Papers

Model selection and hypothesis testing are important tasks on networks. A key challenge lies in the inherent dependence in network data, as well as the fact that typically only a single realization is observed. As a result, many existing methods must be carefully tailored to specific models and only come with asymptotic theoretical guarantees. In this work, however, we propose a general model selection framework using Universal Inference, making our method widely applicable to various testing scenarios. Since Universal Inference requires two sets of data, we employ edge sampling to obtain proper networks with tractable dependence. We prove that the proposed statistic is an e-value, thus controlling the type I error rate in finite samples under nearly any hypothesis test. To our knowledge, this is the first Universal Inference-type statistic constructed from dependent splits of data as well as the first finite-sample testing guarantee for hypothesis testing on networks. We also prove that the logarithm of the test statistic diverges to positive infinity under various alternative models. On simulated and real-world networks, the proposed method performs well on tasks such as choosing the random graph model and the number of communities.

Epidemiologists increasingly use machine learning to adjust for high-dimensional confounding. Augmented inverse probability weighting (AIPW) and targeted maximum likelihood estimation (TMLE) are most widely used but may yield different results and both can become unstable under weak positivity violations. Residual-on-residual regression is a stable alternative that estimates an exposure effect encoded in a partially linear model by fitting confounder adjusted models for the outcome and exposure, then regressing outcome residuals against exposure residuals using ordinary least squares. We illustrate the approach using data from the Nulliparous Pregnancy Outcomes Study: Monitoring Mothers-to-Be (nuMoM2b; $n = 7{,}923$), estimating the association between high vegetable intake density and preeclampsia. Residual-on-residual regression, AIPW, and TMLE yielded concordant estimates, indicating a modest reduction in preeclampsia risk. In simulations, residual-on-residual regression was unbiased with near-nominal confidence interval coverage, performing comparably to AIPW and TMLE and substantially better than a misspecified parametric model when the exposure effect is approximately constant. However, in simulation settings with positivity violations, residual on residual regression outperformed AIPW and TMLE when the true effect was coded in a partially linear model. When the exposure effect is approximately constant, residual-on-residual regression is interpretable, computationally simple, and provides a triangulation strategy for observational causal inference.

With the emergence of various tensor data, tensor completion from partial measurements has attracted widespread attention in data science and signal processing. Total Variation (TV) has been widely used as an effective regularization technique for tensor completion; however, theoretical studies on TV regularization in this context remain limited. In this work, we present a rigorous theoretical analysis of TV regularization for tensor completion. Specifically, we consider tensor completion under exponential-family noise, which generalizes the standard settings such as Gaussian and Poisson tensor completion. To handle exponential-family tensor completion, we propose a family of dual-TV (DTV) regularizers based on the transformed L1 function, which simultaneously capture sparsity and low-rank structures in the gradient tensor. Moreover, we establish the theoretical upper bounds on the recovery error of the proposed estimator. In certain cases, these upper bounds can attain the convergence order of $\mathcal{O}\big( n_3 r_t\big(\max_{k} s_k^2\big) \log\big((n_1+n_2)n_3\big) /n \big)$, and the minimax lower bound analysis is further presented to show that the upper-bounds can approach the lower bound with the gap of order $\mathcal{O}(\max_k s_k^2/max(n_1, n_2))$ up to a logarithmic factor. Finally, multiple groups of experiments on synthetic, image and video tensor data sets are conducted to support our theoretical results and demonstrate the effectiveness of our method.

Two-sided marketplaces connect distinct user groups whose interests often conflict – improving outcomes on one side could degrade the other side's experience. To address this challenge, we deploy an integrated framework for personalizing free-value thresholds – a policy governing the scope of complimentary services for job listings – across a two-sided job marketplace connecting millions of employers and job seekers. Our personalized policy delivers statistically significant and economically sizable lift in the target metric while respecting engagement guardrail constraints. Direct application of standard uplift methods proves insufficient here for two reasons. First, cross-side externalities demand multi-objective optimization: maximizing employer-side metrics risks harming job seeker engagement, with effects varying substantially across job segments. Second, marketplace interference necessitates cluster-level randomization, limiting us to few discrete treatment levels – effectively a form of positivity violation that rules out methods designed for continuous treatments. We contribute an integrated framework with three components. Our ensemble-based hybrid ranking models target and guardrail metrics separately, cutting guardrail risk by over 10% for equivalent target gains compared to single-objective approaches. A treatment effect extrapolation method extends our estimates from limited experimental variation to untested policy levels, relying on monotonicity assumptions that we validate empirically. Finally, we present production deployment, where post-launch data confirms both extrapolation accuracy and guardrail compliance. Our deployed system demonstrates that principled methodology can enable meaningful personalization even when experiments are severely constrained and different objectives compete – common conditions that characterize many real-world marketplaces.

