5D Library › 五节点循环网络动力学、黄经历表与多尺度振子扰动 · English

第八章 四个问题与两项纯软件的预注册检验

面对一个「五节点」或「节气」计算之前的四个问题

本章把全书收成四个问题,可以向任何关于循环网络或历表的计算提出——包括本书自己的计算。

一、这个结果依赖哪一个约定? 节气是时刻,日期是约定:时区与经度一变,约十分之一的节气就换一天(第一章)。平气与定气是两种合法的划分,最多相差两天多(第六、七章)。先问用的是哪一种。

二、这个性质来自结构,还是来自参数? 循环对称直接给出全部特征值、行波与黄金比;但守恒量只属于零和环,Hopf 振荡需要自身饱和,瞬态放大需要不均的权重(第二至四章)。把「结构使然」与「参数使然」分开。

三、系统是被锁定的,还是自由的? 受周期强迫的阻尼系统,其极值时刻跟着强迫走,远期与近期一样可预报;自由振荡的时刻误差随距离线性增长(第四、五章)。在问「能预报多远」之前,先问是哪一类系统。

四、数据能分辨什么? 单节点的记录能精确确定两个参数组合,却几乎分辨不了第三个;那个方向可以由对称性事先算出(第五章)。报告参数时,报告它们的相关,而不只是各自的误差棒。

两项预注册检验

两项检验都只用软件与公开的历表,不涉及任何人。它们在打开参考历表或新的合成数据之前写成。

假说 · 尚未验证

检验一:历表核对。 声称:第一章所用的低精度太阳理论,对 2027 至 2036 年的全部 240 个节气时刻,与一份事先指定的高精度参考历表相差不超过 15 分钟。参考历表须在比较之前指定(例如国家天文年历或公开的数值历表),并写明其时间尺度(TT 或 UTC)及转换方法。这是全体比较,不是抽样,因此规则是确定的:最大误差不超过 15 分钟为「支持」;任一时刻超过 30 分钟为「否定」;其余为「部分支持」。

检验二:合成环的盲预报。 声称:对受强迫的阻尼环,100 年后最大能量峰的日期,等于最后一个已观测强迫周期内最大峰的日期加 100 年,误差不超过 0.25 年。在 40 个随机参数的受强迫阻尼环上检验,以 40 个自由振荡环作阴性对照。规则:受强迫组成功率的 Wilson 95% 下限不低于 0.8,且对照组的上限低于 0.5,才判为「支持」。

推导

# Chapter 8. p16: Two pre-registered, software-only checks.
# A. Ephemeris check. Claim: the low-precision solar theory of p03 dates
#    the 240 solar-term instants of 2027-2036 to within 15 minutes of a
#    reference ephemeris named in advance. The comparison is exact, not a
#    sample, so the rule is deterministic: SUPPORTED if the largest error
#    is at most 15 min; REFUTED if any exceeds 30 min; otherwise
#    PARTLY. Shown here on a stand-in reference (the theory plus a
#    simulated error: an annual wave of 2 min and noise of sd 1 min) -
#    the real reference is to be fetched only after this rule is fixed.
# B. Blind peak-date check on synthetic rings. Claim: for a forced,
#    damped ring, the date of the largest energy peak 100 years ahead
#    equals the date of the largest peak in the last observed 20-year
#    forcing cycle plus 100 years, to within 0.25 yr. 40 rings with
#    random parameters (c1 in [0.8, 1.2], c2 in [0.3, 0.7], mu = mu_c +
#    U(0.05, 0.2)), forced as in p08, observed with noise for 40 years.
#    Negative control: 40 free rings (no forcing, mu = mu_c - U(0.05,
#    0.2)), started off equilibrium. Rule: SUPPORTED if the Wilson 95 %
#    lower bound of the forced success rate is at least 0.8 AND the
#    control's upper bound is below 0.5.
import math
import random

SCALE = 10.0

def ring_series(rng, forced):
    c1, c2 = rng.uniform(0.8, 1.2), rng.uniform(0.3, 0.7)
    muc = c1 * math.cos(math.radians(72)) + c2 * math.cos(math.radians(36))
    mu = muc + (1 if forced else -1) * rng.uniform(0.05, 0.2)
    def f(u, t):
        m = max(u)
        e = [math.exp(a - m) for a in u]
        s = sum(e)
        x = [a / s for a in e]
        ax = [c1 * x[(i - 1) % 5] - c2 * x[(i - 2) % 5] - mu * x[i]
              for i in range(5)]
        mm = sum(p * q for p, q in zip(ax, x))
        g = [SCALE * (a - mm) for a in ax]
        if forced:
            w = 0.05 * ((1.0 if int(t // 10) % 2 == 0 else -1.0)
                        + math.sin(2 * math.pi * t))
            g[0] += w
            g[2] -= w
        return g
    u = [0.3 * math.cos(2 * math.pi * i / 5 + rng.uniform(0, 6))
         for i in range(5)]
    h, t, out = 0.01, 0.0, []
    while t < 145.0:
        k1 = f(u, t)
        k2 = f([a + h / 2 * b for a, b in zip(u, k1)], t + h / 2)
        k3 = f([a + h / 2 * b for a, b in zip(u, k2)], t + h / 2)
        k4 = f([a + h * b for a, b in zip(u, k3)], t + h)
        u = [a + h / 6 * (b + 2 * c + 2 * d + e)
             for a, b, c, d, e in zip(u, k1, k2, k3, k4)]
        t += h
        m = max(u)
        e = [math.exp(a - m) for a in u]
        s = sum(e)
        out.append((t, sum((a / s - 0.2) ** 2 for a in e)))
    return out

