Papers

Understanding the dependence structure of asset returns is fundamental in risk assessment and is particularly relevant in a portfolio diversification strategy. We propose a clustering approach where evidence accumulated in a multiplicity of classifications is achieved using classical hierarchical procedures and multiple copula-based dissimilarity measures. Assets that are grouped in the same cluster are such that their stochastic behavior is similar during risky scenarios, and riskaverse investors could exploit this information to build a risk-diversified portfolio. An empirical demonstration of such a strategy is presented by using data from the EURO STOXX 50 index.

Joint Embedding Predictive Architectures (JEPAs) learn representations able to solve numerous downstream tasks out-of-the-box. JEPAs combine two objectives: (i) a latent-space prediction term, i.e., the representation of a slightly perturbed sample must be predictable from the original sample's representation, and (ii) an anti-collapse term, i.e., not all samples should have the same representation. While (ii) is often considered as an obvious remedy to representation collapse, we uncover that JEPAs' anti-collapse term does much more–it provably estimates the data density. In short, any successfully trained JEPA can be used to get sample probabilities, e.g., for data curation, outlier detection, or simply for density estimation. Our theoretical finding is agnostic of the dataset and architecture used–in any case one can compute the learned probabilities of sample $x$ efficiently and in closed-form using the model's Jacobian matrix at $x$. Our findings are empirically validated across datasets (synthetic, controlled, and Imagenet) and across different Self Supervised Learning methods falling under the JEPA family (I-JEPA and DINOv2) and on multimodal models, such as MetaCLIP. We denote the method extracting the JEPA learned density as JEPA-SCORE.

A function on a topological space is called unimodal if all of its super-level sets are contractible. A minimal unimodal decomposition of a function $f$ is the smallest number of unimodal functions that sum up to $f$. The problem of decomposing a given density function into its minimal unimodal components is fundamental in topological statistics. We show that finding a minimal unimodal decomposition of an edge-linear function on a graph is NP-hard. Given any $k \geq 2$, we establish the NP-hardness of finding a unimodal decomposition consisting of $k$ unimodal functions. We also extend the NP-hardness result to related variants of the problem, including restriction to planar graphs, inapproximability results, and generalizations to higher dimensions.

Subsampling is a computationally efficient and scalable method to draw inference in large data settings based on a subset of the data rather than needing to consider the whole dataset. When employing subsampling techniques, a crucial consideration is how to select an informative subset based on the queries posed by the data analyst. A recently proposed method for this purpose involves randomly selecting samples from the large dataset based on subsampling probabilities. However, a major drawback of this approach is that the derived subsampling probabilities are typically based on an assumed statistical model which may be difficult to correctly specify in practice. To address this limitation, we propose to determine subsampling probabilities based on a statistical model that we acknowledge may be misspecified. To do so, we propose to evaluate the subsampling probabilities based on the Mean Squared Error (MSE) of the predictions from a model that is not assumed to completely describe the large dataset. We apply our subsampling approach in a simulation study and for the analysis of two real-world large datasets, where its performance is benchmarked against existing subsampling techniques. The findings suggest that there is value in adopting our approach over current practice.

The Wald-Wolfowitz runs test can assess the correctness of a regression curve fitted to a data set with one independent parameter. The assessment is performed through examination of the residuals, where the signs of the residuals would appear randomly if the regression curve were correct. We propose extending the test to the case where multiple data points were measured for specific independent parameter values. By randomly permutating the data points corresponding to each independent parameter value and treating their residuals as occurring in their permutated sequence and then executing the runs test, results are shown to be equivalent to those of a data set containing the same number of points with no repeated measurements. This approach avoids the loss of points, and hence loss of test sensitivity, were the means at each independent parameter value used. It also avoids the problem of weighting each mean differently if the number of data points measured at each parameter value is not identical.