Modern deep learning has been shown to operate at the edge of stability, routinely using learning rates far larger than those justified by classical optimization theory. Most prior analyses of the edge of stability phenomenon focus on deterministic gradient descent, leaving the stochastic setting largely unexplored. In this work, we provide sharp convergence guarantees for Stochastic Gradient Descent (SGD) applied to the multiclass cross-entropy loss, for both linear classifiers and two-layer neural networks. We show that the stochasticity of SGD may cause the dynamics to alternate between an edge-of-stability regime that is dominated by curvature-driven oscillations, and a stable regime in which the expected loss decreases at a controlled rate. Despite that, we prove that SGD self-stabilizes the dynamics, ensuring that the iterates return to stability in a fixed number of iterations and allowing convergence in the best-iterate sense even with large learning rates. Experiments validate our theoretical findings and illustrate the benefits of SGD in the large-stepsize regime.

Imitation Learning is a natural framework for learning in sequential decision-making systems and has emerged as the dominant paradigm through which we understand language model training. A central puzzle is that, while in theory offline IL can be horizon-free and optimal, in practice online methods such as on-policy distillation often outperform offline methods such as supervised fine-tuning. We propose a noisy expert model to explain this gap, in which the learner only has access to a noisy version of the expert's policy, but wishes to compete against the reward achieved by a clean expert, motivated by the fact that in many applications, e.g. training language models to perform long chains of thought, the expert is often imperfect. In this setting, we show a sharp separation between offline and online IL. Offline learning from noisy trajectories is fundamentally hard: to compete with the clean expert, the sample complexity must grow exponentially, in contradistinction to the clean expert setting where no explicit horizon dependence exists. In contrast, we prove that online interaction with the noisy expert via a novel variant of OPD enables polynomial dependence on the horizon in general. We further show that, under a natural condition on the expert noise distribution, which we show to be necessary for any horizon-free sample complexity, one can obtain such a guarantee, although our proposed algorithm sacrifices statistical efficiency in its dependence on the size of the policy class. Our analysis leads to an alternative loss function that is commonly considered empirically for LM training. We further provide algorithms and lower bounds, and extend our results to the more realistic setting of unknown corruption when the clean expert is deterministic, thereby providing a theoretical foundation for why OPD can outperform SFT when training language models from imperfect teachers.

Cross-fitting is not a refinement of survey-weighted causal machine learning but, once the nuisances are flexible, what restores valid inference. We study the population average treatment effect under a stratified multistage design, estimated by a survey-aware targeted maximum likelihood estimator (TMLE) whose variance is obtained by Taylor-series linearization of the influence function, treating the primary sampling unit as the replication unit. Our central result is that this validity turns on cross-fitting at the cluster level: sufficiency is established in theory, and the failure without it is shown in simulation. Once flexible learners cross a complexity (Donsker) boundary, single-fit survey TMLE can severely under-cover, and internal cluster-aware cross-validation does not substitute for cross-fitting; among the estimators we evaluate, only out-of-fold fitting at the cluster level restores valid coverage. In simulations spanning a many-PSU and an NHANES-like design, on a diverse ensemble the single-fit and internal cross-validation estimators cover at about 0.89-0.91 and 0.85-0.88 while the cross-fitted estimator holds at 0.93-0.95, and an aggressively grown learner drives single-fit coverage to 0.18-0.22. Two scope choices are deliberate: survey-weighted point estimation is prior work, and the nuisance product-rate condition is assumed and probed empirically. Within these conditions we prove asymptotic normality and design-consistency of the linearization variance. Four NHANES analyses and open-source software illustrate the method.

Alternating recurrent events – event-times of a specific nature that trigger a secondary refractory period – occur in a wide-range of fields, including behavioral science, criminal justice, and biostatistics. Analysis of these events requires careful attention to the statistical nuance, including correlated observations and repeated outcomes subject to potential censoring. We develop an online dynamic prediction framework appropriate for predicting subsequent alternating recurrent events, by developing neural network theory for a statistical audiences and applying inverse probability weighted pseudo-observations. The proposed model is applied to dynamically predict alternating recurrent event-free time, showing good performance in simulation, and outstanding capability in application to predicting periods of low mood for first-year medical residents. We close with a discussion.

