PyTorchKR Papers

업데이트된 논문

총 544편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 카테고리 구독

arXiv:2611.24254v2cs.AI

FedE2E: 폐루프 평가를 갖춘 종단간 자율주행을 위한 원샷 연합학습

FedE2E: One-Shot Federated Learning for End-to-End Autonomous Driving with Closed-Loop Evaluation

연합학습은 데이터셋을 공유하지 않고도 협력적 모델 학습을 가능하게 한다. 그러나 자율주행을 위한 기존 연합학습 방법은 일반적으로 다중 라운드 통신을 요구하고 폐루프 평가가 부족하며, 원샷 방법은 주로 인식 과제에 초점을 둔다. 이러한 한계를 해결하기 위해 우리는 종단간 자율주행을 위한 원샷 연합학습 방법 \method를 제안한다. 이 방법은 이질적인 모델로부터 지식을 증류하기 위한 다중 클라이언트 보정, 앵커 기반 특징 정렬, 토큰 단위 합의 증류로 구성된다. Bench2Drive에서의 폐루프 평가는 \method가 원샷 및 다중 라운드 연합학습 모두에서 효과적임을 보여 준다. 원샷 설정에서 Driving Score 55.85를 달성하며, 이는 기존 연구보다 4.07점 높다. 다중 라운드 설정에서는 다섯 번의 통신 라운드 후 60.35에 도달하여, 동일한 통신 라운드 수 조건에서 비교된 모든 방법 중 최고 성능을 보인다.

arXiv:2611.12466v2stat.ML

배포 예산 하에서의 파운데이션 모델 기반 예측

Prediction with Foundation Models under Deployment Budgets

파운데이션 모델은 방대한 사전 학습에서 얻은 지식을 활용하여 하위 과제에 대한 예측을 제공할 수 있으나, 과제별 정확도와 질의 비용이 서로 다르다. 본 연구는 입력당 평균 질의 비용에 대한 예산 하에서 과제의 예측을 개선하는 프레임워크를 개발하며, 학습 데이터를 이용해 어떤 모델에 질의할지, 그리고 그 출력을 질의 없이 사용할 수 있는 특징과 어떻게 결합할지를 학습한다. 첫 번째 방법은 모든 입력에 대해 동일한 모델들에 질의하고, 두 번째 방법은 입력에 따라 질의 모델이 달라지도록 하며 Q-러닝을 사용하여 다음에 질의할 모델과 중단 시점을 결정한다. 첫 번째 방법에 대해, 개별적으로는 부정확한 여러 출력이 함께 사용되면 최선의 단일 출력보다 우수할 수 있음을 보인다. 이 방법은 회귀 함수가 예산 내의 모델 출력에만 의존하는 경우 미니맥스 비율을 달성한다. 그렇지 않은 경우에는 예산 내의 어떤 방법이든 더 느린 비율로 강제될 수 있으며, 첫 번째 방법은 이 비율을 달성한다. 두 번째 방법에 대해, 입력에 따라 질의 모델을 달리하는 것이 모든 입력에 동일한 모델을 질의하는 것보다 엄밀하게 개선되는 경우를 보이고, 유한 표본에서의 예측 위험과 평균 비용에 대한 한계를 제시한다. 시뮬레이션과 멀티모달 노트북 가격 응용에서 두 방법 모두 기존 방법보다 우수한 성능을 보인다.

