Papers

The inverse Potts problem for estimating evolutionary single-site fields and pairwise couplings in homologous protein sequences from their single-site and pairwise amino acid frequencies observed in their multiple sequence alignment would be still one of useful methods in the studies of protein structure and evolution. Since the reproducibility of fields and couplings are the most important, the Boltzmann machine method is employed here, although it is computationally intensive. In order to reduce computational time required for the Boltzmann machine, parallel, persistent Markov chain Monte Carlo method is employed to estimate the single-site and pairwise marginal distributions in each learning step. Also, stochastic gradient descent methods are used to reduce computational time for each learning. Another problem is how to adjust the values of hyperparameters; there are two regularization parameters for evolutionary fields and couplings. The precision of contact residue pair prediction is often used to adjust the hyperparameters. However, it is not sensitive to these regularization parameters. Here, they are adjusted for the fields and couplings to satisfy a specific condition that is appropriate for protein conformations. This method has been applied to eight protein families.

This paper provides an overview of recent developments in quickest change detection (QCD) for high-dimensional multi-sensor systems, with an emphasis on settings involving structural constraints and limited sensing resources. Classical QCD methodologies, while well understood in low-dimensional and fully observed regimes, face significant challenges when extended to modern applications characterized by large-scale data, constrained sampling or communication, and heterogeneous signal structures. We review key approaches for handling high dimensionality, including methods that exploit sparsity, and other forms of signal heterogeneity. Additionally, we discuss sampling constraints, where observations must be selected or acquired sequentially under resource limitations. Multi-stream applications can require making multiple detections, for example when detecting changes separately in different streams. The underlying assumptions on probability models, the types of changes taking place, commonly used decision-making criteria, performance indices, and error types are described. We also briefly discuss the application of machine learning in cases where the underlying probability models are not known or there is a need to select which sensors should monitor the phenomena because of the large scale of the system.

Uncertainty quantification is crucial in safety-critical systems, where decisions must be made under uncertainty. In particular, we consider the problem of online uncertainty quantification, where data points arrive sequentially. Online conformal prediction is a principled online uncertainty quantification method that dynamically constructs a prediction set at each time step. While existing methods for online conformal prediction provide long-run coverage guarantees without any distributional assumptions, they typically assume a full feedback setting in which the true label is always observed. In this paper, we propose a novel learning method for online conformal prediction with partial feedback from an adaptive adversary-a more challenging setup where the true label is revealed only when it lies inside the constructed prediction set. Specifically, we formulate online conformal prediction as an adversarial bandit problem by treating each candidate prediction set as an arm. Building on an existing algorithm for adversarial bandits, our method achieves a long-run coverage guarantee by explicitly establishing its connection to the regret of the learner. Finally, we empirically demonstrate the effectiveness of our method in both independent and identically distributed (i.i.d.) and non-i.i.d. settings, showing that it successfully controls the miscoverage rate while maintaining a reasonable size of the prediction set.

Machine learning has become integral to medical research and is increasingly applied in clinical settings to support diagnosis and decision-making; however, its effectiveness depends on access to large, diverse datasets, which are limited within single institutions. Although integrating data across institutions can address this limitation, privacy regulations and data ownership constraints hinder these efforts. Federated learning enables collaborative model training without sharing raw data; however, most methods rely on complex architectures that lack interpretability, limiting clinical applicability. Therefore, we proposed a federated RuleFit framework to construct a unified and interpretable global model for distributed environments. It integrates three components: preprocessing based on differentially private histograms to estimate shared cutoff values, enabling consistent rule definitions and reducing heterogeneity across clients; local rule generation using gradient boosting decision trees with shared cutoffs; and coefficient estimation via $\ell_1$-regularized optimization using a Federated Dual Averaging algorithm for sparse and consistent variable selection. In simulation studies, the proposed method achieved a performance comparable to that of centralized RuleFit while outperforming existing federated approaches. Real-world analysis demonstrated its ability to provide interpretable insights with competitive predictive accuracy. Therefore, the proposed framework offers a practical and effective solution for interpretable and reliable modeling in federated learning environments.

