Papers

We investigate a generalized stochastic fractional neuronal model combining fractional dynamics with correlated stochastic inputs. The proposed framework is described by a fractional differential equation driven by a latent stochastic process with stationary increments and mean-reverting structure. This formulation allows the inclusion of both short-range and long-range dependence structures and naturally produces non-exponential relaxation phenomena. The main goal is the development of a feasible parameter estimation procedure based on discrete observations of the neuronal state process. We propose a two-step methodology. First, the parameters governing the fractional dynamics are estimated by exploiting the asymptotic behavior of Mittag-Leffler functions near the origin. Subsequently, the latent stochastic input is reconstructed through fractional differentiation techniques, allowing the estimation of the parameters governing the hidden noise dynamics. We derive quantitative error bounds for the estimators and analyze the reconstruction error of the latent process under suitable regularity assumptions on the driving noise. In particular, the interplay between the order of the fractional derivative and the Hölder regularity of the noise process naturally emerges in the stability analysis of the reconstruction procedure. Finally, simulation studies illustrate the applicability of the proposed methodology and highlight the influence of memory effects and noise regularity on the quality of statistical inference. The results support the relevance of fractional stochastic analysis for the modeling and inference of neuronal systems with memory and correlated inputs.

Retrieval algorithms are used to estimate atmospheric concentrations of greenhouse gases (GHGs), such as carbon dioxide (CO2) and methane (CH4), by solving inverse problems from high-spectral-resolution satellite radiance measurements. However, these algorithms are computationally expensive, which makes real-time estimation at scale difficult. Machine-learning models have therefore been proposed as fast emulators of retrieval algorithms. Most existing studies, however, evaluate them only on test data from the same period as the training data. We study the stability over time of such emulators using data from the Greenhouse Gases Observing SATellite (GOSAT). We show that prediction accuracy generally deteriorates when the test period moves away from the training period. We also show that including time as an input feature substantially improves XCH4 prediction for Lasso and neural-network models. Among the methods considered, a simple Lasso model performs as well as or better than more complex methods such as neural networks, and yields more stable predictions over time. We further validate the results using the Total Carbon Column Observing Network (TCCON), a ground-based observation network. On the TCCON-matched dataset, the time-augmented Lasso achieves errors against TCCON that are comparable to the disagreement between GOSAT and TCCON for both XCO2 and XCH4.

Multivariate regression with many correlated responses and predictors commonly violates Gaussian error assumptions due to heavy tails, outliers, and asymmetry. Gaussian procedures then lose efficiency in coefficient estimation and produce biased estimates of conditional dependence graphs. We develop a robust Bayesian framework using a scale-location mixture error distribution and horseshoe+ global-local priors on both the regression coefficients and off-diagonals of the error precision matrix, coupling sparsity in the regression map with sparsity in the residual dependence structure. Theoretical contributions include joint posterior contraction, selection consistency for both supports, a Kullback-Leibler risk bound showing the dominance of horseshoe+ over horseshoe, and bounded sensitivity, ensuring that a single large outlier has vanishing influence under t errors. Simulations across four error regimes, contamination, and varying dimensions show that our estimator matches Gaussian procedures under normality and dominates them under heavy tails and skewness. Applications to FRED-MD macroeconomic data and S&P 500 daily returns recover interpretable sparse coefficient maps and residual dependence graphs while automatically down-weighting crisis-period observations.

For approximately half of the individuals receiving mental health care, the results are suboptimal, even when treatments align with evidence-based guidelines. These limited effects may partly stem from how clinical decisions on treatment focus are made in mental health care. Typically, treatment strategy is guided by the diagnostic classification combined with the individualized case conceptualization. While standard, this approach may fall short for several reasons such as biases on the part of both the patient and therapist, and treatment guidelines being based on average effects that may not (exactly) suit the individual patient. To address these challenges, we propose a novel framework that reduces biases in clinical decision-making and makes it genuinely possible to tailor treatment focus to the individual patient. This framework involves (a) constructing causal graphs and estimating causal effects from intensively collected, longitudinal patient data, (b) simulating new time series based upon the causal relationships, and (c) using these simulations to identify the most effective treatment focus for the individual patient. By simulating and comparing different intervention strategies and examining both the estimated individual's responsiveness and its long-term effectiveness, this approach may generate useful insights to guide treatment focus and strategy, which can lead to a significant improvement of treatment outcomes in mental health care.

