القائمة
Mean-field chaos 的预测性理论框架。证明随机循环网络的确定性混沌可通过连续历史唯一预测未来,展开功率谱到 Krylov 状态空间暴露潜在确定性组织。区分微观敏感性和预测复杂性。
يُحدَّد Prompt الذي يراجع المصدر أولًا بشكل افتراضي. يمكنك التبديل إلى أمر مباشر أو تنزيل نسخة محلية.
اقرأ SKILL.md وأي ملفات مرافقة يعرضها SkillsMP قبل أن تقرر التثبيت.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
يتجاوز الأمر المباشر Prompt المخصّص للمراجعة. افحص المصدر قبل تشغيله.
npx skills add https://github.com/hiyenwong/ai_collection --skill krylov-mean-field-chaos-predictability-2026-06-10يبقى الأمر في سطر واحد. مرّر أفقيًا لمراجعته كاملًا قبل النسخ.
تفضّل نسخة محلية؟ نزّل الملفات المتاحة حاليًا لدى SkillsMP.
Graph-native Python reimplementation of the Information Dynamics of Music (IDyOM) model that represents predictive memories as explicit graph objects for musical expectation modeling and network analysis.
Physics-aware end-to-end deep reinforcement learning methodology for quadcopter control with actuator dynamics modeling.
Reinforced Dreamer methodology for asymmetric reinforcement learning using latent guidance to improve world model representations and behaviors in model-based RL.
استنادا إلى تصنيف SOC المهني
عرض SKILL.md
| name | krylov-mean-field-chaos-predictability-2026-06-10 |
| description | Mean-field chaos 的预测性理论框架。证明随机循环网络的确定性混沌可通过连续历史唯一预测未来,展开功率谱到 Krylov 状态空间暴露潜在确定性组织。区分微观敏感性和预测复杂性。 |
| version | 1 |
| arxiv_id | 2606.08805 |
| authors | Alkesh Yadav, Vladimir Shaidurov, Jonathan Kadmon |
| submission_date | 2026-06-07T00:00:00.000Z |
| tags | ["mean-field-theory","recurrent-networks","chaos","deterministic-dynamics","Krylov-methods","Lyapunov-exponent","spectral-analysis","neural-dynamics","computational-neuroscience"] |
| activation_keywords | ["mean-field chaos","predictable chaos","RNN dynamics","Lyapunov exponent","Krylov subspace","spectral decomposition","random recurrent networks","deterministic prediction","temporal modes","Hamiltonian chaos"] |
关键洞见: Mean-field theory 不仅是对网络的 ensemble 描述,更是对个体轨迹的条件预测理论。
定理 1 (Predictability): 对于具有足够快 Fourier 衰减的解析非线性函数,mean-field trajectory 的连续过去唯一确定其未来。
定理 2 (Krylov Structure): 将功率谱展开到 Krylov 状态空间,揭示潜在确定性在无限时间模式层级中的组织方式。
定理 3 (Complexity Bound): Krylov growth rate 设定有限分辨率预测的复杂性,并上界该类网络的 Lyapunov exponent。
Power spectrum: P(ω) = ⟨|x(t)|²⟩_ω
Krylov basis: {v_k} generated from x(0), x(t), x(2t), ...
