Papers

The prediction of the extremely large values of a response variable $Y$ in terms of a vector of covariates $X=(X_i)_{i=1}^d$ is a fundamental problem arising in many scientific and engineering domains. The scarcity of data in the extremes makes the optimal solution of this problem of particular importance. The optimal predictors of such events can be explicitly characterized in just a few cases and it is of fundamental practical and theoretical interest to develop optimal estimators over large classes of models and predictors. In this work, the focus is on the case where $(Y,X)$ have a multivariate regularly varying distribution and one seeks an optimal predictor expressed as a positive homogeneous function $h(X)$ of the covariates. The asymptotic prediction precision in this setting coincides with the tail-dependence coefficient $\lambda(Y,h(X))$ and it can be expressed as an integral functional of the associated angular measure of $(Y,X)$. Thus, finding asymptotically optimal homogeneous predictors amounts to solving a variational problem. We obtain a general solution to this problem, which is expressed in terms of a non-extreme conditional quantile of a tilted distribution derived from the angular measure. This leads to a general inference methodology for the optimal predictors in the peaks-over-threshold framework form extreme value theory. We establish the universal consistency for these estimators over large classes of angular measures. A general-purpose implementation of the resulting inference procedure is shown to work remarkably well against optimal oracle estimators, as well as in the challenging problem of extreme solar flare prediction.

Several theoretical works have tried to explain the adversarial vulnerability of deep neural networks through properties of high-dimensional geometry. However, the assumptions underlying these works are rarely examined empirically, and systematic evidence remains limited. In this work, we present a systematic study of the role of input dimensionality in both the emergence and the targeted control of adversarial examples. We first analyse the scope and limitations of existing theoretical frameworks based on concentration of measure, showing that real image classes exhibit strong empirical localization, beyond what such theories typically assume. We then conduct an extensive empirical evaluation across hierarchical image datasets spanning a wide range of input dimensionalities and diverse neural architectures. Our results consistently show that adversarial examples become easier to construct as dimensionality increases. We also investigate how input dimensionality affects the additional difficulty of crafting targeted adversarial examples. In particular, we provide theoretical arguments showing that high-dimensional geometry implies that enforcing a specific target label entails only a limited additional distortion compared to untargeted attacks. We corroborate this insight through extensive experiments, demonstrating that the gap between targeted and untargeted perturbations remains small and further narrows as input dimensionality increases. While, taken together, our findings establish high input dimensionality as a fundamental factor underlying the emergence and targeted control of adversarial examples, whether this phenomenon primarily arises from the interplay between high-dimensional geometry and data distributions or from the architectural properties of deep neural networks remains an open question.

The relationship between crime, mental health service access, and socioeconomic deprivation in publicly-funded healthcare systems allowing impactful policy interventions offers an alternative lens to crime prevention that remains underexplored. We address this critical gap through an analysis of street-level crime data, mental health referral information, and socioeconomic metrics across Greater London, using both traditional statistical methods and machine learning techniques to identify relevant relationships and spatial patterns to reveal a persistent positive association between crime rates and mental health referrals as a proxy for service access. The prevailing prevention hypothesis is contrasted with a nuanced U-shaped relationship suggesting a contrast between preventive effects at lower service levels and demand-driven responses to crime exposure for higher referral rates. Subsequent analyses, focussing on explainable artificial intelligence, show distinct crime category patterns, with a cluster analysis identifying four borough typologies with distinct combinations of crime rates, mental health service access, and deprivation levels, requiring multifaceted approaches rather than universal solutions. This research provides one of the first comprehensive studies on this topic for the UK's publicly-funded healthcare system and introduces interpretation-oriented approaches to uncover the patterns essential to evidence-based policies.

