5D Library › 亚知觉色胺药代动力学、5-HT2A 受体脱敏与神经可塑性窗口的生物数学模型 · English

第五章 连接熵:与噪声不可区分

熵脑假说

证据

熵脑(理论与影像学)。 Carhart-Harris 等人 2014 年提出「熵脑」假说:意识状态可以用脑活动某种熵的高低来刻画,而致幻剂把系统推向更高熵的状态 (Carhart-Harris et al. 2014);2018 年的修订版整理了后续证据与批评 (Carhart-Harris 2018)。相关的影像学结果包括 Daws 等人报告赛洛西宾治疗抑郁后全脑整合度升高 (Daws et al. 2022)。第 8 本第三章讨论过这一假说与它的限度,这里不重复。

本章只处理一个狭窄的方法学问题:如果有人测量「全脑功能连接的熵」并报告它升高了,这个测量能支持什么?

高斯图模型中的熵

对多元高斯分布,微分熵为 H=12logdet(2πeΣ)H = \tfrac12\log\det(2\pi e\,\Sigma),所以对相关矩阵而言,熵完全由 logdet\log\det 决定。这使问题变得可以精确处理。

推导

# Chapter 5. "Entropy of whole-brain functional connectivity", and what it
# measures. For a multivariate Gaussian the differential entropy is
#   H = 0.5 log det(2 pi e Sigma)
# so for a correlation matrix the only thing entropy depends on is log det.
# The program builds a schematic correlation matrix with a single global
# factor of loading g plus independent noise, and shows that entropy rises
# monotonically as g falls. Then it shows that TWO quite different things
# lower g: a genuine reduction in global coupling, and added independent
# measurement noise (head motion, physiological noise, a shorter window). Both
# raise the entropy by the same amount, and no entropy value distinguishes
# them. Exact linear algebra, no simulation; dimensionless.
import math

P = 20                      # number of regions

def corr_from_factor(g, noise=0.0):
    """One global factor with loading g, plus independent noise of variance
    `noise`."""
    out = []
    for i in range(P):
        row = []
        for j in range(P):
            if i == j:
                row.append(1.0)
            else:
                row.append(g * g / (1.0 + noise))
        out.append(row)
    return out

def logdet(M):
    """Cholesky; M must be symmetric positive definite."""
    n = len(M)
    L = [[0.0] * n for _ in range(n)]
    for i in range(n):
        for j in range(i + 1):
            s = M[i][j] - math.fsum(L[i][k] * L[j][k] for k in range(j))
            if i == j:
                if s <= 0:
                    return float("-inf")
                L[i][i] = math.sqrt(s)
            else:
                L[i][j] = s / L[j][j]
    return 2.0 * math.fsum(math.log(L[i][i]) for i in range(n))

def entropy_bits(M):
    """Entropy of the standardised multivariate Gaussian, relative to
    independence."""
    return logdet(M) / (2.0 * math.log(2.0))

if __name__ == "__main__":
    print(
        "{} regions, one global factor of loading g. Entropy is reported"
        " relative to".format(P)
    )
    print("the independent case (g = 0), in bits; 0 is the maximum.\n")
    print("   g      mean off-diagonal r   relative entropy (bits)")
    for g in (0.0, 0.2, 0.4, 0.6, 0.8, 0.9):
        M = corr_from_factor(g)
        print(
            "   {:5.2f} {:21.4f} {:24.3f}".format(g, g * g, entropy_bits(M))
        )

    print(
        "\nEntropy rises as coupling falls, which is the result the"
        " entropic-brain"
    )
    print(
        "literature reports as an 'increase in entropy'. Now the confound."
        " Adding"
    )
    print("independent noise of variance v to every channel attenuates every")
    print("correlation by 1/(1+v) and so raises entropy too:")
    print(
        "\n   true g   added noise v   observed mean r   relative entropy"
        " (bits)"
    )
    rows = []
    for g, v in ((0.8, 0.0), (0.8, 0.25), (0.8, 0.78), (0.6, 0.0)):
        M = corr_from_factor(g, v)
        rows.append((g, v, g * g / (1 + v), entropy_bits(M)))
        print("   {:6.3f} {:15.2f} {:17.4f} {:24.3f}".format(*rows[-1]))