Missing data is a common challenge in observational studies. Another challenge stems from the observational nature of the study itself. Here, propensity score analysis can be used as a technique to replicate conditions similar to those found in clinical trials. With regard to the missing data, a majority of studies only analyze the complete cases, but this has several pitfalls. In this review, we investigate which methods are used for the handling of missing data in the context of propensity score analyses. Therefore, we searched PubMed for the keywords propensity score and missing data, restricting our search to the time between January 2010 and February 2024. The PRISMA statement was followed in this review. A total of 147 articles were included in the analyses. A major finding of this study is that although the usage of multiple imputation (MI) has risen over time, only a limited number of studies describe the mechanism of missing data and the details of the MI algorithm. Keywords Missing data, Propensity Score, Observational Data, Multiple Imputation, Systematic Review

Guiding pretrained flow-based generative models for conditional generation or to produce samples with desired target properties enables solving diverse tasks without retraining on paired data. We present ESS-Flow, a gradient-free method that leverages the Gaussian prior of the source distribution in flow-based models to perform Bayesian inference directly in the source space using elliptical slice sampling. ESS-Flow only requires forward passes through the generative model and observation process, making it applicable even when gradients are unavailable, such as with non-differentiable simulation-based observations or quantization in the generation process. We demonstrate its effectiveness on scientific applications including generating materials with target properties, protein backbone structures with improved designability, and synthesizable, drug-like molecules.

Inference-Time Scaling (ITS) improves language models by allocating more computation at generation time. Particle Filtering (PF) has emerged as a strong ITS method for complex mathematical reasoning tasks, but it is vulnerable when guided by process reward models, which often assign overconfident scores early in the reasoning process. This causes PF to suffer from premature exploitation: it myopically commits to locally promising trajectories, prunes potentially correct hypotheses, and converges to suboptimal solutions. This failure mode, known as particle impoverishment, is especially severe under constrained computational budgets. To address this, we analyze the problem and identify two root causes: a lack of diversity in the particle set due to overconfident resampling and consequent inability to assess the potential of a reasoning path. We introduce Entropic Particle Filtering (ePF), an algorithm that integrates two new techniques to solve these issues. The first technique, Entropic Annealing (EA), directly mitigates particle impoverishment by monitoring search diversity via entropy; when diversity drops, it intervenes by dynamically annealing the resampling distribution to preserve exploration. The second, an enhancement called Look-ahead Modulation (LaM), adds a predictive guide to evaluate a state's potential based on its successors. On several challenging math benchmarks, ePF significantly outperforms strong baselines and achieves up to a 50% relative improvement in task reward. Together, these methods improve PF's resilience by balancing the exploration of diverse solution spaces with the exploitation of high-reward regions, ultimately leading to higher-quality solutions.

We develop a statistical framework for risk estimation, inspired by the axiomatic theory of risk measures. Coherent risk estimators – functionals of P&L samples inheriting the economic properties of risk measures – are defined and characterized through robust representations linked to $L$-estimators. The framework provides a canonical methodology for constructing estimators with sound financial and statistical properties, unifying risk measure theory, principles for capital adequacy, and practical statistical challenges in market risk. Numerical illustrations based on simulated and market data demonstrate that coherence of a risk measure does not necessarily carry over to its estimators and show that alternative admissible weight structures within the CRE representation can lead to substantially different capital adequacy outcomes.

Simulations are valuable tools for empirically evaluating the properties of statistical methods and are primarily employed in methodological research to draw general conclusions about methods. In addition, they can often be useful to applied statisticians, who may rely on published simulation results to select an appropriate statistical method for their application. However, on the one hand, applying published simulation results directly to practical settings is frequently challenging, as the scenarios considered in methodological studies rarely align closely enough with the characteristics of specific real-world applications to be truly informative. Applied statisticians, on the other hand, may struggle to construct their own simulations or to adapt methodological research to better reflect their specific data due to time constraints and limited programming expertise. We propose bridging this gap between methodological research and statistical practice through a translational approach by developing dedicated software along with simulation studies, which should abstract away the coding-intensive aspects of running simulations while still offering sufficient flexibility in parameter selection to meet the needs of applied statisticians. We demonstrate this approach using two practical examples, illustrating that the concept of translational simulation can be implemented in practice in different ways. In the first example - simulation-based evaluation of power in two-arm randomized clinical trials with an ordinal endpoint - the solution we discuss is a Shiny web application providing a graphical user interface for running informative simulations in this context. For the second example - assessing the impact of measurement error in multivariable regression - a less labor-intensive approach is suggested, involving the provision of user-friendly, well-structured, and modular analysis code.

