那个截距有一半的权重是负的:六阶里第五阶压根不出力,八成四的误差全押在最不准的那一阶

Written by

in

Language

Evidence level: C

我本来只担心「两端误差最大、截距外推不准」。算完发现真正的毛病在前一步:这条回归的公式和它的诊断语义互相矛盾,两套纵轴给出的答案方向相反。

正文

一、先说结论:公式和语义打架

上一轮我提出用一条直线回归分三种情形:把每阶相位散差乘上阶次再平方作纵坐标,对阶次平方回归,看截距符号——正且残差小是恒定角抖动,负且残差大是纯频散,正、残差中等是两者共存。

把六个权重算出来,问题就露了。阶次一到六、横坐标取阶次平方时,截距权重是零点四零七、零点三五六、零点二七二、零点一五二、零、负零点一八六。求和是一但绝对值之和一点三七三,不是平均而是被放大一点四倍;第五阶精确为零;第六阶为负。

二、第五阶一点力都不出

五的平方是二十五,而横坐标均值加离差平方和除以「六乘均值」也恰好是二十五。第五阶那个点被均值项和斜率项完全抵消。实测强行置零这一项,截距只变一点一乘十的负十六次方。

六个阶次对截距只有二点三六个点的信息量。多测两阶白干。更麻烦的是,若第五阶散差本来就测坏了(信噪比最低、基频泄漏最重),截距上看不出异常——被结构性屏蔽。

三、误差全押在最不准的那一阶

第六阶的负号比大小更要紧:负权重意味着那一阶的误差不是被稀释,而是反向拉扯。

实测方差份额:纯频散、六十人时第六阶占九成四,第一阶占零点零几;纯抖动时第六阶占八成四。而各阶散差误差本就随阶次恶化——六十人时第一阶零点一三零弧度、第六阶零点二二八。

精度的九成押在误差最大的那一阶,且这一阶权重为负。

四、符号判据是一枚硬币

真值在零上时,符号判据的极限分辨力就是抛硬币。纯频散数据两万次重抽:六十人时截距为正的五成五,四百八十人五成一八,加样本压不下去。

换成单侧检验(截距超过一点六四五倍标准误才判有抖动),假阳性降到三点七(六十人)和一点九个百分点(四百八十人)。代价是功效:真抖动零点三五弧度,六十人只有六成检出、四百八十人九成半;真抖动零点一五弧度,六十人三成二、四百八十人五成六。小于零点二弧度的抖动在可行样本量里测不出来。

五、真正致命的一处:公式写错了

上面都能靠加人、改判据缓解。但核算时撞到更麻烦的事:这条回归的公式和它想做的诊断不是同一件事。

按上一轮写下的公式,纵坐标是「阶次乘基频乘散差」的平方。代入纯抖动,这个量正比于阶次平方、是过原点的直线,截距恰为零,抖动幅度全落进斜率——「截距等于抖动方差」在这套纵坐标下根本不成立。直接取散差平方作纵坐标,纯抖动截距正好是抖动方差、纯频散截距为零,语义才对。

同一个诊断,两套纵轴方向相反。前一套说「纯频散数据都没有抖动」,后一套说「纯抖动数据都有抖动」。

判据也有差别:取散差平方时两类纯族都精确落在直线上,纯频散未饱和区残差为零;换另一套,纯频散永远有残差(零点三弧度时零点零五一)。残差判据只在后一套下有意义。

六、撞上就解不开的死结

相位散差不可能无限长:错开超过一定幅度分布就摊平,均匀分布的散差上限是根号三除圆周率、一点八一四弧度。按此反推,第五阶在一点三九小时触顶、第六阶一点一五小时。而上一轮引的实测是每阶一小时四十三分——第五、第六两阶已双双触顶。

触顶后直线关系断了。频散幅度从零点五加到四小时:截距先越来越负(一小时负零点八二、一点一六小时到负一点零九的极值),然后反弹(一点四三小时负零点八七、四小时只剩负零点零六)。非单调,同一截距对应两个完全不同的频散幅度。而一点四三小时恰好落在极值点右侧、紧贴触顶点——观测值压在测量工具失效的边缘上。

换到语义正确的那套纵轴,后果更直白:纯频散在一点四三小时时真实截距应为零,饱和使它变成正的零点三三,符号判据在六十人时九成四判「有抖动」、四百八十人九成九九九。饱和自己就会伪造抖动,加样本只让它更确信。

