Evidence level: B
岁差反演把”观测-模型之差”(残差)吐出来,残差不是垃圾,而是一把尺子——它量的正是古人手里那件仪器到底有多准。但只有先分清”残差里哪些是仪器、哪些是星表本身断层、哪些是模型没算对”,这把尺子才敢用。
正文
上一篇(《全二十八宿距星统一岁差反演的”聚类结构”判别检验》)把”28 颗距星到底散成几簇”变成了一道统计模型选择题。本篇往下问一步:反演残差——也就是每颗星的观测坐标与岁差模型预测坐标之差——能不能倒推出古代仪器的真实测量误差?
直觉上很诱人:仪器越准,残差越小;残差大,就是古人手抖。但残差是个”三合一”的混合物,直接取标准差会把三种东西搅在一起。
第一层:仪器随机误差。这是最干净的一层。如果误差只是”读数时差一点”,残差应近似高斯分布,可用标准差直接刻画精度。但真实观测从来不是纯高斯的。Posekany 等人对大量数据的系统评估发现,无论怎么预处理,厚尾噪声模型都显著优于单纯高斯,用高斯会把离群观测当”不可能的怪值”误删,反而低估真实误差(PMID 21252077)。Park 与 Lim 对多元高斯的厚尾扩展做了系统梳理,指出存在离群点时会选错分布族,连带改变后续所有参数估计(PMID 36213771)。也就是说,”残差标准差”这个数字取决于你先假定它是高斯还是厚尾——两者能差出可观的一截。
第二层:仪器系统误差的空间模式。这才是最有考古价值的一层。古人的仪器(浑仪、简仪)是机械结构,刻度误差、环面偏心、观测者站位都会留下方向性痕迹,而非随机散布。历史上的星表对比研究反复呈现这一点:托勒密与乌鲁伯格星表的黄经系统误差都在约 1° 量级,且误差本身还有约 1° 的散布;直到第谷才把中位标准误差压到 60 角秒。如果二十八宿距星残差也存在”东宫偏左、西宫偏右”的系统空间模式,那它就是仪器机械特性的直接指纹。现代天体测量学处理这类问题的标准做法,是把残差按天区分块、在每块内取均值当作该区的系统偏置,从而把”信号”和”噪声”分开(ICRF 编目残差分解;小行星 O−C 残差反演星表系统误差)。中国二十八宿覆盖黄道/赤道整圈,天然具备做这种分宫(东/西宫)空间自相关检验的条件。
第三层:模型误差与星表断层。残差大不一定是仪器差,也可能是模型没算对(岁差常数取值、ΔT 时间尺度)或星表本身是多次观测拼盘(见上一篇)。三者贡献互相纠缠,这是最难的地方。Stephenson 等人用 720BC–AD2015 的日食、掩星记录把地球自转的 ΔT 反演到公认精度(PMID 28119545),其更早工作系统展示了历史天象如何约束地球自转长期变化(PMID 10800374)——这为”模型误差”提供了独立的外部锚。反过来,Meyers 与 Malinverno 对 14 亿年前韵律岩层的贝叶斯反演给出 85.79±2.72 角秒/年的岁差常数与 18.68±0.25 小时的日长(PMID 29866837),Wu 等人把岁差频率反演推到 6.5 亿年前(PMID 39413189)——两项共同证明:只要模型正确、误差刻画诚实,从带噪周期记录里恢复”时间”是成熟技术,输出格式就是”最可能值 ± 严格不确定度框”。星表定年应当追求同样的输出格式。
那么具体怎么分?答案是引入一致性分析的标准工具。Bland-Altman 限界一致性方法专门解决”两种测量方法差多少算可换用”:它不看相关系数(相关系数度量的是”有没有关系”,不是”一不一致”),而是直接画差值-均值图,算出偏倚与 95% 限界(PMID 10702443 的报告标准综述)。但 Taffé 特别警告:当两种方法的精度不同、或系统差不是常数时,标准 LoA 方法会失效,此时必须用重复测量才能无偏分离(PMID 33892090)。二十八宿反演恰恰是”每颗星只测一次、不同星精度可能不同”的场景——这一条件的前置检验不能省。若要做跨朝代、跨星表的汇总,Tipton 与 Shuster 给出了 Bland-Altman 研究的随机效应 meta 分析框架(PMID 28664537),可作为”多代星表残余误差演化趋势”的统计骨架。
于是这道题的可执行路线是三步:①先定分布族——残差到底是高斯还是厚尾,用 AIC/BIC 与拟合优度裁决,不预先假定;②再做空间分解——按东/西宫、按二十八宿分区检验残差空间自相关,把系统模式与随机散布分离;③拿外部锚对表——用 ΔT 重建与史载仪器刻度(浑仪/简仪)作独立标尺,判断残差中哪一部分可归因于仪器、哪一部分必须记在模型或星表断层的账上。
三种可能的结局,每一种都对应一个不同的历史判断:残差近似高斯且无空间模式——说明系统误差已被模型吸收,残差给出的是纯随机精度上限;残差厚尾且有稳定分宫模式——说明存在真实的仪器机械误差指纹,可据此反推浑仪/简仪的实测精度,并与史载刻度对接;残差被模型误差主导——说明以现有定年精度,仪器误差还测不出来,只能给出上界。