This paper develops a methodological framework for reverse stress testing (RST) in which a multivariate stress scenario, coherent with the empirical dependence structure of a market, is reconstructed from a single exogenous shock prescribed on one asset class. The problem is formulated as the maximisation of the conditional density given the imposed shock, and is solved under three progressively weaker distributional assumptions. In the parametric setting, joint Gaussianity of the returns yields a closed-form modal scenario coinciding with the conditional mean of the non-shocked components. In the semiparametric setting, the modal scenario is estimated nonparametrically through the empirical likelihood methodology and the surrounding stressed trajectories are generated via a Gaussian or Student-t local sampling scheme. In the fully nonparametric setting, stressed trajectories are obtained by inverse-distance resampling of the historical observations within a Mahalanobis neighbourhood of the estimated scenario. The three variants are validated on real market data. The simulated scenarios prove to be economically coherent and capable of reproducing the standard risk-reward asymmetry observed in stressed market regimes.

High-Dimensional Low-Sample Size (HDLSS) tabular domains (e.g., omics) are characterized by $n \ll m$, where $n$ = number of samples, and $m$ = number of features. Such domains often exhibit strong local correlation groups, sparse cross-group dependencies, heavy-tailed non-Gaussian marginals, heteroscedastic noise, and structured missingness, making direct density learning in $\mathbb{R}^m$ ill-conditioned since $n \ll m$. We propose BSTabDiff, a block-subunit generative framework that partitions the $m$ observed features into $M$ latent blocks ($M \ll m$) and generates each block via a shared low-dimensional subunit variable, concentrating global dependence learning in the compact block-latent space $\mathbb{R}^M$ while decoding to the full feature space with copula-driven dependence, flexible per-feature marginals, and explicit missingness mechanisms. BSTabDiff supports modern deep priors on block latents, including diffusion and normalizing flows, enabling stable synthesis and controllable benchmark generation in the HDLSS regime. Empirically, BSTabDiff produces more realistic and stable high-dimensional synthetic data when compared with unstructured tabular generators on HDLSS data.

We prove that $\rho\text{-}\mathrm{NPTS}_{\mathrm{SG}}$, an anchor-free nonparametric Thompson Sampling algorithm for risk-averse bandits, achieves regret matching the instance-dependent lower bound to leading order in $\log n$, establishing it as asymptotically optimal for any continuous risk functional $\rho$ (CVaR, mean-variance, Sharpe ratio, distortion risk measures, and more) on the class of distributions with bounded density and sub-Gaussian tails, including Gaussian arms. Both this result and its bounded-support counterpart require only continuity of $\rho$: strictly weaker than the dominance condition of prior parametric Thompson Sampling results, and strictly weaker than the Lipschitz condition of UCB-type algorithms, yielding the first instance-optimal guarantees for non-Lipschitz functionals such as the Sharpe ratio without parametric reward assumptions. The bounded-support case is developed first as a stepping stone sharing the same proof structure. The key technical contributions are a discretisation lemma (bounded support) and a truncated discretisation lemma (sub-Gaussian tails), each projecting the growing-alphabet Dirichlet posterior onto a fixed grid via the Dirichlet aggregation property, holding all polynomial prefactors at fixed degree independent of sample size and breaking the super-exponential barrier that blocked prior proofs.

This paper studies the asymptotic behavior of sample canonical directions in a finite-rank spiked high-dimensional canonical correlation analysis model under a Gaussian population assumption. Under the asymptotic regime in which the dimensions of the two data blocks grow proportionally with the sample size, sample canonical directions are generally not consistent estimators of their population counterparts, even when the corresponding sample canonical correlations separate from the bulk spectrum. To quantify directional recovery, we investigate the squared alignment between a sample canonical direction and its associated population direction. For each simple population spike, we first establish a deterministic first-order limit for this squared alignment, which gives an explicit measure of the population-level directional information retained by the sample direction. We then prove a central limit theorem for its fluctuations around the deterministic limit, with an explicit asymptotic variance expressed through deterministic limits of resolvent trace functionals. To make the theoretical quantities computable from data, we further construct plug-in estimators for both the limiting mean and the asymptotic variance by inverting the deterministic outlier eigenvalue map, and prove their consistency. Numerical simulations and a real-data illustration support the theoretical results and demonstrate how the proposed estimators assess the recovery quality of sample canonical directions.