Modern machine learning systems have outgrown their origins as isolated predictive constructs, evolving into complex socio-technical architectures that actively mediate human opportunity. As algorithms increasingly determine access to economic and social opportunities, it has become widely recognized that these systems are deeply embedded with the structural inequalities and prejudices of their environments. The field of algorithmic fairness emerged in response to the growing recognition that models optimized for predictive accuracy can systematically disadvantage marginalized groups. Early mitigation strategies, however, rested on fragile simplifications that limited their effectiveness in complex socio-technical environments. This thesis identifies and addresses two fundamental limitations of contemporary fairness paradigms: the reliance on deterministic point estimates for auditing and the treatment of individuals as isolated entities devoid of structural context.

Finding all modes of a multimodal black-box function is a fundamental challenge in optimization, Bayesian inference, and scientific computing. Existing approaches – basin-hopping, CMA-ES, multistart gradient descent – operate sequentially and cannot exploit the massive parallelism of modern GPU hardware. We introduce \chisao{} (Convergence-Halt-Invert-Stick-And-Oscillate), a GPU-native population optimizer that runs an entire sample batch simultaneously and exploits a deliberate convergence-anticonvergence oscillation cycle to escape local traps while freezing confirmed modes. The structural move is asymmetric: samples that reach true peaks are frozen (“stuck”) and preserved, while the rest keep exploring via momentum-based anti-convergence and stochastically smoothed gradients. Adaptive reseeding via two complementary strategies (Repulse Monkey and Golden Rooster) maintains population diversity throughout. On all 42 functions of the Simon Fraser University optimization benchmark suite across dimensions $d \in \{2, 4, 8, 16, 32, 64\}$, \chisao{} achieves 100% mode recovery where all CPU baselines collapse at $d \geq 8$ on the hardest multimodal functions, at up to \textbf{$34\times$} speedup over basin-hopping on functions where all methods succeed (Michalewicz $d=64$) and up to \textbf{$39\times$} on unimodal functions (Rotated Hyper-Ellipsoid $d=64$, pure GPU dividend). All benchmarks evaluate the objective by value alone – gradients come from finite differences – so the reported speedups are a derivative-free worst case. Under substantial likelihood noise ($\sigma_{\mathrm{noise}}$ up to 1.0), mode detection remains 100% reliable. The algorithm is available as a standalone open-source Python package on PyPI.

For piecewise deterministic samplers such as Randomized Hamiltonian Monte Carlo (RHMC), Bouncy Particle Sampler (BPS) or Zig-Zag Process (ZZP), long-time exponential convergence rates have been established in previous works using Harris or $L^2$ hypocoercivity approaches. In particular, in the $L^2$ framework, a so-called diffusive-to-ballistic speedup was known for log-concave targets, according to which the convergence rates of these samplers, with suitable parameters, are quadratically improved with respect to the standard overdamped Langevin diffusion process. A recent work by Jianfeng Lu showed that this speedup also holds for the kinetic Langevin diffusion process when the convergence is stated in terms of relative entropy, raising the question whether this also holds for piecewise deterministic samplers. The present work provides a positive and a negative answer to this: first, we show that the speedup holds in entropy for RHMC; second, we show that for BPS or ZZS, even for a standard Gaussian target, a similar result cannot hold, and even that exponential convergence (at any rate) in entropy fails.