def wilson(k, n, z=1.96):
    p = k / n
    c = (p + z * z / (2 * n)) / (1 + z * z / n)
    hw = (z * math.sqrt(p * (1 - p) / n + z * z / (4 * n * n))
          / (1 + z * z / n))
    return max(c - hw, 0.0) + 0.0, min(c + hw, 1.0)

if __name__ == "__main__":
    rng = random.Random(60016)
    errs = [2.0 * math.sin(2 * math.pi * k / 24) + rng.gauss(0, 1.0)
            for k in range(240)]
    worst = max(abs(e) for e in errs)
    verdict = ("SUPPORTED" if worst <= 15 else
               "REFUTED" if worst > 30 else "PARTLY")
    print("A. stand-in reference: largest error {:.2f} min, mean {:+.2f},"
          " rms {:.2f}".format(worst, sum(errs) / 240,
                               math.sqrt(sum(e * e for e in errs) / 240)))
    print("   rule: {}".format(verdict))
    print("\nB. blind peak dates, 40 rings each")
    for label, forced in (("forced, damped", True),
                          ("free (control)", False)):
        hits, offs = 0, []
        for _ in range(40):
            s = ring_series(rng, forced)
            obs = [(t, e * math.exp(rng.gauss(0, 0.05))) for t, e in s
                   if 20 <= t < 40]
            t_obs = max(obs, key=lambda r: r[1])[0]
            fut = [(t, e) for t, e in s if 120 <= t < 140]
            t_fut = max(fut, key=lambda r: r[1])[0]
            off = t_fut - (t_obs + 100)
            hits += abs(off) <= 0.25
            if abs(off) > 0.25:
                offs.append(off)
        lo, hi = wilson(hits, 40)
        print("   {:<16} {:2d}/40 within 0.25 yr; Wilson 95 % [{:.3f},"
              " {:.3f}]".format(label, hits, lo, hi))
        if forced:
            f_lo = lo
            print("   misses (yr): {}".format(
                ", ".join("{:+.1f}".format(o) for o in sorted(offs))))
        else:
            c_hi = hi
    ok = f_lo >= 0.8 and c_hi < 0.5
    print("   rule: {}".format("SUPPORTED" if ok else "NOT SUPPORTED"))
    print("\nReading: the ephemeris check has no sampling error - one")
    print("instant beyond 30 minutes refutes the claim. The blind test")
    print("separates the two kinds of system of Chapter 5: free peaks")
    print("drift and never hit. But the forced rule as written fails its")
    print("own bar: most misses are not drift but a swap - noise makes")
    print("the cycle's second-highest peak look highest. The rule, not")
    print("the entrainment, is at fault, and it is reported as failed;")
    print("a revised rule must be registered afresh on new rings.")
# A. stand-in reference: largest error 3.95 min, mean +0.07, rms 1.79
#    rule: SUPPORTED
#
# B. blind peak dates, 40 rings each
#    forced, damped   33/40 within 0.25 yr; Wilson 95 % [0.680, 0.913]
#    misses (yr): -10.5, -10.4, -10.4, +9.5, +9.7, +10.4, +10.5
#    free (control)    0/40 within 0.25 yr; Wilson 95 % [0.000, 0.088]
#    rule: NOT SUPPORTED
#
# Reading: the ephemeris check has no sampling error - one
# instant beyond 30 minutes refutes the claim. The blind test
# separates the two kinds of system of Chapter 5: free peaks
# drift and never hit. But the forced rule as written fails its
# own bar: most misses are not drift but a swap - noise makes
# the cycle's second-highest peak look highest. The rule, not
# the entrainment, is at fault, and it is reported as failed;
# a revised rule must be registered afresh on new rings.

检验一的规则在一份替身参考上演示:替身由理论值加上模拟误差构成,最大误差 3.95 分钟,判为「支持」。真正的参考历表在这条规则写定之后才取用;演示只说明规则如何运作,不说明结果会是什么。

检验二给出了一个否定结果,本书如实报告。阴性对照按预期失败:40 个自由环无一命中(0/40)。但受强迫组只有 33/40 命中,Wilson 下限 0.680,低于事先定下的 0.8,所以规则判为「不支持」。七次失误的偏移都在 ±10 年左右:不是漂移,而是换位——一个强迫周期里有两个高度相近的峰(第四章 p08 的 4.130 与 3.788),观测噪声让第二高的峰看上去最高。

这说明的是规则的缺陷,而不是锁定本身不成立:峰的时刻确实被强迫锁定,但「取最高峰」这个操作对噪声不稳健。按照预注册的原则,这个结果按「不支持」报告;修订后的规则(例如同时登记一个周期内的两个峰)必须在新的合成环上重新登记与检验,不能在已经看过的 40 个环上改写判定。

注

结果意味着什么、不意味着什么。 检验一若被支持,说明这套低精度理论足以给出节气的时刻,但仍不足以为午夜附近约 2% 的节气定日(第一章)。检验二的否定只针对这一条规则;它提醒的是,一个在理论上成立的性质(锁定),落实为一条操作规则时可能失败。两项检验都不涉及任何真实的「五节点」系统。

本书没有做的事

本书没有把五节点环与任何医学、生理、命理或社会系统对应,没有用模型对真实世界作出预报,也没有评价任何以「五行」为名的实践。书中的环与参数都是数学设定,所有时间序列都是程序生成的合成数据;与真实世界相关的只有太阳的位置与历法的数字,它们都可以用公开历表独立核对。

Fan, Y. (2026). Five-Node Cyclic Network Dynamics, Solar-Longitude Ephemeris and Multi-Scale Oscillator Perturbations: Rock-Paper-Scissors-Lizard-Spock Lotka-Volterra Graphs, Equation-of-Time Keplerian Mechanics and Jacobian Sensitivity Bounds [五节点循环网络动力学、黄经历表与多尺度振子扰动] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090168 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).