arXiv:2610.14988v2cs.LG

Mixing Regularization for Average-Reward $Q$-Learning and $Q$-Planning

The average-reward setting for reinforcement learning is challenging because, without discounting, the span-contraction of the Bellman operator depends on the mixing properties of the underlying Markov dynamics. Whereas existing work typically assumes contraction, for instance through environment mixing, we study regularization toward a given reward-agnostic, $m$-step mixing policy to control the mixing properties of the learned policy. Our central mixing lemma quantifies how the reference policy's mixing properties transfer to the resulting learned policies. In particular, a single regularizer-dependent quantity simultaneously controls the contraction of the regularized Bellman operator and the exploration of the resulting policies. A second regularizer-dependent quantity controls the bias introduced by regularization, revealing a trade-off between convergence speed and approximation accuracy. Building on this result, we establish finite-time guarantees for regularized $Q$-iteration, synchronous and single-trajectory $Q$-learning, and $Q$-planning. Under one-step mixing, our bounds for approximating the regularized fixed point match the classical model-query bound for relative value iteration and, up to logarithmic factors, an existing sample-complexity bound for synchronous $Q$-learning. We also establish finite-time guarantees under multi-step mixing. For $Q$-planning, we characterize how the planning horizon trades off boundary bias against sample complexity. Finally, we present numerical studies validating the analysis: (i) knowledge of a strong mixing policy improves convergence and exploration, with larger benefits under more severe bottlenecks; and (ii) the observed rankings of $\alpha$-divergence regularizers by bias and convergence speed are consistent with our theoretical analysis.

arXiv:2610.01935v3econ.EM

AI 학습 표현을 활용한 실용적 DML

Pragmatic DML with AI-Learned Representations

Text, images, and other rich objects increasingly enter causal analysis through AI-learned representations used as controls. For a broad class of estimands, replacing the raw object by its representation shifts the target by the inner product of two errors: one in the outcome regression and one in the balancing weight. The bias is therefore bounded by the product of their sizes. Three consequences follow. With a fixed representation, the usual cross-fitted debiased machine learning (DML) interval is valid for a representation-specific target, and for the causal target when the approximation bias is negligible relative to sampling error. When this bias may not be negligible, a sensitivity analysis bounds the causal target and supplies confidence bounds for the endpoints. When the embedding is tuned on the sample, three approaches preserve valid inference: fixed-menu fine-tuning, joint-loss tuning, and double representation selection. A multimodal demand application illustrates the workflow.

arXiv:2609.35628v2cs.LG

Arbitrary-Accuracy Neural Approximation with Optimal Neuron Count and Optimal Bit Complexity

We study the minimum number of hidden neurons required for arbitrary-accuracy approximation of continuous functions on $[0,1]^d$ and the associated encoding complexity. For $d\geq 2$, we construct two fixed, explicitly defined activation functions and, for each, a closed-form network with two hidden layers of widths $d$ and $1$ that achieves arbitrary accuracy in the uniform norm. We prove that $d+1$ is the exact minimum total number of hidden neurons among standard feedforward networks with locally integrable activations and affine outputs. We further give a simpler construction using a single elementary activation that combines the floor and exponential functions. This construction requires three hidden layers of widths $d$, $1$, and $2$. If a skip connection is allowed, widths $d$, $1$, and $1$ suffice. For a bounded $\alpha$-H\"older class with $\alpha\in(0,1]$, the networks with the digit-extraction activation and with the elementary activation require $O(\varepsilon^{-d/\alpha}\log(1/\varepsilon))$ bits, matching the metric-entropy lower bound up to a logarithmic factor. The other $(d,1)$ network, which stores differences of neighboring values along a spanning tree, needs only $\Theta(\varepsilon^{-d/\alpha})$ bits and is optimal in both neuron count and bit complexity.

arXiv:2609.32478v2stat.ML

Agnostic Smoothed Online Regression with Adversarial Responses

We study smoothed online prediction with bounded adversarial responses. This widely studied framework bridges i.i.d. sampling and adversarial covariate selection through a smoothness parameter $\textsf{C}_{\textsf{cov}}$, which bounds conditional covariate densities relative to a fixed, unknown base measure. We propose Hedge-Cover, an information-theoretic algorithm that achieves sublinear regret $\widetilde{O}(\sqrt{\text{Pdim}(\mathcal{F}) \textsf{C}_{\textsf{cov}} T})$ for function classes with bounded pseudo-dimension. The algorithm aggregates a carefully constructed family of experts using Hedge, with a prior that links regret to the number of disagreements between a consistent selector and a target function. We bound this number by exploiting covariate smoothness. This answers an open problem posed in \cite{blanchard2025agnostic} on the minimax optimal adaptive regret of the smoothed online regression problem. We establish a matching lower bound for the class of linear predictors. The main intricacy of the lower bound lies in explicitly constructing a challenging sequential covariate distribution supported on mutually orthogonal hyperplanes. This construction may be of independent technical interest. Finally, we revisit the well-specified setting and quantify the effect of response noise. For conditionally $\nu^2$-subGaussian responses, we extend the existing lower bound under realizable responses by showing that the minimax expected regret is $\Omega((1\vee \nu)\sqrt{(\textsf{C}_{\textsf{cov}}-1)dT})$ for a function class of VC dimension $d$. A corresponding upper bound for ERM matches this dependence on $\nu$.