Synthetic data generation is a powerful approach to privacy-preserving statistical analysis, where data-release mechanisms are governed by a privacy-utility tradeoff: they should provide privacy guarantees while preserving the statistical utility of confidential data. We develop a Bayesian nonparametric framework for private synthetic data generation tailored to discrete data. Specifically, the confidential data are modeled as a random sample from an unknown discrete distribution endowed with a Pitman-Yor process prior, and synthetic data are generated from the corresponding posterior-predictive distribution. Since the Pitman-Yor process defines an almost surely discrete random probability measure, the resulting mechanism is naturally suited to data with ties and settings involving a potentially large, unknown, or growing number of categories. We study differential privacy guarantees of the Pitman-Yor posterior-predictive mechanism across the three regimes of the discount parameter $\sigma\in(-\infty,1)$. For $\sigma\in(0,1)$, we establish an instance-level $(\varepsilon,\delta)$-differential privacy guarantee. For $\sigma=0$ and $\sigma<0$, corresponding respectively to the Dirichlet process prior and to a parametric Dirichlet-Multinomial model, stronger guarantees are obtained, under suitable conditions on the released sample size. We also investigate statistical utility, or informativity, of the released data via the expected $1$-Wasserstein distance between the empirical distribution of the synthetic data and the "true" data-generating distribution. For $\sigma<0$ and $\sigma=0$, we prove consistency of the empirical distribution in this metric and derive explicit convergence rates, making precise the privacy-utility tradeoff: stronger privacy guarantees impose more restrictive choices of the released sample size, slowing down convergence to the "true" data-generating distribution.

We introduce a Dirichlet–multinomial (DM) deviance residualization for sparse, jointly overdispersed count matrices, the regime that dominates sequencing-based biochemical assays. The DM null treats each sample's count vector as a fixed-total composition with a single scalar concentration $\alpha_0$ governing overdispersion, and arises exactly by conditioning independent negative-binomial feature counts on the observed sample total – making the DM the joint conditional analogue of standard feature-wise overdispersed count models. The resulting transform preserves exact sparsity, evaluates in constant time per nonzero entry, agrees with multinomial residuals on singleton counts, shrinks repeated-count residuals according to the overdispersion the null tolerates, and recovers the multinomial residual as $\alpha_0\to\infty$. The same fixed-dispersion comparison principle extends to ordered and tree-structured features via the generalized DM and the Dirichlet-tree multinomial, giving a single residual family that subsumes joint and feature-wise count nulls under a common compositional logic and is computationally lightweight enough to drop into existing sparse pipelines.

Synthetic data augmentation is widely used to mitigate class imbalance, but its theoretical effects on score-based classification remain poorly understood. This paper develops a framework for characterizing when synthetic minority augmentation can improve threshold-integrated and threshold-optimized metrics, including AUROC, AUPRC, best-threshold balanced accuracy, and best-threshold \(\F_1\) score. We separate the effect of augmentation into two components: a change in effective class weighting and a discrepancy between the synthetic and true minority distributions. Under well-specified score models, the raw estimator already targets the likelihood-ratio ordering, which is population-optimal for the metrics considered. Consequently, augmentation cannot provide a fundamental population-level improvement beyond possible finite-sample variance reduction, and may introduce additional bias through synthetic distributional error. We further establish minimax lower bounds showing that the raw estimator already achieves the optimal metric-regret rate in the well-specified regime. Under misspecification, however, augmentation can play a qualitatively different role: by changing the effective class balance, it can alter the restricted-class projection and correct ranking errors induced by the raw imbalanced objective. We provide explicit improvement bounds quantifying the roles of approximation error, finite-sample estimation error, and synthetic distributional error. Simulation studies corroborate the theory, demonstrating limited gains under well-specification and nontrivial but nonmonotone improvements under misspecification.

Federated learning has emerged as the foremost approach for decentralized model training with privacy preservation. The global class imbalance and cross-client data heterogeneity naturally coexist, and the mismatch between local and global imbalances exacerbates the performance degradation of the aggregated model. The agnosticism of global class distribution poses significant challenges for data-level methods, especially under extreme conditions with severe class absence across clients. In this paper, we propose FedReLa, a novel data-level approach that tackles the coexistence of data heterogeneity and class imbalance in federated learning. By re-labeling samples with a feature-dependent label re-allocator, FedReLa corrects biased global decision boundaries without requiring knowledge of the global class distribution. This modular, model-agnostic approach can be integrated with algorithmic methods to deliver consistent improvements without additional communication overhead. Through extensive experiments, our method significantly improves the accuracy of minority classes and the overall accuracy on stepwise-imbalanced and long-tailed datasets, outperforming the previous state of the art.

