5D Library › Power-Law Weighting and Bottom-Up Controllability in Multi-Tier Conditional Excitation Systems · 中文
Chapter 5 Bottom-Up Controllability
From a pyramid to a control system
The second direction in the author’s framework is bottom-up: knowledge investment is the only input one can apply directly. Can it, through the couplings between tiers, push upwards tier by tier and finally “unlock the top in reverse”? Control theory offers a precise language for such questions. This chapter writes the five tiers as a linear control system and asks three questions: which tiers can the input reach? How much “energy” does reaching them take? What do the answers depend on?
Calculation
The author’s model (linearized). Let be the deviations of the five tiers from some baseline and the input applied to :
The matrix describes the couplings: means that a change in tier pushes tier . The chapter compares three coupling structures (all numbers illustrative):
- TD (top-down only): tier pushes tier with gain 0.5; tiers 2–5 each decay at rate 1; does not change.
- UP (with bottom-up feedback): as TD, plus tier pushes tier with gain 0.4 (): knowledge improves reputation, reputation improves one’s network, the network brings better timing; still does not change.
- PL (“plastic prior”): as UP, plus with gain and decay 0.1 in . This corresponds to the assumption that the prior itself can be slowly rewritten by experience.
“ does not change” is the author’s own assumption: the endogenous prior is an unchangeable quantity. As we shall see, that assumption alone settles the answer mathematically.
Kalman’s rank condition
Evidence
Controllability. Kalman showed in 1960 that the linear system can be steered from any initial state to any final state in finite time (complete controllability) if and only if the controllability matrix has rank equal to the state dimension (Kalman 1960). The equivalent Popov–Belevitch–Hautus (PBH) test says there is no left eigenvector with and . Sontag’s textbook gives a systematic account (Sontag 1998). Lin proved “structural controllability”: controllability depends largely on which entries of and are non-zero rather than on their values, and almost all values with the same zero pattern give the same rank (Lin 1974). Liu, Slotine and Barabási applied these ideas to complex networks, asking how many driver nodes are needed to control a whole network (Liu et al. 2011).
Calculation
Proposition ( is uncontrollable under the author’s assumption). If the first row of is zero ( is influenced by no tier and does not decay) and the first entry of is zero (the input does not act on directly), then for every input .
Proof. , so is constant. In the language of the PBH test, and , so is an uncontrollable mode.
The proposition holds whatever the strength of the couplings and however the other tiers are coupled: as long as “endowment cannot change”, it is uncontrollable in the model. The program checks the three structures with the Kalman matrix.
# Bottom-up controllability of the five layers (author's model, Chapter 5).
# x' = A x + B u, x = deviations of layers C1..C5, one input u on C5: B = e5.
# Couplings (illustrative): TD C_i -> C_(i+1) gain 0.5, decay 1 in layers 2-5, C1 frozen;
# UP TD plus feedback C_(i+1) -> C_i gain 0.4 for i = 2, 3, 4, C1 still frozen;
# PL UP plus a "plastic prior": C2 -> C1 gain eps = 0.05, decay 0.1 in C1.
# Kalman: controllable iff rank [B, AB, A^2 B, A^3 B, A^4 B] = 5. PBH test for C1: if
# e1' A = 0 and e1' B = 0, then c1' = 0 whatever u does, so C1 cannot be moved.
def structure(name, k=0.5, h=0.4, eps=0.05, d1=0.1):
A = [[0.0] * 5 for _ in range(5)]
for i in range(1, 5): A[i][i] = -1.0
for i in range(4): A[i + 1][i] = k
if name in ("UP", "PL"):
for i in range(1, 4): A[i][i + 1] = h
if name == "PL": A[0][0], A[0][1] = -d1, eps
return A
def rank(cols, tol=1e-9):
M = [c[:] for c in cols]; r = 0; scale = max(abs(x) for c in M for x in c)
for j in range(len(M[0])):
p = max(range(r, len(M)), key=lambda i: abs(M[i][j]), default=None)
if p is None or abs(M[p][j]) < tol * scale: continue
M[r], M[p] = M[p], M[r]
for i in range(len(M)):
if i != r: M[i] = [a - M[i][j] / M[r][j] * b for a, b in zip(M[i], M[r])]
r += 1
return r
def krylov(A, b):
cols = [b]
for _ in range(4):
cols.append([sum(a * v for a, v in zip(row, cols[-1])) for row in A])
return cols
if __name__ == "__main__":
B = [0, 0, 0, 0, 1.0]
for name in ("TD", "UP", "PL"):
A = structure(name); K = krylov(A, B)
reach = [f"C{i + 1}" for i in range(5) if any(abs(c[i]) > 1e-12 for c in K)]
print(f"{name}: Kalman rank {rank(K)}; reached by u: {', '.join(reach)}")
print(f" e1'A = ({', '.join(f'{x:g}' for x in A[0])}), e1'B = {B[0]:g}"
+ (" -> C1 is an uncontrollable mode" if not any(A[0]) else ""))
print(" (A^j B)_1 for j = 0..4: " + " ".join(f"{c[0]:.4f}" for c in K))
# TD: Kalman rank 1; reached by u: C5
# e1'A = (0, 0, 0, 0, 0), e1'B = 0 -> C1 is an uncontrollable mode
# (A^j B)_1 for j = 0..4: 0.0000 0.0000 0.0000 0.0000 0.0000
# UP: Kalman rank 4; reached by u: C2, C3, C4, C5
# e1'A = (0, 0, 0, 0, 0), e1'B = 0 -> C1 is an uncontrollable mode
# (A^j B)_1 for j = 0..4: 0.0000 0.0000 0.0000 0.0000 0.0000
# PL: Kalman rank 5; reached by u: C1, C2, C3, C4, C5
# e1'A = (-0.1, 0.05, 0, 0, 0), e1'B = 0
# (A^j B)_1 for j = 0..4: 0.0000 0.0000 0.0000 0.0000 0.0032With top-down couplings only (TD), the Kalman matrix has rank 1: the input can change only itself, since no channel carries it upwards. Adding bottom-up feedback (UP) raises the rank to 4: to can be reached, but remains uncontrollable, as the proposition says. Only in PL, that is, assuming that can slowly rewrite the prior, is the rank 5; even then the first entry of is zero until , where it is 0.0032: the input has to pass through four couplings to reach , shrinking at each.