    print(
        "\n   Compare the last two rows. True coupling 0.8 with added"
        " noise, and"
    )
    print(
        "   genuinely weaker coupling 0.6 with no noise, give observed"
        " correlations"
    )
    print(
        "   of {:.4f} and {:.4f} and entropies of {:.3f} and {:.3f} bits: the"
        " same".format(rows[2][2], rows[3][2], rows[2][3], rows[3][3])
    )
    print(
        "   measurement to within {:.3f} bits, from two completely different"
        " states".format(abs(rows[2][3] - rows[3][3]))
    )
    print(
        "   of the brain. An entropy increase is therefore not evidence of a"
        " change"
    )
    print(
        "   in coupling unless the noise is shown not to have changed - and"
        " head"
    )
    print("   motion, breathing and heart rate all change with a drug.")

    print(
        "\nHow much noise would be needed to fake a given entropy rise,"
        " starting"
    )
    print("from true coupling g = 0.8?")
    base = entropy_bits(corr_from_factor(0.8))
    print("   target entropy rise (bits)   required added noise v")
    need = {}
    for target in (0.5, 1.0, 2.0, 4.0):
        lo, hi = 0.0, 50.0
        for _ in range(60):
            mid = 0.5 * (lo + hi)
            if entropy_bits(corr_from_factor(0.8, mid)) - base < target:
                lo = mid
            else:
                hi = mid
        need[target] = hi
        print("   {:26.1f} {:22.3f}".format(target, hi))
    print(
        "\n   A rise of 1 bit across {} regions needs added noise of only"
        " {:.1%} of".format(P, need[1.0])
    )
    print(
        "   the signal variance, and 4 bits needs {:.0%} - both well inside"
        " the range".format(need[4.0])
    )
    print(
        "   that a change in head motion or physiological noise can"
        " produce. The"
    )
    print(
        "   measure is far more sensitive to nuisance variance than its use"
        " as a"
    )
    print(
        "   marker of 'brain entropy' implies. This is why Chapter 5 asks for"
        " motion"
    )
    print(
        "   and physiological regressors, a matched-length window and a"
        " pre-specified"
    )
    print(
        "   estimator, and treats any connectivity entropy as exploratory"
        " only."
    )
# 20 regions, one global factor of loading g. Entropy is reported relative to
# the independent case (g = 0), in bits; 0 is the maximum.
#
#    g      mean off-diagonal r   relative entropy (bits)
#     0.00                0.0000                    0.000
#     0.20                0.0400                   -0.152
#     0.40                0.1600                   -1.382
#     0.60                0.3600                   -4.631
#     0.80                0.6400                  -12.143
#     0.90                0.8100                  -20.744
#
# Entropy rises as coupling falls, which is the result the entropic-brain
# literature reports as an 'increase in entropy'. Now the confound. Adding
# independent noise of variance v to every channel attenuates every
# correlation by 1/(1+v) and so raises entropy too:
#
#    true g   added noise v   observed mean r   relative entropy (bits)
#     0.800            0.00            0.6400                  -12.143
#     0.800            0.25            0.5120                   -8.121
#     0.800            0.78            0.3596                   -4.622
#     0.600            0.00            0.3600                   -4.631
#
#    Compare the last two rows. True coupling 0.8 with added noise, and
#    genuinely weaker coupling 0.6 with no noise, give observed correlations
#    of 0.3596 and 0.3600 and entropies of -4.622 and -4.631 bits: the same
#    measurement to within 0.009 bits, from two completely different states
#    of the brain. An entropy increase is therefore not evidence of a change
#    in coupling unless the noise is shown not to have changed - and head
#    motion, breathing and heart rate all change with a drug.
#
# How much noise would be needed to fake a given entropy rise, starting
# from true coupling g = 0.8?
#    target entropy rise (bits)   required added noise v
#                           0.5                  0.022
#                           1.0                  0.046
#                           2.0                  0.101
#                           4.0                  0.248
#
#    A rise of 1 bit across 20 regions needs added noise of only 4.6% of
#    the signal variance, and 4 bits needs 25% - both well inside the range
#    that a change in head motion or physiological noise can produce. The
#    measure is far more sensitive to nuisance variance than its use as a
#    marker of 'brain entropy' implies. This is why Chapter 5 asks for motion
#    and physiological regressors, a matched-length window and a pre-specified
#    estimator, and treats any connectivity entropy as exploratory only.