A popular technique for reducing the variance of importance sampling (IS) estimators is to modify the weights of some importance points. One approach is to truncate the largest weights, which reduces variance but can introduce substantial bias. Pareto smoothed importance sampling (PSIS), by contrast, reduces the variance of the weights by fitting a generalised Pareto distribution to the upper tail of the weight distribution and replacing the weights in this tail with the corresponding expected quantiles from the fitted distribution. PSIS can therefore also reduce variance, but typically with less bias, and has been used successfully in IS-based approximations for Bayesian cross-validation. This paper explores the use of PSIS steps within sequential Monte Carlo (SMC) samplers, with a particular focus on approximate Bayesian computation (ABC)-SMC algorithms, where we aim to use Pareto smoothing to reduce the use of Markov chain Monte Carlo (MCMC) moves, each of which requires simulation from a model that is often computationally expensive. Our empirical investigation suggests that there are only minimal benefits to using Pareto smoothing in SMC, since the variance reduction through using a sequence of targets dominates the impact of the weight adjustment.

The bootstrap is a versatile method for quantifying statistical uncertainty. Among its variants, a popular approach, the studentized bootstrap, provably achieves higher-order coverage error reduction compared to other benchmarks. However, its implementation typically requires an analytical form of the standard error, or otherwise an additional layer of resampling effort which can be computationally expensive. In this paper, we introduce what we call the studentized cheap bootstrap that achieves the same higher-order coverage accuracy as the conventional studentization, but substantially thinning the computational effort in the additional resampling layer to only very few Monte Carlo replications. Intriguingly, while conventional wisdom views "studentization" as an informal link between the bootstrap and t-distribution, we provide a first recognition that this link is in fact formal, notably with a distinct insight that the degree of freedom in the t-distribution corresponds to the Monte Carlo computation effort in the additional resampling layer, rather than the data size as in traditional thinking. Moreover, our desirable higher-order coverage accuracy builds crucially on this insight, as well as explicit calculations and geometric analyses of higher-order terms in the Edgeworth and Cornish-Fisher expansions tailored to limiting t-distributions.

Numerical integration involving expensive target functions is a common bottleneck in Bayesian inference and simulation. When a cheap surrogate is available, standard approaches such as reweighting or importance sampling often suffer from high variance and inefficient use of function evaluations. We introduce Slice Monte Carlo integration (S$\ell$MC), a method that leverages a Nested Sampling-like procedure on the surrogate to partition the space into informative strata, or slices, while generating samples in the parameter space drawn from the prior within each slice. This enables stratified Monte Carlo integration of the expensive target function over the surrogate-induced partition, yielding an efficient estimate of the target integral. The surrogate level sets therefore make the induced ordering of the parameter space the relevant information for S$\ell$MC, rather than the pointwise target-to-surrogate ratios governing importance-sampling efficiency. Another key advantage of S$\ell$MC is the decoupling of slice volume estimation from the evaluation of the target, allowing the refinement of the slice-volume estimates without additional target evaluations. We investigate the properties of S$\ell$MC, demonstrate how to efficiently generate posterior samples, and assess its robustness in controlled Gaussian benchmark families with continuously tunable surrogate-target mismatch.

We study the sample complexity of regularized optimal transport for general convex regularizations including the Kullback–Leibler divergence and $L^p$ penalties. Our main results are non-asymptotic bias and variance bounds for the empirical cost, with explicit dependence on the regularization parameter and on the intrinsic dimension of the marginals. Our approach simultaneously improves, unifies, and extends existing finite-sample bounds. In particular, we improve the state of the art for entropic optimal transport, and we obtain the first fully quantitative results for $L^p$ regularization with $1<p<\infty$. For the quadratic transport cost, we deduce that quadratically regularized optimal transport (i.e., $L^2$ regularization) estimates the unregularized optimal transport cost at rate $n^{-2/(d+4)}$, the fastest non-asymptotic rate currently available for any estimator based on regularized optimal transport.