从频率科学的角度看,这件事的意义在于:二十八宿是 28 个”频率基准点”,岁差是让整条基准线缓慢移位的载波,残差就是解调后的相位噪声谱。噪声谱的形状(高斯/厚尾)与空间结构(有无分宫模式),直接告诉我们这套古代测量系统是”一台好仪器”还是”一套拼凑的观测习惯”。
证据与局限
已验证: 岁差是精确已知的天文常数;贝叶斯反演已在 6.5 亿年(PMID 39413189)与 14 亿年(PMID 29866837)尺度上稳定重构出岁差频率与日长;古代日食/掩星记录可反推日长变化(PMID 28119545, PMID 10800374);高通量数据中厚尾噪声模型显著优于高斯(PMID 21252077),分布族选择会改变参数估计(PMID 36213771);Bland-Altman 一致性检验及其报告标准、方法适用边界已成熟(PMID 10702443, PMID 33892090, PMID 28664537)。
部分验证: 二十八宿距星残差是否呈现可识别的分宫空间模式,尚无系统性公开研究;古代仪器精度(浑仪/简仪刻度)与史载数字的定量对接仍依赖零散史料;残差三成分(仪器/模型/星表断层)的方差分解缺乏约束充分的实测标定。
待验证: 二十八宿距星反演残差的分布形态是高斯还是厚尾?残差是否存在稳定的东/西宫系统模式?能否用残差离散度反推古代仪器的真实测量精度,并与史载刻度对齐?跨朝代的残差是否呈现可建的演化趋势?
相关阅读
证据等级: B
参考文献: PMID 40273251, PMID 28119545, PMID 10800374, PMID 29866837, PMID 39413189, PMID 21252077, PMID 36213771, PMID 25378893, PMID 10702443, PMID 33892090
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。
Can the residuals of the retrograde analysis of the twenty-eight mansions’ pole stars measure the true error of ancient instruments?
Evidence level: B
Precession inversion reveals the “observation-model difference” (residuals); residuals are not garbage but a ruler—they measure exactly how accurate the instrument in the ancient hands was. However, this ruler can only be used if we first distinguish which part of the residuals stems from the instrument, which from discontinuities in the star catalog itself, and which from incorrect model calculations.
Main Text
The previous article (“Cluster Structure Discrimination Test for Unified Precession Inversion of All 28 Mansion Distance Stars”) transformed the question of “into how many clusters do the 28 distance stars scatter” into a statistical model selection problem. This article proceeds one step further: can the inversion residuals—the difference between observed coordinates for each star and the predicted coordinates from the precession model—be used to back-calculate the true measurement error of ancient instruments?