We study finite expected-signature information for mixed-fBm paths with Hurst indices above $1/4$. Up to level three, the only parameter-dependent expected features are the variance transform $q_\theta$ and the time-ordered transform $R_\theta$. We prove the scale tradeoff $2K$ level-two scales versus $K$ selected level-two/three scales, together with separation and local inverse bounds.

Ultra-high-dimensional array-based CpG methylation studies require statistical frameworks that simultaneously provide supervised structure discovery, interpretability, scalable latent-dimension identification, and computational feasibility. We propose SOLAR (Supervised Orthogonal Low-rank Adaptive Regression), a supervised low-rank latent-factor framework for identifying CpG-level methylation structure associated with residualized DNAm age. SOLAR combines orthogonal low-rank regression with a penalized maximum a posteriori formulation, dimension-adaptive BIC-type penalization, and a trans-dimensional simulated-annealing strategy for automatic latent-rank selection, together with theoretical guarantees including identifiability, fixed-rank recovery, and rank-selection consistency under suitable regularity conditions. The framework additionally incorporates computationally and memory-efficient optimization strategies demonstrating scalability up to $p=10^7$, while analyses at $p=10^6$ remain feasible on standard desktop computing environments. Simulation studies demonstrate stable rank recovery, competitive supervised signal recovery, and strong scalability across moderate-, high-, and ultra-high-dimensional regimes. Using longitudinal EPIC-array CpG methylation data from the GUSTO birth cohort, comprising $n=1051$ methylation profiles collected across infancy and early childhood with approximately 860,000 assayed CpGs per sample, SOLAR identifies heterogeneous supervised methylation structure associated with residualized DNAm age beyond chronological age alone, together with biologically coherent CpG signatures and enrichment patterns.

Self-evolution offers a scalable path to stronger reasoning: a pretrained language model improves itself with only minimal external supervision. Yet existing methods either depend on extensively curated or teacher-generated training data, or, when the generator runs unsupervised, reward it by a difficulty heuristic that need not improve the solver. We introduce INFUSER, an iterative co-training framework with two co-evolving roles: a Generator that drafts questions and reference golden answers from a pool of unstructured, automatically collected documents, and a Solver that improves by training on them. The solver is trained with standard correctness rewards against the generator-provided answers, while the generator is rewarded by an optimizer-aware influence score that measures whether each proposed question would actually improve the solver on the target distribution. Because this continuous, noisy influence score is poorly served by standard GRPO, we propose DuGRPO, a dual-normalized variant of GRPO, for generator training. Together, these turn the document pool into an adaptive curriculum that favors questions useful to the current solver, not just hard ones. On Qwen3-8B-Base, INFUSER outperforms strong self-evolution baselines with over 20% relative improvement on Olympiad and SuperGPQA benchmarks, and an 8B INFUSER co-evolving generator outperforms a frozen 32B thinking generator on math and coding. Ablations confirm each design choice is necessary, and two extensions, applying INFUSER to an instruction-finetuned anchor and augmenting it with rule-verifiable RLVR data, further demonstrate the flexibility and generalizability of the framework. Code is available at https://github.com/FFishy-git/INFUSER.

