5D Library › 关系断裂生存分析、断联期记忆衰减与复联博弈窗口 · English

第二章 以两种速度消退的两种记忆

模型

跟踪人们分手后几周情绪的研究发现,悲伤与愤怒通常在几周内缓和,人与人之间的差异很大 (Sbarra and Emery 2005)。关系积极的一面,也就是日后以怀旧形式回来的共同回忆 (Wildschut et al. 2006),似乎消退得更慢。把这一点写下来,最简单的办法是两个指数函数。防御性的负面情绪 N(t)=Ae−t/τnN(t) = A e^{-t/\tau_n},τn\tau_n 约 14 至 21 天;积极的依恋记忆 P(t)=Be−t/τpP(t) = B e^{-t/\tau_p},τp\tau_p 约 60 至 90 天。对前任的净牵引力是

f(t)=Be−t/τp−Ae−t/τn.f(t) = B e^{-t/\tau_p} – A e^{-t/\tau_n}.

令 f′(t)=0f'(t) = 0,得到峰值出现的那一天

t*=τnτpτp−τnln⁡AτpBτn,t^* = \frac{\tau_n \tau_p}{\tau_p – \tau_n} \ln\frac{A \tau_p}{B \tau_n},

只有当 Aτp>BτnA \tau_p > B \tau_n 时它才存在。如果负面一开始比正面大(A>BA > B),净牵引力在第 t0t_0 天才第一次转正,其中

t0=τnτpτp−τnln⁡AB.t_0 = \frac{\tau_n \tau_p}{\tau_p – \tau_n} \ln\frac{A}{B}.

推导

# Chapter 2. p03: The two-exponential model and its peak. After a
# rupture, negative defensive arousal N(t) = A exp(-t / tn) fades fast
# and positive attachment memory P(t) = B exp(-t / tp) fades slowly.
# Net pull f(t) = P(t) - N(t). Setting f'(t) = 0 gives the peak
#   t* = tn tp / (tp - tn) * ln(A tp / (B tn)),
# which exists only when A tp > B tn. Net pull turns positive at
#   t0 = tn tp / (tp - tn) * ln(A / B)   (when A > B).
# Grid: tn 14 to 21 days, tp 60 to 90 days, A / B = 1.5, B = 1.
from math import exp, log

def peak(tn, tp, a, b=1.0):
    if a * tp <= b * tn:
        return None
    return tn * tp / (tp - tn) * log(a * tp / (b * tn))

def pull(t, tn, tp, a, b=1.0):
    return b * exp(-t / tp) - a * exp(-t / tn)

if __name__ == "__main__":
    a = 1.5
    print("   tn   tp     t0 (days)   t* (days)   f(t*)   P(t*)")
    for tn in (14, 17.5, 21):
        for tp in (60, 75, 90):
            t0 = tn * tp / (tp - tn) * log(a)
            ts = peak(tn, tp, a)
            print("   {:4.1f} {:3d}   {:9.1f}   {:9.1f}   {:5.3f}   {:5.3f}"
                  .format(tn, tp, t0, ts, pull(ts, tn, tp, a),
                          exp(-ts / tp)))
    grid = [peak(tn, tp, a) for tn in (14, 17.5, 21) for tp in (60, 75, 90)]
    print("   t* across the grid: {:.1f} to {:.1f} days".format(
        min(grid), max(grid)))
#    tn   tp     t0 (days)   t* (days)   f(t*)   P(t*)
#    14.0  60         7.4        34.0   0.435   0.568
#    14.0  75         7.0        35.9   0.504   0.620
#    14.0  90         6.7        37.6   0.556   0.659
#    17.5  60        10.0        40.5   0.361   0.510
#    17.5  75         9.3        42.5   0.435   0.568
#    17.5  90         8.8        44.4   0.492   0.611
#    21.0  60        13.1        47.0   0.297   0.457
#    21.0  75        11.8        49.0   0.375   0.521
#    21.0  90        11.1        51.0   0.435   0.568
#    t* across the grid: 34.0 to 51.0 days

当 A/B=1.5A/B = 1.5、τn=17.5\tau_n = 17.5、τp=75\tau_p = 75 时,净牵引力在第 9.3 天转正,在第 42.5 天达到峰值;那时积极记忆仍有起初的 0.568,净牵引力为 0.435。在典型参数的网格上,峰值落在第 34.0 至 51.0 天之间。两个时间常数都会让它推迟:愤怒消退得慢、好记忆消退得慢,都会延后最好的时刻。

负面情绪、积极记忆与两者之差,标出峰值(p03)。

窗口有多确定?