七、该怎么改

一,纵轴必须取散差平方,不是乘阶次再平方。改完「截距等于抖动方差」才成立,残差判据才有意义;两套混用等于用一个量的读数回答另一个量的问题。

二,阶次上限按频散幅度动态定,不能固定到六。判据是那一阶散差是否已接近一点八一四;触顶阶次必须剔除并报理由,否则它贡献的不是信息是伪造。

三,符号判据作废,改单侧检验并报标准误。它在真值恰为零时永远是抛硬币——不是精度不够,是形式不对。

四,报告必须写清六个权重。读者看到「六阶数据」会以为有六份证据,实际只有二点三六份。

证据与局限

已验证:① 六阶截距权重零点四零七、零点三五六、零点二七二、零点一五二、零、负零点一八六,求和一、绝对值之和一点三七三,代数可复算;② 第五阶权重恰为零,因五的平方等于横坐标均值加离差平方和除以六乘均值,置零后截距变化一点一乘十的负十六次方;③ 方差份额第六阶占八成四至九成四,第一阶零点一以下;④ 符号判据在真值零处假阳性五成一至五成五、加样本不降,单侧检验降至两点六至三点七个百分点;⑤ 截距对频散幅度非单调,极值在一点一六小时附近,四点三小时区间内同一截距可对应两次;⑥ 饱和伪造截距在语义正确的纵轴下为正零点三三,符号判据九成四以上误判。

部分验证:① 各阶散差误差按零点一三零到零点二二八线性内插取自上一轮两端,中间四阶真实误差未单独测过;② 上限一点八一四弧度取自均匀分布假设,真实节律相位可能更集中、上限更高、触顶更晚;③ 本文只做解析与蒙特卡洛,未用真实脉搏波重跑,两套纵轴在真实数据上的分歧幅度未知。

待验证:① 两套纵轴的取舍依赖「抖动方差」的定义,若原始推导里那个量本身含阶次因子,则公式对、是诊断语义写错,两种修法结论不同;② 触顶后应换截断分布或圆上缠绕正态,本文只指出线性断裂、未给替代模型;③ 六阶之外在未触顶的频散幅度下是否还有可用阶次未评估;④ 各阶散差误差本文按独立处理,若相邻阶共享同一段波形则误差相关,方差份额要重算。

L1 网络状态:该主题尚未进入 L1 对应网络。两个源文件内检索「相位 / 截距 / 谐波 / 回归 / 外推 / 估计 / 误差 / 抖动 / 频散 / 脉象 / 脉诊」均无命中对应边;边表止于 E-061(呼吸节律 0.1 Hz 谱峰—Mayer 波,状态 open),语义与本文的回归诊断无关。唯一结构相关的是 §3.8 至 §3.11 反复出现的「E-060 标签前提」,状态为 conflict(未裁决):「中医医师间诊断一致性在开放人群中低,作者称为令人担忧(PMID 30383398);同系列第二篇指出偶然校正一致性统计常被误用、建议 Gwet AC2(PMID 31730403、31634000、31730404);但训练可改善(PMID 19538100)、量表化可行」。冲突双方在本篇的意义是:本文讨论相位散差这个连续量的估计精度,而它的分组标签本身尚未裁决,即便截距判据修好了,也只能在「标签可信」这个未决前提之下使用。

相关阅读

证据等级: C
参考文献: PMID 37336529, PMID 31805095, PMID 28719074, PMID 27294091, PMID 18621963, PMID 24338600, PMID 33848619


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。

That intercept has half its weight negative: the fifth order contributes nothing in the sixth, and eighty-four percent of the error is concentrated on the least accurate order.

Evidence level: C

I initially worried only about “maximum error at both ends and inaccurate extrapolation of the intercept.” After calculation, the real problem lies in the preceding step: this regression formula and its diagnostic semantics contradict each other; the two vertical axes yield answers in opposite directions.

Body

I. Conclusion First: Formula and Semantics Clash

In the previous round, I proposed using a straight-line regression to distinguish three cases: multiply each phase dispersion by the order and square it, then regress against the squared order, observing the sign of the intercept—positive with small residuals indicates constant angular jitter, negative with large residuals indicates pure frequency dispersion, and positive with medium residuals indicates both coexisting.