We establish explicit, universal, and distribution-free bounds for the $n$-th cumulant, $\kappa_n(X)$, of a scalar random variable, controlled solely by an $n$-th order absolute moment functional $M_n(X)$. The bounds take the form $\lvert\kappa_n(X)\rvert \le C_n M_n(X)$. Our principal contribution is the derivation of coefficients satisfying $C_n \sim (n-1)!/\rho^{\,n}$, which offers an exponential improvement over classical bounds where the coefficients grow superexponentially (on the order of $n^n$). We present a hierarchy of refinements where the rate parameter $\rho$ increases as the functional $M_n(X)$ incorporates more structural information. The most general bound uses the raw moment $M_n(X)=\mathsf{E}[\lvert X\rvert^n]$ with rate $\rho=\ln 2 \approx 0.693$. Using the central moment $M_n(X)=\mathsf{E}[\lvert X-\mathsf{E}[X]\rvert^n]$ improves the rate to $\rho_{\mathrm{cen}} \approx 1.146$, while assuming symmetry yields even higher rates. The proof is elementary, combining the moment-cumulant partition formula with a uniform moment-product inequality. We further prove that while these bounds are not attainable whenever the relevant coefficient is positive, they are asymptotically efficient given the limited information of a single moment. The utility of the bounds is demonstrated through an application to standardized cumulants of independent sums.

Bougerol (1993) and Straumann and Mikosch (2006) gave conditions under which there exists a unique stationary and ergodic solution to the stochastic difference equation $Y_t \overset{a.s.}{=} \Phi_t (Y_{t-1}), t \in \mathbb{Z}$ where $(\Phi_t)_{t \in \mathbb{Z}}$ is a sequence of stationary and ergodic random Lipschitz continuous functions from $(Y,|| \cdot ||)$ to $(Y,|| \cdot ||)$ where $(Y,|| \cdot ||)$ is a complete subspace of a real or complex separable Banach space. In the case where $(Y,|| \cdot ||)$ is a real or complex separable Banach space, Straumann and Mikosch (2006) also gave conditions under which any solution to the stochastic difference equation $\hat{Y}_t \overset{a.s.}{=} \hat{\Phi}_t (\hat{Y}_{t-1}), t \in \mathbb{N}$ with $\hat{Y}_0$ given where $(\hat{\Phi}_t)_{t \in \mathbb{N}}$ is only a sequence of random Lipschitz continuous functions from $(Y,|| \cdot ||)$ to $(Y,|| \cdot ||)$ satisfies $\gamma^t || \hat{Y}_t - Y_t || \overset{a.s.}{\rightarrow} 0$ as $t \rightarrow \infty$ for some $\gamma > 1$. In this note, we give slightly different conditions under which this continues to hold in the case where $(Y,|| \cdot ||)$ is only a complete subspace of a real or complex separable Banach space by using close to identical arguments as Straumann and Mikosch (2006).

We prove a new sample complexity result for divergence regularized optimal transport. Our bound holds for probability measures on~$\mathbb{R}^d$ with exponential tail decay and for radial cost functions that satisfy a local Lipschitz condition. It is sharp up to logarithmic factors, and captures the intrinsic dimension of the marginal distributions through a generalized covering number of their supports. Examples that fit into our framework include subexponential and subgaussian distributions and radial cost functions $c(x,y)=|x-y|^p$ for $p\ge 1$ with logarithmic entropy or polynomial $\alpha$-divergence.