We consider one-sample testing of a high-dimensional location parameter under elliptically symmetric distributions with heavy tails and pervasive cross-sectional dependence. We propose an elliptical regularized Hotelling test with Cauchy combination (ERHT–CC), based on the sample spatial median and the spatial-sign covariance matrix centered at that median. We derive its null asymptotic normality, consistent estimators of the centering and variance, and an explicit local power function. Since the power-optimal ridge parameter depends on the unknown alternative, we aggregate fixed-ridge $p$-values over a deterministic grid using the Cauchy rule. We establish a finite-grid joint Gaussian limit, justify the analytic combined $p$-value without estimating cross-ridge correlations, and characterize its local power. Simulation studies and an empirical analysis demonstrate the favorable finite-sample performance of ERHT–CC under heavy tails and pervasive dependence.

As machine learning models and datasets continue to grow, developing complex models has become increasingly computationally demanding. Knowledge distillation reduces deployment cost by compressing a large, well-trained teacher model into a compact student model, but it does not address settings where constructing the teacher itself is the bottleneck. Motivated by this challenge, we introduce Knowledge Cascade (KCas), a reverse knowledge distillation framework that uses information from a small, inexpensive student model to guide the development of a more complex teacher model. Although this direction is counterintuitive because the teacher typically has greater representational capacity, we show that student-to-teacher transfer can be principled when supported by statistical scaling relationships. We first develop KCas for nonparametric multivariate functional estimation in reproducing kernel Hilbert spaces via smoothing splines, where selecting multiple smoothing parameters is a major computational bottleneck. KCas transfers student-selected smoothing parameters to the full-sample regime through asymptotic scaling laws, substantially reducing computational cost for high-dimensional and large-scale datasets while retaining theoretical guarantees. Beyond smoothing splines, we illustrate the same principle through kernel density estimation and deep learning hyperparameter transfer. Simulations and real-data experiments show that KCas achieves substantial computational savings while maintaining strong statistical performance, and can sometimes outperform the corresponding full-sample procedure.

Malaria remains a major public health concern in many urban regions of India, where timely prediction of malaria incidence is essential for effective surveillance and resource allocation. This study examines count data approaches for understanding and predicting malaria incidence in the Mumbai region. The analysis used monthly $\textit{Plasmodium vivax}$ surveillance data from the Health Management Information System (HMIS) collected between 2012 and 2019, together with meteorological variables. Initial Poisson regression models suggested strong associations between malaria incidence and environmental factors; however, diagnostic assessment revealed substantial overdispersion, indicating that the Poisson model did not adequately capture the data's variability. Negative binomial regression provided a better representation of the data and indicated that seasonal effects were more strongly associated with malaria incidence than individual climatic covariates. Residual analyses further identified significant serial dependence not captured by baseline regression models. To address this limitation, a Generalized Linear Autoregressive Moving Average (GLARMA) framework was implemented to model temporal correlation explicitly. Forecasts were generated using simulation-based methods and evaluated through rolling time series cross-validation. The GLARMA Negative binomial model consistently demonstrated superior predictive performance and greater predictive stability than competing regression and time series approaches. These findings highlight the importance of jointly accounting for overdispersion and serial dependence in malaria surveillance data and demonstrate the value of count time series models for supporting early warning systems in urban settings.

Many statistical models have natural symmetries described by a group action. We study how such symmetries affect the comparison of two distributions. We work with a transformation model in which a group acts on both the sample space and the parameter space, and the densities transform with a multiplier. Under this assumption, we show that every $f$-divergence is invariant under the group action. As a consequence, an invariant divergence depends only on a maximal invariant of the pair of parameters. When the action on the parameter space is transitive, this maximal invariant is given by a double coset. We apply this result to multidimensional location-scale families, and we show that the same reduction applies to the Fisher–Rao distance.

