话:残差的双峰性能否被一个统一的年代平移消掉?能消 → 单期观测 + 系统误差;消不掉 → 两次真实观测/修订。这正是《石氏星经》定年之争(广义霍夫变换反演出”约前 355 年一批 + 约公元 125 年一批修订”,PMID 40273251)的破局点——反对者主张计入 1° 系统误差后两批可重新自洽(”能被平移吸收”),支持者主张双期真实(”消不掉”),两种假说被逼成同一个可判定的问题。
这套”单一序列 vs 多期事件”的判别,考古年代学里有现成的成熟做法可借。Monongahela 传统的贝叶斯放射性碳年代模型,就拿一堆年代点去问:同一遗址的多次聚居是一个连续相位,还是中间有间断的离散相位?判据是相位区间下限是否大于零、各相位年代估计是否重叠(校准曲线有平台/反转时会出现多峰后验密度,须用区间约束收窄)(PMID 36288354)。Manning 与 Hart 更给出教训:他们用贝叶斯年代建模推翻了”欧洲金属器到 1525–1550 年才流通”这一原被当成年代表的区间(PMID 31841538)——先验区间一旦被当成事实,后续定年就会系统性偏。星表统一反演同理:年龄模型先验、误差框方向,都会决定你看到单峰还是双峰。外部锚是最终裁判——两千年不间断的日食/掩星记录已把地球自转 ΔT 反演到每世纪 +1.8 毫秒的公认精度(PMID 28119545),星表残差必须与它对得上。
完整判别是五步:① 距星坐标与误差框全量数字化;② 统一岁差模型逐星反演年代(带误差);③ 前置”有没有簇”的 Dip/带宽检验;④ 三法并行选模型(BIC/ICL 合并解 + 贝叶斯数存活成分 + ICAL),并用似然比守”加成分”的关;⑤ 用外锚(ΔT、独立定年)交叉校验。三种结局对应三种历史叙述:单成分胜出(残差可被单一平移吸收)→ 星表是一次观测,”现存最早星表”站得住;双成分胜出且平移吸收不掉 → 确有两次观测/修订,叙述降级为”最早星表之一 + 后世修订版”;功效不足(28 点分不开两成分)→ 只能给区间,请外锚入场。
最后一句提醒:28 个点是很小的样本,此时”单峰 vs 双峰”的分辨力会急剧下降——要多少颗距星、误差框压到多窄才能把两者分开,本身就该先用模拟标定,否则你看到的”簇”很可能只是噪声的起伏。
证据与局限
已验证: 岁差是精确已知的天文常数,逐星反演在数学上可行;”成分数 ≠ 簇数”的高估问题已被系统解决(PMID 20953302);BIC 自动判决在已知真类的模拟中常非最优,须配合 ICL/质量系数(PMID 34548917);贝叶斯”数非空成分”判据与过拟合混合的渐近性质已被刻画并工具化(PMID 28081166, PMID 26177375, PMID 26900266);”先验区间被当事实会导致系统性偏”有真实考古案例(PMID 31841538);二千年日食记录已把 ΔT 反演到公认精度(PMID 28119545)。
部分验证: 集簇可检验性(Dip/带宽)在”高维、小样本”下的功效仍偏弱,簇彼此靠得近时几种检验都会漏检(PMID 37003995);多模态判别需要多指标联合、单指标易假阳性(PMID 39837577);《石氏星经》双期 vs 单期之争尚无定论(PMID 40273251)。
待验证: 全 28 宿距星统一反演后,模型选择(单成分 vs 双成分)在 BIC/ICL/贝叶斯三法下是否一致收敛?反演残差的双峰性能否被单一全局年代平移吸收?28 颗距星的样本量与误差框能否支撑”1 次 vs 2 次观测”的统计分辨?须以模拟标定最小可辨条件。
相关阅读
证据等级: B
参考文献: PMID 20953302, PMID 34548917, PMID 28081166, PMID 26177375, PMID 26900266, PMID 26461845, PMID 37003995, PMID 39837577, PMID 36288354, PMID 31841538, PMID 28119545, PMID 40273251
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。
Discriminating the \”Cluster Structure\” of Retrograde Ages for the Twenty-Eight Mansions: Is a Star Table a Single Observation or a Stack of Assemblages?
