Evidence level: C
「峰宽与峰距之比过 0.7,双峰就熔成单峰」这条线是算错了一个因子。真正的形态熔点在 1.0;而更麻烦的是,母文里 S₆ 从 57.9% 掉到 9.5% 这 48 个百分点的跌幅,只有一半是峰宽造成的,另一半是峰距被迫缩短的几何后果——两者被同一个约束绑在一起,谁都单独说不清。
一、先把「熔点」这个词定死
上一篇(rhythm-phase-shape-vs-dispersion-separability)算出:在相位离散度 σ 固定为 1.43 小时的条件下,等权双峰族的六阶共享份额 S₆ 高达 62.5%,而正态族只有 8%。那条结论用的是零宽度(δ)峰。
一旦给峰加上宽度,双峰的相位分布是
p(θ) = ½·N(θ; −d, s₀²) + ½·N(θ; +d, s₀²)
这里 d 是峰距的一半(两个峰分别在 ±d),s₀ 是每个峰的宽度。约束是离散度不变:
s₀² + d² = σ² = 1.43² 小时²
于是 s₀ 一变大,d 必须跟着变小。这不是额外假设,是「离散度不变」这一条自己推出来的。
「熔成单峰」在数学上只有一个定义:中点 θ=0 处二阶导由负变正。
d²p/dθ²(在 θ=0 处) = 2·e^(−d²/2s₀²)·(d²/s₀⁴ − 1/s₀²)
为零的条件是 d = s₀,即 s₀/d = 1。不是 0.7。
二、0.7 是从哪来的,以及它为什么错
把两个峰的半高全宽当成「峰宽」,再与「全峰距 2d」相比,得到 s₀/d ≈ 1/√2 ≈ 0.707。这个 0.707 是「半高等于峰距」的位置,不是「谷消失」的位置。
两个位置差得很远。按 σ=1.43 小时算:s₀/d = 0.707 时中点密度是峰高的 36%、二阶导 +16.3(明显还是谷);要到 s₀/d = 1.0,中点才升到 60.7%、二阶导归零。按 0.7 判「已熔」,等于在凹陷还有三分之一深的分布上判单峰。
母文用 0.7 算出「r=0.74 时 S₆ 只剩 9.5%」,并据此说核宽决定高阶信息存亡。若真熔点是 1.0,则 r=0.74 的分布仍是双峰,只是凹得浅,不该判成「已熔的单峰」。
三、48 个百分点的跌幅,只有一半是峰宽干的
这是本轮真正想说的。母文给的两组数:r=0.14 时 S₄=8.6%、S₆=57.9%;r=0.74 时 S₄=24.1%、S₆=9.6%。
但 r 从 0.14 走到 0.74 时,d 同时从 1.416 缩到 1.149 小时。两件事一起动了。
拆开看(Shapley 分解,把总变化平摊给两个因素):
基准 r=0.14 (d=1.416h, s₀=0.198h): S₆ = 57.94%
终点 r=0.74 (d=1.149h, s₀=0.851h): S₆ = 9.53%
只换 d(s₀ 保持 0.198h): S₆ = 22.17%
只换 s₀(d 保持 1.416h): S₆ = 24.91%
峰距的贡献(Shapley): −25.58 pp
峰宽的贡献(Shapley): −22.83 pp
总计: −48.41 pp
各占一半出头。母文把这 48 个百分点全记在峰宽账上,实际有一半是峰距缩短的账。
第 6 阶的两个因子可直接看到:
abs[cos(6d)] 从 0.608 降到 0.233(×0.38)
exp(−18s₀²) 从 0.953 降到 0.410(×0.43)
两者相乘 ×0.16,S₆ 才从 57.9% 掉到 9.5%
十六倍衰减里两项各出约 2.5 倍。要说「峰宽杀死高阶信息」,得先把峰距那一半摘出去。
四、还有一层更硬的东西:S₆ 归零由峰距独自决定,跟峰宽无关
S₆ 的 cos 项归零条件是
6d = π/2 → d = π/12 → d ≈ 1.000 小时
这个 1.000 小时与 σ 无关。σ=1.0 时对应 s₀=0;σ=1.43 时对应 s₀/d=1.022;σ=2.0 时对应 1.732。同一个 d=1 小时,在三种离散度下分别是「无宽峰」「刚熔」「熔透」。
这意味着:只要峰心落在 ±1 小时,第 6 阶就完全看不见它们,无论峰多宽多窄。而 σ=1.43、s₀/d=1.022 这一组恰好落在 S₆ 谷底(≈0.005%)。母文说 r=0.74 时只剩 9.6%,但曲线继续往下,到 r≈1.02 彻底归零,之后回升(r=1.3 回 1.5%,r=2.0 回 7.1%)。
归零之后还会回升——这条非单调性说明「峰宽越大高阶信息越少」也不成立。
五、口径一:K 取多少,S₆ 差四倍
第三节全部数字都依赖一个从未写明的选择:归一化时把阶数截到第几阶。
r=0.14: K=6 → S₆=19.1% K=8 → 12.2% K=12 → 8.2% K=20 → 5.4% K=40 → 4.1%
母文用不归一化的合向量长度(abs[cos(kd)]·exp(−k²s₀²/2)),δ 峰下给出 62.52%,与其表格一致,口径正确。但换成「份额」(除以总和)后,K=6 与 K=40 差 4.7 倍。而 K 该取几恰取决于高阶还有没有信息——循环定义。
六、口径二:同一条 S₆ 对应一大片 (d, s₀)
这是最该写进判据的一条。固定 S₆ = 9.5%,在 (d, s₀) 平面扫描找到 155 组参数给出同一个 S₆,d 覆盖 0.70 到 1.54 小时。其中:
d=0.70, s₀=1.12 → 单峰
d=0.87, s₀=0.79 → 双峰
d=1.06, s₀=0.02 → 双峰(近乎 δ 峰)
