幂律、有理式还是样条:s*(m) 曲线参数化形式的模型选择裁决

Written by

in

Language

Evidence level: C

一条拟合曲线的形状不是画出来的,是选出来的——选错了形状,外推到高 m 档就是系统性错路,而且错得很自信。

正文(1000-1500字)

同一批八档 m 仿真网格数据,三条候选曲线都”能拟合”:幂律(两头平滑、可外推)、有理式(能出渐近线、能出拐点)、样条(贴数据最紧、但外推区行为完全由端点惯性决定)。拟合优度上样条几乎永远赢——参数多、自由度大,把每个网格点的残差都吃掉了。但本系列前一篇已经说过:我们关心的不是网格内的贴合度,是外推到 m=24、m=32 之外的 α 预告值。样条在训练区赢来的每一分贴合度,都可能在外推区变成一份没有约束的任性。

所以裁决标准必须换成:留一外推误差。做法是同一套规则跑三遍:每次从八档里留出一档(优先留高端档,因为外推误差在高端最要命),用剩下七档拟合,再去预告留出档,记录误差。三个候选各跑完,比谁的外推误差小。这不是我们发明的招——生态学里 548 个物种的大规模模型选择就是这么做的,逐观测留一交叉验证(LOOCV)被证明比 WAIC 这类廉价替代品更可靠(PMID 32324930);交叉验证驱动的模型选择算法本身就是为了防”过拟合数据”这一件事而设计的(PMID 25505354)。

但有三个坑必须提前写进规则里。

坑一:留一验证在小样本上自己也会乐观。 方法学研究用多个临床数据集实打实比过:样本小时,leave-1-out 交叉验证给出的校正估计有偏,带重复的交叉验证和 bootstrap 表现更稳(PMID 24966219)。八档 m 是很小的数据,所以裁决不能只跑一轮 LOOCV 就拍板——须在仿真母体上做重复生成(不同随机种子重抽网格),报外推误差的分布而不是单点值。哪条曲线的误差分布右尾更短,谁才配当默认形式。

坑二:交叉验证的切法必须贴着外推任务来。 空气质量预测领域有现成教训:随机切分的十折交叉验证把 RMSE 报成 2.10,换成按空间距离留出远处站点的切法,RMSE 立刻涨到 3.11——随机切分高估了外推能力约 48%(PMID 33122961)。对应到我们的网格:如果随机留档,等于测”插值”;要想测”外推”,必须只留高端档(m=24、32),让模型用低中档去够高端。这是本裁决与一般模型选择最大的不同点。

坑三:拐点检验先于形式选择。 幂律和有理式的关键分歧在 m=12/24 附近有没有拐点。判法用残差符号游程检验:三候选各自拟合后,把残差按 m 排序看符号序列——若幂律拟合的残差符号出现”先正后负再正”的系统性游程,就说明数据里有幂律装不下的弯曲,拐点存在,裁决应倒向有理式或样条。游程检验是每个统计软件都有的老工具,成本为零。

三条规则合起来,裁决流程就定型了:同一网格数据、三形式、高端留档的重复 LOOCV、残差游程检验拐点、误差分布右尾定胜负。胜者成为 s*(m) 的默认参数化形式,写进后续所有外推预算(包括前一篇的几何因子传播带宽表)的底层假设;败者不删,留作敏感性分析的对照组。

对整条”气功外气”测量线来说,这一步看着枯燥,其实是地基:曲线形式选错,后面所有 α 预告、所有带宽、所有”m=32 档 0.02 传到 0.3″的算术都建立在一个错误形状上。测量学文献反复验证的那句话在这里同样成立——模型选择的正确做法不是看谁拟合得漂亮,是看谁在没见过的数据上错得少(PMID 32324930、25505354)。外推世界里,谦虚的曲线比自信的曲线活得久。

证据与局限

已验证: 留一交叉验证作为模型选择裁判在大规模真实数据上可行且优于 WAIC 等替代品(PMID 32324930);小样本下 LOOCV 存在乐观偏差、重复 CV 与 bootstrap 更稳(PMID 24966219);随机切分交叉验证会系统性高估外推能力,须用与外推任务同构的切法(PMID 33122961);CV 驱动的模型选择可有效抑制过拟合(PMID 25505354)。

部分验证: 高端留档切法在空间外推领域有直接先例(空间 CV),移植到”高档 m 留出”属同类结构,但未在本系列网格上实测。

待验证: 三候选(幂律/有理式/样条)在同一八档网格上的外推误差头对头;m=12/24 拐点的残差符号游程检验;重复生成下误差分布右尾比较。全部为本系列提出的仿真裁决方案,结果未出。

相关阅读

证据等级: C
参考文献: PMID 32324930, PMID 24966219, PMID 25505354, PMID 33122961

L1 网络状态:该主题尚未进入 L1 对应网络。最近邻边 E-061(呼吸节律 0.1 Hz 谱峰 ↔ Mayer 波/自主神经谱,层级 L2+,状态 open)——本文为纯测量学方法学课题,不依赖任何传统对应边,故按规则取 C 级。


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。

Power law, rational function or spline: Model selection for the parameterization of the s*(m) curve

Evidence level: C

The shape of a fitted curve is not drawn; it is selected. Selecting the wrong shape leads to systematic error when extrapolating to high m values, and this error is often made with confidence.

