5D Library › 五维灵商量化架构、先验条件化生成与检索增强认知体 · English

第一章 五维还是一维:灵商量表的特征值谱与信息增益

一个老问题

把一种能力拆成几个维度来测,是心理测量学从一开始就有的争论。Spearman 认为各种智力测验之间普遍正相关,背后有一个共同的「一般因素」(Spearman 1904);后来的验证性因子分析则允许研究者预先指定几个维度及其相关,再用数据检验 (Jöreskog 1969)。判断「有几个维度」的一个常用经验法则是数相关矩阵中大于 1 的特征值(Kaiser 准则)(Kaiser 1960)。

作者的灵商(SQ)框架有五个维度:C 生理代谢(引擎 I)、D 神经情绪(引擎 II)、G 博弈契约(引擎 III)、F 周期风险(引擎 IV)、E 执行成长(引擎 V)。提纲认为它们正交,五维比一维解释更多方差、携带更多信息。本章检验这一主张在什么条件下成立。

特征值谱

设一份 20 题的问卷,每个维度 4 题,每题在自己的维度上的载荷为 0.7;任意两个维度之间的相关为 ϕ\phi。题目的相关矩阵为 R=LΦL′+ΨR = L \Phi L' + \Psi。程序精确地构造这个矩阵,并用 Jacobi 旋转求出全部特征值 (Golub and Van Loan 2013)。

推导

# Chapter 1. p01: Five dimensions or one? The spectrum of a 20-item scale.
# A spiritual-quotient (SQ) questionnaire has 20 items, 4 for each of
# five dimensions: C physiology and metabolism (engine I), D neural
# emotion (II), G game contracts (III), F cycle risk (IV) and E execution
# and growth (V). Each item loads 0.7 on its own dimension (a CFA model):
#     R = L Phi L' + Psi,
# with every pair of dimensions correlated phi. For phi = 0, 0.3, 0.6
# and 0.9 the program computes the item correlation matrix exactly and
# its eigenvalues by Jacobi rotations. Reported: the six largest, the
# share of variance in the first component and in the first five, and
# how many eigenvalues exceed 1 (the Kaiser count).

LOAD, ITEMS, DIMS = 0.7, 4, 5

def corr_matrix(phi):
    n = ITEMS * DIMS
    r = [[0.0] * n for _ in range(n)]
    for i in range(n):
        for j in range(n):
            if i == j:
                r[i][j] = 1.0
            else:
                same = i // ITEMS == j // ITEMS
                r[i][j] = LOAD * LOAD * (1.0 if same else phi)
    return r

def jacobi_eigs(m, sweeps=60):
    n = len(m)
    a = [row[:] for row in m]
    for _ in range(sweeps):
        off = sum(a[i][j] ** 2 for i in range(n) for j in range(n) if i != j)
        if off < 1e-20:
            break
        for p in range(n - 1):
            for q in range(p + 1, n):
                if abs(a[p][q]) < 1e-15:
                    continue
                th = (a[q][q] - a[p][p]) / (2 * a[p][q])
                t = (1 if th >= 0 else -1) / (abs(th) + (th * th + 1) ** 0.5)
                c = 1 / (t * t + 1) ** 0.5
                s = t * c
                for k in range(n):
                    akp, akq = a[k][p], a[k][q]
                    a[k][p], a[k][q] = c * akp - s * akq, s * akp + c * akq
                for k in range(n):
                    apk, aqk = a[p][k], a[q][k]
                    a[p][k], a[q][k] = c * apk - s * aqk, s * apk + c * aqk
    return sorted((a[i][i] for i in range(n)), reverse=True)

if __name__ == "__main__":
    n = ITEMS * DIMS
    print("   phi  " + "six largest eigenvalues".ljust(35)
          + "    1st    1-5  >1")
    for phi in (0.0, 0.3, 0.6, 0.9):
        ev = jacobi_eigs(corr_matrix(phi))
        print("   {:.1f}  {}  {:.3f}  {:.3f}  {}".format(
            phi, " ".join("{:5.2f}".format(e) for e in ev[:6]),
            ev[0] / n, sum(ev[:5]) / n, sum(e > 1 for e in ev)))
    print("\nReading: with orthogonal dimensions five components carry")
    print("the common variance and one carries a fifth of it. As the")
    print("dimensions correlate, the first component swallows the rest")
    print("and the Kaiser count drops from 5 to 1: whether SQ needs five")
    print("numbers or one is an empirical question about phi, not a")
    print("matter of design.")
#    phi  six largest eigenvalues                1st    1-5  >1
#    0.0   2.47  2.47  2.47  2.47  2.47  0.51  0.124  0.618  5
#    0.3   4.82  1.88  1.88  1.88  1.88  0.51  0.241  0.618  5
#    0.6   7.17  1.29  1.29  1.29  1.29  0.51  0.359  0.618  5
#    0.9   9.53  0.71  0.71  0.71  0.71  0.51  0.476  0.618  1
#
# Reading: with orthogonal dimensions five components carry
# the common variance and one carries a fifth of it. As the
# dimensions correlate, the first component swallows the rest
# and the Kaiser count drops from 5 to 1: whether SQ needs five
# numbers or one is an empirical question about phi, not a
# matter of design.
四种维度相关下的前十个特征值。虚线为 1。