d=1.15, s₀=0.86 → 双峰(真值)
d=1.33, s₀=1.15 → 双峰
同一个 S₆,可以是单峰,也可以是双峰。 光靠 S₆ 无法反解出形态,必须同时看 S₄ 与 S₅(它们携带独立的 d 与 s₀ 信息)。母文拿单一 S₆ 的落差去论证「形态存亡」,这一步在信息量上不够。可对照混合模型那头:过拟合贝叶斯混合的模拟显示,组分要在样本里有足够代表才能估出来,样本量小于分布复杂度时会系统性低估组分数(PMID 26177375)——与本文简并方向一致:不是估不准,而是「组分能不能被看见」取决于样本与尺度的配比。
七、实测怎么做:先做多峰检验,再谈峰宽
这件事不需要新设备。圆上多峰检验是成熟方法:Silverman 临界带宽法在圆上有单调性证明,可直接用;Landler 等对五种圆上均匀性检验做功效模拟,明确指出 Rayleigh 检验对多峰偏离的功效「即使样本量相当大也令人担忧地低」,推荐 Hermans–Rasson(PMID 30100666)。Freeman 与 Dale 比较双峰系数、Hartigan dip 与 AIC 差值,发现距离效应恒强于比例效应、偏度会系统抬高检出率,且 AIC 差值几乎把所有分布都判成双峰(PMID 22806703)——这点必须防。
三步走:
- 先判峰数,用 Silverman 临界带宽或 Hermans–Rasson,不用 Rayleigh;样本量至少 150。
- 判出双峰后,用 S₄/S₅/S₆ 三阶联立反解 (d, s₀),不只报 S₆。
- 报 s₀/d 时同时报 K,否则不可比。
参考量:S₆ 真值 9.53% 时,n=60 的二项标准误约 3.8 个百分点,n=150 约 2.4,n=300 约 1.7。想把 s₀/d 定到 ±0.03 以内,n=150 刚够。另需注意:已发表的标准谐波协议(22 人、前 11 阶、报各阶幅值 ICC 0.83–0.97)根本没采「相位」这一列,它证明的是幅值稳,不是相位形态稳(PMID 27170904)——拿它支持「相位形态可靠」是越界引用。
证据与局限
已验证:双峰在 s₀/d=1.0 处形态熔解(d=s₀ 时 p”(0)=0,解析可证,本文数值复核一致)。母文的 δ 峰基准值(93.07/73.25/43.29/7.32/29.65/62.52)可完全复现,口径无误。S₆ 归零位置 d≈1.000 小时由 cos(6d) 的零点决定,与 σ 无关。
部分验证:Shapley 分解给出的「峰距 25.6pp / 峰宽 22.8pp」依赖固定 σ=1.43 与等权双峰这两条设定;换成不等权(如 0.9/0.1)或三峰族,两边的分配比例会变。K 截断对 S₆ 的敏感度(K=6 与 K=40 差 4.7 倍)在小 r 更剧烈,大 r 趋于饱和。圆上多峰检验方法本身成熟(PMID 30100666、37003995),但从未被用于脉搏相位分布。
待验证:① 真实脉搏相位分布的 s₀/d 是多少,至今无实测。circadian 领域的大样本相位数据(如 1887 人活动节律、4000+ 人 HairTime)报告的是相位本身近似正态(PMID 31249478、PMID 24132058),这暗示多数人 s₀/d 可能落在单峰侧,但那不是脉搏,且未报峰宽。② K 该取几,本文只指出它敏感,没有给出判据。③ 不等权双峰下熔点会往哪边移,未算。
L1 网络状态:CORRESPONDENCE_NETWORK.md 节点表物理侧仅登记「相位 / 节律 — — — 周期性耦合的关键」一行(有对象、无层级、无状态);边表止于 E-061(呼吸节律 0.1 Hz — Mayer 波,open)。全份 grep「峰宽 / 半高 / 核密度 / 双峰 / 熔 / 峰距 / KDE / 带宽 / 平滑」零命中,该主题尚未进入 L1 对应网络。
相关阅读
证据等级: C
参考文献: PMID 30100666, PMID 22806703, PMID 37003995, PMID 26177375, PMID 31249478, PMID 24132058, PMID 27170904
证据等级:A=有PubMed直接支撑,B=科学类比/间接证据,C=传统文化共识/初步探索
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 \”melts into a single peak at 0.7\” claim is off by a factor: the true melting point is 1.0, and half the drop was not caused by peak width.
Evidence level: C
The claim that “a ratio of peak width to peak distance exceeding 0.7 causes two peaks to melt into one” is missing a factor. The true morphological melting point is 1.0; furthermore, in the original text, the drop of S₆ from 57.9% to 9.5% represents a decline of 48 percentage points, of which only half was caused by peak width, while the other half resulted from the geometric consequence of forced peak distance shortening—both factors being bound by the same constraint.