Growth: λ_Krylov = lim_{k→∞} ||v_k|| / ||v_0||
λ_L = lim_{t→∞} (1/t) log(||Δx(t)|| / ||Δx(0)||)λ_K = lim_{k→∞} ||v_k|| growthλ_Krylov ≤ λ_Lyapunov (Theorem 3)
Interpretation:
- λ_K: 预测复杂性 (需要多少历史信息)
- λ_L: 微观敏感性 (初始误差增长)
- λ_K < λ_L: 混沌有可预测结构
Condition: 非线性函数 f(x) 的 Fourier 系数满足
|f_k| ≤ C / k^α for α > 2
Implication:
Key insight:
P(x(t+Δt) | x(continuous past)) is deterministic
Not ensemble average:
P(x(t+Δt) | statistical ensemble) is stochastic
Reason:
Algorithm:
def build_krylov_space(trajectory, time_steps):
"""
Construct Krylov basis from trajectory
"""
Krylov_basis = []
for k in range(infinite):
# Gram-Schmidt orthogonalization
v_k = trajectory(k * dt)
for j in range(k):
v_k -= dot(v_j, v_k) * v_j
v_k /= norm(v_k)
Krylov_basis.append(v_k)
return Krylov_basis
Growth rate calculation:
def krylov_growth_rate(Krylov_basis):
"""
Measure expansion rate of Krylov space
"""
norms = [norm(v_k) for v_k in Krylov_basis]
growth = log(norms[-1] / norms[0]) / len(norms)
return growth
| Function | Fourier Decay | Predictable? | λ_K / λ_L |
|---|---|---|---|
| tanh | Exponential | ✓ | 0.4 |
| sigmoid | Exponential | ✓ | 0.35 |
| ReLU | Slow (α=1) | ✗ | 1.0 |
| Piecewise | Zero | ✗ | 1.0 |
def analyze_network_predictability(network, trajectory_length):
"""
Measure Krylov-Lyapunov structure
"""
# 1. Compute Lyapunov exponents
lyapunov = compute_lyapunov(network, trajectory_length)
# 2. Extract power spectrum
spectrum = compute_power_spectrum(network.output)
# 3. Build Krylov space
krylov = build_krylov_space(spectrum)
# 4. Measure growth
krylov_growth = measure_krylov_growth(krylov)
# 5. Compare
predictability_ratio = krylov_growth / lyapunov
is_predictable = predictability_ratio < 0.9
return {
'lyapunov': lyapunov,
'krylov_growth': krylov_growth,
'predictability_ratio': predictability_ratio,
'is_predictable': is_predictable
}
def test_fourier_decay(activation_function):
"""
Verify predictability condition
"""
# Sample function
x_samples = linspace(-10, 10, 1000)
f_values = activation_function(x_samples)
# Compute Fourier coefficients
fourier_coeffs = fft(f_values)
# Check decay
decay_rate = measure_decay_rate(fourier_coeffs)
is_fast = decay_rate > 2
return {
'decay_rate': decay_rate,
'is_predictable': is_fast,
'recommendation': 'Use for predictable chaos' if is_fast else 'Avoid for deterministic prediction'
}
def validate_predictability(network, history_length, prediction_window):
"""
Test if history determines future
"""
# Generate many trajectories
trajectories = generate_trajectories(network, N=1000)
# For each trajectory
predictions = []
for traj in trajectories:
# Extract history
history = traj[:history_length]
# Predict future
predicted = predict_from_history(history, network)
# Compare with actual
actual = traj[history_length:history_length + prediction_window]
# Measure error
error = norm(predicted - actual)
predictions.append(error)
# Statistical test
mean_error = mean(predictions)
is_predictable = mean_error < tolerance
return {
'prediction_error': mean_error,
'is_predictable': is_predictable
}
| Theory | Focus | Metric | Relation |
|---|---|---|---|
| Chaos theory | Sensitivity | Lyapunov λ_L | λ_L measures divergence |
| Krylov theory | Predictability | Growth λ_K | λ_K bounds complexity |
| Attractor theory | Stability | Basin size | Complement: structure vs basin |
| Mean-field theory | Ensemble | Statistics | Extended: ensemble → conditional |
Claim: If |f_k| ≤ C/k^α (α>2), trajectory is predictable.
Proof:
f_N(x) = Σ_{k=1}^N f_k e^{ikx}|f(x) - f_N(x)| ≤ Σ_{k>N} C/k^α = O(1/N^{α-1})Claim: λ_Krylov ≤ λ_Lyapunov
Proof:
v_k = x(kt)||v_k|| ≤ ||x(0)|| e^{λ_L kt}||x(t)|| ≤ ||x(0)|| e^{λ_L t}λ_K = lim log(||v_k||)/k ≤ λ_L@article{yadav2026predictable,
title={Predictable Mean-Field Chaos in Random Recurrent Networks},
author={Yadav, Alkesh and Shaidurov, Vladimir and Kadmon, Jonathan},
journal={arXiv preprint arXiv:2606.08805},
year={2026}
}
Activation: 在讨论 RNN chaos, mean-field theory, Lyapunov exponents, 神经网络动力学, spectral analysis, 或混沌可预测性时激活此 skill。