We introduce an estimator for the pointwise, time-varying Hölder exponent (Hurst parameter) of a stochastic process, based on the geometry accumulation integral G_Lambda(t) = integral from Lambda to 1 of |eth_s X(t)| s^{-1} ds, where eth_s X(t) = (X(t+s)-X(t))/s is the scale derivative at resolution s. We prove consistency, noise robustness with explicit threshold Lambda* = sigma^{1/H}, and a CLT at rate (log Lambda)^{-1/2}. The estimator is pointwise in time, defined at finite resolution, and eliminates microstructure noise by scale separation. Existing estimators give a global H from integrated variance; this gives a time-varying H(t) directly from the price path.

Longitudinal data collected in clinical trials are almost always incomplete due to some of the participants dropping out from the study during the planned follow-up. A common strategy to handle nonresponse expresses missingness in terms of a dropout process, which is jointly analysed with the outcome process to facilitate the formulation of the missingness assumptions. However, when the outcome is multivariate, the identification of the dropout process becomes problematic, especially when individuals have different dropout times for each type of response, and sensitivity analysis is difficult. The modelling task may be also be complicated by data complexities (e.g. skewness and spikes) which are difficult to capture through standard parametric methods. An example of this analysis framework occurs in trial-based economic evaluations, where a longitudinal bivariate response, formed by suitably-defined measures of effectiveness and costs, is analysed to inform policymakers about the cost-effectiveness of alternative interventions. We present a novel Bayesian nonparametric approach to handle a missing bivariate longitudinal outcome by jointly modelling the dropout process associated with each type of response while also taking into account the complexities of the data. We specify a flexible nonparametric model for the observed data and partially identify the distribution of the missing data with identifying restrictions conditional on the dropout indicators and sensitivity parameters. We explore alternative nonignorable scenarios through different priors for the sensitivity parameters. Our approach is motivated by, and applied to, data from a trial assessing the cost-effectiveness of a new treatment for intellectual disability and challenging behaviour.

Mean Field Variational Inference (MFVI) is widely understood to underestimate posterior variance. By analysing conjugate Bayesian Linear Regression (BLR), we show that this characterization is incomplete: while MFVI underestimates the variance in parameter space, it can overestimate the predictive variance compared to the exact posterior. We show that if the MFVI posterior underestimates predictive variances in some directions, it necessarily overestimates them in others. Crucially, this overestimation occurs in directions where the training data concentrates. This leads to the surprising result that, for a test point drawn from the training distribution, MFVI's expected predictive variance exceeds that of the exact posterior. We demonstrate a pathological case of this effect, where the MFVI posterior fails to reduce predictive variance compared to the prior on in distribution data. We connect these results to the Cold Posterior Effect, arguing that varying the temperature can correct this overestimation, yielding predictions closer to those of the exact posterior. We validate our theory on synthetic and real-world regression tasks.

Generalized posteriors replace the likelihood by an exponentiated empirical criterion, but their credible sets generally lack asymptotic justification for frequentist coverage. General posterior calibration selects a scalar learning rate by estimating coverage with the bootstrap. Using Edgeworth expansions under regular fixed-dimensional asymptotics, we derive higher-order coverage expansions and analyze the stochastic approximation step used in the implemented algorithm. For a fixed nominal level, the root of the bootstrap coverage equation is consistent under a uniform coverage approximation and local identification. The higher-order expansions separate two sources of coverage error: the sampling Edgeworth correction for the estimator and the posterior Edgeworth correction for credible set boundaries, centres, and shapes. A scalar learning rate can calibrate all nominal levels in the Gaussian limit only when the posterior covariance and the sampling covariance are proportional. Hence, bootstrap calibration is a level-specific scale correction, not a remedy for general shape misspecification.