维度正交(ϕ=0\phi = 0)时,前五个特征值相等(2.47),第一主成分只占全部方差的 0.124,五个成分合占 0.618;Kaiser 准则数出 5 个维度。维度相关升高时,第一个特征值迅速变大:ϕ=0.6\phi = 0.6 时为 7.17,占 0.359;ϕ=0.9\phi = 0.9 时为 9.53,其余四个降到 0.71,Kaiser 准则只数出 1 个维度。五个维度合计解释的方差(0.618)在四种情形下都不变,变的是这些方差在「一个共同因素」与「五个独立维度」之间如何分配。

所以「五维正交」不是设计出来的,而是一个关于 ϕ\phi 的经验问题:如果真实数据中五个维度相关很高,五个分数实际上是一个分数的五种写法。

五个分数比一个总分多告诉我们多少

即使维度相关,五个分数也可能比一个总分更有用,这取决于我们要预测什么。程序设一个结果变量 yy(例如日后的某种生活指标),它由五个真实维度决定一半的方差,并比较三种预测方式:五个维度分数之和(总分)、最好的单个维度分数、全部五个分数。对联合正态的变量,一个预测携带的关于 yy 的信息为 −12log⁡2(1−R2)-\tfrac{1}{2}\log_2(1-R^2) 比特 (Shannon 1948; Cover and Thomas 2006)。

推导

# Chapter 1. p02: What five scores add over one: R-squared and bits.
# Same scale as p01 (5 dimensions x 4 items, loading 0.7, dimension
# correlation phi). Each dimension score s_k is the sum of its 4 items.
# An outcome y (some later life measure) depends on the true dimensions,
# y = sum_k w_k f_k + e, scaled so that the dimensions explain half of
# its variance. Two cases: SPREAD (equal weights on all five) and ONE
# (only F, cycle risk, matters). Exact covariance algebra gives the
# share of y explained by the total score T = sum_k s_k, by the best
# single dimension score, and by all five scores (R2 = c' S^-1 c). For
# jointly Gaussian variables the information an estimate carries about
# y is I = -1/2 log2(1 - R2) bits.
import math

LOAD, ITEMS, DIMS = 0.7, 4, 5

def solve(a, b):
    n = len(a)
    m = [row[:] + [b[i]] for i, row in enumerate(a)]
    for c in range(n):
        p = max(range(c, n), key=lambda r: abs(m[r][c]))
        m[c], m[p] = m[p], m[c]
        for r in range(n):
            if r != c:
                f = m[r][c] / m[c][c]
                m[r] = [x - f * y for x, y in zip(m[r], m[c])]
    return [m[i][n] / m[i][i] for i in range(n)]

def case(phi, weights):
    Phi = [[1.0 if i == j else phi for j in range(DIMS)] for i in range(DIMS)]
    v = sum(weights[i] * weights[j] * Phi[i][j]
            for i in range(DIMS) for j in range(DIMS))
    w = [x * math.sqrt(0.5 / v) for x in weights]
    S = [[ITEMS + ITEMS * (ITEMS - 1) * LOAD ** 2 if i == j
          else ITEMS * ITEMS * LOAD ** 2 * phi for j in range(DIMS)]
         for i in range(DIMS)]
    c = [ITEMS * LOAD * sum(w[j] * Phi[j][k] for j in range(DIMS))
         for k in range(DIMS)]
    r5 = sum(ci * xi for ci, xi in zip(c, solve(S, c)))
    r1 = sum(c) ** 2 / sum(sum(row) for row in S)
    rb = max(c[k] ** 2 / S[k][k] for k in range(DIMS))
    return r1, rb, r5

def bits(r2):
    return -0.5 * math.log2(1 - r2)