网格掩盖了最大的变异来源:分手时有多少怨气。p04 让负面与正面的起始比例在人与人之间大幅变化。

推导

# Chapter 2. p04: How sure is the "golden window"? Draw tn uniform on
# 14 to 21 days, tp uniform on 60 to 90, and the ratio A / B from a
# lognormal law with median 1.5 and log-sd 0.8 (people differ a lot in
# how bitter the rupture was). Report the spread of t*, how often the
# peak of net pull is still negative (N still stronger than P at its
# best moment), how often it falls before day 21 or after day 60,
# and how often A tp <= B tn (no interior peak at all:
# net pull is highest on day 0 and only falls). 20,000 draws, seed 92.
import random
from math import exp, log

def one(rng):
    tn, tp = rng.uniform(14, 21), rng.uniform(60, 90)
    a = 1.5 * exp(rng.gauss(0, 0.8))
    if a * tp <= tn:
        return None, None
    ts = tn * tp / (tp - tn) * log(a * tp / tn)
    return ts, exp(-ts / tp) - a * exp(-ts / tn)

if __name__ == "__main__":
    rng = random.Random(92)
    res = [one(rng) for _ in range(20000)]
    peaks = sorted(t for t, _ in res if t is not None)
    q = lambda p: peaks[int(p * (len(peaks) - 1))]
    print("   draws with no interior peak: {:.4f}".format(
        sum(1 for t, _ in res if t is None) / len(res)))
    print("   t* median {:.1f} days; 10% {:.1f}; 90% {:.1f}".format(
        q(0.5), q(0.1), q(0.9)))
    neg = sum(1 for t, f in res if t is not None and f < 0) / len(peaks)
    print("   peaks where net pull is still negative: {:.3f}".format(neg))
    print("   peaks before day 21: {:.3f}; after day 60: {:.3f}".format(
        sum(1 for t in peaks if t < 21) / len(peaks),
        sum(1 for t in peaks if t > 60) / len(peaks)))
    inside = sum(1 for t in peaks if 30 <= t <= 60) / len(peaks)
    print("   peaks inside days 30 to 60: {:.3f}".format(inside))
#    draws with no interior peak: 0.0103
#    t* median 41.9 days; 10% 19.5; 90% 67.0
#    peaks where net pull is still negative: 0.000
#    peaks before day 21: 0.119; after day 60: 0.174
#    peaks inside days 30 to 60: 0.571

峰值的中位数是第 41.9 天,与网格接近,但十分之八的峰值散布在第 19.5 至 67.0 天之间;0.119 在第 21 天之前,0.174 在第 60 天之后。只有 0.571 落在流行的「第 30 至 60 天」之内。在 0.0103 的抽样中根本没有内部峰值:分手时怨气很少,净牵引力一开始就最高,之后只会下降,等待没有任何好处。「黄金窗口」是一个带着假设数字的模型的性质,不是日历上的日期。

假说 · 尚未验证

作者的诠释(尚未验证): 在分手后的每日评分中,对前任的负面情绪消退的时间常数比对关系的积极记忆短好几倍,因此两者之差在几周之后达到峰值。什么会证明它错: 在一项预先登记的日记研究中,两种评分以相近的速度消退,或者两者之差在第一周之后没有峰值。

小结

两种速度不同的消退过程会让净牵引力出现峰值,它的日期可以写成公式。但这个日期取决于三个因人而异的量,而且散布很宽:模型给出的是一段几周的范围,不是一个约定的日子。

参考文献

Sbarra, David A., and Robert E. Emery. 2005. “The Emotional Sequelae of Nonmarital Relationship Dissolution: Analysis of Change and Intraindividual Variability over Time.” Personal Relationships 12 (2): 213 to 232.
Wildschut, Tim, Constantine Sedikides, Jamie Arndt, and Clay Routledge. 2006. “Nostalgia: Content, Triggers, Functions.” Journal of Personality and Social Psychology 91 (5): 975 to 993.

Fan, Y. (2026). Relationship Rupture Survival Analysis, No-Contact Memory Decay and Reconciliation Game Windows: Negative-Affect Dissipation Rates, Dual-Process Nostalgia Peaks and Signaling Equilibria [关系断裂生存分析、断联期记忆衰减与复联博弈窗口] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090241 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).