We propose the data augmented bootstrap (DAB), a framework for constructing confidence intervals from approximately invariant transformations of the data. As special cases, DAB recovers popular methods that rely on exact group symmetries, such as conformal prediction, wild bootstrap for Maximum Mean Discrepancy U-statistics and the recently proposed SymmPI. Meanwhile, DAB also recovers the classical bootstrap method, which exploits the dataset's approximate invariance under uniform sampling of data indices as the dataset size grows. For all DAB methods, we establish theoretical coverage results that interpolate between finite-sample and asymptotic guarantees according to the strength of the invariance, and without assuming a group structure. The approximate invariance is measured in the Kolmogorov distance and, for statistics that satisfy Gaussian universality, reduces to conditional mean and variance matching. This allows us to incorporate data augmentation (DA), a widely used machine learning heuristic based on approximate invariances, into known statistical methods. We empirically test the performance of incorporating DA into bootstrap, wild bootstrap and conformal prediction for simulated settings as well as for image, language and scientific data.

We study high-dimensional linear regression under a general symmetric convex constraint. Rather than imposing a specific sparsity-inducing penalty, we start from an arbitrary sign-symmetric and permutation-invariant convex body $K\subseteq \mathbb R^p$ and construct the sparse convexification hierarchy \[ K^{(s)} = \operatorname{conv}\{v\in K:\|v\|_0\le s\}. \] We propose a penalized least-squares estimator that searches over this hierarchy and adapts to the best sparse convex approximation of the target. Under standard sub-Gaussian assumptions on the random design and noise, we prove an oracle inequality showing that the estimator adapts to the best sparse convex approximation of the target. For an $s$-sparse target, the result yields a squared-error rate governed by the noise level $\sigma$, and the Gaussian width of the sparse convexification $K^{(s)}$. The method applies broadly to symmetric norm balls and can be implemented using oracle access to the Minkowski functional of $K$. As a special case, the framework yields a consistency result for the constrained Lasso.

Post-training quantization (PTQ) converts a trained full-precision model into low-bit weights without task-level retraining, while quantization-aware training (QAT) incorporates quantization into the training loop. Although PTQ is efficient and often accurate at moderate bitwidths, it can fail sharply at aggressive bitwidths; QAT is more expensive but can often recover the lost accuracy. We propose a unified geometric framework that explains both PTQ failure and QAT recovery. We model full-precision training as following a low-loss river inside a wider valley: a normal neighborhood of the river forms a nearly flat basin, while leaving this basin incurs a sharp loss increase. When the quantization grid is comparable to the basin width, local PTQ objectives, including rounding and Hessian-based second-order reconstruction, can select a high-loss deployed quantized point outside the basin even when nearby low-loss quantized points exist. In this regime, straight-through-estimator-based QAT has a useful bias: it evaluates gradients at the deployed quantized weights while updating latent full-precision weights, causing the gradient to sense the valley wall and acquire an inward component that steers subsequent quantized iterates back into the basin. We formalize this mechanism through a local landscape model, construct a geometric PTQ failure mode, and prove finite-time QAT recovery under local quantizer-compatibility assumptions. Experiments across vision and language models under multiple neural-network quantization schemes corroborate the predicted basin-crossing failure of PTQ and the corresponding recovery mechanism of QAT.

We study a stochastic multi-armed bandit problem in which the set of available arms expands over time. This setting arises in sequential experimentation when new actions or treatments become available during an ongoing study, making regret against a single best arm in hindsight inappropriate. We instead evaluate performance relative to the best arm currently available, leading to a dynamic-regret criterion for arriving-arm environments. To address the resulting challenges of arrival information discrepancy (AID) and a drifting benchmark (DB), we propose UCB for Arriving Arms (UCB-AA), an elimination-based procedure with an aiding preliminary screening step for newly arrived arms before full competition with incumbent arms. We show that UCB-AA attains regret bounds that depend explicitly on the arrival process, achieves sublinear dynamic regret under regularity conditions on gap evolution, and admits an online extension for unknown horizons. Simulation results show that UCB-AA reduces wasted pulls and maintains a smaller active arm set while preserving competitive regret performance.

