arXiv subjects
Publications and source records attributed to Hengzhi He.
We prove that the Esseen's constant $0.4097...$ is sharp in the iid Berry--Esseen inequality as long as the sample sizes $n\ge N$. The threshold $N$ is universal over all real summand distributions with positive variance and a finite third absolute moment. In particular, one may take $N=2\left\lceil\exp(10^{17})\right\rceil. $
Let $X_1,\ldots,X_n$ be independent Gaussian tensors in $\mathbb{R}^{d_1}\otimes\cdots\otimes\mathbb{R}^{d_k}$ with a common covariance matrix given by the Kronecker product of $k$ unknown positive-definite factors, and let $D=\prod_{a=1}^k d_a$ and $d_{\max}=\max_a d_a$. Franks et al. (2026) established condition-number-free guarantees for the tensor-normal maximum likelihood estimator under the sample-size condition $nD\gtrsim k^2 d_{\max}^3$ and asked whether the cubic dependence on $d_{\max}$ could be reduced to a quadratic one. We answer this question affirmatively. For $t\geq 1$, if $nD\geq C k^2 d_{\max}^2 t^2$, then with high probability the maximum likelihood estimator exists, is unique, and satisfies $d_{\rm FR}(\widehatΘ,Θ)\leq C t \sqrt{k} d_{\max}/\sqrt{n}$ and $d_{\rm FR}(\widehatΘ_a,Θ_a)\leq C t\sqrt{k d_a} d_{\max}/\sqrt{nD}$ for every mode $a$. For every mode $a$ with $d_a=d_{\max}$, we further establish the sharp Thompson-metric bound $d_{\rm op}(\widehatΘ_a,Θ_a)\leq C t d_{\max}/\sqrt{nD}$. These guarantees are uniform over the unknown covariance factors and require neither condition-number bounds nor sparsity assumptions. Gaussian submodel lower bounds match the full and largest-factor Fisher--Rao rates up to a factor of $\sqrt{k}$ and the largest-factor Thompson rate up to universal constants. Consequently, for fixed $k$, the quadratic dependence of the sample-size threshold on $d_{\max}$ is optimal. GPT-5.6 Sol and Claude Fable 5 were used to assist with proof development, verification, and manuscript preparation.
We consider a mixture of at most $k$ unit-covariance Gaussians in $\mathbb{R}^d$ whose means belong to a fixed-radius ball, with no separation or minimum-weight condition. Doss, Wu, Yang and Zhou (2023) proved that the minimax Hellinger risk is of order $\sqrt{d/n}\wedge 1$ and constructed a proper polynomial-time estimator with the slower general bound $(d/n)^{1/4}$; obtaining the sharp rate in polynomial time for fixed $k\geq 3$ was left open. We resolve this question. The key device is a moment-fiber range finder. A second-moment subspace controls the energy missed by projection. We then estimate finitely many one-free-index Hermite contractions. These vector-valued contractions recover every tensor component containing exactly one missed direction at the sharp $\sqrt{d/n}$ scale. Every remaining term contains at least two missed factors and is therefore controlled by the residual second-moment energy. The resulting subspace has dimension depending only on $k$. Exhaustive moment fitting in this constant-dimensional space produces a proper mixture and, together with the dimension-free moment characterization of Gaussian mixtures, achieves the optimal Hellinger rate in polynomial arithmetic time for every fixed $k$.
Watermarking has recently emerged as a crucial tool for protecting the intellectual property of generative models and for distinguishing AI-generated content from human-generated data. Despite its practical success, most existing watermarking schemes are empirically driven and lack a theoretical understanding of the fundamental trade-off between detection power and generation fidelity. To address this gap, we formulate watermarking as a statistical hypothesis testing problem between a null distribution and its watermarked counterpart. Under explicit constraints on false-positive and false-negative rates, we derive a tight lower bound on the achievable fidelity loss, measured by a general $f$-divergence, and characterize the optimal watermarked distribution that attains this bound. We further develop a corresponding sampling rule that provides an optimal mechanism for inserting watermarks with minimal fidelity distortion. Our result establishes a simple yet broadly applicable principle linking hypothesis testing, information divergence, and watermark generation.