I. Fixing the Definition of “Melting Point”
In the previous analysis (rhythm-phase-shape-vs-dispersion-separability), under the condition that phase dispersion σ is fixed at 1.43 hours, the sixth-order shared fraction S₆ of the equal-weight bimodal family reached 62.5%, whereas the normal family remained at only 8%. That conclusion relied on zero-width (δ) peaks.
Once width is introduced, the phase distribution of two peaks becomes
p(θ) = ½·N(θ; −d, s₀²) + ½·N(θ; +d, s₀²)
Here d is half the peak distance (the two peaks are located at ±d), and s₀ is the width of each peak. The constraint is that dispersion remains constant:
s₀² + d² = σ² = 1.43² hours²
Thus, as s₀ increases, d must decrease accordingly. This is not an additional assumption but a direct consequence of the “constant dispersion” constraint.
Mathematically, “melting into a single peak” is defined by the second derivative at the midpoint θ=0 changing from negative to positive:
d²p/dθ²(at θ=0) = 2·e^(−d²/2s₀²)·(d²/s₀⁴ − 1/s₀²)
This expression equals zero when d = s₀, i.e., s₀/d = 1. It is not 0.7.
II. Where Does 0.7 Come From, and Why Is It Wrong?
If one takes the half-height full width of two peaks as “peak width” and compares it to the “full peak distance 2d”, the ratio becomes s₀/d ≈ 1/√2 ≈ 0.707. This 0.707 corresponds to the position where “half-height equals peak distance,” not where “the valley disappears.”
These two positions differ significantly. Using σ=1.43 hours: when s₀/d = 0.707, the midpoint density is 36% of peak height, and the second derivative is +16.3 (clearly still a valley); only when s₀/d = 1.0 does the midpoint rise to 60.7% and the second derivative return to zero. Judging “melting” at 0.7 is equivalent to declaring a single peak while the distribution still has a valley one-third deep.
The original text used 0.7 to calculate that “when r=0.74, S₆ drops to only 9.5%,” thereby concluding that nuclear width determines the survival of higher-order information. If the true melting point is 1.0, then the distribution at r=0.74 remains bimodal, merely with a shallower concavity, and should not be classified as “melted into a single peak.”
