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

第六章 两个静止状态:戈特曼的差分方程

把伴侣看作动力系统

戈特曼的观察研究发现,一对伴侣怎样处理冲突,尤其是冲突中有多少负面情绪,可以预测日后的离婚 (Gottman 1994; Gottman and Levenson 2000)。他与默里等人把这些互动写成一对差分方程 (Gottman et al. 2002)。在一次冲突对话中,每一方下一轮的得分取决于自己不受影响时的心情 aa、惯性 rr,以及对方上一轮得分的影响函数 II:

Wt+1=a+rWt+I(Ht),Ht+1=a+rHt+I(Wt+1).W_{t+1} = a + r W_t + I(H_t), \qquad H_{t+1} = a + r H_t + I(W_{t+1}).

p11 使用带阈值的影响函数:小的得分把自身的五分之一传给对方;得分高于 0.8 时把对方提高 0.6;得分低于负面阈值 TT 时把对方压低 1.2,力度是两倍。

推导

# Chapter 6. p11: Gottman-Murray difference equations. Each partner's
# score in a conflict talk is W(t+1) = a + r W(t) + I(H(t)) and
# H(t+1) = a + r H(t) + I(W(t+1)), with a threshold influence function:
# I(x) = 0.2 x between the thresholds, +0.6 above 0.8, and -1.2 below
# the negativity threshold T (negativity counts double). Uninfluenced
# start a = 0.3, inertia r = 0.5. Steady states solve x = a + r x + I(x):
# above 0.8 at (a + 0.6) / (1 - r), below T at (a - 1.2) / (1 - r);
# between the thresholds a / (1 - r - 0.2) = 1.0 lies above 0.8, so
# there is no third resting point.
# Two couples share every parameter and differ only in the first
# exchange; then the negativity threshold T is lowered (made harder to
# cross) from -1 to -2.5.
def make(t_neg):
    def infl(x):
        if x > 0.8:
            return 0.6
        if x < t_neg:
            return -1.2
        return 0.2 * x
    return infl

def path(w, h, t_neg=-1.0, a=0.3, r=0.5, n=40):
    infl = make(t_neg)
    out = [(w, h)]
    for _ in range(n):
        w = a + r * w + infl(h)
        h = a + r * h + infl(w)
        out.append((w, h))
    return out

if __name__ == "__main__":
    print("   upper steady state {:+.2f}; lower {:+.2f}".format(
        (0.3 + 0.6) / 0.5, (0.3 - 1.2) / 0.5))
    for t_neg in (-1.0, -2.5):
        for w0, h0 in ((1.0, 1.0), (-1.5, -2.0)):
            p = path(w0, h0, t_neg)
            print("   T {:+.1f} start ({:+.1f}, {:+.1f}):".format(
                t_neg, w0, h0), end="")
            for t in (1, 3, 10, 40):
                print("  t{} {:+.2f}".format(t, p[t][0]), end="")
            print()
#    upper steady state +1.80; lower -1.80
#    T -1.0 start (+1.0, +1.0):  t1 +1.40  t3 +1.70  t10 +1.80  t40 +1.80
#    T -1.0 start (-1.5, -2.0):  t1 -1.65  t3 -1.76  t10 -1.80  t40 -1.80
#    T -2.5 start (+1.0, +1.0):  t1 +1.40  t3 +1.70  t10 +1.80  t40 +1.80
#    T -2.5 start (-1.5, -2.0):  t1 -0.85  t3 +0.11  t10 +1.68  t40 +1.80

存在两个静止状态:温暖的 +1.80 与敌对的 -1.80。阈值为 T=−1T = -1 时,一对开头顺利的伴侣停在温暖状态;同一对伴侣开头不顺,就停在敌对状态并留在那里。两人没有任何不同,只有第一轮的交流不同。如果把负面阈值降到 T=−2.5T = -2.5,也就是需要更多才会引发强烈的负面反应,敌对状态就不再存在:开头不顺的对话在第三轮经过 +0.11,到第十轮达到 +1.68。在这个模型里,提高一对伴侣对对方负面情绪的承受度,比让任何一方更开心更重要。