1. Introduction: The s*(m) Framework and Model Selection

The study of the relationship between spatial frequency $s$ and magnitude $m$—denoted as $s^*(m)$—is central to understanding the underlying architecture of complex systems, from neural networks to urban planning. This relationship is often modeled using parametric curves, but the choice of model significantly impacts the interpretation of results.

Three primary models are frequently considered:
1. Power Law: $s^(m) \propto m^\alpha$
2.
Rational Function: $s^(m) = \frac{P(m)}{Q(m)}$
3. Spline: A piecewise polynomial interpolation.

Each model offers a different structural analogy for how the system organizes itself. The power law implies scale invariance; the rational function suggests a saturation or asymptotic behavior; and the spline allows for arbitrary local variations. However, these are not merely mathematical conveniences; they represent competing hypotheses about the frequency epistemology of the system.

2. The Power Law Model: Scale Invariance and Its Limits

The power law model posits that $s^(m)$ follows a simple scaling relationship:
$$ s^
(m) = A \cdot m^\alpha $$
where $A$ is a scaling constant and $\alpha$ is the exponent.

Traditional Consensus: In many natural and social systems, power laws are observed as a signature of self-organized criticality or fractal geometry.

Evidence Level: A (Strong empirical support in physics and biology; mixed in social sciences).

Limitations:
While elegant, the power law model often fails when extrapolating beyond the observed range of $m$. If the underlying process is not truly scale-invariant, the power law fit may appear good over a limited range but diverge significantly at higher $m$ values. This divergence is not a minor deviation; it represents a systematic error in the model’s assumptions.

3. The Rational Function Model: Saturation and Asymptotes

The rational function model proposes that $s^(m)$ is the ratio of two polynomials:
$$ s^
(m) = \frac{P(m)}{Q(m)} $$
where $P(m)$ and $Q(m)$ are polynomials of degrees $n$ and $d$, respectively.

Traditional Consensus: Rational functions are useful for modeling systems with saturation effects, such as enzyme kinetics or traffic flow.

Evidence Level: B (Moderate support; often fits data better than power laws in constrained environments).

Limitations:
The rational function model can capture asymptotic behavior, which the power law cannot. However, it requires more parameters and may overfit if the degrees of $P$ and $Q$ are not justified by the underlying pattern and timing (Qi) of the system.

4. The Spline Model: Flexibility and Overfitting

The spline model uses piecewise polynomials to interpolate the data points, allowing for arbitrary curvature.

Traditional Consensus: Splines are flexible and can fit any set of points exactly.

Evidence Level: C (Weak; high risk of overfitting).

Limitations:
While splines can fit the observed data perfectly, they lack predictive power for unobserved $m$ values. Extrapolating a spline often leads to unrealistic behavior, such as oscillations or unbounded growth, which contradicts the physical or logical constraints of the system.

5. Model Selection Criteria: Beyond Goodness-of-Fit

Choosing the correct model for $s^*(m)$ requires more than just minimizing the sum of squared errors. We must consider:

  • Extrapolation Behavior: How does the model behave at high $m$?
  • Parameter Parsimony: Is the added complexity justified by the evidence?
  • Structural Analogy: Does the model reflect the underlying frequency epistemology of the system?

Table 1: Comparison of Models for $s^*(m)$

Model Equation Extrapolation Behavior Evidence Level Structural Analogy
Power Law $s^*(m) = A \cdot m^\alpha$ Unbounded growth or decay; no saturation A Scale invariance, fractal geometry
Rational Function $s^*(m) = \frac{P(m)}{Q(m)}$ Asymptotic convergence; saturation possible B Saturation, finite capacity
Spline Piecewise polynomial Arbitrary; often oscillatory or unbounded C Local variation, no global structure

6. Conclusion: The Importance of Model Selection

The choice of model for $s^*(m)$ is not trivial. A power law fit may be statistically adequate over a limited range but fail catastrophically when extrapolated. A rational function may capture saturation but require unjustified complexity. A spline may fit perfectly but predict nonsense.

Therefore, model selection must be guided by both empirical evidence and theoretical plausibility. The frequency epistemology of the system—its underlying pattern and timing—must inform the choice of model, not just the fit to observed data.

In summary, the shape of $s^*(m)$ is not drawn; it is selected. And selecting the wrong shape leads to systematic error, often with great confidence.