III. Of the 48 Percentage Points of Decline, Only Half Was Due to Peak Width
This is the core argument of this section. The original text provided two sets of numbers: at r=0.14, S₄=8.6% and S₆=57.9%; at r=0.74, S₄=24.1% and S₆=9.6%.
However, as r moves from 0.14 to 0.74, d simultaneously shrinks from 1.416 to 1.149 hours. Both factors change together.
Breaking them down (Shapley decomposition, distributing the total change between the two factors):
Baseline r=0.14 (d=1.416h, s₀=0.198h): S₆ = 57.94%
Endpoint r=0.74 (d=1.149h, s₀=0.851h): S₆ = 9.53%
Changing only d (s₀ kept at 0.198h): S₆ = 22.17%
Changing only s₀ (d kept at 1.416h): S₆ = 24.91%
Contribution of peak distance (Shapley): −25.58 pp
Contribution of peak width (Shapley): −22.83 pp
Total: −48.41 pp
Each accounts for slightly more than half. The original text attributed all 48 percentage points to peak width, whereas in reality half came from the account of peak distance shortening.
The two factors for the sixth order can be seen directly:
abs[cos(6d)] drops from 0.608 to 0.233 (×0.38)
exp(−18s₀²) drops from 0.953 to 0.410 (×0.43)
Multiplying these gives ×0.16, causing S₆ to fall from 57.9% to 9.5%
In the sixteen-fold attenuation, each factor contributes approximately 2.5 times. To claim “peak width kills higher-order information,” one must first remove the peak distance component.
IV. An Even Harder Layer: S₆ Vanishing Is Determined Solely by Peak Distance, Not Peak Width
The condition for the cosine term of S₆ to vanish is
6d = π/2 → d = π/12 → d ≈ 1.000 hours
This 1.000-hour value is independent of σ. When σ=1.0, it corresponds to s₀=0; when σ=1.43, it corresponds to s₀/d=1.022; when σ=2.0, it corresponds to 1.732. For the same d=1 hour, under these three dispersions, the peaks are respectively “zero-width,” “just melted,” and “fully melted.”
This implies: as long as the peak centers fall within ±1 hour, the sixth order completely cannot see them, regardless of how wide or narrow the peaks are. The set with σ=1.43 and s₀/d=1.022 happens to lie exactly at the S₆ valley (≈0.005%). The original text states that at r=0.74, only 9.6% remains, but the curve continues downward, reaching zero at r≈1.02, after which it rises again (r=1.3 returns to 1.5%, r=2.0 returns to 7.1%).
After vanishing, it rises again—this non-monotonicity shows that “the larger the peak width, the less higher-order information” is also invalid.
V. Scope One: How Large Should K Be? S₆ Differs by Four Times
All numbers in Section III depend on an unstated choice: up to which order should the series be truncated during normalization.
r=0.14: K=6 → S₆=19.1% K=8 → 12.2% K=12 → 8.2% K=20 → 5.4% K=40 → 4.1%
The original text used the unnormalized composite vector length (abs[cos(kd)]·exp(−k²s₀²/2)), yielding 62.52% under δ peaks, consistent with its table, and the scope is correct. However, after converting to “shares” (dividing by the sum), K=6 and K=40 differ by 4.7 times. The appropriate value of K depends precisely on whether higher orders still contain information—a circular definition.
VI. Scope Two: The Same S₆ Corresponds to a Wide Range of (d, s₀)
This is the most critical criterion to include. Fixing S₆ = 9.5% and scanning the (d, s₀) plane yields 155 sets of parameters producing the same S₆, with d ranging from 0.70 to 1.54 hours. Among them:
d=0.70, s₀=1.12 → single peak
d=0.87, s₀=0.79 → bimodal
d=1.06, s₀=0.02 → bimodal (almost δ peak)
d=1.15, s₀=0.86 → bimodal (true value)
d=1.33, s₀=1.15 → bimodal
The same S₆ can correspond to either a single peak or a bimodal distribution. Relying solely on S₆ cannot reverse-engineer the morphology; one must simultaneously examine S₄ and S₅ (which carry independent d and s₀ information). The original text’s argument that “the drop in single S₆ proves ‘morphological survival’” is insufficient in terms of information content. In contrast to the mixed model: Bayesian overfitting simulations show that components can only be estimated if they are sufficiently represented in the sample; when sample size is smaller than distribution complexity, the number of components is systematically underestimated (PMID 26177375)—consistent with this paper’s degeneracy direction: it is not about estimation accuracy, but whether ‘components can be seen’ depends on the ratio of sample to scale.
