5D Library › 双人相位耦合、双线性相容张量与时空网格博弈 · English

第四章 分离的停时与「高危窗口」

锁相会被打断

第二、三章的锁相是确定性的:一旦锁住,永远锁住。真实的节律受到无数小扰动——生病、出差、家事——相位差在锁定点附近随机抖动,偶尔被推过一道势垒,滑掉整整一圈。这一刻称为一次「相位滑移」。在类比的意义上,它对应锁定关系的一次断裂。本章问两件事:这一刻在时间上如何分布;它会不会集中在某个特定的「窗口」里。

推导

带噪声的阿德勒方程与势垒。 dϕ=(Δω−Ksin⁡ϕ)dt+σdW,K=kAB+kBA.d\phi = \big(\Delta\omega – K\sin\phi\big)\,dt + \sigma\, dW, \qquad K = k_{AB} + k_{BA}. 漂移项是倾斜周期势 V(ϕ)=−Δωϕ−Kcos⁡ϕV(\phi) = -\Delta\omega\,\phi – K\cos\phi 的负梯度。锁定点 ϕ*=arcsin⁡(Δω/K)\phi^{*} = \arcsin(\Delta\omega/K) 在势阱底部,通往下一个势阱的势垒高度为 ΔV=2Kcos⁡ϕ*−Δω(π−2ϕ*).\Delta V = 2K\cos\phi^{*} – \Delta\omega\,(\pi – 2\phi^{*}). 记 D=σ2/2D = \sigma^2/2。克莱默斯公式给出 ΔV≫D\Delta V \gg D 时的平均等待时间 T≈2πKcos⁡ϕ*eΔV/DT \approx \frac{2\pi}{K\cos\phi^{*}}\, e^{\Delta V/D} (Kramers 1940);一维扩散的平均首达时间还有一个精确的二重积分 (Hänggi et al. 1990)。

第七段程序用三种方法计算第一次滑移的平均时间:克莱默斯近似、精确积分与 300 次随机模拟。

推导

# Chapter 4. How long a locked pair stays locked when life is noisy.
#   dphi = (dw - K sin phi) dt + sigma dW      (K = kAB + kBA)
# The locked phase sits in a well of the tilted potential
#   V(phi) = -dw phi - K cos phi,
# at phi* = asin(dw / K); the barrier towards the next well has height
#   dV = 2 K cos(phi*) - dw (pi - 2 phi*).
# A "slip" is the first time phi reaches phi* + 2 pi. Three estimates of
# the mean time to the first slip, with D = sigma^2 / 2:
#   Kramers   2 pi / (K cos phi*) * exp(dV / D)   (valid for dV >> D)
#   integral  the exact mean first-passage time of 1-D diffusion,
#             T = (1/D) int dy exp(V(y)/D) int_{-inf}^{y} exp(-V(z)/D) dz
#   simulated Euler-Maruyama, step 0.01, 300 runs
# K = 1, dw = 0.5, in model time units.
import math
import random

K, DW = 1.0, 0.5
PHI = math.asin(DW / K)
DV = 2 * K * math.cos(PHI) - DW * (math.pi - 2 * PHI)

def V(x):
    return -DW * x - K * math.cos(x)

def integral_T(d, n=1500):
    # (lower limit truncated at PHI - 2*pi, where exp(-V/D) is negligible)
    lo, hi = PHI - 2 * math.pi, PHI + 2 * math.pi
    h = (hi - lo) / n
    xs = [lo + i * h for i in range(n + 1)]
    inner, acc, total = 0.0, [], 0.0
    for x in xs:
        inner += math.exp(-V(x) / d) * h
        acc.append(inner)
    for x, a in zip(xs, acc):
        if x >= PHI:
            total += math.exp(V(x) / d) * a * h
    return total / d

def simulate(d, runs, rng, dt=0.01):
    s = math.sqrt(2 * d * dt)
    times = []
    for _ in range(runs):
        x, t = PHI, 0.0
        while x < PHI + 2 * math.pi:
            x += (DW - K * math.sin(x)) * dt + s * rng.gauss(0, 1)
            t += dt
        times.append(t)
    return times