We present a bivariate vector valued discrete autoregressive model of order $1$ (BDAR($1$)) for discrete time series. The BDAR($1$) model assumes that each time series follows its own univariate DAR($1$) model with dependent random mechanisms that determine from which component the current status occurs and dependent innovations. The joint distribution of the random mechanisms which are expressed by Bernoulli vectors are proposed to be defined through copulas. The same holds for the joint distribution of innovation terms. Properties of the model are provided, while special focus is given to the case of bivariate ordinal time series. A simulation study is presented, indicating that model provides efficient estimates even in case of moderate sample size. Finally, a real data application on unemployment state of two countries is presented, for illustrating the proposed model.

This article focuses on the use of Geographically Weighted Regression (GWR) method to correct air quality low-cost sensors measurements. Those sensors are of major interest in the current era of high-resolution air quality monitoring at urban scale, but require calibration using reference analyzers. The results for NO2 are provided along with comments on the estimated GWR model and the spatial content of the estimated coefficients. The study has been carried out using the publicly available SensEURCity dataset in Antwerp, which is especially relevant since it includes 9 reference stations and 34 low-cost sensors collocated and deployed within the city.

We investigate the asymptotic behavior of parametric Bayes estimators under a broad class of loss functions that extend beyond the classical translation-invariant setting. To this end, we develop a unified theoretical framework for loss functions exhibiting locally polynomial structure. This general theory encompasses important examples such as the squared Wasserstein distance, the Sinkhorn divergence and Stein discrepancies, which have gained prominence in modern statistical inference and machine learning. Building on the classical Bernstein–von Mises theorem, we establish sufficient conditions under which Bayes estimators inherit the posterior's asymptotic normality. As a by-product, we also derive conditions for the differentiability of Wasserstein-induced loss functions and provide new consistency results for Bayes estimators. Several examples and numerical experiments demonstrate the relevance and accuracy of the proposed methodology.

The Monte Carlo-type Neural Operator (MCNO) introduces a framework for learning solution operators of one-dimensional partial differential equations (PDEs) by directly learning the kernel function and approximating the associated integral operator using a Monte Carlo-type approach. Unlike Fourier Neural Operators (FNOs), which rely on spectral representations and assume translation-invariant kernels, MCNO makes no such assumptions. The kernel is represented as a learnable tensor over sampled input-output pairs, and sampling is performed once, uniformly at random from a discretized grid. This design enables generalization across multiple grid resolutions without relying on fixed global basis functions or repeated sampling during training, while an interpolation step maps between arbitrary input and output grids to further enhance flexibility. Experiments on standard 1D PDE benchmarks show that MCNO achieves competitive accuracy with efficient computational cost. We also provide a theoretical analysis proving that the Monte Carlo estimator yields a bounded bias and variance under mild regularity assumptions. This result holds in any spatial dimension, suggesting that MCNO may extend naturally beyond one-dimensional problems. More broadly, this work explores how Monte Carlo-type integration can be incorporated into neural operator frameworks for continuous-domain PDEs, providing a theoretically supported alternative to spectral methods (such as FNO) and to graph-based Monte Carlo approaches (such as the Graph Kernel Neural Operator, GNO).

Continual learning, the ability of a model to adapt to an ongoing sequence of tasks without forgetting earlier ones, is a central goal of artificial intelligence. To better understand its underlying mechanisms, we study the limitations of continual learning in a tractable yet representative setting. Specifically, we analyze one-hidden-layer quadratic neural networks trained by gradient descent on a sequence of XOR-cluster datasets with Gaussian noise, where different tasks correspond to clusters with orthogonal means. Our analysis is based on a tight characterization of gradient descent dynamics for the training loss, which yields explicit bounds on the rate of train-time forgetting as functions of the number of iterations, sample size, number of tasks, and hidden-layer width. We then leverage an algorithmic stability framework to bound the generalization gap, leading to corresponding guarantees on test-time forgetting. Together, our results provide the first closed-form guarantees for forgetting in continual learning with neural networks and show how key problem parameters jointly govern forgetting dynamics. Numerical experiments corroborate our theoretical results.