Evidence level: B
After retrodating the observation years for all 28 lunar stations, the question of whether “they belong to the same year” cannot be resolved by visual inspection of scatter plots—it is a standard statistical modeling choice: should the data be explained by a single population (unimodal) or by two, three, or more populations (multimodal)? This is precisely the step most easily misinterpreted in modern mixture modeling.
Main Text
The previous article, Unified Precession Retrodating for All 28 Lunar Stations, framed the problem as: retrodate the observation year for each of the 28 lunar stations using a single precession model, then observe whether the results converge. Convergence implies a single observation; dispersion into clusters implies multiple observations and revisions. However, the judgment of “dispersing into clusters” itself is an independent statistical problem, and it is where failures are most likely to occur.
Where does the pitfall lie? The most convenient approach is to fit a Gaussian mixture model and select the number of components using the Bayesian Information Criterion (BIC)—the point with the lowest BIC corresponds to the “number of clusters.” The issue is: BIC selects “how many Gaussians can fit the density well,” not “how many subgroups truly exist in the data.” A non-Gaussian single population (e.g., a skewed distribution) is often approximated by two or three Gaussians; BIC then readily declares “3 components,” leading you to believe you have discovered 3 observation events. Baudry et al. specifically addressed this overestimation problem where “component count ≠ cluster count” (PMID 20953302): first determine the total number of components $K$ using BIC, then perform hierarchical merging based on an entropy criterion; only the merged cluster count can be meaningfully compared with historical records. Gergely and Vargha’s empirical findings are even more direct—in simulations with “known true class counts,” BIC’s automatic decision rarely yields the optimal solution, requiring a quality coefficient to choose between candidates (PMID 34548917).
The second pitfall is even deeper: determining “how many classes exist” is notoriously difficult in statistics. The Bayesian finite mixture field has long observed that when the component count is set too high (overfitting), extra classes with small proportion parameters tend to asymptotically become empty classes under a Dirichlet prior with small concentration parameters. Nasserinejad et al. proposed a criterion to “count only non-empty classes” and systematically compared it with mainstream methods like DIC, marginal likelihood, and Bayes factors (PMID 28081166); van Havre et al. implemented this as a ready-made tool using Zmix (PMID 26177375); Malsiner-Walli et al. added sparse priors to component means, allowing direct reading of the posterior for non-empty component counts from MCMC outputs (PMID 26900266). Therefore, “1 vs. 2 observations” cannot be judged solely by the lowest point on a BIC curve, but should instead employ three parallel methods: BIC/ICL merged solutions, Bayesian “count of surviving components,” and the ICAL criterion incorporating external grouping information (PMID 26461845).
Third step: Before asking “how many clusters,” one must first ask “do clusters actually exist.” Are the 28 points genuinely possessing a clustered structure, or should they be explained by a single population (plus error)? Clusterability testing addresses this—running the Hartigan Dip test and Silverman’s critical bandwidth test on a one-dimensional projection after dimensionality reduction first judges “unimodal vs. multimodal” (PMID 37003995). This is the first gate for the “single observation” hypothesis: if multimodality cannot even be detected, all subsequent narratives of “multiple observations” are built on sand. Furthermore, adding a component should not be done merely because residuals look like bumps—Tuttle et al. established rules in multimodal mass spectrometry analysis: use an F-test (likelihood ratio) to determine whether “adding a second subpopulation is legitimately supported by the data,” and relying on any single metric alone leads to false positives (PMID 39837577). Applied to star tables: adding a component for a “second observation” must pass the likelihood ratio test.
The most interesting point emerges: systematic errors and traces left by real multi-period observations have distinctly different shapes. If ancient instruments had a systematic scale error of approximately 1°, they would not randomly scatter the 28 stars but rather translate them as a whole—which can be absorbed by a “single observation plus a global offset.” Real existence of two observation periods would leave a bimodality that cannot be absorbed by a single global translation. Thus, the discrimination reduces to an actionable statement: Can the bimodality in the residuals be eliminated by a unified date translation? If yes → single-period observation + systematic error; if no → two real observations/revisions. This is the breakthrough point for the Shi Jixing dating controversy (generalized Hough transform retrodates “a batch around 355 BCE + a revision around 125 CE,” PMID 40273251)—opponents argue that after accounting for a 1° systematic error, both batches can be reconciled (“can be absorbed by translation”), while supporters argue for two real periods (“cannot be absorbed”), forcing both hypotheses into a single determinable problem.