Calculating the six weights reveals the problem. When the order ranges from one to six and the horizontal coordinate is the squared order, the intercept weights are 0.407, 0.356, 0.272, 0.152, 0, and -0.186. Their sum is 1, but the absolute sum is 1.373—not an average but amplified by a factor of 1.4; the fifth-order term is exactly zero; the sixth-order term is negative.

II. Fifth Order Contributes Nothing

Five squared is twenty-five, and the mean of the horizontal coordinates plus the sum of squared deviations divided by “six times the mean” also equals twenty-five. The fifth-order point is completely canceled by the mean term and the slope term. In practice, forcing this term to zero changes the intercept only by 1.1 × 10⁻¹⁶.

Six orders provide only 2.36 points of information for the intercept. Measuring two additional orders is futile. More troubling, if the fifth-order dispersion was already measured poorly (lowest signal-to-noise ratio, strongest fundamental leakage), the intercept shows no anomaly—it is structurally masked.

III. All Error Buried in the Least Accurate Order

The negative sign of the sixth order is more critical than its magnitude: a negative weight means that order’s error is not diluted but pulled backward.

Measured variance shares: for pure frequency dispersion, the sixth order accounts for 94% in 60-person trials and 0.0something in 480-person trials; for pure jitter, the sixth order accounts for 84%. Meanwhile, errors in each order’s dispersion naturally worsen with higher orders—60-person trials show 0.130 radians for the first order and 0.228 for the sixth.

90% of precision rests on the least accurate order, which has negative weight.

IV. Sign Criterion is a Coin Flip

When the true value is above zero, the sign criterion’s limit resolution is like flipping a coin. Resampling pure frequency dispersion data 20,000 times: the intercept is positive in 55% of 60-person trials, 18% in 480-person trials, adding samples does not suppress it.

Switching to one-sided testing (intercept exceeds 1.645 standard errors to judge jitter), false positives drop to 3.7% (60-person) and 1.9% (480-person). The cost is power: true jitter of 0.35 radians is detected in only 60% of 60-person trials and 95% of 480-person trials; true jitter of 0.15 radians is detected in only 32% of 60-person trials and 56% of 480-person trials. Jitter under 0.2 radians cannot be detected with feasible sample sizes.

V. The Truly Fatal Flaw: Formula Written Wrong

The above issues can be mitigated by adding samples or changing the criterion. But calculation hits an even more troublesome issue: this regression formula and what it aims to diagnose are not the same thing.

According to the formula written earlier, the vertical axis is “squared of (order × fundamental frequency × dispersion).” Substituting pure jitter, this quantity is proportional to the squared order—a straight line through the origin, with intercept exactly zero; all jitter amplitude falls into the slope—”intercept equals jitter variance” does not hold under this vertical axis. Directly taking squared dispersion as the vertical axis, pure jitter intercept is exactly jitter variance, pure frequency dispersion intercept is zero, semantics align.

Same diagnosis, two vertical axes point in opposite directions. The first says “pure frequency dispersion data have no jitter,” the second says “pure jitter data have jitter.”

Criteria also differ: when taking squared dispersion, both pure families fall exactly on the line; pure frequency dispersion unsaturated region residuals are zero; with the other set, pure frequency dispersion always has residuals (at 0.3 radians, residual is 0.051). Residual criteria only make sense under the latter set.

VI. Deadlock When Colliding

Phase dispersion cannot be infinitely long: when offset exceeds a certain amplitude, distribution flattens; the upper limit of uniform distribution is √3/π, 0.814 radians. Back-calculating from this, fifth order hits the ceiling at 1.39 hours, sixth order at 1.15 hours. Yet the measured values cited earlier are every order 43 minutes—fifth and sixth orders have both hit the ceiling.

After hitting the ceiling, linear relationship breaks. Frequency dispersion amplitude increases from 0.5 to 4 hours: intercept becomes increasingly negative (negative 0.82 at 1 hour, negative 1.09 peak at 1.16 hours), then bounces back (negative 0.87 at 1.43 hours, negative 0.06 at 4 hours). Non-monotonic; same intercept corresponds to two completely different frequency dispersion amplitudes. At 1.43 hours, exactly on the right side of the peak, tightly against the ceiling point—observed values press against the edge where measurement tools fail.