This paper develops a framework for identification, estimation, and inference on the causal mechanisms driving endogenous social network formation. Identification is challenging because of unobserved confounders and reverse causality; inference is complicated by questions of equilibrium and sampling. We leverage repeated observations of a network over time and random variation in initial ties to address challenges to causal identification. Our design-based approach sidesteps questions of sampling and asymptotics by treating both the set of nodes (individuals) and potential outcomes as non-random. We apply our approach to data from a large professional services firm, where new hires are randomly assigned to project teams within offices. We estimate the causal effect on tie formation of indirect ties, network degree, and local network density. Indirect ties have a strong and significant positive effect on tie formation, while the effects of degree and density are smaller and less robust.

Generative modeling within constrained sets is essential for scientific and engineering applications involving physical, geometric, or safety requirements (e.g., molecular generation, robotics). We present a unified framework for constrained diffusion models on generic nonconvex feasible sets $\Sigma$ that simultaneously enforces equality and inequality constraints throughout the diffusion process. Our framework incorporates both overdamped and underdamped dynamics for forward and backward sampling. A key algorithmic innovation is a computationally efficient landing mechanism that replaces costly and often ill-defined projections onto $\Sigma$, ensuring feasibility without iterative Newton solves or projection failures. By leveraging underdamped dynamics, we accelerate mixing toward the prior distribution, effectively alleviating the high simulation costs typically associated with constrained diffusion. Empirically, this approach reduces function evaluations and memory usage during both training and inference while preserving sample quality. On benchmarks featuring equality and mixed constraints, our method achieves comparable sample quality to state-of-the-art baselines while significantly reducing computational cost, providing a practical and scalable solution for diffusion on nonconvex feasible sets.

Discussion on “Regression by Composition” by Farewell, Daniel, Stensrud, and Huitfeldt

Discussion on "Regression by Composition" by Farewell, Daniel, Stensrud, and Huitfeldt.

In this paper, we study the stability of the shadow, a projection of a measure onto the set of couplings with respect to the Wasserstein distance. The shadow was introduced by \citet{Eckstein_Nutz_2022} to analyze the stability of the Sinkhorn algorithm, and was recently revisited by \citet{kim2026extensioncouplingprojectionoptimal} for statistical applications. Under mild conditions, we establish the bi-Hölder continuity of the shadow. As a consequence, we also derive the sample complexity of the shadow by combining smoothing techniques with recent results on the rate of convergence of empirical measures in Wasserstein distance. The key idea of the proof is twofold: first, a contraction property of the $L^p$ projection, recently used independently by \citet{kim2025stabilitywassersteinprojectionsconvex} and \citet{alfonsi2025wassersteinprojectionsconvexorder} to study the stability of projections onto the convex order cone in Wasserstein space; and second, the Hölder continuity of optimal transport maps established by \citet{Quantitative_stability_duke2023}, together with its recent extension by \citet{mischler2025quantitativestabilityoptimaltransport}.

In this paper, we survey results on the asymptotic behavior of the variance of the best linear unbiased estimator (BLUE) for the mean of stationary processes. This behavior is influenced by the regularity and memory structures of the observed models. The results show that the asymptotic behavior of the variance of the BLUE is determined solely by the behavior of the spectrum near the origin. For nondeterministic models, the variance of the BLUE exhibits hyperbolic behavior, similar to the power function, while for purely deterministic models, the variance decreases at an exponential rate. Specifically, a necessary condition for the variance of the BLUE to approach zero exponentially is that the spectral density of the model vanishes on a set of positive Lebesgue measure in any neighborhood of zero. We also present results on the asymptotic efficiency of various unbiased linear estimators in comparison to the BLUE.