Intuitively, this is appealing: the more accurate the instrument, the smaller the residuals; large residuals imply ancient hand tremors. However, residuals are a “three-in-one” mixture, and taking the standard deviation directly would conflate these three components.
Layer 1: Instrument Random Error. This is the cleanest layer. If errors were merely “slight reading deviations,” residuals should approximate a Gaussian distribution, characterizable directly by the standard deviation. However, real observations are never purely Gaussian. A systematic evaluation by Posekany et al. on large datasets found that, regardless of preprocessing, thick-tailed noise models significantly outperform simple Gaussian models; using a Gaussian model would treat outlier observations as “impossible anomalies” for deletion, thereby underestimating the true error (PMID 21252077). Park and Lim conducted a systematic review of thick-tailed extensions to the multivariate Gaussian, pointing out that outliers lead to selecting the wrong distribution family, which in turn alters all subsequent parameter estimates (PMID 36213771). Thus, the “residual standard deviation” figure depends entirely on whether one assumes a Gaussian or thick-tailed distribution—both can differ by a significant margin.
Layer 2: Spatial Patterns of Instrument Systematic Error. This is the most archaeologically valuable layer. Ancient instruments (armillary spheres, simplified armillary spheres) are mechanical structures; scale errors, ring eccentricities, and observer positioning leave directional traces rather than random scatter. Historical star catalog comparisons repeatedly demonstrate this: the ecliptic longitude systematic errors in Ptolemy’s and Ulugh Beg’s star catalogs were both around 1°, with an additional scatter of about 1°; it was Tycho Brahe who compressed the median standard error to 60 arcseconds. If the 28 mansion distance star residuals also exhibit a systematic spatial pattern such as “East Palace biased left, West Palace biased right,” it constitutes a direct fingerprint of the instrument’s mechanical characteristics. The standard practice in modern astrometry for handling such issues involves partitioning residuals by day, taking the mean within each block as the systematic bias for that region, thereby separating “signal” from “noise” (ICRF catalog residual decomposition; asteroid O−C residual inversion of star catalog systematic errors). The Chinese 28 mansions cover the entire ecliptic/equatorial circle, naturally providing the conditions for performing such intra-palace spatial autocorrelation tests (East/West Palace).
Layer 3: Model Error and Star Catalog Discontinuities. Large residuals do not necessarily imply poor instruments; they may also stem from incorrect model calculations (precession constant values, ΔT time scale) or discontinuities in the star catalog itself (as noted in the previous article). The contributions of these three factors are intertwined, which is the most difficult aspect. Stephenson et al. inverted Earth rotation ΔT to a recognized precision using eclipse and occultation records spanning 720 BC to AD 2015 (PMID 28119545); their earlier work systematically demonstrated how historical celestial phenomena constrain long-term Earth rotation changes (PMID 10800374)—providing an independent external anchor for “model error.” Conversely, Meyers and Malinverno provided a Bayesian inversion of rhythmic rock layers 1.4 billion years ago yielding a precession constant of $85.79 \pm 2.72$ arcseconds/year and a day length of $18.68 \pm 0.25$ hours (PMID 29866837); Wu et al. extended the precession frequency inversion to 650 million years ago (PMID 39413189). Together, these two studies prove: provided the model is correct and error characterization is honest, recovering “time” from noisy periodic records is a mature technique, with output formatted as “most likely value ± strict uncertainty bounds”. Star catalog dating should pursue the same output format.
So, how to specifically separate these factors? The answer is to introduce standard tools for consistency analysis. The Bland-Altman limit-of-agreement method specifically addresses “how much difference between two measurement methods counts as interchangeable”: it does not look at correlation coefficients (which measure “whether there is a relationship,” not “how inconsistent they are”), but directly plots difference-versus-mean graphs, calculating bias and 95% limits (PMID 10702443 report standard review). However, Taffé specifically warns: when the precision of two methods differs or the systematic difference is not constant, standard LoA methods fail, and repeated measurements must be used to unbiasedly separate them (PMID 33892090). The 28 mansion inversion precisely fits the scenario of “each star measured only once, with potentially different precisions for different stars”—this prerequisite validation cannot be skipped. If conducting a summary across dynasties and star catalogs, Tipton and Shuster provided a random-effects meta-analysis framework for Bland-Altman studies (PMID 28664537), which can serve as the statistical skeleton for “multi-generational star catalog residual error evolution trends.”