Knowledge transfer across data sources holds great promise for improving the estimation of target population parameters by leveraging the growing availability of data from different sources. However, the effectiveness of knowledge transfer is often challenged by the complex and pervasive heterogeneity between data sources and the lack of access to individual-level data. This paper proposes the divide-and-shrink (dShrink) method, a transfer estimation method that estimates target population parameters in a closed form using summary statistics from a target population and some external source populations while accounting for population heterogeneity. The dShrink estimator is guaranteed to outperform the estimator based solely on the target population in terms of expected quadratic error under arbitrary population heterogeneity. The gain can be substantial when the target and source populations are similar, or the underlying true parameter values are near zero. Notably, dShrink is model-free, requires no user-specified tuning parameters, robust to various types of heterogeneity between data sources, and applies to a broad range of parameter estimation problems. dShrink remains effective even when the covariance matrix is not accessible for the external summary statistics and offers flexibility in incorporating side information and summary statistics from multiple source populations. Simulations and real data analyses demonstrate the superior performance of the dShrink estimator and its potential as a robust tool for transfer estimation.

The strong likelihood principle (SLP) is conventionally derived from the sufficiency principle and a conditionality principle in an argument due to Birnbaum, and much of the literature contests whether the derivation is sound. We take a different approach. We ask what the SLP says when its terms are read carefully, and argue that the principle as ordinarily stated reflects a confusion about the domain of the likelihood function. The likelihood is naturally defined as a function on a family of distributions $M$, not on a parameter space, and once it is so defined the SLP collapses into its weak counterpart, the weak likelihood principle. The diagnosis is illustrated by analogy with monetary value, developed concretely through a comparison of the binomial and negative binomial families that share a parameter, and connected to the geometric structure of $M$ through the Fisher information metric. The same standardization emerges from a statistical argument about comparing measurements across populations and from a geometric argument about manifold distance; this convergence supplies the positive content of the weak likelihood principle.

Class imbalance is common when developing clinical prediction models (CPMs) and is often assumed to lead to poor predictive performance. Several methods have been proposed to correct data imbalance during CPM development. However, it remains unclear whether correcting class imbalance improves or harms CPM performance. This study investigated how imbalance correction affects classification performance and prediction stability. We simulated the development and internal validation of CPMs using penalised logistic regression under different imbalance-correction strategies, including algorithm-level rebalancing, data-level rebalancing by oversampling, and combined over- and under-sampling. The simulation dataset was derived from the GUSTO-I trial, which included 40,830 patients and 2,851 events. All imbalance-correction strategies were evaluated across sample-size scenarios ranging from 500 to 40,830. Model performance and prediction stability were assessed using 200 bootstrap resamples, including discrimination, calibration, calibration stability, mean absolute prediction error (MAPE), and classification instability index (CII). Class imbalance correction did not meaningfully improve model discrimination. Both data-level and algorithm-level correction led to miscalibration, risk overestimation, and increased prediction instability, as shown by prediction stability, MAPE, and CII plots, compared with models developed without correction. These findings suggest that class imbalance correction does not necessarily improve CPM performance and may compromise calibration and prediction stability. Class imbalance should not be treated as a pathology that automatically requires correction. In clinical prediction modelling, routine imbalance correction by default is generally not advisable.

Diffusion Monte Carlo (DMC) and Monte Carlo for particle transport with importance sampling both involve simulations of weighted walkers that undergo birth and death processes (splitting and Russian Roulette). The established implementations of these methods are quite different: Particle simulation Monte Carlo employs a stack to handle the splitting history whereas in traditional DMC one follows a swarm of walkers. The particle simulation Monte Carlo approach involves a depth first traversal of the visited configurations whereas the traditional DMC approach may be seen as a breadth first traversal. In the present work the implementation of a depth first, stack based approach to DMC is described and a complete code is presented. The depth first approach, called DMCD here, can be more memory efficient than the breadth first approach, both for total memory and for use of a memory hierarchy and of co-processors. The implementation appears very natural for population control and for descendant weighting and it unifies algorithmic treatment of the eigenvalue problem (DMC) with the linear equation problem (particle transport). A concern with DMCD that is not present in the breadth first approach, and that is successfully addressed here, is the need to maintain a pool of starters for use when a new walker is required and the stack is empty. The DMCD approach appears to have the potential to become the preferred implementation for many DMC applications.