Reliable deployment of language models requires two capabilities that appear distinct but share a common geometric foundation: predicting whether a model will accept targeted behavioral control, and detecting when its internal structure degrades. We show that geometric stability, the consistency of a representation's pairwise distance structure, addresses both. Supervised Shesha variants that measure task-aligned geometric stability predict linear steerability with near-perfect accuracy ($\rho = 0.89$-$0.97$) across 35-69 embedding models and three NLP tasks, capturing unique variance beyond class separability (partial $\rho = 0.62$-$0.76$). A critical dissociation emerges: unsupervised stability fails entirely for steering on real-world tasks ($\rho \approx 0.10$), revealing that task alignment is essential for controllability prediction. However, unsupervised stability excels at drift detection, measuring nearly $2\times$ greater geometric change than CKA during post-training alignment (up to $5.23\times$ in Llama) while providing earlier warning in 73% of models and maintaining a $6\times$ lower false alarm rate than Procrustes. Together, supervised and unsupervised stability form complementary diagnostics for the LLM deployment lifecycle: one for pre-deployment controllability assessment, the other for post-deployment monitoring.

Predictions from machine learning algorithms can vary across random seeds, inducing instability in downstream debiased machine learning estimators. We formalize random seed stability via a concentration condition and prove that subbagging guarantees stability for any bounded-outcome regression algorithm. We introduce a new cross-fitting procedure, adaptive cross-bagging, which simultaneously eliminates seed dependence from both nuisance estimation and sample splitting in debiased machine learning. Numerical experiments confirm that the method achieves the targeted level of stability whereas alternatives do not. Our method incurs a small computational penalty relative to standard practice whereas alternative methods incur large penalties.

This paper studies a structural failure of subsample-based estimation in dynamic time series models. Even under oracle knowledge of contamination locations, removing contaminated observations does not restore the uncontaminated objective. In such settings, contamination propagates through the residual filter and distorts the estimation criterion. As a result, subsample-based estimators are generically inconsistent for the clean-data parameter. We characterise this failure as a structural incompatibility between pointwise subsampling and residual propagation. More generally, the failure arises whenever contamination propagates through transformations that enter the estimation criterion, with dynamic time series models as a leading example. To address it, we propose a propagation-compatible transformation of index sets via a patch removal operator. Under general high-level conditions, this transformation leaves the estimator asymptotically unchanged under the uncontaminated model while restoring consistency under contamination. The results apply to a broad class of residual-based estimators and do not rely on modelling the contamination process.

Random-effects meta-analysis summarizes heterogeneous trials by estimating an average effect over the observed evidence base, which may not represent the clinically relevant target population. In cardiovascular medicine, treatment effects vary systematically across era, endpoint definitions, background therapy, and case-mix, making the historical average often misaligned with current decision-making. We propose stable transport meta-analysis (AMT-MA), a nuisance-anchor estimator that models anchor-aligned variation but does not transport it to the target population. The method combines a weighted-average loss with a scale-normalized softmax regime loss, and incorporates a precision-weighted sign-stability diagnostic with a two-condition abstention rule to avoid reporting a single pooled estimate when stability is not supported. AMT-MA is not intended to minimize RMSE relative to random-effects models, but to redefine the estimand as a stable target-population effect. In a pre-specified ADEMP simulation across six scenarios, AMT-MA (rho = 0.2) showed reduced bias relative to unadjusted pooling and improved coverage in adversarial settings where classical Wald intervals fail (dominant trial: 0.85 vs 0.01; confounded anchor: 0.86 vs 0.34; anchor shift: 0.91 vs 0.60). WLS meta-regression remained competitive when correctly specified. Under sign-flip heterogeneity, the abstention rule triggered in ~84% of replications, compared with ~28-30% in stable regimes. Applications to post-myocardial infarction streptokinase trials and primary-prevention aspirin trials illustrate how AMT-MA quantifies transport uncertainty and provides a clinically interpretable alternative to averaging heterogeneous effects.