arXiv:2609.24422v3cs.LG

일반화 선형 혼합효과 모델을 위한 사전분포 상각(Prior-Amortized) 인컨텍스트 베이지안 추론

Prior-Amortized In-Context Bayesian Inference for Generalized Linear Mixed-Effects Models

Hierarchical data is ubiquitous in the empirical sciences and most commonly analyzed with generalized linear mixed-effects models (GLMMs). Bayesian inference for GLMMs yields calibrated uncertainty estimates but requires Markov Chain Monte Carlo (MCMC). The gold-standard No-U-Turn Sampler (NUTS) is slow and must restart from scratch for every new dataset, model and prior specification. We introduce MetaBeta, a pretrained amortized model for in-context Bayesian inference over GLMMs. The full prior and model specification are supplied as context at test time, and the model returns an asymptotically exact posterior at the speed of a forward pass. Two set transformers and two conditional neural spline flows mirror the GLMM's hierarchical structure and are trained jointly on millions of realistic simulated datasets with continuous, binary and count outcomes. The flow posterior is refined by Independence Metropolis-Hastings against the unnormalized posterior. This yields tuning-free asymptotically correct inference. Alternatively, MetaBeta can warm-start NUTS. This gives nearly identical inference 3.5-13x faster than cold-started NUTS and removes over 98% of NUTS' divergent transitions. On controlled benchmarks with ground-truth parameters, MetaBeta matches NUTS in parameter recovery, posterior calibration and out-of-sample prediction. On real datasets with unknown parameters, priors and model structure, its posteriors closely agree with those of NUTS. Furthermore, it remains faithful where amortized models typically struggle: under misspecified likelihoods, misspecified priors, out-of-distribution predictors, collinear designs and data-poor regimes. On real data, a full prior sensitivity analysis including model comparison takes a few seconds with MetaBeta against 14-18 core-hours for NUTS, tracing its posteriors throughout. Our model is open-source and open-weights.

arXiv:2609.22226v3cs.CL

Swiss-Knife: 디코딩 시점의 재구성 가능한 외부화 다목적 정렬 프레임워크

Swiss-Knife: A Framework for Reconfigurable Externalised Multi-Objective Alignment at Decode Time

Decode-time alignment methods steer a frozen language model by scoring candidate continuations with an external reward and selecting the maximiser. We argue that this shared design is a single degenerate point in a much larger space. We introduce Swiss-Knife, a framework for externalised multi-objective alignment in which the alignment specification is a first-class runtime object: hot-swappable scoring blades, a batch normaliser, a pairwise aggregation operator, and a selection rule. Six axioms characterise the admissible aggregation operators, and we prove a representation theorem: every operator satisfying them has the form $R_i = \sum_{j \neq i} g((\mu_i - \mu_j)/s(\sigma_i,\sigma_j))$, a two-parameter family containing probit and logistic comparison rules and pointwise argmax as named coordinates. Within it, pairwise aggregation is Lipschitz-stable under adversarial reward contamination while argmax is not, and Candidate-Batch Normalization (CBN) makes the weight simplex invariant to the rescalings under which reward models are only ever identified. Our reference instantiation pairs DPO-LoRA blades with an uncertainty-aware pairwise tournament. Sweeping the helpfulness/honesty/harmlessness simplex, it attains the best balanced frontier of six decode-time methods (harmonic $F_1$ 0.797 vs. 0.750 for the strongest baseline, $p < 10^{-14}$) with the lowest refusal rate and highest helpfulness of any arm, and reconfigures its objectives in 0.05 ms with no gradient computation.

