Skip to main content Skills Marketplace Découvrez et explorez les compétences IA créées par la communauté.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Afficher les détails du prompt Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
npx skills add https://github.com/hiyenwong/ai_collection --skill krylov-mean-field-chaos-predictability-2026-06-10La commande reste sur une seule ligne. Faites défiler horizontalement pour la vérifier avant de la copier.
Vous préférez une copie locale ? Téléchargez les fichiers actuellement disponibles dans SkillsMP.
Télécharger Zip Téléchargement... Métiers associés SOC
Basé sur la classification professionnelle SOC
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"]
Predictable Mean-Field Chaos in Random Recurrent Networks
核心发现
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。
核心定理
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。
理论框架
1. Mean-Field Theory 重释
传统观点
Ensemble description : Mean-field 描述大量网络的平均行为Stochastic approximation : 混沌视为有效随机过程Unpredictable : 无法预测个体轨迹
新观点
Conditional prediction : 历史完全确定未来Deterministic chaos : 随机性只是表象Predictable : 可预测(有限分辨率)
2. Krylov State Space
定义
Krylov space : 由功率谱构造的状态空间Temporal modes : 无限层级的时间模式Growth rate : Krylov expansion 的增长率
数学表述
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||
组织结构
Mode hierarchy : Mode k 对应时间尺度 τ_kInformation encoding : 每个模式编码历史片段Determinism exposure : 层级揭示潜在秩序
3. Lyapunov vs Krylov
传统 Lyapunov Exponent
Definition : λ_L = lim_{t→∞} (1/t) log(||Δx(t)|| / ||Δx(0)||)Mean : 微观敏感性(初始条件敏感性)Unpredictability : 正 Lyapunov → 混沌
新 Krylov Growth Rate
Definition : λ_K = lim_{k→∞} ||v_k|| growthMean : 预测复杂性(预测未来所需信息)Predictability : λ_K < λ_L → 可预测部分
关系
λ_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
数学推导
1. Fourier Decay Condition
Condition : 非线性函数 f(x) 的 Fourier 系数满足
|f_k| ≤ C / k^α for α > 2
Rapid decay → finite approximation
Analytic f → exponential decay
Predictability preserved
2. Conditional Probability Structure P(x(t+Δt) | x(continuous past)) is deterministic
Not ensemble average:
P(x(t+Δt) | statistical ensemble) is stochastic
Continuous past contains infinite information
Fourier coefficients uniquely encode history
Future determined by Fourier representation
3. Krylov Construction def build_krylov_space (trajectory, time_steps ):
"""
Construct Krylov basis from trajectory
"""
Krylov_basis = []
for k in range (infinite):
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
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
实验验证
1. Simulation Protocol
Network : N=1000 neurons, random connectivityNonlinearity : tanh, sigmoid (analytic with fast decay)Measurement :
Lyapunov exponents (standard methods)
Krylov growth (spectrum-based)
Predictability (conditional probability)
2. Results
λ_Lyapunov ≈ 0.8 (chaotic regime)λ_Krylov ≈ 0.3 (predictable structure)λ_K < λ_L (confirmed bound)Conditional prediction accuracy : >90% (finite resolution)
3. Comparative Tests 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
理论贡献
1. 重新定义混沌
Old : Chaos = unpredictable randomnessNew : Chaos = deterministic structure with two metrics
Sensitivity (Lyapunov)
Predictability (Krylov)
2. Hamiltonian → Dissipative
Hamiltonian systems : Krylov methods establishedNeural networks : First extension to dissipative chaosBridge : Classical chaos theory ↔ neural dynamics
3. Spectral Predictability
Power spectrum → Predictability :
Spectral shape encodes determinism
Decay rate ↔ predictability
Mode hierarchy ↔ information organization
应用场景