We prove a version of the Dvoretzky-Kiefer-Wolfowitz inequality for Markov chains with a regenerative structure. Suppose we have a regenerative Markov chain with stationary distribution $\pi$. Given a functional $\theta$ on the state space and a confidence level $1-\delta$, our result provides a uniform $1-\delta$ confidence band for the CDF of $\theta$ under $\pi$ based on the empirical CDF. By inversion, we get a $1-\delta$ confidence band for the quantile function of $\theta$ under $\pi$. Our bounds are fully explicit and nearly optimal. In addition, they are data-dependent in the following sense: in the formula for the width of the confidence band, the leading term can be computed directly from the sample path without any a priori information about the convergence rate of the chain. A convergence bound is required, but it contributes to the width of the confidence band only through a lower-order term. For this reason, our result is attractive for Markov chains whose convergence rate is much quicker in practice than what can be proved in theory. Data-dependent bounds of this type are called empirical concentration inequalities in the literature. Thus, our result is an empirical concentration inequality for the empirical CDF of $\theta$ given the sample path.

Consider a nonparametric regression model with one-dimensional covariates and a continuous regression function. Assume that the regression function from the left of the covariate support starts equal to zero and then changes at some unknown point. Our aim is to estimate this gradual change point. We define and compare various consistent estimators based on a new general optimization method in the case where the aim is to estimate the largest minimization point of some objective function. We discuss rates of convergence and estimating the regression function based on the gradual change structure. Bootstrap bias approximation is discussed. Further applications in a two sample case are considered, where two continuous regression functions first equal and then change at some point of interest.

The number of trees is a central computational parameter in Random Forests: increasing it reduces finite-ensemble variability but increases training and prediction cost. Plateau-based tuning adapts this parameter through local comparisons of out-of-bag scores at a geometric triplet of tree counts. After the remaining hyperparameters have stabilized, however, the central triplet point need not converge to a deterministic value; instead, it fluctuates around a stationary regime. This paper develops a stationary-distribution theory for this process. The central ensemble size $B_t$ is modeled as a birth-death Markov chain on a geometric grid, and its stationary distribution is derived through local balance. Under a leading centered folded-normal approximation, equilibrium equations are obtained for the original update rule and a symmetric modified variant, implying that the stationary center $B_*=O(\varepsilon^{-2})$ as $\varepsilon\downarrow 0$. The stationary spread is also characterized. A local Gaussian approximation and a Fokker-Planck interpretation give grid-level variance constants. After conversion to the ensemble-size scale, $\sigma_{B,*}=O(\varepsilon^{-2})$, while the variance is $O(\varepsilon^{-4})$. The leading relative spread is independent of $\varepsilon$ and controlled by the scale factor and update rule. These results interpret plateau-based Random Forest tuning as a stochastic process rather than a deterministic stopping rule.

We study the squared singular value spectrum of a non-square product of independent real Gaussian matrices, equivalently the feature covariance spectrum of a deep linear neural network at initialization. Starting from the fixed-$m$ covariance diffusion previously obtained in the proportional depth-width limit, we record an equivalent matrix realization, describe its affine invariance, and derive the interacting diffusion satisfied by its eigenvalues. We then take a second limit, sending $m\to\infty$ on the accelerated spectral clock $\tau=mt$, which corresponds in this sequential construction to the relation $dm/n\to\bar\tau$. We establish convergence of the empirical spectral measure path to a deterministic mean-field limit and derive a closed Burgers equation for its $T$-transform. Together with the proportional depth-width limit, these results give a rigorous sequential route from the deep non-square Gaussian product to the free log-normal limit of its feature covariance spectrum; for more general initial laws, the transform yields a free multiplicative convolution form. We further analyze the support of the free log-normal law, give a fixed point iteration for numerical evaluation and a formal Marchenko–Pastur approximation at small time, and use the limiting spectrum to predict the risk in a toy random feature model.