arXiv:2608.07385v2cs.LG

옴니모달 분해 오토인코더를 통한 풀스택 웨어러블 얽힘 풀린 표현 학습

Omni-modal decomposition autoencoders learn full-stack wearable disentangled representations

Learning disentangled representations is key for general-purpose multi-modal wearable models, yet existing approaches rarely combine classification, disentanglement, fusion, and generation of heterogeneous multi-modal time series in a single model. We introduce Omni-modal Variational Decomposition Autoencoders (OmniDecVAEs), a framework that learns multi-purpose representations from arbitrarily many modalities with a single shared encoder-decoder. OmniDecVAEs extend DecVAEs by learning time-frequency latent subspaces per modality through a multi-view self-supervised decomposition loss and a component-wise conditional decoder. On the omni-modal human activity recognition (HAR) HARWE dataset (up to thirty channels) and the public UCI HAR and WESAD datasets, OmniDecVAEs learn the most disentangled representations for unseen subjects (DCI disentanglement up to 3.0 percentage points (pp) higher). On HARWE, they improve subject-dependent HAR and identity recognition by 1.1 pp and 5.8 pp, and outperform supervised end-to-end models by +31.6 pp and +22.3 pp on unseen subjects of multi-device datasets. Moreover, their generative capacity improves reconstruction mean squared error over multi-modal VAEs (34-74% improvement) and their synthetic data improve affect recognition on unseen subjects (2 pp). With 4.1M parameters, invariant to the number of modalities (21xfewer than a multi-branch VAE), OmniDecVAEs are scalable, single-model full-stack learners for wearable representations.

arXiv:2606.19230v3cs.LG

A human-in-the-loop Bayesian optimization framework for constraint-aware bioprocess development

We present Pareto Front Guided Sampling (PFGS), a human-in-the-loop (HiL) Bayesian Optimization (BO) framework that reformulates surrogate-model predictions as a multi-objective optimization problem, allowing domain knowledge to beadded intuitively to any BO problem. PFGS addresses constrained optimization by treating the posterior probability of constraint satisfaction as an explicit Pareto objective, and robust optimization via a Monte Carlo estimate of expected lower-confidence performance under input perturbations, capturing likely degradation. The multi-dimensional Pareto front makes trade-offs between predicted performance, model uncertainty, constraint satisfaction, and robustness visible through pairwise two-dimensional projections on an interactive dashboard, allowing selection criteria to be refined as the model and objectives evolve. The framework is showcased on a eight-dimensional in-silico fed-batch Chinese Hamster Ovary cell culture case study. We showed improved performance over BO when the expert has perfect domain knowledge; however, PFGS is not designed to outperform BO, but rather to present a HiL acquisition function framework.

arXiv:2606.03736v4stat.ML

자원 제약 동적 가격 책정을 위한 적응형 추론

Adaptive Inference for Resource-Constrained Dynamic Pricing

We study dynamic pricing over a finite selling horizon when limited resource capacity determines revenue and the observations available for inference at a prespecified price. Resource depletion can remove the target neighborhood from the feasible price set, changing the experiment generated by the pricing policy. We develop inference-aware re-solving controllers that check target-band feasibility before current covariates arrive and log the pricing mixture. Target-reserved and smooth controllers take population mean-pair geometry as a predeployment input; learned barycentric re-solving instead estimates stationary mean-consumption vectors of predeclared component kernels. On an affine binding-capacity family, an exact-input target-reserved controller assigning mass $t^{-\gamma}$ obtains an information clock of order $T^{1-\gamma}$ in probability, radius $O_p\{T^{-(1-\gamma)/2}\}$, and, under an exposed-face reward identity, a signed fluid-benchmark gap bounded above by $O(\log T+T^{1-\gamma})$. Under the exogenous affine-face condition, predeclared target support, and polynomial error spending with exponent greater than one, learned barycentric re-solving has a linear information clock in probability and an $O(\log T)$ signed-gap upper bound; centered local pricing has the same orders under slack capacity and global target optimality. An exact-input, target-compatible smooth alternative without reservation gives a linear clock in probability, an $O_p(T^{-1/2})$ deterministic-envelope radius with unconditional coverage and reporting probability tending to one, and an $O(\log^2 T)$ signed-gap upper bound. Boundary results show when physical support is lost and why a $1/t$ target branch yields only $O_p(1)$ information if it is the sole target-local source. The policy reports an interval when its prespecified support and information conditions hold and otherwise abstains.