Thus, the executable roadmap for this problem is three steps: ① First determine the distribution family—whether residuals are Gaussian or thick-tailed, decided by AIC/BIC and goodness-of-fit without prior assumption; ② Then perform spatial decomposition—partition by East/West Palace and by 28 mansions to test residual spatial autocorrelation, separating systematic patterns from random scatter; ③ Cross-validate with external anchors—use ΔT reconstruction and historical instrument scale records (armillary spheres/simplified armillary spheres) as independent rulers to judge which part of the residuals can be attributed to instruments and which must be recorded under model or catalog discontinuity accounts.
There are three possible outcomes, each corresponding to a different historical judgment: Residuals approximate Gaussian with no spatial pattern—indicating systematic errors have been absorbed by the model, and residuals provide a pure random precision upper limit; Residuals are thick-tailed with stable intra-palace patterns—indicating real instrument mechanical error fingerprints exist, allowing back-calculation of actual precision for armillary spheres/simplified armillary spheres and docking with historical scale records; Residuals dominated by model error—indicating that with current dating precision, instrument errors cannot yet be measured, only an upper bound can be provided.
From the perspective of frequency science, the significance lies in this: the 28 mansions are 28 “frequency reference points,” precession is the carrier wave slowly shifting the entire reference baseline, and residuals are the demodulated phase noise spectrum. The shape of the noise spectrum (Gaussian/thick-tailed) and its spatial structure (presence/absence of intra-palace patterns) directly tell us whether this ancient measurement system represents “a good instrument” or “a set of assembled observational habits.”
Evidence and Limitations
Verified: Precession is a precisely known astronomical constant; Bayesian inversion has stably reconstructed precession frequencies and day lengths on scales of 650 million years (PMID 39413189) and 1.4 billion years (PMID 29866837); ancient eclipse/occultation records can infer day length variations (PMID 28119545, PMID 10800374); thick-tailed noise models significantly outperform Gaussian models in high-throughput data (PMID 21252077), and distribution family selection alters parameter estimates (PMID 36213771); Bland-Altman consistency testing, its reporting standards, and method applicability boundaries are mature (PMID 10702443, PMID 33892090, PMID 28664537).
Partially Verified: Whether 28 mansion distance star residuals exhibit identifiable intra-palace spatial patterns lacks systematic public research; quantitative docking of ancient instrument precision (armillary sphere/simplified armillary sphere scales) with historical numerical records still relies on scattered historical data; variance decomposition of the three residual components (instrument/model/catalog discontinuity) lacks sufficient constraint from actual calibration.
To Be Verified: Is the distribution morphology of 28 mansion distance star inversion residuals Gaussian or thick-tailed? Do residuals exhibit stable East/West Palace systematic patterns? Can residual dispersion back-calculate the true measurement precision of ancient instruments and align with historical scale records? Do residuals across dynasties exhibit constructible evolution trends?
Related Reading
- Unified Precession Inversion of All 28 Mansion Distance Stars: Which Year Did Ancients Observe the Stars?
- “Cluster Structure” Discrimination Test for 28 Mansion Distance Star Inversion
- Chronology of 28 Mansion Distance Stars
- Ancient Eclipse Records and Long-Term Earth Rotation
- Research Pool: Spatiotemporal Frequency Topology
Evidence Level: B
References: PMID 40273251, PMID 28119545, PMID 10800374, PMID 29866837, PMID 39413189, PMID 21252077, PMID 36213771, PMID 25378893, PMID 10702443, PMID 33892090
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