In this paper, we attempt to enhance the theoretical understanding of convolutional neural networks (CNNs) as feature extractors in classification tasks by analyzing them through the lens of Cover's function-counting theory. Specifically, our focus lies on the notion of separation capacity, a combinatorial quantity derived from counting the number of realizable dichotomies (i.e., binary label assignments). Our contributions are threefold. First, we extend Cover's framework by establishing a conceptually insightful and practically useful formulation for the separation capacity. Second, leveraging this formulation, we identify the factors governing the separation capacity of feature extractors that employ a specific CNN architecture, so-called scattering networks, in terms of their network building blocks. Third, we provide practical insights for scattering network design.

The U.S.\ Census Bureau's Low Response Score (LRS) is a central planning instrument for identifying places likely to require additional self-response outreach and nonresponse follow-up. The published LRS is intentionally interpretable: it is built from tract-level covariates using an ordinary least squares specification. That transparency, however, leaves open an important question for official statistics: how much spatial structure remains after the own-tract covariates have done their work, and what form does that structure take? Using the observed 2010 Census mail non-return rate for 71,076 U.S. census tracts and the twenty-five Erdman–Bates LRS predictors, this paper compares the full spatial autoregressive model family under queen-contiguity weights and validates the leading candidates with both random and spatial-block cross-validation. OLS leaves strong residual spatial autocorrelation ($I=0.399$). Formal diagnostics and model comparisons indicate that the remaining dependence is primarily error-type rather than a global endogenous lag process. Although the spatial Durbin model minimizes in-sample AIC, spatial-block validation reverses that ranking: the error-family models (SEM/SDEM) generalize best, while the AIC-best SDM is weakest out of sample. The SDEM provides an interpretable middle ground, absorbing residual spatial dependence while representing neighborhood demographic effects as local spillovers. Robustness checks show that these conclusions are invariant to the weights definition and are not an artifact of tract-size-driven heteroskedasticity. The results suggest that LRS-style response models should be evaluated with spatial validation, not only in-sample fit, and that local neighborhood context can be operationally meaningful without invoking a global response-contagion mechanism.

We develop a first-principles reduced-order model of these dynamics. Under a single mean-field assumption that summarizes the policy by its expected reward, we reduce the GRPO update to a stochastically-forced damped oscillator whose mass, damping, and stiffness are fixed in closed form by the optimizer hyperparameters together with a single measured curvature scale – momentum supplies the inertia, off-policy lag erodes the damping, and the group size enters, to leading order, as a noise temperature. The reduction has three consequences. First, it subsumes the empirical single-exponential saturation law as its overdamped limit, recasting the fitted plateau, timescale, and size exponent as the fixed point, inverse stiffness, and curvature-scaling exponent of the underlying potential, and adding, through the retained inertial term, the slow-start phase the single exponential cannot represent. Second, it yields predictions tied to independently measurable quantities rather than fitted ones: group-size invariance of the deterministic trajectory with a $1/G$ stationary fluctuation, a sharp stability threshold in the refresh interval, and an overdamped-to-oscillatory transition. Third, it furnishes diagnostics that separate failure modes a reward curve alone conflates – reward hacking, advantage degeneracy, policy concentration, and dynamical instability. Across three models and two group sizes, the closed-form trajectory fits training reward to $R^2 \geq 0.91$ and the mean trajectory is group-size invariant to leading order – on both the reward curve and out-of-distribution transfer to eight math benchmarks – while the within-group reward spread retains a residual $G$-dependence that the leading-order temperature picture does not capture.

Structured tabular data dominates clinical medicine, yet existing benchmarks fail to reflect real-world properties like complex survey sampling, demographic oversampling, and subgroup fairness. We introduce the NHANES Accelerometry Cardiometabolic Benchmark, derived from NHANES 2003-2006, comprising 1,381 adults with hip-worn accelerometry, fasting laboratory biomarkers, dietary intake, and anthropometrics. We evaluate three tabular learning methods – ridge regression, XGBoost, and the foundation model TabPFN v2 – to predict glycated haemoglobin (HbA1c), fasting triglycerides, and C-reactive protein (CRP) from activity phenotypes and lifestyle covariates. TabPFN v2 achieves the best overall performance (HbA1c R^2=0.156, CRP R^2=0.383), while triglycerides remain largely unpredictable (R^2 < 0.05), consistent with known genetic dominance. We apply split conformal prediction to generate distribution-free 90% prediction intervals and evaluate demographic coverage equity across sex and race/ethnicity subgroups. Marginal coverage aligns with the 90% target for CRP and HbA1c but falls below for triglycerides. At the subgroup level, we observe localized undercoverage (e.g., HbA1c for Mexican American participants), illustrating the gap between marginal guarantees and the conditional coverage required for clinical fairness. Code and data are at https://github.com/felizzi/nhanes-accel-cardiometabolic-benchmark.