第一张表给出预期的方向:耦合越弱,熵越高。这就是文献报告为「熵升高」的量。

第二张表是本章的要点。给每个通道加上独立噪声会使每一个相关系数衰减,从而也使熵升高。于是真实耦合 0.8 加噪声与真实耦合 0.6 无噪声给出相同的观测相关(0.3596 对 0.3600)与相同的熵(相差 0.009 比特)。两个完全不同的大脑状态,同一个测量值。

而所需的噪声量很小:程序最后给出,跨 20 个脑区使熵升高 1 比特只需约 4.6% 的额外噪声方差,升高 4 比特需约 25%。

证据

头动会造成系统性的连接伪迹。 Power 等人证明受试者运动在功能连接中产生虚假但系统性的相关,并且这些伪迹不是随机的——它们在短距离连接上使相关升高、在长距离连接上使相关降低 (Power et al. 2012)。如何回归混淆因素本身是一个有基准研究的课题,Ciric 等人系统比较了多种策略并指出没有一种能完全消除伪迹 (Ciric et al. 2017)。

把这两件事放在一起就得到本章的结论:

「连接熵升高」不是大脑耦合改变的证据,除非证明噪声没有改变。而头动、呼吸与心率都会随药物改变。

左:高斯图模型中的相对熵随全局耦合 g 的变化(20 个脑区)。右:真实耦合 0.8 加独立噪声(实心点)与真实耦合 0.6 无噪声(空心点)落在同一位置:相同的观测相关、相同的熵,两个不同的状态。

估计量的偏差

第二个问题与第 53 本第三章的表象流熵是同一个数学问题:短序列(这里是短窗口)上的熵估计有系统偏差。

推导

# Chapter 5. The other half of the entropy problem: the estimator.
# Correlations estimated from a window of T samples are noisy, and the noise
# inflates log det downward (toward apparent lower entropy) or upward
# depending on the regime, so the measured "entropy" depends on scan length.
# The program estimates the correlation matrix of a P-region Gaussian from
# windows of several lengths and reports the bias, then shows the matched-
# length comparison that removes it. Synthetic data; fixed seed; standard
# library only.
import math
import random

P, G_TRUE = 12, 0.7

def sample_window(T, g, rng):
    """T samples of P regions with one global factor of loading g."""
    cols = [[0.0] * T for _ in range(P)]
    for t in range(T):
        f = rng.gauss(0.0, 1.0)
        for i in range(P):
            cols[i][t] = g * f + math.sqrt(1.0 - g * g) * rng.gauss(0.0, 1.0)
    return cols

def corr(cols):
    T = len(cols[0])
    means = [math.fsum(c) / T for c in cols]
    sds = [math.sqrt(math.fsum((v - m) ** 2 for v in c) / (T - 1))
           for c, m in zip(cols, means)]
    M = [[0.0] * P for _ in range(P)]
    for i in range(P):
        for j in range(P):
            if i == j:
                M[i][j] = 1.0
            else:
                cov = math.fsum(
                    (cols[i][t] - means[i]) * (cols[j][t] - means[j])
                    for t in range(len(cols[0]))
                ) / (T - 1)
                M[i][j] = cov / (sds[i] * sds[j])
    return M

def logdet(M):
    n = len(M)
    L = [[0.0] * n for _ in range(n)]
    for i in range(n):
        for j in range(i + 1):
            s = M[i][j] - math.fsum(L[i][k] * L[j][k] for k in range(j))
            if i == j:
                if s <= 0:
                    return float("-inf")
                L[i][i] = math.sqrt(s)
            else:
                L[i][j] = s / L[j][j]
    return 2.0 * math.fsum(math.log(L[i][i]) for i in range(n))

def ent(M):
    return logdet(M) / (2.0 * math.log(2.0))

def truth(g):
    M = [[1.0 if i == j else g * g for j in range(P)] for i in range(P)]
    return ent(M)

if __name__ == "__main__":
    rng = random.Random(5209)
    t0 = truth(G_TRUE)
    print(
        "{} regions, true global loading {:.1f}; true relative entropy {:.3f}"
        " bits.".format(P, G_TRUE, t0)
    )
    print(
        "\n   window T   mean estimated entropy (bits)   bias     s.d. over"
        " runs"
    )
    for T in (30, 60, 120, 300, 1200):
        vals = [ent(corr(sample_window(T, G_TRUE, rng))) for _ in range(120)]
        m = math.fsum(vals) / len(vals)
        sd = math.sqrt(
            math.fsum((v - m) ** 2 for v in vals) / (len(vals) - 1)
        )
        print("   {:8d} {:30.3f} {:+8.3f} {:16.3f}".format(T, m, m - t0, sd))