In applied health research, Markov cohort models are built on a precisely specified transition probability matrix. However, in many applications, the available evidence – transition counts, structural constraints, and treatment-effect data – identifies a set of admissible matrices rather than one uniquely justified matrix. This paper formulates an imprecise-probability extension in which inference yields lower and upper expectations over an evidence-compatible set of precise Markov cohort models. The contribution differs from existing imprecise Markov-chain work by focusing on finite-horizon cohort trajectories, additive accumulated outcomes, and transition matrices constructed from empirical transition counts. Under non-empty compact separately specified outgoing-row sets, the lower and upper accumulated outcomes are computed exactly by Bellman-style lower and upper transition operators. We prove the envelope theorem, reduction to the classical model, coherence properties of the lower transition operator, and algebraic conditions under which a single selected matrix yields a non-robust decision. We then show how multinomial transition counts induce admissible matrix sets through the Imprecise Dirichlet Model. A real-world cost-effectiveness example of patent foramen ovale closure after cryptogenic stroke illustrates the practical consequence: the empirical transition matrix slightly favors closure, whereas the imprecise analysis yields an incremental net monetary benefit interval crossing zero. The method provides both a rigorous lower-expectation formulation and a practical diagnostic for decisions that depend on transition probabilities not fully resolved by the evidence.

We extend the cone-ray exact-NMF pipeline of Ramteke (arXiv:2606.22451) from the uniform-support regime r_+ = r to the gap regime r_+ > r, and classify recoverable nonnegative factorisations by the rank of the W-factor into a three-regime taxonomy. Regime A (rank(W) = r_+, full column rank): a two-sided SVD-gauge cone-ray pipeline W = U_{r+}(G) Q, H = P V_{r+}(K)^T with G, K on Stiefel manifolds and square consistency Q P = diag(S_r, 0). On 10x10 dense random gap matrices it gives 100/100 recovery at r_+ = 5 and 6. We explain this via two geometric facts: slack enclosure (the data cone has codimension r_+ - r in the outer cone) and NRF-variety thickness (valid gauges form a positive-measure set, so the blind SVD lands on one with probability one). Regime B (rank(W) = r, W a column subset of M): a rank-deficient-W branch enumerating r_+-subsets of M's columns with per-column LP tests. On the block-diagonal family diag(C, J_k), where additivity of nonnegative rank collapses the valid gauges to a single point and the blind SVD pipeline fails, the column-subset branch restores recovery in milliseconds. Regime C (r < rank(W) < r_+, W not a column subset): exposed by the regular octagon's slack matrix. An exact size-6 NRF exists and is reachable by the symmetric formulation at an oracle gauge derived from a known factorisation (residual 1.5e-10), but the blind problem is open: 50 Haar random restarts and Riemannian gradient descent on the Stiefel/Grassmann gauge all stall, because the alt-LP residual is piecewise constant on cells of gauge-space, so local descent cannot cross cell walls. A combined toolkit (Regime B then A) covers regimes A and B with no regression on dense draws; Regime C remains open, with the regular octagon as the cleanest unsolved test case.

This work addresses the issue of partially missing correlations within the framework of bivariate and surrogate meta-analyses. While restricting the analysis to complete-case studies may appear to constitute the most straightforward analytical strategy, such an approach has been demonstrated to yield substantial inefficiencies and potential bias in the resulting estimates. Current methodological contributions in the literature circumvent this limitation either through aggregate estimation procedures grounded in likelihood-based frameworks under simplifying assumptions, or by resorting to deterministic imputation strategies, such as the empirical mean derived from observed units. In the present paper, we propose a multiple imputation framework in which imputation is performed via stochastic procedures based on a Beta regression model, thereby explicitly accounting for the missing at random (MAR) assumption underlying the observed missingness mechanism. We demonstrate the effectiveness of our method through an extensive simulation study across various scenarios, comparing our proposal with simple mean imputation and complete-case analysis under the MAR assumption.