Conservative offline training is widely advocated as a safe foundation for subsequent online adaptation: if a policy stays close to well-supported behaviour, the argument goes, it is less likely to exploit imperfections in a learned reward model. We challenge this intuition empirically and mechanistically. We train a Qwen3-14B policy under Direct Preference Optimisation (DPO) with three levels of conservatism ($\beta \in \{\beta_{\mathrm{lo}}, \beta_{\mathrm{mid}}, \beta_{\mathrm{hi}}\}$ derived from empirical log-ratio percentiles), then adapt each checkpoint online against a learned reward ensemble (3\,$\times$\,Qwen3-1.7B) while measuring true performance on GSM8K exact-answer accuracy. We find that higher offline conservatism monotonically increases reward-hacking damage, measured by the Goodhart gap and its area under the curve (AUGC), with Spearman $\rho = 1.0$ across all three conditions. Mechanistic analysis reveals a three-link causal chain: (i) high-$\beta$ DPO compresses policy entropy, (ii) Low-entropy policies generate responses with reduced diversity, concentrating in a narrow region of the reward model's training distribution (lower pairwise cosine distance), and (iii) despite this proximity, ensemble disagreement (epistemic uncertainty) increases with $\beta$ and is exploited faster during online optimisation. We further fit a power-law curve to the $(\beta, \augc)$ data and identify a practical optimal conservatism level $\beta^{\star}$ that balances alignment fidelity against hacking vulnerability. Our results suggest that the field needs calibrated, not maximal, conservatism.

Contrastive embedding models trained with scale-invariant losses are typically paired with distance metrics like cosine similarity, effectively ignoring embedding magnitudes. However, surprisingly, empirical studies reveal that despite this, these "discarded" norms seem to correlate with semantic properties such as concept specificity, token frequency, and human uncertainty. In this work, we provide a formal theoretical framework explaining this phenomenon. By analyzing the optimization dynamics, we derive an analytic formula demonstrating that embedding length naturally encodes this information as a byproduct of the training process. We also show how this gives rise to signals that can serve as "free" calibration tools in specific models and retrieval tasks, providing a grounded explanation for a previously heuristic observation.

The problem of the mean-square optimal estimation of the linear functionals which depend on the unknown values of a multidimensional homogeneous random field from observations of the field with noise is considered. The minimax (robust) method of estimation is applied in the case where the spectral densities of the fields are not known exactly while some sets of admissible spectral densities are given. Formulas that determine the least favourable spectral densities and the minimax spectral characteristics are derived for some special sets of admissible densities.

Modern statistical learning problems often involve multiple related data sets, in which learning efficiency on a target set can be improved by borrowing information from related sources, while source heterogeneity may introduce bias. Existing approaches are limited by suboptimal performance in multi-source settings, insufficient use of covariance information, or the computational burden of tuning procedures. We propose a covariance-aware shrinkage framework that constructs shrinkage directions using covariance information to improve efficiency. We establish finite-sample risk bounds that yield an explicit risk-improving interval for the shrinkage size, making the transfer strength fully data-driven. When multiple source sets are available, we further propose a sequential algorithm that applies source-specific shrinkage updates according to a priority score. Under a mixed-source separation condition, the estimator asymptotically attains the oracle risk and, when the pooled discrepancy is large, strictly improves over single-step shrinkage. The framework is further extended to a broad class of smooth M-estimation problems. Numerical studies show substantial gains over competing methods, especially when the source data sets are highly heterogeneous.

Many modern generative modeling methods, including diffusion models, normalizing flows, and flow matching, estimate transport maps or plans between distributions without explicitly targeting an optimal transport (OT) map. In applications like generative modeling, the transport cost itself is irrelevant, and this makes it natural to target maps which are more tractable from either a statistical or computational standpoint. In this short note, we formalize the task of estimating any valid transport map in a rigorous minimax framework. One consequence of this framing is that it yields sample complexity lower bounds for any method whose learned object is evaluated as a transport map or plan, including flow matching and diffusion-based generative models, in settings where direct analysis would be challenging due to the analytic complexity of the methods and their target maps. We observe that, under standard, though strong, stability assumptions from the OT literature, estimating any valid transport map is statistically as hard as estimating the OT map. We complement these results with some examples showing that when these stability assumptions fail, alternative transport maps can be learned substantially more accurately than the OT map. Our minimax framing provides a rigorous foundation for understanding the statistical limits of modern transport-based generative methods and clarifies when targeting sub-optimal maps can provide real statistical advantages.

