Evidence level: C
两套算法,两个排名。吃亏的偏偏是练得最稳的那一组——因为偏差的大小,本身就是由「谁更稳」决定的。
一、先把切点说清:这篇拆的不是 MDC,是排名
531 那一节把「修炼者呼吸更稳」挂在 φ 的离散度上。此前已有两篇动过这把尺子:circular-mdc-cross-zero-arithmetic-artifact-calibration 拆的是单个差值的 MDC 公式在跨零时失效,判的是「测得出测不出」;dantian-phase-dispersion-stability-vs-freezing 拆的是离散度大到底算好还是算坏,判的是「解读方向」。这一篇问第三件事:两套算法给出的数,会不会把两组的排名对调。这是三个不同的问题。前两个都判好了,排名照样可能是反的——排名只看次序,而偏差恰好与「谁更同步」共线。
二、偏差不是随机的,它和 φ 均值同向
算术 SD 把 −179° 与 +179° 当成相隔 358°,实际只隔 2°。关键在于偏差方向与大小取决于 μ 落在哪:μ 越靠近 0°(几乎同相),越容易有几次相位从正翻负,样本 SD 就被这几次翻转撑大。于是偏差在「同步最好的人」身上最大。这不是噪声,是系统性的、与自变量共线的偏移:你要比较的正是「谁更同步」,而偏差强度恰好由同步程度决定。方向也对不上——偏差让人显得更不稳,实际更稳。圆形侧同样不干净:1−R(即相位锁定值、Kuramoto 序参量)在完全随机时只有样本量趋于无穷才趋于零,有限样本下它的期望值随 n 变化,跨样本量比较必然带偏(PMID 37756609)。该文的重归一化版本正为此设计:让期望值不再依赖 n。
三、更狠的一层:样本量恰好和「呼吸慢」锁死
第二节的两条偏差本来可以分开处理,偏偏在这里被同一个设计绑死。静息记录通常固定时长(如 10 分钟),呼吸次数 n 与呼吸周期成反比。练得深的人呼吸慢,10 分钟里凑不满次数。而 1−R 的负偏在 n 小时更大——于是「呼吸慢」这个本该是功力的标志,通过 n 变成了「离散度偏大」的伪信号。算术 SD 侧同理:n 小则单次跨零翻转的权重更大。这就构成了双重共线:偏差既随 μ 变、又随 n 变,而 μ 与 n 在修炼者身上同向变动。两条偏差叠起来,方向一致地指向「练得越深越显得不稳」。这比单纯说「算法有偏」严重:既往任何用固定时长窗口、未报 φ 均值与呼吸次数的组间比较,结论方向都不可预判。
四、RIP 这层还得再剥一次
上面都在讨论「拿到 φ 序列之后怎么算」,可 φ 本身怎么来的也有讲究。以 Lissajous 图算相位角是 TAA 的常规做法,但该法有数学限制:2290 个周期的分析显示,相位角只在 −86.90° 到 +88.40° 之间取值,而全局相位延迟覆盖 −179.75° 到 +178.54°;相位角与环面积无法正确分类严重偏移与矛盾运动(PMID 38829281)。即当不同步严重到出现矛盾呼吸时,算术相位角自己饱和,把「很不同步」压成上限值。对做组间比较的人,这是方向明确的坑:组间差被压扁在最不同步那一端。再叠一个实测参照:儿童 PSG 中 θavg 在 N3 期为 13.7°、REM 期为 33.1°,差 2.4 倍(PMID 29991414)——同一个人换个睡眠分期就能翻倍,跨组比较若不同分期,根本谈不上可比。
五、还有一层:误差本身可以「分组不同」
即便离散度算对了,组间差仍可能偏。差异测量误差(differential measurement error)指误差的性质随组别或时点而变,模拟研究显示它会让处理效应估计产生偏倚、并降低检出功效、影响置信区间覆盖率(PMID 34530859)。放到这里:新手戴着带子紧张、动作多、基线漂移大;老手躺下不动。误差结构天然分组不同,而 φ 的噪声底本来就不小——早产儿安静睡眠期 φ 的个体内 CV 达 11%–123%,且经机械模型排除后判定源于相位关系本身不稳,而非算法(PMID 24716708)。分母这么大,再叠一层分组误差,组间差里有多少是「稳」、多少是「谁更会配合测量」,分不开。
六、判据与做法