arXiv:2605.20143v3stat.AP

An interpretable semiparametric Bayesian machine learning model for high-dimensional circulating epigenetic signatures of multiple myeloma

Genome-wide 5-hydroxymethylcytosine (5hmC) profiles of circulating cell-free DNA (cfDNA) can reveal epigenetic changes associated with multiple myeloma (MM), but their analysis requires adjusting for known risk factors while selecting a few informative genes from thousands. We propose semiparametric probit Bayesian Additive Regression Tree (spBART), in which the covariates enter a linear component with interpretable coefficients and the genes enter a sparse tree ensemble whose posterior inclusion probabilities select genes conditional on the covariates. The analysis selects eight candidate loci, \textit{IL1RAP}, \textit{ST5}, \textit{HERC6}, \textit{KL}, \textit{MYO1E}, \textit{ELK3}, \textit{CAPN2} and \textit{UBR4}. Selection of \textit{IL1RAP} connects the circulating signature to the \textit{IL-1/IL-6} signaling axis implicated in MM. The final model achieves an area under the receiver operating characteristic curve of 0.96 and a Brier score of 0.09 in the held-out validation set. Our simulation studies support the modeling strategy when adjustment variables act approximately linearly, but false discoveries exceed the nominal target in application-sized settings with correlated features. Overall, the proposed model identifies a compact set of circulating epigenetic signatures that are highly associated with MM.

arXiv:2605.08963v2stat.ML

Survey-aware Machine Learning: A Guideline for Valid Population Health Inference based on Scoping Review

Machine Learning (ML) models trained on complex health surveys such as the National Health and Nutrition Examination Survey (NHANES) often ignore primary sampling units, stratification variables, and sampling weights. This practice violates the independence assumptions of standard evaluation methods. As a result, estimates become biased, uncertainty is underestimated, and fairness assessments fail to reflect population-level disparities. We propose Survey-aware Machine Learning (SaML), a nine-step guideline that incorporates survey design metadata across the ML lifecycle. Through a scoping review of 16 methodological papers, we summarize existing work on weighted model training, design-based cross-validation, and survey-adjusted performance evaluation. We also identify gaps in hyperparameter tuning and deployment. We provide task-specific guidance that clarifies which steps are required for different analytical objectives. SaML provides a checklist for valid population inference from survey data.

arXiv:2605.02989v3cs.IT

정보 이론과 통계적 학습

Information Theory and Statistical Learning

This manuscript contains preprint of a chapter in the forthcoming third edition of Cover and Thomas's Elements of Information Theory, posted with permission from Wiley. For feedback, please contact abbas@ee.stanford.edu. Learning and information theory intersect in both model training and the characterization of fundamental performance limits. This manuscript provides a concise and accessible treatment of the first intersection, requiring only basic background in information theory and statistics at the senior undergraduate / first-year graduate level. End-of-chapter exercises make the material well suited for classroom use as well as self-study. The chapter focuses on the role of divergence measures in model training, with examples ranging from linear and logistic regression to autoregressive models, variational autoencoders, diffusion models, and score-based models. In particular, the treatment of the generative diffusion model provides a more systematic and explicit derivation than is typical in the literature.