三次冲突对话的得分:同一对伴侣、两种开头、两种负面阈值(p11)。

修复与那个著名的比例

戈特曼研究常被概括为:稳定的伴侣在冲突中每一次负面互动大约对应五次正面互动。p12 加入随机冲击与修复尝试(以机会 ρ\rho 抵消强烈的负面推力),并记录正面与负面交流之比。

推导

# Chapter 6. p12: Repair attempts and the 5 to 1 ratio. The p11 couple
# (T = -1) talks for 60 exchanges with random shocks (sd 1.0) added to
# each score. When a score falls below T, a repair attempt cancels the
# -1.2 influence with chance rho. Start (+0.5, +0.5). Report the share
# of talks that end in the negative basin, and the ratio of positive to
# negative exchanges (score > 0 versus score < 0) in talks that end
# well and in talks that end badly. 4,000 talks per row, seed 92.
import random

def talk(rng, rho, a=0.3, r=0.5, n=60):
    def infl(x):
        if x > 0.8:
            return 0.6
        if x < -1:
            return 0.0 if rng.random() < rho else -1.2
        return 0.2 * x
    w = h = 0.5
    pos = neg = 0
    for _ in range(n):
        w = a + r * w + infl(h) + rng.gauss(0, 1.0)
        h = a + r * h + infl(w) + rng.gauss(0, 1.0)
        for x in (w, h):
            pos += x > 0
            neg += x < 0
    return (w + h) / 2 < 0, pos, neg

if __name__ == "__main__":
    rng = random.Random(92)
    print("   repair rho   ends negative   pos:neg (good)   pos:neg (bad)")
    for rho in (0.0, 0.3, 0.6, 0.9):
        res = [talk(rng, rho) for _ in range(4000)]
        bad = [x for x in res if x[0]]
        good = [x for x in res if not x[0]]
        rg = sum(x[1] for x in good) / max(1, sum(x[2] for x in good))
        rb = sum(x[1] for x in bad) / max(1, sum(x[2] for x in bad))
        print("   {:10.1f}   {:13.3f}   {:14.2f}   {:13.2f}".format(
            rho, len(bad) / len(res), rg, rb))
#    repair rho   ends negative   pos:neg (good)   pos:neg (bad)
#           0.0           0.184             4.02            2.75
#           0.3           0.128             5.19            3.91
#           0.6           0.100             6.05            4.57
#           0.9           0.096             6.82            5.10

没有修复时,0.184 的对话以敌对状态结束;修复三成成功时 0.128;六成时 0.100;九成时 0.096。比例的表现没有那么简单。结局良好的对话,没有修复时比例是 4.02,修复频繁时 6.82;结局不好的对话分别是 2.75 与 5.10。于是,修复频繁的坏对话可以显示出比没有修复的好对话更高的比例。这个比例是一对伴侣动态的结果,并且取决于怎样计数;它不是一个本身就决定伴侣命运的门槛。

假说 · 尚未验证

作者的诠释(尚未验证): 在观察到的冲突讨论中,一对伴侣从一次负面交流中恢复的速度,比正负交流的总比例更能预测日后的稳定。什么会证明它错: 在一项对冲突讨论编码的预先登记纵向研究中,总比例预测分手至少与恢复速度一样好,而在已知总比例之后恢复速度没有增加任何预测力。

小结

在戈特曼与默里的模型里,同一对伴侣可以停在温暖或敌对的状态,而负面情绪在什么阈值引发强烈反应,决定了敌对状态是否存在。修复有帮助,但好坏时刻的单一比例只是概括,不是定律。

参考文献

Gottman, John M. 1994. What Predicts Divorce? The Relationship between Marital Processes and Marital Outcomes. Lawrence Erlbaum Associates.
Gottman, John M., and Robert W. Levenson. 2000. “The Timing of Divorce: Predicting When a Couple Will Divorce over a 14-Year Period.” Journal of Marriage and Family 62 (3): 737 to 745.
Gottman, John M., James D. Murray, Catherine C. Swanson, Rebecca Tyson, and Kristin R. Swanson. 2002. The Mathematics of Marriage: Dynamic Nonlinear Models. MIT Press.

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).