做法分两段。合成段:Monte Carlo 造 φ 序列,真实集中度 κ 已知,μ 从 0° 扫到 ±60°,n 取 20/50/200(对应不同呼吸速率),含与不含跨零两情形,比较算术 SD 与圆形离散度(1−R、圆方差、von Mises κ)对真实 κ 的恢复偏差。真实段:≥90 人同体位 ≥10 分钟 RIP,双算法重算,看排名是否翻转,并报 φ 均值与呼吸次数。判据三格:① 合成中算术 SD 在 μ→0 时相对真实离散度高估 >30%、随 μ→0 单调放大,且 n=20 与 n=200 之间差异 >10 个百分点 ⇒ 算法伪迹与样本量偏差双重确认,未报 φ 均值与 n 的 SD 结论一律不可解释;② 真实数据上双算法排名出现翻转(哪怕只一对组)⇒ 既往以算术 SD 论证「修炼者更稳/更不稳」的文献全部须重算;③ 两算法排名一致且偏差 <10% ⇒ 算术 SD 在成人静息 φ 上可继续使用,本文的方法学担忧撤回。必须自认:30% 与 10 个百分点均为本轮自定、无文献出处,执行前须改成以 MDC95 为锚的阈值。
七、这一篇能带来什么
价值不在「又发现一个算法缺陷」,而在把「稳」的可比性问题推到排名层面。离散度是相对量,排名是它唯一用法;偏差与自变量共线,「更稳」这个比较本身就在比偏差。更朴素的教训是:圆形数据的无偏度量早就存在(成对相位一致性即为一例,PMID 20114076),但呼吸协调领域大量沿用的仍是算术 SD。不是不知道,是没人在比较前算过这笔账。对丹田研究的直接含义是:在把 φ-SD 当终点之前,先做一次纯算法的偏差标定,比再招一批受试者便宜。
证据与局限
已验证:圆形量的集中度指标(1−R/相位锁定值/序参量)在有限样本下期望值随 n 变化、跨样本量比较带偏,已有重归一化方案(PMID 37756609);Lissajous 相位角存在取值饱和与分类失效(PMID 38829281);θavg 随睡眠分期显著变化(PMID 29991414);差异测量误差可致效应估计偏倚与功效下降(PMID 34530859)。
部分验证:φ 的噪声底量级有实测支撑(CV 11%–123%,PMID 24716708),但该数据来自早产儿,成人静息是否同量级未知;圆形统计的无偏替代度量存在且有工程可用性(PMID 20114076),但在呼吸协调领域尚未被系统采用。
待验证:算术 SD 在 μ→0 时的高估量级、以及该高估与 n 的交互大小,全部是本文推论,尚无任何实测标定;真实 RIP 数据上双算法排名是否真的会翻转,完全没有数据;「呼吸慢 ⇒ n 小 ⇒ 显得不稳」这条三重共线的实际效应量,也未测。
相关阅读
- 差值的 MDC 不能直接判圆形量:φ 跨零那一刻,算术公式先把最好的人判成「测不出来」
- 呼吸稳不稳,不能只看晃得小不小:丹田修炼者胸腹相位角逐次变异的「冻结式稳定」与「功能性变异」裁决
- 这道缝到底量不量得出来:Δ 自身的最小可检测变化与年限斜率可检出性的功效预算
- 环境-生命频率接口(池4)
证据等级: C
参考文献: PMID 37756609, PMID 38829281, PMID 29991414, PMID 34530859, PMID 24716708, PMID 20114076
L1 网络状态:CORRESPONDENCE_NETWORK.md2.2 物理侧仅列「相位 / 节律 | — | 周期性耦合的关键」,2.3 生理侧仅列「呼吸率 / 呼吸模式 | 可穿戴 | 与节律耦合最直接」;逐词检索「离散度/圆形统计/算术/排名/样本量/偏差/方差/算法」全部 0 命中,边表 3.1–3.5 止于 E-050(层级 L1、状态 open,现代证据栏原文「尚无任何可测来源」),无任何以「离散度 / 算法 / 排名」为源或目标的边 —— 该主题尚未进入 L1 对应网络。
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。
The same \”steady\” character, two algorithms yielding opposite rankings: deviation calibration of φ’s circular dispersion versus arithmetic SD across zero-breathing
Evidence level: C
Two algorithms, two rankings. The group that trained the most stably loses out—because the magnitude of the bias itself is determined by “who is more stable.”