arXiv:2605.00581v2stat.ML

Gradient Regularized Newton Boosting Trees with Global Convergence

Gradient Boosting Decision Trees (GBDTs) dominate tabular machine learning, with modern implementations like XGBoost, LightGBM, and CatBoost being based on Newton boosting: a second-order descent step in the space of decision trees. Despite its empirical success, the global convergence of Newton boosting is poorly understood compared to first-order boosting. In this paper, we introduce Restricted Newton Descent, a framework for convex optimization with Newton's method on Hilbert spaces with inexact iterates, based on the concepts of cosine angle and weak gradient edge. Within this framework, we recover Newton boosting with GBDTs and classical finite-dimensional theory as special cases. We first prove that vanilla Newton boosting achieves a linear rate of convergence for smooth, strongly convex losses that satisfy a Hessian-dominance condition. To handle general convex losses with Lipschitz Hessians, we extend a recent gradient regularized Newton scheme to the restricted weak learner setting, establishing the first global $\mathcal{O}(\frac{1}{k^2})$ convergence rate for second-order GBDTs. This scheme minimally modifies the classical algorithm by introducing an adaptive $\ell_2$-regularization term proportional to the square root of the gradient norm at each iteration. In numerical experiments, we show that this scheme converges while vanilla Newton boosting may diverge.

arXiv:2604.23212v2stat.ML

Learning Curves and Prediction Persistence to Under-Regularization of Spectral Algorithms in Large Dimensions

We study how prediction risk changes as regularization is reduced for spectral algorithms when the sample size $n$ and dimension $d$ satisfy $n\asymp d^\gamma$ for fixed $\gamma>0$. For inner-product kernels on the sphere, under a source condition with smoothness $s\geq0$ and a signal nondegeneracy condition, we characterize the conditional excess risk up to constant factors at each fixed admissible scale $\lambda\asymp d^{-u}$, $u>0$, and at interpolation. The resulting risk orders describe the effects of source smoothness, filter qualification, and transitions between spectral blocks. They identify when the conditional excess risk retains the minimax risk order as regularization is weakened from an optimal scale to interpolation, and when an interval before interpolation exhibits a polynomially larger risk. We also compare the kernel risk with a population sequence risk for the same spectral filter, using a deterministic regularization scale bounded below by order $n^{-1}$. Finally, for noninteger $\gamma$, we extend the risk characterization for kernel ridge regression to kernels on general domains under spectral scaling, hyper-contractivity, and tail diagonal concentration conditions, together with the source and signal conditions.

arXiv:2603.24974v3math.OC

The Value of Information in Resource-Constrained Pricing

Dynamic pricing with resource constraints requires balancing exploration of unknown demand patterns against exploitation of known information to maximize revenue. We develop algorithms for this problem across different information regimes. Products share perishable resources, and the firm posts prices and observes the realized requests, which may exceed the remaining capacity. In the full-information setting, a residual-correction policy that solves the fluid problem once attains $O(\log T)$ regret with bounded prices and arbitrary-rank resource matrices. In the no-information setting, deterministic exploration followed by a certainty-equivalent policy attains $O(\sqrt T)$ regret when the limiting fluid solution is nondegenerate. In the informed-price setting, given an initial price-demand estimate, a feedback controller learns the implementing prices without estimating the slope matrix if the slope matrix is symmetric, a set containing it is known in advance, the informed price is aligned with the resource constraints, and an inverse-image condition holds. A selector chooses between the informed-price algorithm and the no-information algorithm from historical data. Its regret bound includes observation error, misalignment, inventory exits, and periods on the fixed price. We further show how offline surrogate models reduce the regret of the learning policy through variance reduction, even when misspecified. The bound includes calibration and screen failures, and the regret is reduced for all sufficiently large horizons. We finally demonstrate the empirical performances of our policies through complete numerical experiments.

arXiv:2603.24427v2stat.ML

Continuous-time learning of probability distributions: a case study in a digital trial of young children with type 1 diabetes