    print(
        "\nShorter windows give systematically HIGHER entropy estimates:"
        " sampling"
    )
    print(
        "noise pushes the estimated correlations toward zero, which is"
        " exactly what"
    )
    print(
        "a genuine decoupling would do. The bias at T = 30 is larger than the"
        " whole"
    )
    print("difference between two quite different coupling strengths:")
    print(
        "   true g = 0.7 -> {:.3f} bits; true g = 0.5 -> {:.3f} bits;"
        "\n   difference {:.3f}".format(
            truth(0.7), truth(0.5), truth(0.5) - truth(0.7)
        )
    )

    print(
        "\nThe fix is not a better estimator but a matched design."
        " Comparing two"
    )
    print("conditions using windows of the SAME length cancels the bias:")
    print(
        "   T      E[entropy | g=0.7]   E[entropy | g=0.5]   difference  "
        " truth"
    )
    for T in (30, 60, 120, 300):
        a = [ent(corr(sample_window(T, 0.7, rng))) for _ in range(120)]
        b = [ent(corr(sample_window(T, 0.5, rng))) for _ in range(120)]
        ma = math.fsum(a) / len(a)
        mb = math.fsum(b) / len(b)
        print("   {:5d} {:20.3f} {:20.3f} {:12.3f} {:7.3f}"
              .format(T, ma, mb, mb - ma, truth(0.5) - truth(0.7)))

    print(
        "\n   The estimated DIFFERENCE is close to the truth at every window"
        " length,"
    )
    print(
        "   even where each estimate separately is badly biased. So a"
        " connectivity-"
    )
    print(
        "   entropy comparison is usable if and only if the windows are"
        " matched in"
    )
    print(
        "   length and the data are matched in quality; an entropy value"
        " reported on"
    )
    print(
        "   its own, or compared across scans of different length or"
        " different"
    )
    print("   motion, carries a bias of the same size as the effect.")
# 12 regions, true global loading 0.7; true relative entropy -4.005 bits.
#
#    window T   mean estimated entropy (bits)   bias     s.d. over runs
#          30                         -5.983   -1.978            1.191
#          60                         -4.853   -0.848            0.699
#         120                         -4.362   -0.357            0.493
#         300                         -4.197   -0.192            0.317
#        1200                         -4.069   -0.064            0.171
#
# Shorter windows give systematically HIGHER entropy estimates: sampling
# noise pushes the estimated correlations toward zero, which is exactly what
# a genuine decoupling would do. The bias at T = 30 is larger than the whole
# difference between two quite different coupling strengths:
#    true g = 0.7 -> -4.005 bits; true g = 0.5 -> -1.329 bits;
#    difference 2.676
#
# The fix is not a better estimator but a matched design. Comparing two
# conditions using windows of the SAME length cancels the bias:
#    T      E[entropy | g=0.7]   E[entropy | g=0.5]   difference   truth
#       30               -6.119               -3.354        2.765   2.676
#       60               -4.747               -2.248        2.499   2.676
#      120               -4.409               -1.779        2.630   2.676
#      300               -4.195               -1.477        2.718   2.676
#
#    The estimated DIFFERENCE is close to the truth at every window length,
#    even where each estimate separately is badly biased. So a connectivity-
#    entropy comparison is usable if and only if the windows are matched in
#    length and the data are matched in quality; an entropy value reported on
#    its own, or compared across scans of different length or different
#    motion, carries a bias of the same size as the effect.

