5D Library › 下丘脑垂体肾上腺轴耦合振荡、皮质醇觉醒反应扁平化与倦怠恢复动力学 · English

第七章 以妄为常

一句关于方差的抱怨

《素问》第一篇把上古之人与今时之人对举:前者有节度,后者以酒为浆、以妄为常,而且不知持满 (黄帝内经素问 1963)。当作修辞读,它是一句关于世风日下的抱怨。窄着读,它是两个不同的论断,而其中一个是关于方差的论断,不是关于任何单次过度的论断。

原典

那两句抱怨。 《素问》首篇这段文字把若干过失并举,其中有两句值得分开。以妄为常并不是说某一个晚睡的夜晚有害;它说的是不规律本身,一旦成为习惯,就是那处损伤。不知持满是另一个论断,讲的是动用储备而不顾其中还剩多少,这正是本章后半的主题。前者可以写成一句关于作息离散度的陈述;后者可以写成一句关于水库的陈述 (黄帝内经素问 1963)。

第一个论断就按它的原样即可检验。一个每天被拉回锚点、而每夜又被「这个人实际几点上床」所扰动的相位,有一个稳态离散度,而落在可容忍窗口之外的机会由它推出。

推导

# Chapter 7. p13: Taking the irregular as the rule. The Suwen's
# complaint about people who take the reckless as their norm is, read
# narrowly, a claim about variance rather than about any single bad
# night. Model the phase as pulled back towards its anchor each day by a
# fraction of its error, with a fresh random disturbance added by that
# night's bedtime. The error then has a steady-state spread that depends
# on the disturbance and on the pull, and the chance of being outside a
# tolerable window follows from it. Four thousand people, 400 days each,
# seed 97, so the figures repeat exactly.
import random

SEED, PEOPLE, DAYS = 97, 4000, 400
PULL = 0.28                  # share of the error corrected each day
WINDOW = 1.0                 # hours of error a person tolerates
SDS = (0.25, 0.50, 1.00, 1.50, 2.00)     # bedtime spread in hours

def run(rng, sd):
    phase = 0.0
    out = []
    for d in range(DAYS):
        phase = phase * (1.0 - PULL) + rng.gauss(0.0, sd)
        if d > DAYS // 2:
            out.append(phase)
    return out

def spread(values):
    m = sum(values) / len(values)
    return (sum((v - m) ** 2 for v in values) / (len(values) - 1)) ** 0.5

if __name__ == "__main__":
    print("   the pull back towards the anchor is {:.2f} a day".format(PULL))
    print("   predicted spread is the bedtime spread divided by the")
    print("   square root of one minus the square of one minus the pull")
    factor = (1.0 - (1.0 - PULL) ** 2) ** 0.5
    print("   that divisor is {:.4f}".format(factor))
    print("   bedtime sd h   phase sd h   predicted   days outside"
          " {:.1f} h".format(WINDOW))
    rng = random.Random(SEED)
    for sd in SDS:
        allv = []
        outside = 0
        for _ in range(PEOPLE):
            v = run(rng, sd)
            allv.extend(v)
            outside += sum(1 for x in v if abs(x) > WINDOW)
        n = len(allv)
        print("   {:<15.2f}{:<13.4f}{:<12.4f}{:.4f}"
              .format(sd, spread(allv), sd / factor, outside / n))
    print("   the spread grows in proportion to the bedtime spread, so")
    print("   the chance of being outside the window is set by the ratio")
    print("   of the two and nothing else")
    print("   bedtime sd that keeps 0.95 of days inside the window:")
    lo, hi = 0.05, 3.0
    for _ in range(26):
        mid = 0.5 * (lo + hi)
        r2 = random.Random(SEED + 1)
        bad = tot = 0
        for _ in range(300):
            v = run(r2, mid)
            bad += sum(1 for x in v if abs(x) > WINDOW)
            tot += len(v)
        if bad / tot < 0.05:
            lo = mid
        else:
            hi = mid
    print("   {:.3f} hours, which is {:.0f} minutes"
          .format(0.5 * (lo + hi), 0.5 * (lo + hi) * 60.0))
    print("   so a bedtime that wanders by more than about that much")
    print("   cannot be anchored by anything done in the morning, and")
    print("   the classical complaint is about the wandering and not")
    print("   about any one late night")
#    the pull back towards the anchor is 0.28 a day
#    predicted spread is the bedtime spread divided by the
#    square root of one minus the square of one minus the pull
#    that divisor is 0.6940
#    bedtime sd h   phase sd h   predicted   days outside 1.0 h
#    0.25           0.3602       0.3602      0.0054
#    0.50           0.7197       0.7205      0.1647
#    1.00           1.4398       1.4410      0.4864
#    1.50           2.1622       2.1615      0.6437
#    2.00           2.8786       2.8820      0.7274
#    the spread grows in proportion to the bedtime spread, so
#    the chance of being outside the window is set by the ratio
#    of the two and nothing else
#    bedtime sd that keeps 0.95 of days inside the window:
#    0.352 hours, which is 21 minutes
#    so a bedtime that wanders by more than about that much
#    cannot be anchored by anything done in the morning, and
#    the classical complaint is about the wandering and not
#    about any one late night