We characterize weakly regularized continual classification in homogeneous models as sequential projections onto task margin sets. This result generalizes prior analyses restricted to either stationary (single-task) deep models or continual linear models. We show that global convergence generally fails, even for simple models linear in data but nonlinear in parameters. Nevertheless, by leveraging results from nonconvex projection theory, we identify regularity properties of homogeneous deep networks that guarantee local linear convergence under random and cyclic task sequences. Finally, we extend our analysis to continual regression, unifying the framework for homogeneous models.

Covariance matrices serve as compact descriptors of feature distributions in many machine-learning pipelines, including domain adaptation and Gaussian embeddings. Under a centered Gaussian approximation, the unregularized Wasserstein-2 optimal-transport (OT) discrepancy admits a closed form on covariances given by the Bures-Wasserstein (BW) objective on the symmetric positive definite (SPD) cone. We propose ITSPACE (Iterative Transport for Stable Proximal Alignment of Covariance Embeddings), a proximal majorization-minimization method that directly optimizes this exact BW objective through closed-form updates in a square-root factorization. In exact arithmetic, each iteration satisfies a sufficient-decrease inequality for the BW objective; under inexact polar computations, we provide an explicit certificate-gap bound controlling deviations from exact descent. The resulting iterations preserve PSD structure by construction and naturally support rank-restricted factors, making ITSPACE well-suited as a lightweight inner-loop primitive in settings where adaptation must be performed from unlabeled target batches under strict step and compute budgets. Across real-world covariance-alignment benchmarks, ITSPACE reaches low-BW-gap solutions substantially faster than BW-gradient descent, methods based on other covariance geometries, and entropically regularized sample-OT baselines.

Why overparameterised deep networks generalise so remarkably well remains one of the most stubborn open questions in machine learning theory. Classical frameworks like VC dimension and Rademacher complexity predict catastrophic overfitting in modern models, leaving a massive theoretical gap between theory and reality. In this paper, we bridge this divide by introducing a unified framework that links information theory, topology, and statistical mechanics to map the hard limits of deep learning. Central to our approach is the Entropic Learnability Horizon (ELH): a fundamental law stating that a network can only truly learn a target function if the Shannon entropy of the data manifold outpaces the topological entropy of the function's decision boundary, balanced by the von Neumann entropy of the network's weight space. We establish the Shannon-Topological Bottleneck Theorem, proving that when a target boundary's geometric complexity exceeds this informational horizon, the system undergoes a sudden entropic phase transition. It falls into a state of Informational Frustration - a glassy, rigid memorization phase where generalization becomes thermodynamically impossible. Using this lens, we show that the enigmatic phenomenon of "grokking" is actually an Entropic Release, where weights abruptly reorganise to unlock the bottleneck. Finally, we translate this theory into practice with Entropic Gradient Descent (EGD), an optimization algorithm that dynamically manages weight entropy to keep learning on track. Ultimately, this work repositions entropy not just as a tool for tracking uncertainty but as the fundamental physical currency that dictates whether a machine can learn.

We propose doubly robust adaptive conformal inference (DR-ACI), which constructs prediction intervals for doubly robust pseudo-outcomes under temporal dependence.

Normalizing Flows excel at modeling a single fixed density, yet many problems across the sciences, such as high energy physics, instead require modeling how that density deforms as a function of continuous parameters: the strength of a physical effect, a calibration constant, or a source of systematic uncertainty. Learning a separate flow for every parameter configuration quickly becomes intractable, since the number of joint settings grows exponentially with the number of parameters. We introduce Factorizable Normalizing Flows (FNFs), which represent the parameter-dependent density as a fixed, high-fidelity flow for a reference configuration composed with a learnable transformation that is polynomial in the parameters and factorized over them. This structure has a practical consequence: each parameter's effect is learned in isolation, from samples in which that parameter alone is varied. The combined response of many parameters is then recovered by summation at inference, without ever sampling their combinatorially large joint space. On a controlled problem with two interpretable deformations applied jointly to the data, the learned transformation reproduces the true deformations and matches the optimal likelihood, while optional interaction terms capture residual correlations when several parameters vary strongly at once. The resulting model is interpretable, scales linearly with the number of parameters, and keeps the likelihood tractable. This provides a general tool for any inference workflow requiring continuous density morphing, and directly enables the next generation of unbinned likelihood fits in high energy physics.