量级值得记下:12 个脑区、窗口 30 个时间点时,偏差为 −1.978-1.978 比特;而真实耦合 0.7 与 0.5 之间的差别是 2.676 比特。偏差与效应同量级。Paninski 对熵估计量的性质有系统分析 (Paninski 2003)。

但程序的第二张表给出解决办法,而它是设计上的而不是统计上的:用相同长度的窗口比较两个条件,偏差大体抵消。估计的差值在每个窗口长度下都接近真值(2.765、2.499、2.630、2.718,真值 2.676),即使每个估计本身都有很大偏差。

注

对第八章的后果。 连接熵类指标在第八章中只作探索性结局,并附三项规定:(一)两个条件的窗口长度必须相同,且扫描长度在方案中固定;(二)必须报告头动指标并作为混淆回归,按已有基准研究选择策略 (Power et al. 2012; Ciric et al. 2017);(三)估计量在方案中预先指定。即使如此,它也不能单独支持任何关于「脑熵」的结论。

假说 · 尚未验证

作者的假说(尚未确立): 在匹配窗口长度、并把头动与生理噪声作为协变量之后,亚知觉暴露不会使功能连接熵产生超过测量变异的变化。什么样的结果能证明它是错的: 在事先登记的双盲交叉研究中,以匹配窗口的连接熵差值为预先指定结局,若活性条件相对安慰剂的差别的 95% 置信区间排除零,且在加入运动与生理协变量后仍然排除零,并在独立样本中重现,则本假说错。

本章小结

熵脑假说与相关影像学结果是这一领域的核心。本章只处理测量问题,得到两个结果。第一,在高斯图模型中,「熵升高」与「噪声增加」不可区分:真实耦合 0.8 加噪声与真实耦合 0.6 无噪声给出相同的观测相关与相同的熵(相差 0.009 比特),而跨 20 个脑区使熵升高 1 比特只需约 4.6% 的额外噪声方差;头动已知会造成系统性的连接伪迹,且没有混淆回归策略能完全消除它。第二,短窗口的熵估计有系统偏差:12 个脑区、30 个时间点时偏差 −1.978-1.978 比特,而两种耦合强度之间的真实差别是 2.676 比特。解决办法是设计上的:匹配窗口长度的差值比较在每个长度下都接近真值。第八章因此把连接熵类指标列为探索性结局并附三项规定。

参考文献

Carhart-Harris, Robin L. 2018. “The Entropic Brain — Revisited.” Neuropharmacology 142: 167–78. https://doi.org/10.1016/j.neuropharm.2018.03.010.
Carhart-Harris, Robin L., Robert Leech, Peter J. Hellyer, Murray Shanahan, Amanda Feilding, Enzo Tagliazucchi, Dante R. Chialvo, and David Nutt. 2014. “The Entropic Brain: A Theory of Conscious States Informed by Neuroimaging Research with Psychedelic Drugs.” Frontiers in Human Neuroscience 8: 20. https://doi.org/10.3389/fnhum.2014.00020.
Ciric, Rastko, Daniel H. Wolf, Jonathan D. Power, David R. Roalf, Graham L. Baum, Kosha Ruparel, Russell T. Shinohara, et al. 2017. “Benchmarking of Participant-Level Confound Regression Strategies for the Control of Motion Artifact in Studies of Functional Connectivity.” NeuroImage 154: 174–87. https://doi.org/10.1016/j.neuroimage.2017.03.020.
Daws, Richard E., Christopher Timmermann, Bruna Giribaldi, James D. Sexton, Matthew B. Wall, David Erritzoe, Leor Roseman, David Nutt, and Robin Carhart-Harris. 2022. “Increased Global Integration in the Brain After Psilocybin Therapy for Depression.” Nature Medicine 28: 844–51. https://doi.org/10.1038/s41591-022-01744-z.
Paninski, Liam. 2003. “Estimation of Entropy and Mutual Information.” Neural Computation 15 (6): 1191–1253. https://doi.org/10.1162/089976603321780272.
Power, Jonathan D., Kelly A. Barnes, Abraham Z. Snyder, Bradley L. Schlaggar, and Steven E. Petersen. 2012. “Spurious but Systematic Correlations in Functional Connectivity MRI Networks Arise from Subject Motion.” NeuroImage 59 (3): 2142–54. https://doi.org/10.1016/j.neuroimage.2011.10.018.

Fan, Y. (2026). Biomathematical Models of Sub-Perceptual Tryptamine Pharmacokinetics, 5-HT2A Receptor Desensitisation and Neuroplastic Window Dynamics: From Michaelis–Menten Kinetics of Prodrug Deacetylation and Whole-Brain Connectivity Entropy to the Microdose Tolerance Ratchet, with a Textual Examination of the Dose Boundary in the Classical Materia Medica [亚知觉色胺药代动力学、5-HT2A 受体脱敏与神经可塑性窗口的生物数学模型] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24019347 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).