在每天 0.28 的拉回强度下,相位的预测离散度等于就寝离散度除以 0.6940,而模拟在所检验的每一个水平上都与那个预测吻合到四位小数。四分之一小时的就寝离散度留下 0.0054 的日子落在一小时宽的窗口之外;半小时留下 0.1647;一小时留下 0.4864;两小时留下 0.7274。使 0.95 的日子落在窗口之内的就寝离散度是 0.352 小时,也就是 21 分钟。

这就是那句古典抱怨用现代单位说出来的样子,而它是一个比大多数人以为的更严格的要求。二十一分钟不是一个宽松的容差。它也解释了为什么晨间干预在就寝时刻游移的人身上失效:拉回锚点的力度是固定的,所以一个足够大的扰动不可能被任何一天只施加一次的东西撤销。损伤在于游移,而不在于任何一个晚睡的夜晚,正如那段文字所说。

一个水库,以及抽在它上面的火

第二句抱怨讲的是储备,而它的古典词汇是张介宾的。他论命门,把它当作同时藏着身体的水与火、而两者必须相称的所在,并特别警告在水已耗尽时还去抽那把火 (张介宾 1991)。

原典

命门水火。 张介宾的论述把命门作为身体之水与之火共同的所在,并坚持两者都不可不顾对方而单独补益。这里要紧的那句警告在于不对称:火是产生可见产出的东西,而水是让火可持续的东西,所以一个以产出被衡量的人会去抽那把火,而在水耗尽之前不会注意到它。写成算术,这是一个两变量系统:所求的产出由一个驱动经由一份能力产生,而这份能力又被这个驱动消耗 (张介宾 1991)。

推导

# Chapter 7. p14: Fire drawn on an empty reservoir. Zhang Jingyue reads
# the gate of life as holding a water and a fire that must be kept in
# proportion, and warns against drawing the fire while the water is
# spent. Written as arithmetic, that is a two-variable system in which
# the output a person demands is produced by the drive acting through a
# capacity, and the capacity is consumed in proportion to the drive. To
# hold the output steady while the capacity falls, the drive must rise,
# which consumes the capacity faster. The steady state has a closed
# form: the refill must equal the consumption, and because the drive is
# the output divided by the capacity, the consumption does not depend on
# the capacity at all. So the steady capacity is one minus the output
# times the ratio of cost to refill, which is linear, and it reaches
# zero at an output of refill over cost. Above that there is no steady
# state at any level, and the reservoir empties.
CAP0, REFILL, COST = 1.0, 1.0 / 90.0, 0.040
DEMANDS = (0.30, 0.50, 0.70, 0.80, 0.90, 1.00, 1.10)
DAYS, FLOOR = 720, 0.05

def drive_for(output, cap):
    # output is the product of drive and capacity, so drive is the ratio
    return output / max(cap, 1e-9)

def run(output):
    cap = CAP0
    for day in range(DAYS):
        d = drive_for(output, cap)
        cap += REFILL * (CAP0 - cap) - COST * d * cap
        if cap <= FLOOR:
            return day + 1, cap
        cap = min(CAP0, cap)
    return None, cap