We introduce Prior-Fitted Functional Flows, a generative foundation model for pharmacokinetics that enables zero-shot population synthesis and individual forecasting without manual parameter tuning. We learn functional vector fields, explicitly conditioned on the sparse, irregular data of an entire study population. This enables the generation of coherent virtual cohorts as well as forecasting of partially observed patient trajectories with calibrated uncertainty. We construct a new open-access literature corpus to inform our priors, and demonstrate state-of-the-art predictive accuracy on extensive real-world datasets.

We propose post-screening portfolio selection (PS$^2$), a two-step framework for high-dimensional mean–variance investing. First, assets are screened by Lasso-type regression of a constant on excess returns without an intercept. Second, portfolio weights are estimated on the selected set using standard low-dimensional methods. Because strong factors can destroy sparsity in real data, we further introduce PS$^2$ with factors (FPS$^2$), which defactors returns before screening and allows factor investing in the final step. We establish theoretical guarantees, and simulations and an empirical application show competitive performance, especially when sparse screening is appropriate or strong factors are explicitly accommodated.

Continual learning (CL) is concerned with learning multiple tasks sequentially without forgetting previously learned tasks. Despite substantial empirical advances over recent years, the theoretical development of CL remains in its infancy. At the heart of developing CL theory lies the challenge that the data distribution varies across tasks, and we argue that properly addressing this challenge requires understanding this variation–dependency among tasks. To explicitly model task dependency, we consider nonlinear regression tasks and propose the assumption that these tasks are dependent in such a way that the data of the current task is a nonlinear transformation of previous data. With this model and under natural assumptions, we prove statistical recovery guarantees (more specifically, bounds on estimation errors) for several CL paradigms in practical use, including experience replay with data-independent regularization and data-independent weights that balance the losses of tasks, replay with data-dependent weights, and continual learning with data-dependent regularization (e.g., knowledge distillation). To the best of our knowledge, our bounds are informative in cases where prior work gives vacuous bounds.

Given only observational data $X = g(Z)$, where both the latent variables $Z$ and the generating process $g$ are unknown, recovering $Z$ is ill-posed without additional assumptions. Existing methods often assume linearity or rely on auxiliary supervision and functional constraints. However, such assumptions are rarely verifiable in practice, and most theoretical guarantees break down under even mild violations, leaving uncertainty about how to reliably understand the hidden world. To make identifiability actionable in the real-world scenarios, we take a complementary view: in the general settings where full identifiability is unattainable, what can still be recovered with guarantees, and what biases could be universally adopted? We introduce the problem of diverse dictionary learning to formalize this view. Specifically, we show that intersections, complements, and symmetric differences of latent variables linked to arbitrary observations, along with the latent-to-observed dependency structure, are still identifiable up to appropriate indeterminacies even without strong assumptions. These set-theoretic results can be composed using set algebra to construct structured and essential views of the hidden world, such as genus-differentia definitions. When sufficient structural diversity is present, they further imply full identifiability of all latent variables. Notably, all identifiability benefits follow from a simple inductive bias during estimation that can be readily integrated into most models. We validate the theory and demonstrate the benefits of the bias on both synthetic and real-world data.

Persistent homology (PH) encodes global information, such as cycles, and is thus increasingly integrated into graph neural networks (GNNs). PH methods in GNNs typically traverse an increasing sequence of subgraphs. In this work, we first expose limitations of this inclusion procedure. To remedy these shortcomings, we analyze contractions as a principled topological operation, in particular, for graph representation learning. We study the persistence of contraction sequences, which we call Contraction Homology (CH). We establish that forward PH and CH differ in expressivity. We then introduce Hourglass Persistence, a class of topological descriptors that interleave a sequence of inclusions and contractions to boost expressivity, learnability, and stability. We also study related families parametrized by two paradigms. We also discuss how our framework extends to simplicial and cellular networks. We further design efficient algorithms that are pluggable into end-to-end differentiable GNN pipelines, enabling consistent empirical improvements over many PH methods across standard real-world graph datasets. Code is available at \href{https://github.com/Aalto-QuML/Hourglass}{this https URL}.