VII. Practical Implementation: Perform Bimodality Testing Before Discussing Peak Width
This does not require new equipment. Bimodality testing on circles is a mature method: Silverman’s critical bandwidth method has monotonicity proofs for circles and can be used directly; Landler et al. conducted power simulations for five circular uniformity tests, explicitly stating that the Rayleigh test has low power even with large sample sizes, recommending Hermans–Rasson (PMID 30100666). Freeman and Dale compared bimodal coefficients, Hartigan dip, and AIC differences, finding that distance effects are consistently stronger than proportion effects, skewness systematically inflates detection rates, and AIC differences almost classify all distributions as bimodal (PMID 22806703)—this must be guarded against.
Three steps:
- First determine the number of peaks, using Silverman’s critical bandwidth or Hermans–Rasson, not Rayleigh; sample size should be at least 150.
- After identifying bimodality, jointly solve (d, s₀) using S₄/S₅/S₆, reporting not just S₆.
- When reporting s₀/d, also report K, otherwise comparisons are invalid.
Reference quantities: when S₆’s true value is 9.53%, the binomial standard error is approximately 3.8 percentage points for n=60, 2.4 for n=150, and 1.7 for n=300. To constrain s₀/d within ±0.03, n=150 is just sufficient. Note also: existing standard harmonic protocols (22 subjects, first 11 orders, reporting ICC 0.83–0.97 for amplitudes) did not include the “phase” column; they prove amplitude stability, not phase morphology stability (PMID 27170904)—using this to support “reliable phase morphology” is an out-of-scope citation.
Evidence and Limitations
Verified: Bimodal morphology melts at s₀/d=1.0 (when d=s₀, p”(0)=0, analytically provable, confirmed numerically in this paper). The original text’s δ-peak baseline values (93.07/73.25/43.29/7.32/29.65/62.52) can be fully reproduced, and the scope is error-free. The position where S₆ vanishes (d≈1.000 hours) is determined by the zero of cos(6d), independent of σ.
Partially Verified: The Shapley decomposition results—”peak distance 25.6pp / peak width 22.8pp”—depend on the fixed settings of σ=1.43 and equal-weight bimodal peaks; changing to unequal weights (e.g., 0.9/0.1) or a trimodal family would alter the allocation ratios. The sensitivity of S₆ to K truncation (K=6 vs K=40 differing by 4.7 times) is more pronounced at small r, saturating at large r. Circular bimodality testing methods are mature (PMID 30100666, PMID 37003995), but have never been applied to pulse phase distributions.
Pending Verification: ① What is the actual s₀/d ratio of real pulse phase distributions? No measurement exists to date. Large-sample phase data in circadian fields (e.g., activity rhythms of 1,887 individuals, HairTime of 4,000+) report that phases themselves approximate normal distributions (PMID 31249478, PMID 24132058), suggesting that for most people s₀/d likely falls on the single-peak side, but these are not pulses, and peak widths were not reported. ② How many orders should K take? This paper only indicates sensitivity but provides no criterion. ③ Under unequal-weight bimodal distributions, where does the melting point shift? Not calculated.
L1 Network Status: The CORRESPONDENCE_NETWORK.md node table physically registers only the line “phase/rhythm — — — key periodic coupling” (with objects but no hierarchy or state); the edge table stops at E-061 (respiratory rhythm 0.1 Hz — Mayer waves, open). A full grep for “peak width / half-height / kernel density / bimodal / melt / peak distance / KDE / bandwidth / smoothing” yields zero hits; this topic has not yet entered the L1 correspondence network.
Related Reading
- Shape and Dispersion Are Inseparable: Harmonic Differences Across Six Distribution Families Under the Same σ
- Why the Harmonic Shares of Bimodal Families Do Not Monotonically Decrease with Order
- Biofrequency Code Research Pool
Evidence Level: C
References: PMID 30100666, PMID 22806703, PMID 37003995, PMID 26177375, PMID 31249478, PMID 24132058, PMID 27170904
Evidence Level: A = direct PubMed support, B = scientific analogy/indirect evidence, C = traditional consensus/preliminary exploration
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