Conditional sequential Monte Carlo (CSMC) algorithms arise in particle Markov chain Monte Carlo and a number of related settings. As in standard sequential Monte Carlo (SMC) algorithms, it is possible to employ a number of approaches to resampling within CSMC, but some additional care is required to arrive at a valid algorithm. We present a simple framework for implementing valid SMC and CSMC algorithms which (a) covers most known resampling schemes including those with a complicated dependence structure like systematic resampling, but also adaptive resampling, and even more `exotic' schemes like a version of chopthin resampling; (b) explains how to implement conditional analogues of these and other well known resampling schemes without randomly permuting/shifting the order of the ancestor indices; (c) requires only very weak assumptions which include neither (marginal) "unbiasedness" nor exchangeability.

Post-training hyperparameter selection is a critical step in the deployment of modern artificial intelligence systems, given the need to tune degrees of freedom of pre-trained models such as inference-time parameters, implementation-level settings, and thresholds driving decision rules. Despite its practical importance, hyperparameter selection is typically performed using best-effort empirical methods such as grid search or Bayesian optimization, which provide no formal statistical guarantees on reliability or safety. This monograph, intended for an audience of signal processing and machine learning researchers, presents a unified statistical framework for reliable post-training hyperparameter selection, centered on the learn-then-test (LTT) paradigm. LTT formulates the hyperparameter selection problem as multiple hypothesis testing over a candidate set of hyperparameters. The framework enables the choice of hyperparameters that provably satisfy application-specific reliability requirements—such as bounds on average risk, quantile risk, or information-theoretic constraints—with explicit, finite-sample control of error probabilities. The supporting statistical machinery, namely p-values, e-values, and concentration inequalities, is developed from first principles.

This paper introduces the Exact Ensemble Kalman Filter (ExEnKF), a novel algorithm for state estimation in discrete-time nonlinear filtering problems with linear observations. Unlike traditional Ensemble Kalman Filters (EnKFs), which approximate the filtering distribution using ensembles of Dirac measures, the ExEnKF employs Gaussian measures, enabling more efficient exploration of the state space and potentially alleviating the curse of dimensionality. We prove the algorithm's asymptotic consistency with the optimal filter (Theorem 3.1), establishing a convergence rate of order 1/ $\sqrt$ N for N particles. Numerical experiments on the Lorenz-96 multiscale model demonstrate that the ExEnKF outperforms the standard EnKF under model misspecification and poor initialization, particularly in highly stochastic regimes. The algorithm's robustness is further highlighted by its ability to track hidden components of the true signal, even when observations are generated from a different model (e.g., multiscale vs. single-scale). This work advances the theoretical understanding of ensemble methods in nonlinear filtering and provides a practical alternative to sequential Monte Carlo methods for high-dimensional systems

In this article, we propose a two-sample test based on the Wasserstein 2-distance between two 1D samples, where the first sample is assumed to be Gaussian based on an approximation of the underlying distribution by a log-normal distribution. A two-sample test is performed to decide whether two samples are drawn from the same distribution or not. We evaluate the proposed method on synthetic data and extend it to the multivariate case. We also show that this test is closely related to the Shapiro-Wilk normality test.

Building on the large-sample analysis of infinitesimal gradient boosting (Dombry and Duchamps, 2024), we study the fluctuations of the process around its deterministic limit and establish a functional central limit theorem: the rescaled deviations converge in distribution to a Gaussian process. The analysis is carried out in a reproducing kernel Hilbert space (RKHS) naturally associated with the softmax gradient tree base learner, in which the boosting process is characterized as the solution of an autonomous ordinary differential equation (ODE). The proof rests on a general stochastic perturbation analysis of ODEs in Banach spaces, which is of independent interest: whenever a sequence of vector fields converges and satisfies a central limit theorem, so does the associated ODE solution. We first illustrate this perturbation approach in the simpler setting of kernel gradient flow, where the Gaussian limit admits an explicit characterization, and then consider the more complicated tree-based gradient boosting setting.