1. Neural Network Design
Activation selection : Choose analytic functions (tanh > ReLU)Predictability engineering : Optimize spectral decayChaos control : Balance sensitivity vs predictability
2. Cognitive Dynamics
Brain chaos : Measure Krylov growth in neural recordingsPredictability hypothesis : Brain exploits λ_K < λ_L structureMemory encoding : Temporal modes as memory traces
3. AI Chaos Analysis
RNN training : Monitor Lyapunov vs Krylov during learningGeneralization : Predictable chaos → better transferRobustness : Sensitivity ≠ unpredictability
方法论工具
1. Krylov Spectrum Analyzer def analyze_network_predictability (network, trajectory_length ):
"""
Measure Krylov-Lyapunov structure
"""
lyapunov = compute_lyapunov(network, trajectory_length)
spectrum = compute_power_spectrum(network.output)
krylov = build_krylov_space(spectrum)
krylov_growth = measure_krylov_growth(krylov)
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
}
2. Spectral Decay Tester def test_fourier_decay (activation_function ):
"""
Verify predictability condition
"""
x_samples = linspace(-10 , 10 , 1000 )
f_values = activation_function(x_samples)
fourier_coeffs = fft(f_values)
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'
}
3. Conditional Prediction Validator def validate_predictability (network, history_length, prediction_window ):
"""
Test if history determines future
"""
trajectories = generate_trajectories(network, N=1000 )
predictions = []
for traj in trajectories:
history = traj[:history_length]
predicted = predict_from_history(history, network)
actual = traj[history_length:history_length + prediction_window]
error = norm(predicted - actual)
predictions.append(error)
mean_error = mean(predictions)
is_predictable = mean_error < tolerance
return {
'prediction_error' : mean_error,
'is_predictable' : is_predictable
}
神经科学启示
1. Brain Chaos Measurement
Hypothesis : Brain exhibits predictable chaos (λ_K < λ_L)Method :
Record neural activity (fMRI, EEG, spiking)
Compute Lyapunov exponents
Build Krylov space from spectral data
Measure predictability ratio
Expected : λ_K / λ_L ≈ 0.3-0.5 in cognitive regions
2. Learning Dynamics
Before learning : λ_L high, λ_K ≈ λ_L (unpredictable)During learning : λ_K decreases (structure emerges)After learning : λ_K << λ_L (predictable)Interpretation : Learning builds Krylov structure
3. Memory Encoding
Temporal modes : Krylov basis vectorsMemory retrieval : Traverse Krylov hierarchyCapacity : Number of usable Krylov modesDecay : Krylov growth → memory fading
与其他理论关联 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
数学附录
A. Fourier Decay Proof Claim : If |f_k| ≤ C/k^α (α>2), trajectory is predictable.
Finite Fourier approximation: f_N(x) = Σ_{k=1}^N f_k e^{ikx}
Error bound: |f(x) - f_N(x)| ≤ Σ_{k>N} C/k^α = O(1/N^{α-1})
For α>2: error → 0 rapidly
History → Fourier coefficients → Future (unique reconstruction)
B. Krylov Bound Derivation Claim : λ_Krylov ≤ λ_Lyapunov
Krylov vectors: v_k = x(kt)
Growth: ||v_k|| ≤ ||x(0)|| e^{λ_L kt}
Lyapunov by definition: ||x(t)|| ≤ ||x(0)|| e^{λ_L t}
Therefore: λ_K = lim log(||v_k||)/k ≤ λ_L
开放问题
Non-analytic functions : ReLU networks 的 predictability?Finite resolution : 实际预测需要多少 Krylov modes?Noise robustness : 噪声如何影响 λ_K?Multi-scale networks : 不同尺度的 Krylov 结构?
引用 @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}
}
研究启发
Predictable RNNs : 设计具有 λ_K << λ_L 的网络Spectral learning : 通过功率谱优化网络结构Krylov memories : 使用时间模式作为记忆表征Chaos measurement : 区分敏感性与不可预测性
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。