Methods for solving scientific computing and inference problems, such as kernel- and neural network-based approaches for partial differential equations (PDEs), inverse problems, and supervised learning tasks, depend crucially on the choice of hyperparameters. Specifically, the efficacy of such methods, and in particular their accuracy, stability, and generalization properties, strongly depends on the choice of hyperparameters. While bilevel optimization offers a principled framework for hyperparameter tuning, its nested optimization structure can be computationally demanding, especially in PDE-constrained contexts. In this paper, we propose an efficient strategy for hyperparameter optimization within the bilevel framework by employing a Gauss-Newton linearization of the inner optimization step. Our approach provides closed-form updates, eliminating the need for repeated costly PDE solves. As a result, each iteration of the outer loop reduces to a single linearized PDE solve, followed by explicit gradient-based hyperparameter updates. We demonstrate the effectiveness of the proposed method through Gaussian process models applied to nonlinear PDEs and to PDE inverse problems. Extensive numerical experiments highlight substantial improvements in accuracy and robustness compared to conventional random hyperparameter initialization. In particular, experiments with additive kernels and neural network-parameterized deep kernels demonstrate the method's scalability and effectiveness for high-dimensional hyperparameter optimization.

Large language models have achieved impressive performance across diverse tasks. However, their tendency to produce overconfident and factually incorrect outputs, known as hallucinations, poses risks in real-world applications. Conformal prediction provides finite-sample, distribution-free coverage guarantees, but standard conformal prediction breaks down under domain shift, often leading to under-coverage and unreliable prediction sets. We propose a new framework called Domain-Shift-Aware Conformal Prediction (DS-CP). Our framework adapts conformal prediction to large language models under domain shift, by systematically reweighting calibration samples based on their proximity to the test prompt, thereby preserving validity while enhancing adaptivity. Our theoretical analysis and experiments on the MMLU benchmark demonstrate that the proposed method delivers more reliable coverage than standard conformal prediction, especially under substantial distribution shifts, while maintaining efficiency. This provides a practical step toward trustworthy uncertainty quantification for large language models in real-world deployment.

Taiwan's high population and heavy dependence on fossil fuels have led to severe air pollution, with the most prevalent greenhouse gas being carbon dioxide (CO2). There-fore, this study presents a reproducible and comprehensive case study comparing 21 of the most commonly employed time series models in forecasting emissions, analyzing both univariate and multivariate approaches. Among these, Feedforward Neural Network (FFNN), Support Vector Machine (SVM), and Random Forest Regressor (RFR) achieved the best performances. To further enhance robustness, the top performers were integrated with Linear Regression through a custom stacked generalization en-semble technique. Our proposed ensemble model achieved an SMAPE of 1.407 with no signs of overfitting. Finally, this research provides an accurate decade-long emission projection that will assist policymakers in making more data-driven decisions.

Randomized controlled trials (RCTs) are widely adopted for causal inference, yet cost and sample-size constraints limit power. We introduce CALM (Causal Analysis leveraging Language Models), a statistical framework that integrates insights generated by large language models (LLMs) into the analysis of RCTs using established causal estimators to increase precision while preserving statistical validity. In particular, CALM treats LLM-generated outputs as auxiliary prognostic information and corrects their potential bias via a heterogeneous calibration step that residualizes and optimally reweights predictions. We prove that CALM remains consistent even when LLM predictions are biased and achieves efficiency gains over augmented inverse probability weighting estimators for various causal estimands. In particular, CALM develops a few-shot variant that aggregates predictions across randomly sampled demonstration sets. The resulting U-statistic-like predictor restores i.i.d. structure and also mitigates prompt-selection variability. Empirically, in simulations calibrated to a mobile-app depression RCT, CALM delivers lower variance relative to other benchmarking methods, is effective in zero- and few-shot settings, and remains stable across prompt designs. By principled use of LLMs to harness unstructured data and external knowledge learned during pretraining, CALM provides a practical path to more precise causal analyses.