if __name__ == "__main__":
    print("   refill rate {:.5f} a day, cost of drive {:.3f}"
          .format(REFILL, COST))
    print("   demanded output   days to the floor   capacity at day {}"
          .format(DAYS))
    for out in DEMANDS:
        days, cap = run(out)
        print("   {:<18.2f}{:<20}{:.4f}"
              .format(out, days if days is not None else "holds", cap))
    ratio = COST / REFILL
    ceiling = REFILL / COST
    print("   the steady capacity is one minus {:.4f} times the output"
          .format(ratio))
    print("   output   steady capacity   steady drive")
    for out in (0.05, 0.10, 0.15, 0.20, 0.25):
        cap = 1.0 - ratio * out
        print("   {:<9.2f}{:<18.4f}{:.4f}".format(out, cap, out / cap))
    print("   the output at which the steady capacity reaches zero:"
          " {:.4f}".format(ceiling))
    lo, hi = 0.01, 2.0
    for _ in range(60):
        mid = 0.5 * (lo + hi)
        if run(mid)[0] is None:
            lo = mid
        else:
            hi = mid
    crit = 0.5 * (lo + hi)
    print("   the output above which the floor of {:.2f} is reached:"
          " {:.4f}".format(FLOOR, crit))
    print("   which matches one minus {:.2f} over {:.4f}, as it must"
          .format(FLOOR, ratio))
    print("   above the ceiling the system does not settle at a lower")
    print("   level; it empties, because holding the output needs a")
    print("   drive that rises as fast as the capacity falls")
    for over in (1.02, 1.05, 1.10, 1.25):
        days, cap = run(crit * over)
        print("   at {:.2f} times that output the floor arrives on day {}"
              .format(over, days))
    print("   and the same arithmetic gives the way out: lower the output")
    print("   demanded, since the capacity cannot be raised directly")
    print("   output cut to   capacity after 90 days")
    for out in (0.90, 0.70, 0.50, 0.30):
        cap = FLOOR
        for _ in range(90):
            d = drive_for(out * crit, cap)
            cap = min(CAP0, cap + REFILL * (CAP0 - cap) - COST * d * cap)
            cap = max(1e-9, cap)
        print("   {:<16.2f}{:.4f}".format(out, cap))
    print("   so each further cut buys more capacity: a cut to 0.70")
    print("   reaches about a quarter of the full level in three months,")
    print("   a cut to 0.30 about half, and holding the output at 1.00")
    print("   recovers nothing, since that is the critical output itself")
#    refill rate 0.01111 a day, cost of drive 0.040
#    demanded output   days to the floor   capacity at day 720
#    0.30              190                 0.0493
#    0.50              68                  0.0420
#    0.70              43                  0.0386
#    0.80              36                  0.0462
#    0.90              32                  0.0260
#    1.00              28                  0.0329
#    1.10              25                  0.0349
#    the steady capacity is one minus 3.6000 times the output
#    output   steady capacity   steady drive
#    0.05     0.8200            0.0610
#    0.10     0.6400            0.1563
#    0.15     0.4600            0.3261
#    0.20     0.2800            0.7143
#    0.25     0.1000            2.5000
#    the output at which the steady capacity reaches zero: 0.2778
#    the output above which the floor of 0.05 is reached: 0.2640
#    which matches one minus 0.05 over 3.6000, as it must
#    above the ceiling the system does not settle at a lower
#    level; it empties, because holding the output needs a
#    drive that rises as fast as the capacity falls
#    at 1.02 times that output the floor arrives on day 351
#    at 1.05 times that output the floor arrives on day 272
#    at 1.10 times that output the floor arrives on day 215
#    at 1.25 times that output the floor arrives on day 144
#    and the same arithmetic gives the way out: lower the output
#    demanded, since the capacity cannot be raised directly
#    output cut to   capacity after 90 days
#    0.90            0.1101
#    0.70            0.2306
#    0.50            0.3511
#    0.30            0.4717
#    so each further cut buys more capacity: a cut to 0.70
#    reaches about a quarter of the full level in three months,
#    a cut to 0.30 about half, and holding the output at 1.00
#    recovers nothing, since that is the critical output itself

这个系统有闭式解,而它值得一看。稳态下补充必须等于消耗,而因为驱动等于产出除以能力,消耗就只取决于产出。所以稳定能力等于一减去 3.6000 乘以所求产出,这是线性的,而它在产出为 0.2778 处降到零。在它之下存在一个能力被压低了的稳态:产出为 0.20 时,能力稳在 0.2800,驱动稳在 0.7143。在它之上,任何水平上都不存在稳态,而水库会被抽空。

非线性在驱动里,不在能力里。产出为 0.05 时所需的驱动是 0.0610;产出为 0.25 时是 2.5000,产出是五倍而驱动是四十倍。这就是「不知持满」所描述的东西:最后那一点产出的代价,是用一个随能力下降而无界上升的驱动来支付的。

出路在同一套算术里。从地板出发,把所求产出削到临界值的 0.70,九十天把能力带到 0.2306;削到 0.30,则带到 0.4717。维持产出而指望着恢复,什么也恢复不了,因为稳定能力是产出的函数,而且只是产出的函数。

假说 · 尚未验证

作者的诠释(尚未验证): 这句古典警告指出的是一个阈值而不是一条梯度,所以从职业性耗竭中恢复,应当强烈取决于所求产出是否被削到某个临界水平之下,而很少取决于削到低于它多少,这与剂量反应正好相反。什么会证明它错: 如果在一个工作负荷有分级削减的队列里,实测恢复随削减幅度平滑上升而不是在某个阈值处跳变,那么这套水库读法就是错的。

小结

关于以妄为常的那句抱怨是一个方差论断:就寝离散度必须在 21 分钟以内,才有 0.95 的日子落在一小时宽的窗口内,而相位离散度等于就寝离散度除以 0.6940。关于不知持满的那句抱怨是一个水库论断:稳定能力等于一减去 3.6000 乘以所求产出,所以产出超过 0.2778 时什么也稳不住,而所需驱动从产出 0.05 时的 0.0610 升到 0.25 时的 2.5000。第八章把依从性与那道闸放到这一切之前。

参考文献

张介宾. 1991. 景岳全书. 人民卫生出版社.
黄帝内经素问. 1963. 人民卫生出版社.

Fan, Y. (2026). HPA-Axis Coupled Oscillations, Cortisol Awakening Response Flattening and Burnout Recovery Kinetics: Glucocorticoid Receptor Downregulation, Ultradian Pulsatility and Allostatic Load Reset [下丘脑垂体肾上腺轴耦合振荡、皮质醇觉醒反应扁平化与倦怠恢复动力学] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090252 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).