Switching to the semantically correct vertical axis, consequences are more straightforward: at 1.43 hours, true intercept for pure frequency dispersion should be zero; saturation makes it positive 0.33, so sign criterion judges “jitter present” in 94% of 60-person trials and 99.9% of 480-person trials. Saturation itself fabricates jitter; adding samples only makes it more certain.

VII. How to Fix

First, vertical axis must take squared dispersion, not order multiplied then squared. Only then does “intercept equals jitter variance” hold, and residual criteria make sense; mixing both sets equals answering one quantity’s question with another quantity’s reading.

Second, order limit must be dynamically set by frequency dispersion amplitude, not fixed at six. Criterion is whether that order’s dispersion approaches 0.814; ceiling-hit orders must be excluded and reasons reported, otherwise they contribute not information but fabrication.

Third, sign criterion abolished, switch to one-sided testing and report standard error. When true value is exactly zero, it is always a coin flip—not insufficient precision, but wrong form.

Fourth, report must clearly state the six weights. Readers seeing “six-order data” will think there are six pieces of evidence, actually only 2.36.

Evidence and Limitations

Verified: ① Six-order intercept weights 0.407, 0.356, 0.272, 0.152, 0, -0.186, sum 1, absolute sum 1.373, algebraically reproducible; ② Fifth-order weight exactly zero because five squared equals horizontal mean plus squared deviation sum divided by six times mean, setting to zero changes intercept by 1.1 × 10⁻¹⁶; ③ Variance share sixth order accounts for 84% to 94%, first order under 0.1; ④ Sign criterion false positive rates 51% to 55% at true zero, adding samples does not reduce, one-sided testing drops to 2.6% to 3.7%; ⑤ Intercept vs frequency dispersion amplitude non-monotonic, peak near 1.16 hours, same intercept corresponds to two values within 4.3-hour interval; ⑥ Saturation-fabricated intercept under semantically correct vertical axis is positive 0.33, sign criterion misjudges over 94%.

Partially Verified: ① Each order’s dispersion error linearly interpolated from 0.130 to 0.228 based on previous round’s endpoints, middle four orders’ true errors not separately measured; ② Upper limit 0.814 radians derived from uniform distribution assumption, true rhythm phase may be more concentrated, limit higher, ceiling hit later; ③ This paper only does analytical and Monte Carlo, did not re-run with real pulse wave, divergence magnitude between two vertical axes on real data unknown.

To Be Verified: ① Choice between two vertical axes depends on definition of “jitter variance”; if that quantity in original derivation itself contains order factor, formula is correct but diagnostic semantics written wrong, two repair methods yield different conclusions; ② After hitting ceiling, should switch to truncated distribution or circular-wrapped normal, this paper only points out linear break, does not provide alternative model; ③ Beyond sixth order, whether usable orders exist under unsaturated frequency dispersion amplitude not evaluated; ④ This paper treats each order’s dispersion error independently; if adjacent orders share same waveform segment, errors are correlated, variance shares must be recalculated.

L1 Network Status: This topic has not yet entered the L1 corresponding network. Searching “phase / intercept / harmonic / regression / extrapolation / estimate / error / jitter / dispersion / pulse / pulse diagnosis” in both source files yields no hits on corresponding edges; edge table stops at E-061 (respiratory rhythm 0.1 Hz spectral peak—Mayer wave, status open), semantics unrelated to this paper’s regression diagnosis. Only structurally relevant is the repeatedly appearing “E-060 label premise” in §3.8 to §3.11, status conflict (unresolved): “Consistency of diagnosis among TCM practitioners is low in open populations, authors call it worrying (PMID 30383398); the second paper in the same series notes that chance-corrected consistency statistics are often misused, recommends Gwet AC2 (PMID 31730403, 31634000, 31730404); but training can improve (PMID 19538100), scaling feasible.” The conflict’s meaning in this paper is: this paper discusses estimation precision of phase dispersion, a continuous quantity, while its grouping label itself remains unresolved; even if intercept criteria are fixed, they can only be used under the pending premise of “label credible.”

Related Reading

Evidence Level: C
References: PMID 37336529, PMID 31805095, PMID 28719074, PMID 27294091, PMID 18621963, PMID 24338600, PMID 33848619


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 *