We consider sparse multivariate stochastic systems that evolve in continuous time according to a causal mechanism and present methodology to recover the system's time-infinitesimal transition mechanism from mere cross-sectional data. This observational paradigm is motivated by applications such as gene expression analysis, where destructive experimental techniques may only allow recording data once over a cell's lifetime. Precisely, we assume the system follows a time-homogeneous diffusion process that has reached an equilibrium distribution at observation time. Further, we assume the causal mechanism is fully described by the diffusion drift, is acyclic, and its causal structure graph is known. In this setting, we prove that the full causal mechanism, i.e., the drift function, can be non-parametrically identified under a weak non-explosion criterion. We derive a non-parametric kernel estimator for this challenging inverse problem and prove its consistency. Moreover, we propose a cross-validation scheme for hyperparameter tuning, illustrate the behavior of our estimator in simulations, and we discuss connections with irreversible generative diffusion models and low-frequency sampled data.

The standard convergence analysis of mini-batch stochastic gradient descent (SGD) models gradient noise using a single variance term that treats all parameter directions equally, ignoring the fact that noise in high-curvature directions has less impact because learning rates are already constrained there. We introduce Curvature-Weighted Gradient Diversity (CWGD), a geometry-aware measure that weights per-sample gradient diversity by the inverse square root of the Hessian, providing a tighter proxy for the effective optimization noise. For strongly convex quadratic objectives with diagonal Hessians and isotropic noise, we prove that a CWGD-modulated cosine learning-rate schedule can reduce the asymptotic optimization error floor by up to a factor of two compared with standard cosine annealing. We implement this idea as CWGD-Cosine using a Hutchinson-based diagonal Hessian estimator that is exact for quadratic objectives. Across a range of condition numbers, batch sizes, and noise structures, CWGD-Cosine consistently achieves approximately 20% lower final optimization error than standard cosine annealing while incurring negligible overhead in the quadratic setting. We also identify and correct a degenerate curvature estimator, analyze the robustness of the proposed estimator, and explicitly discuss the limitations of the method, including Hessian staleness in non-convex optimization. These results establish CWGD as a principled geometry-aware measure of optimization noise and motivate future extensions to more general learning problems.

Neural networks are known to be susceptible to over-reliance on spurious correlations. However, the precise mechanism by which models exploit shortcut features is not fully understood, and algorithms to mitigate this behavior rely on as yet unjustified assumptions about the learned representations. In this work, we provide the first end-to-end theoretical characterization of spurious feature learning for two-layer ReLU neural networks trained by online minibatch SGD on the logistic loss. We consider data drawn from the high-dimensional Boolean hypercube with a quadratic signal function (namely XOR) and a linear spurious correlation. We show that SGD learns the spurious feature first, and exponentially fast. Moreover, the optimization dynamics couple the spurious and signal features, with a stronger spurious component inhibiting signal feature learning. Our analysis reveals precise phase transitions in the learning dynamics. In the first phase, alignment between the signs of the spurious feature and second-layer weight drives rapid growth of the spurious feature. In the second phase, large majority group margin slows learning and the signal feature remains suppressed. When the spurious correlation is maximally strong, we show theoretically that the spurious feature dominates even at the sample complexity threshold where XOR would be learned in isolation (i.e., if the spurious feature was absent). In contrast, when the correlation strength is constant, we provide preliminary empirical evidence that the model can eventually learn the XOR signal, although the spurious feature is not forgotten.

We study a two-dimensional Ornstein-Uhlenbeck system where only the first coordinate is observed, whereas the second coordinate remains hidden. Our goal is the estimation of the coupling parameter in the drift of the observed coordinate. The core novelty lies in accounting for the influence of the observed component on the unobserved one, making the system fully coupled. Using linear filtering, we derive the likelihood under partial observation and establish local asymptotic normality of the statistical model. Within the Ibragimov-Hasminskii framework (1981), we prove consistency, asymptotic normality, convergence of moments and asymptotic efficiency of the MLE under stability and identifiability assumptions as the time horizon tends to infinity.