if __name__ == "__main__":
    rng = random.Random(62007)
    print("   barrier dV = {:.4f}".format(DV))
    print("   dV/D    Kramers   integral   simulated   CV    alive at mean")
    for ratio in (2.0, 3.0, 4.0):
        d = DV / ratio
        kr = 2 * math.pi / (K * math.cos(PHI)) * math.exp(ratio)
        ts = simulate(d, 300, rng)
        m = sum(ts) / len(ts)
        sd = (sum((t - m) ** 2 for t in ts) / (len(ts) - 1)) ** 0.5
        alive = sum(t > m for t in ts) / len(ts)
        print("   {:3.1f}    {:7.1f}   {:8.1f}   {:9.1f}   {:4.2f}   {:5.3f}"
              .format(ratio, kr, integral_T(d), m, sd / m, alive))
    print("\nReading: the simulated mean agrees with the exact integral")
    print("within sampling error (300 runs: about 6 %). Kramers' formula is")
    print("the large-barrier limit and approaches both as dV/D grows. The")
    print("waiting time is close to exponential (coefficient of variation")
    print("near 1, about 1/e still locked at the mean): a constant hazard.")
    print("Noise alone gives no special 'danger window'.")
#    barrier dV = 0.6849
#    dV/D    Kramers   integral   simulated   CV    alive at mean
#    2.0       53.6       65.3        66.1   0.92   0.363
#    3.0      145.7      168.7       160.8   0.95   0.387
#    4.0      396.1      441.0       469.4   0.98   0.367
#
# Reading: the simulated mean agrees with the exact integral
# within sampling error (300 runs: about 6 %). Kramers' formula is
# the large-barrier limit and approaches both as dV/D grows. The
# waiting time is close to exponential (coefficient of variation
# near 1, about 1/e still locked at the mean): a constant hazard.
# Noise alone gives no special 'danger window'.

模拟的平均值与精确积分在抽样误差之内一致;克莱默斯公式是大势垒极限,ΔV/D\Delta V/D 从 2.0 增到 4.0 时,它与精确值之比从 53.6 对 65.3 逐步接近(396.1 对 441.0)。平均等待时间随 ΔV/D\Delta V/D 指数增长:势垒每高出一个 DD,等待时间大约乘以 ee。

更重要的是最后两列。等待时间的变异系数接近 1(0.92 到 0.98),到平均时间时仍保持锁定的比例接近 1/e1/e(0.363、0.387、0.367)——这正是指数分布的特征。指数分布的风险率是常数:已经锁定了多久,对下一刻是否滑移没有影响。只有噪声时,不存在任何特别危险的时间窗口。

加上每年一次的冲击

「流年」一类的说法设想的是另一种机制:每年某个固定的时刻,有一个外来的冲击。第八段程序在同一个系统上加入这样的冲击——每年开始时把相位向势垒推 AA——并记录每次滑移发生在一年中的什么时候。

推导

# Chapter 4. A yearly shock on top of the noise.
# Same locked pair as p07 (K = 1, dw = 0.5, dV/D = 3, mean first slip about
# 165 time units). Let one "year" be 20 time units. At the start of every
# year the phase receives a kick of size A towards the barrier (an outside
# event of fixed size, at a fixed date). For 400 runs each, the program
# records the date of the first slip - the last time phi crossed the top
# of the barrier, pi - phi*, before going on to phi* + 2 pi - and reports
# the mean waiting time in years, the share of slips in the first twelfth
# of a year after a kick (1/12 = 0.083 if the date did not matter), and
# the hazard in that month relative to the rest of the year.
import math
import random