This paper proposes a collapsible method for estimating causal effects that maintains the estimator's consistency before and after marginalization over some variables in completed partially directed acyclic graphs (CPDAGs). We first introduce the estimate collapsibility for CPDAGs and characterize the minimal collapsible sets as strong d-convex hulls. An efficient algorithm is devised to obtain such sets in DAGs and is generalized to CPDAGs. Then, we combine the graph reduction procedure with the IDA framework. Finally, experiments and empirical analysis show the effectiveness of the collapsibility for causal estimations in CPDAGs. Code is available at https://github.com/Jamyang-D/strongly-convex.

Recurrent neural networks maintain a hidden state $h_t$, but its probabilistic meaning is often unclear. We study hidden-state stability through backward coherence: the extent to which $h_t$ can be reconstructed from $h_{t+1}$ by a learned backward projector $g_\phi$. Under contraction and summable backward drift, the hidden-state sequence forms a quasi-reverse-martingale. This yields almost-sure convergence, rates under mixing, an interpretable limiting representation, finite pathwise stopping times, and a theoretical framework for time-uniform confidence sequences. Simulations support the theory. Backward-coherence regularisation reduces the empirical quasi-martingale total $\hat Q$ by $43$–$58%$, reaches stability $28$–$44%$ earlier than an unregularised RNN, and gives tracking-error recovery consistent with geometric bounds. Additional tests confirm echo-state forgetting rates bounded by $\rho$ and verify the increment-sum tube $R_t$ with $100%$ simultaneous coverage, although $R_t$ is conservative; in practice, the defect-tail proxy $\hat Q_t$ is the more useful monitor. The backward-coherence loss is also equivalent to minimising a Kullback–Leibler divergence in a Gaussian backward model, linking the method to variational inference. Extensions cover $\phi$-mixing inputs, change-point tracking, and finite-sample concentration. Three real-data studies further validate the approach. On PhysioNet 2012 ICU data, the Reverse Martingale RNN (RMRNN) matches RNN mortality-prediction AUC while reaching stable representations 13 hours earlier. On FRED-MD, it reduces one-month-ahead forecast error by about fourfold under concept drift. On UCI Human Activity Recognition, it maintains lower post-transition tracking error with geometric decay. The guarantees apply under the stated assumptions; universality is not claimed.

Measuring causal effects in networked two-sided marketplaces is challenging due to treatment interference between market participants on different sides. When treatment is applied to one side (e.g., job seekers), their interactions with the other side (e.g., job posters) introduce spillover effects that violate the Stable Unit Treatment Value Assumption (SUTVA) and bias causal estimates. While cluster-based randomization mitigates this problem, prior approaches struggle with a fundamental trade-off: reducing spillover requires isolated clusters that will reduce the number of qualifying clusters, which decreases statistical power. This paper introduces EgoCluster V3, an iterative clustering algorithm that reduces spillover by 3x compared to prior versions while preserving node coverage and doubling test power. We further introduce MultiEgoCluster, which extends V3 through a two-stage procedure that first groups highly connected egos into multi-ego clusters before applying the iterative clustering algorithm. This achieves an additional ~56% spillover reduction and ~38% increase in sample size. Both methods are deployed in production at LinkedIn and have systematically enabled high-impact two-sided marketplace experiments. Since residual bias cannot be fully eliminated through clustering alone, we derive a theoretical bias correction method for average treatment effect (ATE) estimation based on graph structure and propose an approach to generalize results to the general population.

We study finite-state finite-action Bayesian statistical decision problems. While exact error-exponent characterizations are known for several special cases, including hypothesis testing and hypothesis exclusion, the asymptotic behavior of the optimal Bayes regret is largely unknown for general decision problems. In this paper, we show that the optimal regret always decays exponentially fast and characterize its exact exponent for arbitrary loss functions. The exponent is given by the minimum multivariate Chernoff information over the minimal incompatible subsets of states, where an incompatible subset is a collection of states for which no single action is optimal for all states in the subset. Our result recovers the classical pairwise-minimum Chernoff exponent for symmetric multiple hypothesis testing and the multivariate Chernoff exponent for hypothesis exclusion, while also yielding, to the best of our knowledge, the first exact exponent characterization for list hypothesis testing.