We study convergence rates of the annealed importance sampling algorithm (Neal '01) combined with Langevin Monte Carlo when the target is a multimodal Gibbs measure. The main result shows that for a fixed error threshold, the time complexity is quadratic in the inverse temperature. We identify a simple and useful quantity that controls the sampling error for AIS in a general setting, and then bound this quantity in our setting using spectral estimates. We also study an autonormalized version and obtain bounds for the time complexity in terms of the inverse temperature.

We introduce joint exclusivity (JE), a relaxation of mutual exclusivity (ME) for non-negative random vectors that forbids only the simultaneous positivity of all $n$ components. Whereas ME requires the probabilities of positive outcomes to sum to at most one, a JE random vector with prescribed marginals $F_1,\ldots,F_n$ exists if and only if \[ \sum_{i=1}^n \overline{F}_i(0)\le n-1. \] An inclusion–exclusion identity shows that JE laws are determined by their $(n-1)$-dimensional margins and decomposes aggregate stop-loss premiums into those of sub-portfolios. A single wrap-around construction attains the existence bound, minimises the number of claims in convex order and, in frequency–severity models, the aggregate loss, thereby extending the safest-dependence property of ME beyond its admissible range. We also develop canonical and distortion-based constructions with freely chosen face copulas, and an $m$-exclusivity hierarchy between ME and JE with sharp two-sided stop-loss bounds.

The low-degree polynomial framework has emerged as a powerful tool for providing evidence of statistical-computational gaps in high-dimensional inference. For detection problems, the standard approach bounds the low-degree advantage through an explicit orthonormal basis. However, this method does not extend naturally to estimation tasks, and thus fails to capture the detection-recovery gap phenomenon that arises in many high-dimensional problems. Although several important advances have been made to overcome this limitation \cite{SW22, SW25, CGGV25+}, the existing approaches often rely on delicate, model-specific combinatorial arguments. In this work, we develop a general approach for obtaining conditional computational lower bounds for recovery problems from mild bounds on low-degree testing advantage. Our method combines the notion of algorithmic contiguity in \cite{Li25} with a cross-validation reduction in \cite{DHSS25} that converts successful recovery into a hypothesis test with lopsided success probabilities. In contrast to prior unconditional lower bounds, our argument is conceptually simple, flexible, and largely model-independent. We apply this framework to several canonical inference problems, including planted submatrix, planted dense subgraph, stochastic block model, multi-frequency angular synchronization, orthogonal group synchronization, and multi-layer stochastic block model. In the first three settings, our method recovers existing low-degree lower bounds for recovery in \cite{SW22, SW25} via a substantially simpler argument. In the latter three, it gives new evidence for conjectured computational thresholds including the persistence of detection-recovery gaps. Together, these results suggest that mild control of low-degree advantage is often sufficient to explain computational barriers for recovery in high-dimensional statistical models.

The Mapper algorithm from topological data analysis constructs a graph summarizing the shape of a high-dimensional dataset, and groups of data points identified within this graph are widely interpreted as evidence of distinct subtypes. However, the covariance structure of the data alone can make such groups appear differentiated, even when no subtypes are present. Existing validation approaches do not account for this effect and thus cannot distinguish covariance artifacts from genuine subtypes. We propose a Gaussian null model that generates reference data matching the sample covariance matrix. We pair it with a test statistic that measures mean-level differentiation between communities. In an idealized setting, we prove that covariance geometry alone causes Mapper communities to differ in their average feature profiles, and we show that a simpler label-permutation baseline cannot detect this effect. Simulations confirm well-controlled Type I error under Gaussian data. We apply the framework to four published Mapper analyses spanning breast cancer gene expression, Congressional voting, NBA player performance, and lower-grade glioma genomics. In every case, once outlier singleton communities are accounted for, the observed differentiation does not exceed what the null produces at the {\alpha} = 0.05 level. This result does not rule out subtypes in these datasets, but it does indicate that the observed structure is consistent with what covariance geometry alone can produce. Stronger evidence would be needed to support a subtype claim.

