惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

Blog — PlanetScale
Blog — PlanetScale
B
Blog
A
About on SuperTechFans
大猫的无限游戏
大猫的无限游戏
爱范儿
爱范儿
OSCHINA 社区最新新闻
OSCHINA 社区最新新闻
H
Help Net Security
H
Hackread – Cybersecurity News, Data Breaches, AI and More
博客园 - 三生石上(FineUI控件)
有赞技术团队
有赞技术团队
酷 壳 – CoolShell
酷 壳 – CoolShell
WordPress大学
WordPress大学
IT之家
IT之家
D
Docker
Google DeepMind News
Google DeepMind News
罗磊的独立博客
T
The Blog of Author Tim Ferriss
aimingoo的专栏
aimingoo的专栏
博客园 - 叶小钗
Recent Announcements
Recent Announcements
阮一峰的网络日志
阮一峰的网络日志
D
DataBreaches.Net
博客园 - 司徒正美
Engineering at Meta
Engineering at Meta

cs.IT updates on arXiv.org

Theoretical Limits of Language Model Alignment $f$-Divergence Regularized RLHF: Two Tales of Sampling and Unified Analyses A Unified Measure-Theoretic View of Diffusion, Score-Based, and Flow Matching Generative Models When Can Voting Help, Hurt, or Change Course? Exact Structure of Binary Test-Time Aggregation When Semantic Communication Meets Queueing: Cross-Layer Latency and Task Fidelity Optimization Convexity in Disguise: A Theoretical Framework for Nonconvex Low-Rank Matrix Estimation Conditional Diffusion Under Linear Constraints: Langevin Mixing and Information-Theoretic Guarantees Sharp Capacity Thresholds in Linear Associative Memory: From Winner-Take-All to Listwise Retrieval Expert Routing for Communication-Efficient MoE via Finite Expert Banks Contextual Memory-Enhanced Source Coding for Low-SNR Communications Realizable Bayes-Consistency for General Metric Losses Leveraging Code Automorphisms for Improved Syndrome-Based Neural Decoding A Hierarchical Sampling Framework for bounding the Generalization Error of Federated Learning Dueling DDQN-Based Adaptive Multi-Objective Handover Optimization for LEO Satellite Networks The Causal Description Gap: Information-Theoretic Separations Across Pearl's Hierarchy Optimization of CV-QKD Under Practical Constraints Benchmarking Wireless Representations: High-Dimensional vs. Compressed Embeddings for Efficiency and Robustness Real-Time Text Transmission via LLM-Based Entropy Coding over Fixed-Rate Channels SwiftChannel: Algorithm-Hardware Co-Design for Deep Learning-Based 5G Channel Estimation Evolving Token Communication with Parametric Memory Network Remote Action Generation: Remote Control with Minimal Communication The (Marginal) Value of a Search Ad: An Online Causal Framework for Repeated Second-price Auctions Stabilizing Private LASSO under Heterogeneous Covariates via Anisotropic Objective Perturbation Linear-Readout Floors and Threshold Recovery in Computation in Superposition Soft Graph Diffusion Transformer for MIMO Detection Hierarchical Federated Learning for Networked AI: From Communication Saving to Architecture-Aware Design Exponential families from a single KL identity MIFair: A Mutual-Information Framework for Intersectionality and Multiclass Fairness Diffusion-OAMP for Joint Image Compression and Wireless Transmission Decoupled Descent: Exact Test Error Tracking Via Approximate Message Passing
Relaxation Kernel, Spectral Dissipation, and Global Conve...
Qiao Wang · 2026-04-28 · via cs.IT updates on arXiv.org

We develop a spectral theory for continuous- and discrete-time Blahut--Arimoto (BA) dynamics, centered on the relaxation kernel $ \G = \E_p[K^*_X \otimes K^*_X] $. Five main results are established. (i) Along the continuous-time BA flow, the free energy satisfies the exact $ χ^2 $-dissipation identity $ \dot F_β= -\D(q) $, where $ \D(q)=χ^2(\T q \| q) $ is the Pearson $ χ^2 $-divergence. (ii) The operator $ \G $ admits a threefold identity: it is simultaneously the Gram matrix of the equilibrium Gibbs kernels, the linearised generator of the BA vector field, and the Fisher--Rao Hessian of the free energy at the fixed point. (iii) For the discrete iteration, the one-step Lyapunov dissipation decomposes spectrally as $ Δ\mathcal{L}^{(2)} = \sum_i c_i^2\, d(λ_i) $, where $ d(λ) = -λ+ \tfrac{3}{2}λ^2 - \tfrac{1}{2}λ^3 $. This reveals a double bottleneck at $ λ\approx 0 $ and $ λ\approx 1 $, with optimal dissipation near $ λ\approx 0.423 $. (iv) Global convergence follows a two-stage mechanism: $ χ^2 $-dissipation drives finite-time entry into a local neighbourhood, after which the spectral gap $ \lam = λ_{\min}(\G|_T) $ governs exponential contraction. (v) The KL convergence factor is explicit: $ \KL(q^*\|q_{n+1}) \le (1-\lam)^2\,\KL(q^*\|q_n) + O(\|v_n\|_*^3) $, with per-iteration improvement $ γ= \lam(2-\lam) $. For Gaussian sources, $ \lam = 1/(2βσ^2) $ and the Jacobian is diagonalised by Hermite polynomials. The spectral formula complements Hayashi's global convergence theory with a constructive, computable local rate.