We present IOAH3 (Importance-Oriented Adaptive H3 partitioning), a computational method for constructing data-driven spatial partitions of geo-referenced observation domains. Standard approaches to spatial aggregation adopt fixed areal units, such as administrative boundaries or uniform hexagonal grids at a single resolution, without regard to the informational content of the underlying observations in each region. This leads to the well-known modifiable areal unit problem: statistical and inferential results depend on the arbitrary choice of partition, and spatially concentrated phenomena are averaged out in coarse cells that obscure fine-scale structure. IOAH3 addresses this by constructing an adaptive partition in three stages: multi-source feature extraction and importance scoring via principal component analysis over road density, POI density, building density, and terrain roughness signals, with population and flood-hazard data entering as auxiliary inputs to cell filtering and spatial smoothness; spatial cell selection via Markov Random Field graph-cut optimisation, which jointly maximises per-cell importance while enforcing spatial contiguity; and data-driven hierarchical refinement of high-importance regions to finer H3 resolution levels, with neighbour-propagated support to avoid isolated fine-resolution islands. The resulting partitions serve as input to spatial inference pipelines and provide a principled resolution of the partition-sensitivity problem prior to any modelling step.

The analysis of non stationary signals in complex physical systems often relies on Time Frequency distributions. Among these, the Pseudo Wigner Ville Distribution (PWVD) stands out for its superior resolution but is mathematically challenging due to its inherent quadratic nonlinearity. This nonlinearity generates complex interference artifacts and cross terms in the phase space, potentially obscuring the physical features of the signal, particularly for nonlinear chirps. In this work, we establish a mathematically grounded framework for the PWVD for general windowed nonlinear chirps. By leveraging the theory of oscillatory integrals with coalescing stationary points, we derive a uniform asymptotic expansion that bridges the gap between heuristic signal processing and semiclassical geometric approaches (Berry's chord construction). The resulting closed form representation, expressed in terms of symmetric incomplete Airy functions, provides a unified description of the nonlinear transform's behavior, regularizing the transition across the instantaneous frequency caustics. While the framework is general, we show its power on two illustrative examples: the high precision nonlinear chirps of coalescing binaries in gravitational-wave astronomy and radar nonlinear chirps for pulse compression applications. The analytical results successfully predict the structure of interference patterns and quantify the systematic bias in peak based frequency estimation. Therefore, this study establishes a systematic bridge between nonlinear mathematical analysis and precision experimental physics, validating the PWVD as a robust tool for detailed source characterization in high signal to noise regimes.

Standard Reinforcement Learning with Verifiable Rewards (RLVR) training allocates a fixed rollout budget to every query, without regard for what each query's difficulty means for the current policy. This leads to two symmetric failure modes: easy queries produce near-zero advantage because the policy already solves them, while unsolvable queries produce no signal because the policy never solves them. Both regimes waste training FLOPs without contributing to a learning gradient. We introduce sorted Group Policy Optimization (sGPO), a compute-efficient strategy that trades a small budget of inference FLOPs for a large reduction in wasted training FLOPs. The key insight is that cheap inference compute can serve as a single offline proxy for query difficulty. By generating a small batch of parallel samples per query under the initial policy, we obtain a model-aware empirical success rate. This motivates setting the training rollout group size to the inverse of this success rate, a practical rule that maximizes sample efficiency by extracting the most advantage per generated rollout. This single profiling pass simultaneously drives data filtering (removing trivial queries and sub-sampling unsolvable ones), adaptive group size allocation, and curriculum construction (scheduling queries from easy to hard). sGPO matches or exceeds baseline performance while reducing total training compute by a factor of three, with the upfront inference profiling cost included.