Understanding how biomarker distributions evolve over time is a central challenge in digital health and chronic disease monitoring. For example, in diabetes, changes in the distribution of glucose measurements can reveal patterns of disease progression and treatment response that conventional summary measures miss. Motivated by a 13-week clinical trial comparing the closed-loop insulin delivery system t:slim X2 with standard therapy in children with type 1 diabetes, we propose a probabilistic framework to model the continuous-time evolution of time-indexed distributions using continuous glucose monitoring (CGM) data collected every five minutes. We represent the glucose distribution as a Gaussian mixture, with time-varying mixture weights governed by a neural ODE. We estimate the model parameters using a distribution-matching criterion based on the maximum mean discrepancy. The resulting framework is interpretable, computationally efficient, and sensitive to subtle temporal distributional changes. In the CGM trial data analyzed here, the method detects treatment-related improvements in glucose dynamics that are difficult to capture with traditional analytical approaches.

arXiv:2602.12107v3cs.LG

On the Complexity of Offline Reinforcement Learning with Q*-Approximation and Partial Coverage

We study offline reinforcement learning under $Q^\star$-approximation and partial coverage, a setting that motivates practical algorithms such as Conservative $Q$-Learning (CQL; Kumar et al., 2020) but has received limited theoretical attention. Our work is inspired by the following open question: "Are $Q^\star$-realizability and Bellman completeness sufficient for sample-efficient offline RL under partial coverage?" We answer in the negative via an information-theoretic lower bound. To identify additional structure that enables sample-efficient offline RL under partial coverage, we introduce a general decision-estimation framework, inspired by model-free decision-estimation coefficients (DEC) for online RL (Foster et al., 2023b; Liu et al., 2025b). Our framework decomposes offline RL complexity into decision complexity and value estimation error. This allows modular study of both sub-problems. Our result not only unifies existing results (Chen and Jiang, 2022; Uehara et al., 2023), but further improves and generalizes them. On the decision complexity side, our improvement includes: the first $\epsilon^{-2}$ sample complexity bound for soft $Q$-learning under partial coverage that improves Uehara et al.'s (2023) $\epsilon^{-4}$ bound, the removal of the need for additional online interaction in the value-gap setting of Chen and Jiang (2022), and new learnable settings beyond the above two cases. On the value estimation side, we provide a new characterization of the role of Bellman completeness under partial coverage, and the first characterization of offline learnability for general low-Bellman-rank MDPs (Jiang et al., 2017; Du et al., 2021; Jin et al., 2021). The latter is a canonical online RL setting that has remained unexplored in offline RL except for special cases. As a side contribution, our techniques give the first analysis of CQL in the function approximation setting.

arXiv:2602.05869v4stat.ML

웨지 샘플링: 준선형 샘플 복잡도를 갖는 효율적인 텐서 완성

Wedge Sampling: Efficient Tensor Completion with Nearly-Linear Sample Complexity

We introduce wedge sampling, a non-adaptive sampling scheme for recovering an order-$k$ low-rank tensor of side length $n$ from structured observations of its entries. The scheme samples pairs of entries in the same column of a tensor unfolding, forming length-two paths (wedges) in the associated bipartite sampling graph. The products of the paired entries supply the row correlations needed for spectral initialization in regimes where uniform sampling yields too few wedges. For fixed rank, incoherence, tensor order, and condition number, wedge sampling enables polynomial-time recovery with nearly linear sample complexity in $n$. The spectral initializer can be combined with existing refinement procedures using only an additional $\tilde O(n)$ uniformly sampled entries. Spectral refinement achieves weak recovery to any fixed relative Frobenius error. For symmetric third-order tensors with incoherent canonical polyadic factors, gradient descent achieves exact recovery in the noiseless setting, with error decreasing geometrically. The improvement over the $\tilde O(n^{k/2})$ sample complexity typical of efficient methods under uniform entry sampling shows how the conjectured computational barrier depends on the observation model. A key ingredient for the gradient descent analysis is a concentration bound for sparse random tensors in an incoherent spectral norm. We also give error bounds for both procedures under additive Gaussian noise and explicit signal-to-noise conditions. Beyond tensor completion, the same sampling scheme improves the sample complexity of one-sided matrix completion, where the goal is to recover the left singular subspace of a matrix with many more columns than rows.