This “single sequence vs. multi-period event” discrimination has established, mature practices in archaeology that can be borrowed. The Bayesian radiocarbon dating model of the Monongahela tradition takes a set of date points to ask: are multiple occupations at the same site one continuous phase or discrete phases with interruptions? The criterion is whether the lower bound of the phase interval is greater than zero and whether the date estimates for each phase overlap (calibration curves may show multimodal posterior densities during plateaus/reversals, requiring interval constraints to narrow them) (PMID 36288354). Manning and Hart provided a cautionary lesson: they used Bayesian dating modeling to overturn the originally accepted interval that “European metalware circulated only from 1525–1550” (PMID 31841538)—treating a prior interval as fact leads to systematic bias in subsequent dating. The unified retrodating of star tables is analogous: age model priors and error box directions will determine whether you see a unimodal or bimodal distribution. External anchors serve as the final judge—two millennia of uninterrupted eclipse/transit records have retrodated Earth’s rotation $\Delta T$ to the accepted precision of +1.8 milliseconds per century (PMID 28119545); star table residuals must align with it.
The complete discrimination involves five steps: ① Full digitization of lunar station coordinates and error boxes; ② Unified precession model retrodating years for each star (with errors); ③ Pre-test “existence of clusters” via Dip/bandwidth tests; ④ Parallel selection of models using BIC/ICL merged solutions + Bayesian count of surviving components + ICAL, guarded by likelihood ratios for adding components; ⑤ Cross-validation using external anchors ($\Delta T$, independent dating). Three outcomes correspond to three historical narratives: Single component wins (residuals absorbable by single translation) → star table is a single observation, “earliest existing star table” stands; Double component wins with unabsorbable translation → indeed two observations/revisions, narrative downgraded to “one of the earliest star tables + later revision”; Insufficient power (28 points cannot separate two components) → can only provide an interval, invite external anchors.
Final reminder: 28 points constitute a very small sample; the resolution for distinguishing “unimodal vs. bimodal” drops sharply—how many lunar stations and how narrow the error boxes must be to separate them should first be calibrated via simulation, otherwise the “clusters” you see may merely be noise fluctuations.
Evidence and Limitations
Verified: Precession is a precisely known astronomical constant; retrodating per star is mathematically feasible; the overestimation problem where “component count ≠ cluster count” has been systematically addressed (PMID 20953302); BIC’s automatic decision is often suboptimal in simulations with known true classes and requires pairing with ICL/quality coefficients (PMID 34548917); Bayesian “count of non-empty components” criteria and the asymptotic properties of overfitting mixtures have been characterized and toolized (PMID 28081166, PMID 26177375, PMID 26900266); “treating prior intervals as fact leads to systematic bias” has real archaeological case support (PMID 31841538); two millennia of eclipse records have retrodated $\Delta T$ to accepted precision (PMID 28119545).
Partially Verified: Clusterability (Dip/bandwidth) has relatively low power under “high dimension, small sample,” and several tests may miss clusters when they are close together (PMID 37003995); multimodal discrimination requires joint use of multiple metrics, with single metrics prone to false positives (PMID 39837577); the debate between Shi Jixing‘s double-period vs. single-period remains unresolved (PMID 40273251).
To Be Verified: After unified retrodating of all 28 lunar stations, do model selection criteria (single vs. double component) converge consistently under BIC/ICL/Bayesian methods? Can the bimodality in retrodating residuals be absorbed by a single global date translation? Can the sample size of 28 lunar stations and their error boxes support statistical resolution for “1 vs. 2 observations”? Minimum discernible conditions must be calibrated via simulation.
Related Reading
- Unified Precession Retrodating for All 28 Lunar Stations: Which Year Did Ancients Observe the Stars?
- Chronology of Lunar Station Years
- Star Table “Observation-Literature Lag” Test
- Research Pool: Spatiotemporal Frequency Topology
Evidence Level: B
References: PMID 20953302, PMID 34548917, PMID 28081166, PMID 26177375, PMID 26900266, PMID 26461845, PMID 37003995, PMID 39837577, PMID 36288354, PMID 31841538, PMID 28119545, PMID 40273251
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