Generative AI and large language models can produce realistic predictions of human behavior from rich, unstructured inputs with little to no task-specific training data. Recent work uses these “digital twin” predictions to supplement human responses in surveys and experiments. We study the special case of using AI-generated predictions to reduce variance in randomized experiments. We argue that doing so requires no new estimators and that researchers can simply include AI predictions as covariates in standard regression adjustment, analogous to adjusting for a prognostic score. A benefit of this approach is a “do no harm” property whereby the adjusted estimator reverts to the unadjusted difference in means when predictions are uninformative. Other methods, such as variants of prediction-powered inference, do not have this guarantee. We provide implementation guidance, including how to obtain continuous scores from discrete LLM outputs and how to use LLMs to featurize unstructured inputs as auxiliary covariates. We demonstrate these ideas in simulations and three empirical applications: a survey mega-study, an email marketing A/B test, and a large-scale technology platform experiment. Overall, efficiency gains are real if modest, with greater benefits in studies that contain substantial text and other unstructured data. We also confirm the do no harm property empirically. Given these gains and limited costs, we recommend adjusting for AI-generated predictions as a regular empirical practice.

Inference-Time Scaling (ITS) has largely succeeded in verifiable domains like math and coding, where cheap verification enables scalable output selection. However, extending ITS to tasks prone to systematic failure - driven by faulty initial assumptions or unmet multidimensional constraints - typically relies on costly external solvers or brittle, model-based verifiers. Our key insight is that the intrinsic statistics of parallel sample sets, specifically length-adjusted tail entropy, provide a robust discriminative signal for solution quality without access to ground truth. Crucially, these statistics serve as a difficulty gate for adaptive compute allocation, dynamically routing problems across scaling regimes. First, Intrinsic Selection (iS) ranks candidates post-hoc, matching consensus-based algorithms across three domains and improving engineering design selection by 20% over pass@1 baselines. Second, Intrinsic Particle Filtering (iPF) generalizes this to step-level resampling, guiding generation toward high-confidence reasoning trajectories to improve pass@1 by 6.1 points on average on hard math problems. Finally, Particle Distillation (dPF) injects privileged guidance via early logit blending and KL-guided resampling, steering generation past systematic reasoning errors to satisfy expert rubrics, yielding up to 26.5% gains on complex clinical responses. Our pipeline applies seamlessly across broad-purpose, domain-specialized, and multimodal architectures, successfully extending ITS to open-ended domains without requiring trained reward models or exact ground-truth verification.

Sorted $\ell_1$ Penalized Estimation (SLOPE) models, that perform either variable or group selection, control the false discovery rate (FDR) under orthogonal settings with known noise, but such settings are rare in practice. Under general conditions, cross-validation is the default model selection approach for SLOPE, yet it targets predictive performance rather than FDR control. We address this gap for the SLOPE family of models by proposing new Bayesian approaches, Bayesian Group SLOPE (BGSLOPE) and Bayesian Sparse-group SLOPE (BSGS). BGSLOPE and BSGS embed group-based SLOPE models into a spike-and-slab framework, with BSGS providing a continuous spike-and-slab framework for sparse-group models. We further introduce Two-step Orthogonal (TSO), which transforms a general setting into an orthogonal one to recover SLOPE's FDR control properties. Through extensive synthetic and real data studies comparing all major model selection strategies for SLOPE models, the proposed Bayesian models consistently control FDR, achieve higher power, and outperform competing methods in prediction.

We study the generalization of ridge-regularized nonlinear least-squares models via on-average algorithmic stability, deriving error bounds for local minimizers in terms of a data-dependent effective dimension that reflects the geometry of the gradient model at the trained parameters, through the empirical Jacobian Gram matrix and a residual-curvature term. In the linear case, where the curvature term vanishes, this recovers the classical effective dimension of the Jacobian kernel covariance, but evaluated at the trained model rather than at initialization as is typical in neural tangent kernel analyses. We further bound this effective dimension via covering complexity of the gradient features, leading to guarantees that depend on learned geometry rather than parameter count. In particular, for manifold-supported data and piecewise Lipschitz Jacobians, the bounds scale with intrinsic dimension, while for one-hidden-layer ReLU networks, the mechanism can be made explicit through counts of activation-stable regions. Experiments on synthetic manifolds, clustered distributions, and benchmark datasets illustrate trained-Jacobian compression, the tightness of the residual-curvature linearization, and agreement between the stability bound and observed generalization gaps. A key feature of our bounds is the simplicity of their derivation, which follows from first principles using the Brascamp-Lieb inequality under strongly log-concave noise.