The controllability Gramian and control energy
Rank answers “can it?” but not “how hard?”. The latter is captured by the controllability Gramian
The minimum energy needed to steer the system from the origin to a final state in time is (this requires to be invertible, that is, the system to be completely controllable; when is singular only final states in its range can be reached). If we ask only that tier move by one unit, the other tiers being free, the minimum energy is ; means that the tier cannot be moved at any cost.
Evidence
Control energy and network structure. Yan and colleagues studied the energy needed to control complex networks and found that even when a network is controllable in the rank sense, the energy required can differ by many orders of magnitude with the target direction and the time window (Yan et al. 2012). Pasqualetti, Zampieri and Bullo systematically discussed Gramian-based controllability metrics and their limitations (Pasqualetti et al. 2014).
Calculation
The program computes in two ways: by integrating the differential Lyapunov equation , ; and by first computing (solving ) and then integrating with Simpson’s rule. The two must agree. It then reports the energy needed to move each tier, and how the energy changes with the coupling , the time window and the feedback gain .
# Controllability Gramian and control energy for the couplings of p10 (Chapter 5).
# W(T) = integral_0^T e^(At) B B' e^(A't) dt, computed two ways and compared:
# (1) the differential Lyapunov equation W' = A W + W A' + B B', W(0) = 0 (RK4);
# (2) v(t) = e^(At) B from v' = A v (RK4), then W = integral of v v' (Simpson's rule).
# Least energy (integral of u^2) to move layer i by one unit by T, others free: 1 / W_ii
# (infinite if W_ii = 0). Horizon T = 10; couplings illustrative.
import math
def structure(name, k=0.5, h=0.4, eps=0.05, d1=0.1):
A = [[0.0] * 5 for _ in range(5)]
for i in range(1, 5): A[i][i] = -1.0
for i in range(4): A[i + 1][i] = k
if name in ("UP", "PL"):
for i in range(1, 4): A[i][i + 1] = h
if name == "PL": A[0][0], A[0][1] = -d1, eps
return A
def mm(X, Y): return [[sum(a * b for a, b in zip(r, c)) for c in zip(*Y)] for r in X]
def add(X, Y, s=1.0): return [[a + s * b for a, b in zip(r, q)] for r, q in zip(X, Y)]
def rk4(f, x, h, n):
out = [x]
for _ in range(n):
k1 = f(x); k2 = f(add(x, k1, h / 2)); k3 = f(add(x, k2, h / 2))
k4 = f(add(x, k3, h))
x = add(x, add(add(k1, k4), add(k2, k3), 2.0), h / 6); out.append(x)
return out
def gramian_lyapunov(A, T=10.0, n=2000):
BB = [[1.0 if i == j == 4 else 0.0 for j in range(5)] for i in range(5)]
At = [list(c) for c in zip(*A)]
return rk4(lambda W: add(add(mm(A, W), mm(W, At)), BB), [[0.0] * 5] * 5, T / n, n)[-1]
def gramian_trajectory(A, T=10.0, n=2000):
traj = rk4(lambda v: mm(A, v), [[0.0]] * 4 + [[1.0]], T / n, n)
vs = [[r[0] for r in v] for v in traj]
wt = [1 if i in (0, n) else (4 if i % 2 else 2) for i in range(n + 1)]
return [[sum(w * x[i] * x[j] for w, x in zip(wt, vs)) * T / n / 3 for j in range(5)]
for i in range(5)]
def energies(W): return [1 / W[i][i] if W[i][i] > 1e-14 else math.inf for i in range(5)]
if __name__ == "__main__":
print("energy to move one layer by one unit (T = 10): C1 C2 C3"
" C4 C5")
for name in ("TD", "UP", "PL"):
A = structure(name); W = gramian_lyapunov(A); W2 = gramian_trajectory(A)
gap = max(abs(a - b) for r, q in zip(W, W2) for a, b in zip(r, q))
print(f" {name} (the two methods differ by {gap:.0e}): "
+ " ".join(f"{e:8.3g}" for e in energies(W)))
print("PL: energy to move C1 by one unit, by coupling eps and horizon T")
for eps in (0.05, 0.2):
A = structure("PL", eps=eps)
row = [energies(gramian_lyapunov(A, T=T, n=200 * T))[0] for T in (10, 20, 40)]
print(f" eps = {eps:4.2f}: T = 10: {row[0]:6.0f} T = 20: {row[1]:6.0f}"
f" T = 40: {row[2]:6.0f}")
print("UP, stronger upward feedback h [largest eigenvalue of layers 2-5]")
for h in (0.4, 0.7, 0.9):
lam = -1 + 2 * math.sqrt(0.5 * h) * math.cos(math.pi / 5)
E = energies(gramian_lyapunov(structure("UP", h=h)))
print(f" h = {h}: [{lam:+.3f}] energies C2..C5:"
+ "".join(f"{e:8.3g}" for e in E[1:]))