Body (1000–1500 words)

Using the same set of eight-grid m simulation mesh data, all three candidate curves—power law, rational function, and spline—can “fit”: the power law is smooth at both ends and allows extrapolation; the rational function can produce asymptotes and inflection points; the spline adheres most tightly to the data but its extrapolation behavior is entirely determined by endpoint inertia. In terms of goodness-of-fit, the spline almost always wins—its many parameters and high degrees of freedom absorb residuals at every mesh point. Yet, as the previous article in this series noted, what matters is not fit within the mesh but extrapolation beyond m=24 and m=32 to predict α values. Every point of fit the spline gains in the training region may become unconstrained arbitrariness in the extrapolation region.

Therefore, the decision criterion must shift to leave-one-out extrapolation error. The procedure applies the same rules three times: each run leaves out one of the eight grids (prioritizing high-end grids, because extrapolation error is most severe there), fits the remaining seven, predicts the left-out grid, and records the error. After running all three candidates, we compare whose extrapolation error is smallest. This is not an invention of ours: ecology has used leave-one-out cross-validation (LOOCV) for massive model selection across 548 species, and LOOCV has been shown to be more reliable than cheaper alternatives like WAIC (PMID 32324930). Model selection algorithms driven by cross-validation are themselves designed to prevent “overfitting data” (PMID 25505354).

However, three pitfalls must be built into the rules in advance.

Pitfall 1: LOOCV is inherently optimistic on small samples. Methodological studies using multiple clinical datasets have shown that with few samples, leave-one-out cross-validation yields biased corrected estimates; repeated cross-validation and bootstrap methods perform more stably (PMID 24966219). Eight grids constitute a very small dataset, so the decision cannot rest on a single LOOCV run. Instead, we must generate repeats from the simulation population (re-drawing the mesh with different random seeds) and report the distribution of extrapolation errors rather than a single point estimate. The curve whose error distribution has the shorter right tail earns default status.

Pitfall 2: The cross-validation split must align with the extrapolation task. Air quality prediction offers a clear lesson: random ten-fold cross-validation reported RMSE = 2.10, but when the split respected spatial distance—leaving out distant stations—the RMSE jumped to 3.11; random splitting overestimated extrapolation ability by about 48% (PMID 33122961). Translating to our mesh: random leave-one-out measures “interpolation”; to measure “extrapolation,” we must only leave out high-end grids (m=24, 32), forcing the model to reach for higher values using low and mid grids. This is the key distinction between our decision and general model selection.

Pitfall 3: Inflection point testing precedes form selection. The critical divergence between power law and rational function lies in whether an inflection point exists near m=12/24. The test uses residual sign run tests: after each candidate fits the data, we sort residuals by m and examine the sequence of signs. If the power-law fit shows a systematic run such as “positive, then negative, then positive,” it indicates curvature the power law cannot capture; an inflection point exists, and the decision should favor the rational function or spline. Run tests are a standard statistical tool available in every software package, costing nothing.

Combining these three rules yields the finalized decision workflow: same mesh data, three forms, repeated LOOCV with high-end leave-outs, residual run test for inflection points, and right-tail comparison of error distributions to determine the winner. The winner becomes the default parameterization s*(m) and underpins all subsequent extrapolation budgets—including the geometric factor bandwidth table from the previous article. Losers are not discarded; they serve as controls for sensitivity analysis.

For the entire “qigong external qi” measurement line, this step may appear dry, yet it is foundational: choosing the wrong curve form invalidates all α predictions, all bandwidths, and all arithmetic such as “m=32 grid 0.02 propagates to 0.3.” The measurement literature repeatedly confirms that correct model selection is not about who fits beautifully but about who errs least on unseen data (PMID 32324930, 25505354). In the extrapolation world, humble curves outlive confident ones.

Evidence and Limitations

Verified: Leave-one-out cross-validation serves as a robust model selection criterion on large-scale real-world data, outperforming alternatives such as WAIC (PMID 32324930); under small-sample conditions, leave-one-out cross-validation exhibits optimistic bias, making repeated cross-validation and bootstrap methods more stable (PMID 24966219); random splitting cross-validation systematically overestimates extrapolation capacity and requires splits isomorphic to the extrapolation task (PMID 33122961); cross-validation-driven model selection effectively suppresses overfitting (PMID 25505354).

Partially Verified: High-end leave-one-out splitting has direct precedent in spatial extrapolation (spatial cross-validation); transposing this to “high-end m left-out” constitutes a structurally analogous approach, though it has not been empirically tested on the current series of grids.

Pending Verification: Head-to-head extrapolation error comparison among three candidates (power law, rational function, spline) on the same eight-tier grid; residual sign run tests at m=12/24 inflection points; right-tail distribution comparison under repeated generation. These are all simulation-based adjudication schemes proposed in this series; results are not yet available.

Related Reading

Evidence level: C
References: PMID 32324930, PMID 24966219, PMID 25505354, PMID 33122961

L1 network status: This topic has not yet entered the corresponding L1 network. The nearest neighbor edge E-061 (respiratory rhythm 0.1 Hz spectral peak ↔ Mayer wave / autonomic spectrum, level L2+, status open) — this article is a pure measurement methodology topic and does not rely on any traditional corresponding edge, hence it is assigned level C according to the rules.


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 *