K, DW = 1.0, 0.5
PHI = math.asin(DW / K)
DV = 2 * K * math.cos(PHI) - DW * (math.pi - 2 * PHI)
YEAR, DT = 20.0, 0.01

def first_slip(a, d, rng):
    s = math.sqrt(2 * d * DT)
    steps_per_year = int(round(YEAR / DT))
    top = math.pi - PHI
    x, i, crossed = PHI, 0, 0.0
    while True:
        if a and i % steps_per_year == 0 and i > 0:
            x += a
        prev = x
        x += (DW - K * math.sin(x)) * DT + s * rng.gauss(0, 1)
        i += 1
        if prev < top <= x:
            crossed = i * DT
        if x >= PHI + 2 * math.pi:
            return crossed

if __name__ == "__main__":
    d = DV / 3.0
    rng = random.Random(62008)
    print("   kick A   mean wait (years)   share in month after kick"
          "   hazard ratio")
    for a in (0.0, 0.5, 1.0, 1.5):
        ts = [first_slip(a, d, rng) for _ in range(400)]
        mean_y = sum(ts) / len(ts) / YEAR
        frac = [(t / YEAR) % 1.0 for t in ts]
        early = sum(f < 1 / 12 for f in frac) / len(ts)
        ratio = (early / (1 / 12)) / ((1 - early) / (11 / 12))
        print("    {:3.1f}          {:6.2f}                  {:5.3f}"
              "                {:5.2f}".format(a, mean_y, early, ratio))
    print("\nReading: without kicks the date of a slip is spread evenly over")
    print("the year. A kick both shortens the wait and concentrates slips")
    print("in the month that follows it. The window is a property of the")
    print("assumed kick: the model puts it there, it does not discover it.")
#    kick A   mean wait (years)   share in month after kick   hazard ratio
#     0.0            7.82                  0.090                 1.09
#     0.5            6.32                  0.177                 2.37
#     1.0            4.05                  0.323                 5.24
#     1.5            2.35                  0.505                11.22
#
# Reading: without kicks the date of a slip is spread evenly over
# the year. A kick both shortens the wait and concentrates slips
# in the month that follows it. The window is a property of the
# assumed kick: the model puts it there, it does not discover it.

没有冲击时(A=0.0A = 0.0),滑移在一年中均匀分布:落在冲击后第一个月的比例为 0.090,接近 1/121/12,风险比 1.09。冲击越强,平均等待越短(从 7.82 年缩到 2.35 年),滑移越集中在冲击后的那个月:A=1.0A = 1.0 时该月占 0.323,风险比 5.24;A=1.5A = 1.5 时占 0.505,风险比 11.22。

第一次滑移在一年中的月份分布:没有冲击(灰)与每年冲击 A = 1.0(紫)。虚线是均匀分布的 1/12。

这个结果需要正确地读。窗口存在,是因为模型假设了一个在固定日期、固定大小的冲击;模型把窗口放了进去,并没有发现它。要说一个真实的「流年」窗口存在,必须先在数据里看到它。

检验一个窗口需要多少数据

第九段程序计算,在一个事先指定的月份里,要看见风险比为 hh 的集中,需要多少次带确切日期的分离;再说明事后挑选月份会发生什么。

推导

# Chapter 4. How many break-ups it takes to see a danger window.
# Claim to test: "slips are h times as likely in a pre-named month as in
# the rest of the year". Under no effect a slip falls in the named month
# with probability 1/12; under the claim with p1 = h / (h + 11). With n
# dated events the test counts those in the window and rejects at
# one-sided alpha = 0.05 (exact binomial). The program finds the smallest
# n with 80 % power, and the number of couples to follow for five years if
# 25 % of them separate in that time. It then shows what happens when the
# window is chosen AFTER looking: with no effect at all, picking the
# busiest of 12 months and testing it at 0.05 (20,000 simulated studies).
import math
import random

def tail(n, p, k):
    """P(X >= k) for X ~ Binomial(n, p)."""
    return sum(math.comb(n, j) * p ** j * (1 - p) ** (n - j)
               for j in range(k, n + 1))

def crit(n, p0=1 / 12, alpha=0.05):
    k = 0
    while tail(n, p0, k) > alpha:
        k += 1
    return k

def n_needed(h, power=0.8):
    p1 = h / (h + 11)
    n = 10
    while tail(n, p1, crit(n)) < power:
        n += 1
    return n