if __name__ == "__main__":
    print("                    R2                    bits")
    print("   case    phi   total  best one  five    total  best one  five")
    for name, wts in (("SPREAD", [1, 1, 1, 1, 1]), ("ONE", [0, 0, 0, 1, 0])):
        for phi in (0.0, 0.3, 0.6, 0.9):
            r1, rb, r5 = case(phi, wts)
            print("   {:6s} {:4.1f} {:7.4f} {:9.4f} {:6.4f} {:8.4f} {:9.4f}"
                  " {:6.4f}".format(name, phi, r1, rb, r5, bits(r1),
                                    bits(rb), bits(r5)))
    print("\nReading: no single number does both jobs. When the outcome")
    print("draws on all five dimensions, the total matches the five")
    print("scores and any one dimension falls far short; when it hangs")
    print("on one dimension, that dimension matches them and the total")
    print("dilutes it. Five scores are the only choice that is never")
    print("worse, yet nothing here exceeds half a bit.")
#                     R2                    bits
#    case    phi   total  best one  five    total  best one  five
#    SPREAD  0.0  0.3968    0.0794 0.3968   0.3646    0.0596 0.3646
#    SPREAD  0.3  0.4471    0.1746 0.4471   0.4275    0.1384 0.4275
#    SPREAD  0.6  0.4645    0.2698 0.4645   0.4505    0.2268 0.4505
#    SPREAD  0.9  0.4732    0.3650 0.4732   0.4624    0.3276 0.4624
#    ONE     0.0  0.0794    0.3968 0.3968   0.0596    0.3646 0.3646
#    ONE     0.3  0.1967    0.3968 0.4009   0.1580    0.3646 0.3695
#    ONE     0.6  0.3158    0.3968 0.4128   0.2738    0.3646 0.3840
#    ONE     0.9  0.4354    0.3968 0.4465   0.4123    0.3646 0.4266
#
# Reading: no single number does both jobs. When the outcome
# draws on all five dimensions, the total matches the five
# scores and any one dimension falls far short; when it hangs
# on one dimension, that dimension matches them and the total
# dilutes it. Five scores are the only choice that is never
# worse, yet nothing here exceeds half a bit.

两种情形给出相反的结论。若 yy 平均地依赖全部五维(SPREAD),总分与五个分数一样好(维度正交时 R2R^2 都是 0.3968),单个维度只有 0.0794。若 yy 只依赖一个维度(ONE),那个维度与五个分数一样好(0.3968),总分被其余四维稀释到 0.0794。没有任何一个单一数字能同时胜任两种情形;五个分数是唯一「从不更差」的选择。但所有的增益都不大:信息量最高的一格只有 0.4624 比特,不到半个比特。

假说 · 尚未验证

作者的诠释(尚未验证): 五个灵商维度在真实人群中的相关不高于 0.3,而且不同的人生结果分别依赖不同的维度(例如健康结果主要依赖 C,人际结果主要依赖 G),因此五维报告比一个总分更有用。什么会证明这一判断错:若在一个预先登记的样本中,五个维度分数的平均相关超过 0.6,或者各种结果都主要依赖总分(五个分数的增量 R2R^2 几乎为零),那么五维报告与一个总分没有实质区别。

小结

五个维度是否「正交」,要看数据中的维度相关:相关为 0.9 时,20 题的特征值只剩 1 个大于 1。五个分数相对于一个总分的优势是「从不更差」:结果依赖全部维度时与总分相同,依赖单个维度时远胜总分。但增益都在半个比特以内。

参考文献

Cover, Thomas M., and Joy A. Thomas. 2006. Elements of Information Theory. 2nd ed. Wiley.
Golub, Gene H., and Charles F. Van Loan. 2013. Matrix Computations. 4th ed. Johns Hopkins University Press.
Jöreskog, Karl G. 1969. “A General Approach to Confirmatory Maximum Likelihood Factor Analysis.” Psychometrika 34 (2): 183–202.
Kaiser, Henry F. 1960. “The Application of Electronic Computers to Factor Analysis.” Educational and Psychological Measurement 20 (1): 141–51.
Shannon, Claude E. 1948. “A Mathematical Theory of Communication.” Bell System Technical Journal 27 (3): 379–423.
Spearman, Charles. 1904. “‘General Intelligence,’ Objectively Determined and Measured.” American Journal of Psychology 15 (2): 201–92.

Fan, Y. (2026). Spiritual Quotient Quantification, Prior-Conditioned Generation and Retrieval-Augmented Cognitive Agents: Multidimensional State Vectors, Personalized Prompt Conditioning and Cross-Corpus Synthesis [五维灵商量化架构、先验条件化生成与检索增强认知体] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090203 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).