This paper proposes StrEBM, a structured latent energy-based model for source-wise structured representation learning. The framework is motivated by a broader goal of promoting identifiable and decoupled latent organization by assigning different latent dimensions their own learnable structural biases, rather than constraining the entire latent representation with a single shared energy. In this sense, blind source separation is adopted here as a concrete and verifiable testbed, through which the evolution of latent dimensions toward distinct underlying components can be directly examined. In the proposed framework, latent trajectories are optimized directly together with an observation-generation map and source-wise structural parameters. Each latent dimension is associated with its own energy-based formulation, allowing different latent components to gradually evolve toward distinct source-like roles during training. In the present study, this source-wise energy design is instantiated using Gaussian-process-inspired energies with learnable length-scales, but the framework itself is not restricted to Gaussian processes and is intended as a more general structured latent EBM formulation. Experiments on synthetic multichannel signals under linear and nonlinear mixing settings show that the proposed model can recover source components effectively, providing an initial empirical validation of the framework. At the same time, the study reveals important optimization characteristics, including slow late-stage convergence and reduced stability under nonlinear observation mappings. These findings not only clarify the practical behavior of the current GP-based instantiation, but also establish a basis for future investigation of richer source-wise energy families and more robust nonlinear optimization strategies.

Large Language Models can generate synthetic survey responses at low cost, but their accuracy varies unpredictably across questions. We study the design problem of allocating a fixed budget of human respondents across estimation tasks when cheap LLM predictions are available for every task. Our framework combines three components. First, building on Prediction-Powered Inference, we characterize a question-specific rectification difficulty that governs how quickly the estimator's variance decreases with human sample size. Second, we derive a closed-form optimal allocation rule that directs more human labels to tasks where the LLM is least reliable. Third, since rectification difficulty depends on unobserved human responses for new surveys, we propose a meta-learning approach, trained on historical data, that predicts it for entirely new tasks without pilot data. The framework extends to general M-estimation, covering regression coefficients and multinomial logit partworths for conjoint analysis. We validate the framework on two datasets spanning different domains, question types, and LLMs, showing that our approach captures 61-79% of the theoretically attainable efficiency gains, achieving 11.4% and 10.5% MSE reductions without requiring any pilot human data for the target survey.

Breast cancer is the most prevalent cancer in women worldwide. Histopathology image analysis serves as the gold standard for cancer diagnosis. In this regard, whole-slide imaging (WSI), a revolutionary technology in digital pathology, allows for ultrahigh-resolution tissue analysis. Despite its promise, WSI analysis faces significant computational challenges due to its massive data size and tissue heterogeneity. To address this issue, we present a Gaussian mixture based multiple instance learning (MIL) framework for WSI analysis with partially subsampled instances. Our approach models a WSI as a bag of instances (i.e., randomly cropped sub-images), leveraging a bag-based maximum likelihood estimator (BMLE) to predict metastases. Furthermore, we introduce a subsampling-based maximum likelihood estimator (SMLE) to refine predictions by selectively labeling a subset of instances. Extensive evaluations of the breast carcinoma metastasis prediction demonstrate that BMLE surpasses state-of-the-art methods, while the SMLE further improves the prediction accuracy at both bag and instance levels. We find that our method is fairly robust against various plausible model mis-specifications. Theoretical analyses and simulation studies validate the performance and robustness of our methods.