I. Clarify the Cut-Point First: This Paper Does Not Split MDC, It Splits the Ranking
Section #531 hangs “the practitioner’s breathing is more stable” on φ’s dispersion. Two prior papers have already wielded this ruler: circular-mdc-cross-zero-arithmetic-artifact-calibration dissects the MDC formula for a single difference when crossing zero, judging “measurable vs. unmeasurable”; dantian-phase-dispersion-stability-vs-freezing dissects whether large dispersion is good or bad, judging the “interpretation direction.” This paper asks the third question: Will the numbers from the two algorithms flip the ranking of the two groups. These are three distinct questions. The first two have been judged, yet the ranking can still be reversed—ranking depends solely on order, while the bias is collinear with “who is more synchronized.”
II. The Bias Is Not Random; It Aligns with the φ Mean
Arithmetic SD treats −179° and +179° as separated by 358°, whereas they are actually only 2° apart. The key lies in the direction and magnitude of the bias depend on where μ falls: the closer μ is to 0° (near perfect synchrony), the more likely several phases flip from positive to negative, inflating the sample SD due to these flips. Thus, the bias is largest for “the most synchronized person.” This is not noise; it is a systematic, collinear shift with the independent variable: you are comparing “who is more synchronized,” yet the bias intensity is exactly determined by the degree of synchronization. The direction also mismatches—the bias makes people appear less stable, while they are actually more stable. The circular side is similarly flawed: 1−R (i.e., phase locking value, Kuramoto order parameter) only approaches zero as sample size tends to infinity under complete randomness; under finite samples, its expected value varies with n, so cross-sample-size comparisons inevitably introduce bias (PMID 37756609). The paper’s re-normalized version was designed precisely for this: to make the expected value independent of n.
III. An Even Sharper Layer: Sample Size Is Locked to “Slow Breathing”
The two biases in Section II could have been handled separately, yet here they are bound together by the same design. Resting recordings typically have fixed duration (e.g., 10 minutes); the number of breaths n is inversely proportional to the breathing cycle. Those who train deeper breathe slower, so within 10 minutes they cannot complete enough breaths. Meanwhile, the negative bias of 1−R grows larger as n decreases—thus “slow breathing,” which should be a marker of mastery, becomes via n a false signal of “large dispersion.” The arithmetic SD side is similar: smaller n gives greater weight to single zero-crossings. This creates double collinearity: the bias depends both on μ and on n, yet μ and n move in the same direction for practitioners. Stacked together, these two biases consistently point toward “the deeper the training, the less stable they appear.” This is worse than simply stating “the algorithm is biased”: any prior between-group comparison using fixed-duration windows without reporting φ mean and breathing count has an unpredictable conclusion direction.
IV. RIP Requires One More Peeling Layer
The above discusses “how to compute after obtaining the φ sequence,” but how φ itself is derived also matters. Computing phase angles via Lissajous plots is standard in TAA, yet this method has mathematical limits: analysis of 2,290 cycles shows phase angles range only from −86.90° to +88.40°, while global phase delay spans −179.75° to +178.54°; phase angle and ring area cannot correctly classify severe offsets and contradictory motion (PMID 38829281). That is, when asynchrony becomes so severe that contradictory breathing occurs, the arithmetic phase angle saturates, compressing “very asynchronous” into its upper limit. For those doing between-group comparisons, this is a pitfall with clear direction: between-group differences are compressed at the most asynchronous end. Adding an empirical reference: in pediatric PSG, θavg is 13.7° in N3 stage and 33.1° in REM stage, a difference of 2.4× (PMID 29991414)—the same person can double across sleep stages, making cross-group comparison incomparable if stages differ.
V. Another Layer: Error Structure Itself Can Differ by Group
Even if dispersion is computed correctly, between-group differences may still be biased. Differential measurement error refers to error properties varying by group or time point; simulation studies show it biases effect estimates, reduces detection power, and affects confidence interval coverage (PMID 34530859). Applied here: novices wearing straps feel tense, move more, and have larger baseline drift; experts lie still. The error structure is inherently group-specific, while φ’s noise floor is already substantial—premature infants’ quiet sleep phase φ has an individual CV of 11%–123%, and after mechanical model exclusion, this is judged to stem from inherent phase instability rather than algorithmic issues (PMID 24716708). With such a large denominator, stacking another layer of group-specific error makes it impossible to distinguish how much of the between-group difference reflects “stability” versus “who cooperates better with measurement.”