# energy to move one layer by one unit (T = 10): C1 C2 C3 C4 C5
# TD (the two methods differ by 3e-11): inf inf inf inf 2
# UP (the two methods differ by 4e-11): inf 540 77.6 14.4 1.75
# PL (the two methods differ by 4e-11): 1.18e+04 464 76.1 14.4 1.75
# PL: energy to move C1 by one unit, by coupling eps and horizon T
# eps = 0.05: T = 10: 11819 T = 20: 4448 T = 40: 3445
# eps = 0.20: T = 10: 501 T = 20: 91 T = 40: 17
# UP, stronger upward feedback h [largest eigenvalue of layers 2-5]
# h = 0.4: [-0.276] energies C2..C5: 540 77.6 14.4 1.75
# h = 0.7: [-0.043] energies C2..C5: 5.49 2.33 1.91 1.39
# h = 0.9: [+0.085] energies C2..C5: 0.425 0.267 0.394 0.864The two methods differ by less than . In UP the Gramian is zero in the direction of (); moving by one unit takes energy 1.75, 14.4, 77.6, 540, and is impossible. In PL, becomes controllable but needs about 11,800, some 6,740 times as much as . A longer window or a stronger coupling lowers the energy: with , extending from 10 to 40 lowers it from about 11,800 to about 3,400; with it is about 500 at and only 17 at .
This gives the author’s intuition that “the higher the tier, the harder it is to move” an explanation within the model: the input is attenuated at every coupling, so the further the tier, the more energy it takes. The last three lines of the output show, however, that the ordering depends on the coupling values. Raising the feedback gain from 0.4 to 0.7 lowers the energy for from 540 to 5.49 and shrinks the spread between tiers from more than two orders of magnitude to a factor of four; at 0.9 the subsystem of tiers 2–5 becomes unstable (largest real part of an eigenvalue +0.085) and the ordering even reverses, with upper tiers easier to move than lower ones. An unstable system means deviations amplify themselves without bound, which in reality usually means the model has stopped applying.
Note
Energy is a quantity in the model. Control energy is a mathematical definition, not “hours of effort”. Interpreting it as time, money or strain needs further assumptions and measurements. The conclusions of this chapter are mathematical properties that hold under the stated couplings; they say “if the tiers are coupled like this, then …”, not “this is how reality is”.
Hypothesis · not established
The attenuation hypothesis (the author’s hypothesis, not established). In real careers, sustained investment in a lower tier is passed on to higher tiers with attenuation: it affects most the tier it acts on directly, less the adjacent upper tier, and still less the tiers above. Testable prediction: in studies that randomize some investment (such as training) and pre-define measures of each tier, the effect sizes decline monotonically with distance between tiers. If the effect on a distant tier is as large as on a near one, or does not vary with distance, the hypothesis is rejected. Note that it concerns only –; by the author’s own assumption is uncontrollable and lies outside the test.
Chapter summary
Written as a linear control system, the author’s assumption that “endowment cannot change” implies directly that is an uncontrollable mode: the Gramian is zero in that direction and no input from the bottom can move it. Which tiers an input can reach depends only on the coupling structure: with top-down couplings alone, the input at the bottom goes nowhere; with bottom-up feedback, – are controllable; only under the assumption of a plastic prior is controllable, and then at thousands of times the energy needed for . “The higher, the harder” holds under attenuating couplings but changes with coupling strength and can even reverse under strong feedback. The next chapter separates “potential” from “realization” and asks what the author’s “unlocking” is in the model.
References
Fan, Y. (2026). Power-Law Weighting and Bottom-Up Controllability in Multi-Tier Conditional Excitation Systems: A Mathematical Synthesis of Endogenous Priors, Timing Windows, Network Topology and Reputational Games [多层条件激发系统中的幂律权重与逆向控制可及性] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23810093 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).