Data scarcity challenges the development and implementation of innovative healthcare solutions. In geriatrics, fall-related injuries are a major cause of hospitalization, functional decline, and mortality in older adults. Optimizing post-operative discharge planning can mitigate these outcomes, but limited data hinders predictive model development. Here, we explored generative machine learning approaches to augment data from the SURGE-Ahead project (Supporting SURgery with Geriatric Co-Management and AI), an initiative addressing geriatric perioperative care. Data from the German geriatric trauma register (AltersTraumaZentrum; ATZ) were incorporated using two strategies: (i) combining SURGE-Ahead and ATZ register data with imputation (ComImp) and (ii) generating synthetic data from SURGE-Ahead alone or combined SURGE-Ahead and the ATZ register datasets with Adversarial random forests (ARF). Predictive models, including multinomial logistic regression, random forest, and a prior-fitted transformer (TabPFN), were trained and evaluated using standard performance metrics: accuracy, area under the receiver operating characteristic curve (ROC AUC), Brier score, and the logistic loss. Random forest and TabPFN performed well (accuracy around 0.84 and AUC around 0.94) and were largely unaffected by augmentation. Logistic regression benefited from augmented data, with predictive performance improving from 0.70 to 0.81 for accuracy and 0.85 to 0.92 for AUC. These results highlight generative data augmentation as a viable approach to enhance simpler predictive models in geriatric care and emphasize the importance of method selection when addressing data scarcity in heterogeneous clinical populations.

Double/debiased machine learning (DML) provides a general framework for inference with high-dimensional or otherwise complex nuisance parameters by combining Neyman-orthogonal scores with cross-fitting, thereby circumventing classical Donsker-type conditions in many modern machine-learning settings. Despite its strong empirical performance, bootstrap inference for DML estimators has received little theoretical justification. This is particularly noteworthy since bootstrap methods are suggested ad used for inference on DML estimators, even though bootstrap procedures can fail for estimators that are root-$n$ consistent and asymptotically normal. This paper fills this gap by establishing bootstrap validity for DML estimators under general exchangeably weighted resampling schemes, with Efron's bootstrap as a special case. Under exactly the same conditions required for the validity of DML itself, we prove that the bootstrap law converges conditionally weakly to the sampling law of the original estimator.

In this paper, we develop a finite mixture of convolutional distributions, a statistical model to analyze continuous data distributed approximately on a mixture of low-dimensional affine subspaces. The observations are assumed independent and identically distributed from the mixture of distributions, where each component arises from a convolution of a distribution supported on a low-dimensional subspace with a suitable noise kernel. We discuss theoretical properties of such class of models, including identifiability under very general conditions - in particular, showing that the minimal representation for such mixtures is uniquely identifiable in a semi-parametric setting. We further study the posterior contraction rates for the parameters for a parametrized class of such models where the supports of the component mixing measures are assumed to be convex polytopes under a suitable well-specified Bayesian regime. This still requires developing novel inverse bounds for problems involving a nested mixture structure, where the mixture kernel is itself another continuous mixture. Our approach for both the identifiability theory and posterior contraction rates is to exploit the geometric structure of the underlying support of the latent measures. Apart from applications in end-member analysis, spectral unmixing and topic models, this study provides a grounded framework for subspace clustering with the goal of exploring conditions for learning multiple latent low-dimensional structures. We illustrate our findings through careful simulation study, which also includes developing new algorithms for such class of models

Recursive architectures such as Tiny Recursive Models (TRMs) perform implicit reasoning through iterative latent computation, yet the geometric structure of these reasoning trajectories remains poorly understood. We investigate the activation manifold of TRMs during recursive unrolling and find that activations occupy an effectively linear, low-dimensional subspace whose principal directions can be tracked dynamically with cheap power iterations. This suggests that weight-sharing concentrates iterative computation along a small number of dominant eigendirections, and we find that this concentration varies sharply across computational sites. We exploit this structure through LASER (Low-Rank Activation SVD for Efficient Recursion), a dynamic compression framework that maintains an evolving low-rank basis via matrix-free subspace tracking with a fidelity-triggered reset mechanism, achieving ${\sim}60\%$ activation memory savings with no statistically significant accuracy degradation. Our analysis raises questions about how recursive architectures allocate representational capacity during implicit reasoning, and whether this concentration can be exploited to improve the efficiency and stability of latent computation.