VI. Criteria and Methods
The approach has two segments. Synthetic segment: Monte Carlo generates φ sequences with known true concentration κ, μ sweeps from 0° to ±60°, n takes values of 20/50/200 (corresponding to different breathing rates), including both with and without zero-crossing, comparing arithmetic SD against circular dispersion (1−R, circular variance, von Mises κ) in terms of recovery bias from true κ. Real segment: ≥90 individuals in same posture ≥10 minutes of RIP, recompute with both algorithms to check if rankings flip, and report φ mean and breathing count. Three criteria: ① In synthetic data, arithmetic SD overestimates true dispersion by >30% as μ→0, monotonically increasing, and the difference between n=20 and n=200 exceeds 10 percentage points ⇒ confirms both algorithmic artifact and sample-size bias; any SD conclusion not reporting φ mean and n is uninterpretable; ② In real data, if both algorithms flip rankings (even just one pair of groups) ⇒ all prior literature using arithmetic SD to argue “practitioners are more stable/less stable” must be recalculated; ③ If both algorithms agree and bias <10% ⇒ arithmetic SD can continue for adult resting φ, and this methodological concern is withdrawn. Must self-admit: the 30% and 10 percentage points are self-defined thresholds for this round, with no literature source; before execution, they must be revised to use MDC95 as the anchor.
VII. What This Paper Can Bring
The value lies not in “discovering another algorithmic defect,” but in pushing the comparability of “stability” to the ranking level. Dispersion is a relative quantity, and ranking is its sole usage; bias is collinear with the independent variable, so “more stable” itself compares against the bias. A simpler lesson is: unbiased measures for circular data already exist (paired phase consistency is one example, PMID 20114076), yet the breathing coordination field largely continues using arithmetic SD. They are not unaware; no one has calculated this account before comparison. For dantian research, the direct implication is: before taking φ-SD as the endpoint, perform a pure algorithmic bias calibration—cheaper than recruiting another batch of subjects.
Evidence and Limitations
Verified: Concentration metrics for circular quantities (1−R/phase locking value/order parameter) have expected values varying with n under finite samples, causing cross-sample-size comparison bias, and a re-normalization solution exists (PMID 37756609); Lissajous phase angles exhibit saturation and classification failure (PMID 38829281); θavg varies significantly across sleep stages (PMID 29991414); differential measurement error can bias effect estimates and reduce power (PMID 34530859).
Partially Verified: φ’s noise floor magnitude has empirical support (CV 11%–123%, PMID 24716708), but this data comes from premature infants, leaving unknown whether adults at rest share the same scale; unbiased alternatives for circular statistics exist and are engineering-ready (PMID 20114076), yet not systematically adopted in the breathing coordination field.
To Be Verified: The degree of overestimation by arithmetic SD as μ→0, and the magnitude of interaction between this overestimation and n, are entirely inferred in this paper with no empirical calibration; whether rankings truly flip under both algorithms on real RIP data lacks any data; the actual effect size of “slow breathing ⇒ small n ⇒ appears unstable” remains unmeasured.
Related Reading
- MDC of differences cannot directly judge circular quantities: at φ’s zero-crossing moment, the arithmetic formula first judges the best person as “unmeasurable”
- Whether breathing is stable cannot be judged solely by small swing: ruling on “freezing stability” and “functional variation” of dantian practitioners’ thoraco-abdominal phase competition
- Does this gap even measure itself: power budget of Δ’s own minimum detectable change and tenure slope detectability
- Environment-Life Frequency Interface (Pool 4)
Evidence Level: C
References: PMID 37756609, PMID 38829281, PMID 29991414, PMID 34530859, PMID 24716708, PMID 20114076
L1 Network Status:CORRESPONDENCE_NETWORK.mdSection 2.2 Physical side lists only “phase/rhythm | — | key periodic coupling”; Section 2.3 Physiological side lists only “breathing rate/breathing pattern | wearable | most directly coupled to rhythm”; keyword search for “dispersion/circular statistics/arithmetic/ranking/sample size/bias/variance/algorithm” yields zero hits; edge table 3.1–3.5 stops at E-050 (layer L1, status open, original text states “no measurable source yet“), with no edges originating from or targeting “dispersion/algorithm/ranking” — this topic has not yet entered the L1 corresponding network.
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