Graphon models provide a flexible nonparametric framework for estimating latent connectivity probabilities in networks, enabling a range of downstream applications such as link prediction and data augmentation. However, accurate graphon estimation typically requires a large graph, whereas in practice, one often only observes a small-sized network. One approach to addressing this issue is to adopt a transfer learning framework, which aims to improve estimation in a small target graph by leveraging structural information from a larger, related source graph. In this paper, we propose a novel method, namely GTRANS, a transfer learning framework that integrates neighborhood smoothing and Gromov-Wasserstein optimal transport to align and transfer structural patterns between graphs. To prevent negative transfer, GTRANS includes an adaptive debiasing mechanism that identifies and corrects for target-specific deviations via residual smoothing. We provide theoretical guarantees on the stability of the estimated alignment matrix and demonstrate the effectiveness of GTRANS in improving the accuracy of target graph estimation through extensive synthetic and real data experiments. These improvements translate directly to enhanced performance in downstream applications, such as the graph classification task and the link prediction task.

Decision-making under clustered uncertainty requires models that accommodate discrete overdispersed risk and endogenous control mechanisms. Standard approaches often impose equi-dispersion or Gaussian assumptions and treat control as exogenous or binary, limiting structural realism. In this paper, we developed a stochastic optimization framework in which risk follows a dynamic Negative Binomial process with time-varying success probability, generating persistent overdispersion and temporal dependence. A continuous control variable mitigated non-tail and tail exposure under convex implementation and adjustment costs. Analytical results showed that overdispersion reshapes the optimal policy, generating interior solutions, dispersion-dependent thresholds, and regime-dependent hysteresis effects that do not arise under equi-dispersed specifications. Simulation evidence confirms that overdispersion amplifies tail risk and induces nonlinear responses. An empirical analysis using daily S&P 500 index data documents significant overdispersion and persistent clustering. The results provide a structural basis for integrating discrete risk modeling with dynamic operational decisions under uncertainty.

The massive digital footprints generated by bike-sharing systems in megacities like Shanghai offer a novel perspective on the urban socio-economic fabric. This study investigates whether these daily mobility patterns can quantitatively map the city's underlying social stratification. To overcome the persistent challenge of acquiring fine-grained socio-economic data, we constructed a multi-layered analytical dataset. We annotated 2,000 raw bike trips with local economic attributes, derived from a novel data enrichment methodology that employs a Large Language Model (LLM), and integrated contextual features of the built environment. A Random Forest model was then utilized as an interpretable framework to determine the key factors governing the relationship between mobility behavior and local economic status. The analysis reveals a compelling and unambiguous finding: a neighborhood's economic level, proxied by housing prices, is the single most dominant predictor of its bike-sharing patterns, substantially outweighing other geographic or temporal factors. This economic determinism manifests in three distinct ways: (1) a spatial clustering of resources, a phenomenon we term the club effect, which concentrates mobility infrastructure and usage in affluent areas; (2) a functional dichotomy between necessity-driven, utilitarian usage in lower-income zones and flexible, recreational usage in wealthier ones; and (3) a nuanced inverted U-shaped adoption curve that identifies the urban middle class as the system's primary user base.

Specifications that impose constant treatment effects are common and can be biased under heterogeneity, whereas fully flexible alternatives can be imprecise or infeasible. Under a bound on the target-weighted variance of treatment effects, this paper proposes a generalized ridge estimator, $\texttt{regulaTE}$, that yields heterogeneity-aware confidence intervals (CIs). The ridge penalty is chosen to minimize the heterogeneity-aware CI length in a homoskedastic Gaussian setting; the resulting CIs are asymptotically valid even when overlap fails. Varying the bound enables sensitivity analysis of departures from constant effects, illustrated in applications to unconfoundedness and staggered adoption designs.

Low rank inference on matrices is widely conducted by optimizing a cost function augmented with a penalty proportional to the nuclear norm $\Vert \cdot \Vert_*$. However, despite the assortment of computational methods for such problems, there is a surprising lack of understanding of the underlying probability distributions being referred to. In this article, we study the distribution with density $f(X)\propto e^{-\lambda\Vert X\Vert_*}$, finding many of its fundamental attributes to be analytically tractable via differential geometry. We use these facts to design an improved MCMC algorithm for low rank Bayesian inference as well as to learn the penalty parameter $\lambda$, obviating the need for hyperparameter tuning when this is difficult or impossible. Finally, we deploy these to improve the accuracy and efficiency of low rank Bayesian matrix denoising and completion algorithms in numerical experiments.