if __name__ == "__main__":
    print("   hazard ratio h   events needed   couples followed 5 years")
    for h in (1.5, 2.0, 3.0, 5.0):
        n = n_needed(h)
        print("        {:3.1f}            {:5d}              {:6,d}".format(
            h, n, math.ceil(n / 0.25)))
    rng = random.Random(62009)
    n, hits = 120, 0
    k = crit(n)
    for _ in range(20000):
        counts = [0] * 12
        for _ in range(n):
            counts[rng.randrange(12)] += 1
        hits += max(counts) >= k
    print("\n   no effect, 120 events, busiest month tested at 0.05:")
    print("   false-positive rate {:.3f}".format(hits / 20000))
    print("\nReading: a hazard ratio of 2 in a named month needs about 130")
    print("dated separations; a ratio of 1.5 needs over 400, from more than")
    print("1,600 couples. Choosing the month after seeing the data turns a")
    print("5 % test into one that 'finds' a window in over 40 % of studies.")
#    hazard ratio h   events needed   couples followed 5 years
#         1.5              409               1,636
#         2.0              130                 520
#         3.0               47                 188
#         5.0               17                  68
#
#    no effect, 120 events, busiest month tested at 0.05:
#    false-positive rate 0.432
#
# Reading: a hazard ratio of 2 in a named month needs about 130
# dated separations; a ratio of 1.5 needs over 400, from more than
# 1,600 couples. Choosing the month after seeing the data turns a
# 5 % test into one that 'finds' a window in over 40 % of studies.

风险比为 5 时,17 次带日期的分离就够;风险比为 2 时需要 130 次;风险比为 1.5 时需要 409 次,如果五年内有四分之一的伴侣分开,这意味着要追踪 1,636 对伴侣。下半段是更常见的错误:在完全没有效应的数据里,看过 120 次分离的日期之后挑出最拥挤的那个月再做检验,5% 的检验在 0.432 的研究中「发现」了窗口。

在事先指定的月份中看见风险集中所需的带日期分离数(单侧 0.05、把握度 80%)。纵轴为对数刻度。

注

对「感情决裂高危时间窗」功能的含义。 一个声称能预言分手月份的系统,至少需要三样东西才能被检验:事先写定的窗口规则(不能看过结果再定),每个人各自的确切分离日期,以及足够的数量——效应中等时是几百次分离。没有这三样,任何「应验」的例子都无法与第九段程序下半段的巧合区分开。

假说 · 尚未验证

作者的假说(尚未验证): 用任何一种事先公布的流年窗口规则,对带确切日期的分离数据做检验,窗口月份的风险比与 1 没有可区分的差别。什么结果会证明它错:在预先登记、由独立分析者执行的检验中,窗口月份的风险比在两个独立样本中都显著大于 1。

小结

带噪声的锁相系统中,第一次滑移的等待时间接近指数分布,风险率恒定:噪声本身不产生危险窗口。窗口只在模型假设了固定日期的外部冲击时出现,冲击越强越集中,风险比可以从 1.09 升到 11.22——但这是假设的结果,不是发现。要在真实数据中检验一个窗口,必须事先指定月份并收集大量带确切日期的分离;事后挑选月份会让 5% 的检验在 0.432 的研究中「发现」并不存在的窗口。

参考文献

Hänggi, Peter, Peter Talkner, and Michal Borkovec. 1990. “Reaction-Rate Theory: Fifty Years after Kramers.” Reviews of Modern Physics 62 (2): 251–341.
Kramers, H. A. 1940. “Brownian Motion in a Field of Force and the Diffusion Model of Chemical Reactions.” Physica 7 (4): 284–304.

Fan, Y. (2026). Dyadic Phase Coupling, Bilinear Compatibility Tensors and Spatiotemporal Grid Games: Asymmetric Kuramoto Locking, Stopping-Time Separation Hazards and Magic-Square Permutation Operators [双人相位耦合、双线性相容张量与时空网格博弈] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090172 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).