We study meta-reinforcement learning in finite-horizon MDPs where related tasks share similar structures in their optimal action-value functions. Specifically, we posit a linear representation $Q^*_h(s,a)=\Phi_h(s,a)\,\theta^{(k)}_h$ and place a Gaussian meta-prior $ \mathcal{N}(\theta^*_h,\Sigma^*_h)$ over the task-specific parameters $\theta^{(k)}_h$. Building on randomized value functions, we propose two Thompson-style algorithms: (i) MTSRL, which learns only the prior mean and performs posterior sampling with the learned mean and known covariance; and (ii) $\text{MTSRL}^{+}$, which additionally estimates the covariance and employs prior widening to control finite-sample estimation error. Further, we develop a prior-alignment technique that couples the posterior under the learned prior with a meta-oracle that knows the true prior, yielding meta-regret guarantees: we match prior-independent Thompson sampling in the small-task regime and strictly improve with more tasks once the prior is learned. Concretely, for known covariance we obtain $\tilde{O}(H^{4}S^{3/2}\sqrt{ANK})$ meta-regret, and with learned covariance $\tilde{O}(H^{4}S^{3/2}\sqrt{AN^3K})$; both recover a better behavior than prior-independent after $K \gtrsim \tilde{O}(H^2)$ and $K \gtrsim \tilde{O}(N^2H^2)$, respectively. Simulations on a stateful recommendation environment (with feature and prior misspecification) show that after brief exploration, MTSRL/MTSRL\(^+\) track the meta-oracle and substantially outperform prior-independent RL and bandit-only meta-baselines. Our results give the first meta-regret guarantees for Thompson-style RL with learned Q-priors, and provide practical recipes (warm-start via RLSVI, OLS aggregation, covariance widening) for experiment-rich settings.

We initiate an investigation of learning tasks in a setting where the learner is given access to two competing provers, only one of which is honest. Specifically, we consider the power of such learners in assessing purported properties of opaque models. Following prior work in complexity theory that considers the power of competing provers in various settings, we call this setting refereed learning. After formulating a general definition of refereed learning tasks, we show refereed learning protocols that obtain a level of accuracy that far exceeds what is obtainable at comparable cost without provers, or even with a single prover. We concentrate on the task of choosing the better one out of two black-box models, with respect to some ground truth. While we consider a range of parameters, perhaps our most notable result is in the high-precision range: For all $\varepsilon>0$ and ambient dimension $d$, our learner makes only one query to the ground truth function, communicates only $(1+\frac{1}{\varepsilon^2})\cdot\text{poly}(d)$ bits with the provers, and outputs a model whose loss is within a multiplicative factor of $(1+\varepsilon)$ of the best model's loss. Obtaining comparable loss with a single prover would require the learner to access the ground truth at almost all of the points in the domain. We also present lower bounds that demonstrate the optimality of our protocols in a number of respects, including prover complexity, number of samples, and need for query access.

In graphical models, factor graphs, and more generally energy-based models, the interactions between variables are encoded by a graph, a hypergraph, or, in the most general case, a partially ordered set (poset). Inference on such probabilistic models cannot be performed exactly due to cycles in the underlying structures of interaction. Instead, one resorts to approximate variational inference by optimizing the Bethe free energy. Critical points of the Bethe free energy correspond to fixed points of the associated Belief Propagation algorithm. A full characterization of these critical points for general graphs, hypergraphs, and posets with a finite number of variables is still an open problem. We show that, for hypergraphs and posets with chains of length at most 1, changing the poset of interactions of the probabilistic model to one with the same homotopy type induces a bijection between the critical points of the associated free energy. This result extends and unifies classical results that assume specific forms of collapsibility to prove uniqueness of the critical points of the Bethe free energy.