Recent studies identified an intriguing phenomenon in recursive generative model training known as model collapse, where models trained on data generated by previous models exhibit severe performance degradation. Addressing this issue and developing more effective training strategies have become central challenges in generative model research. In this paper, we investigate this phenomenon within a novel framework, where generative models are iteratively trained on a combination of newly collected real data and synthetic data from the previous training step. To develop an optimal training strategy for integrating real and synthetic data, we evaluate the performance of a weighted training scheme in various scenarios, including Gaussian distribution estimation, generalized linear models, and nonparametric estimation. We theoretically characterize the impact of the mixing proportion and weighting scheme of synthetic data on the final model's performance. Our key finding is that, across different settings, the optimal weighting scheme under different proportions of synthetic data asymptotically follows a unified expression, revealing a fundamental trade-off between leveraging synthetic data and model performance. In some cases, the optimal weight assigned to real data corresponds to the reciprocal of the golden ratio. Finally, we validate our theoretical results on extensive simulated datasets and a real tabular dataset.
Large Language Models (LLMs) struggle with multi-step reasoning over structured tables. The primary reason is the lack of explicit supervision for intermediate reasoning states. Existing learned reward models or executor-based verifiers are either unscalable or rely on answer-checking environments unavailable for many tabular tasks. This leaves no signal that is scalable and grounded in the query. To address this, we introduce TABROUGE, a training-free and deterministic state reward. By adapting the Longest Common Subsequence (LCS) metric from text summarization to evaluate tabular states, TABROUGE assesses the lexical coverage and structural integrity of intermediate tables against the query without requiring learned models or external executors. Built upon this metric, we propose RE-TAB, a plug-and-play, training-free framework. RE-TAB reframes table reasoning as deterministic control over intermediate states, utilizing TABROUGE for stepwise feedback and trajectory-level test-time scaling (TTS) signals. Across six backbones and three benchmarks, RE-TAB improves accuracy by an average of 26.7 pp over no-reward baselines. It also reduces TTS samples by up to 33%. Preliminary GRPO experiments further indicate TABROUGE's viability as a scalable post-training reward, increasing gains by 8.34 pp. We further analyze failure modes of TABROUGE, including paraphrase under-rewarding and echo-column hacking, and identify when structure-aware lexical rewards remain reliable.
Targeted data selection aims to identify training samples from a large candidate pool that improve performance on a specific downstream task. Many recent methods estimate candidate utility by aggregating local attribution scores along a trajectory induced by the candidate pool. When the pool is heterogeneous, however, this reference trajectory may be misaligned with the dynamics of a target-aligned selected subset, creating what we call reference path bias. We propose an alternative reference path: a validation-induced flow obtained from a short, capacity-limited warmup on the available target validation proxy. Along this path, candidates are scored by a normalized endpoint loss drop, yielding a simple zero-order selection rule that requires no candidate gradients or Hessian approximations. Across controlled logistic, vision, and instruction-tuning experiments, this score is competitive with strong dynamic attribution baselines while substantially reducing warmup and storage cost. Moreover, since the reference trajectory is decoupled from any specific candidate pool, the same compact warmup can be reused across additional pools without recomputing the trajectory.
Noise Contrastive Estimation (NCE) has fueled major breakthroughs in representation learning and generative modeling. Yet a long-standing challenge remains: accurately estimating ratios between distributions that differ substantially, which significantly limits the applicability of NCE on modern high-dimensional and multimodal datasets. We revisit this problem from a less explored perspective: the magnitude of the noise distribution. Specifically, we show that with a virtually scaled (\ie, artificially increased) noise magnitude, the gradient of the NCE objective can closely align with that of Maximum Likelihood, enabling a trajectory-wise approximation from NCE to MLE, and faster convergence both theoretically and empirically. Building on this insight, we introduce ``Noisier'' NCE, a simple drop-in modification to vanilla NCE that incurs little to no extra computational cost, while effectively handling density-ratio estimation in challenging regimes where traditional MLE and NCE struggle. Beyond improving classical density-ratio learning, ``Noisier'' NCE proves broadly applicable: it achieves strong results across image modeling, anomaly detection, and offline black-box optimization. On CIFAR-10 and ImageNet64x64 datasets, it yields 10-step and even 1-step samplers that match or surpass state-of-the-art methods, while cutting training iterations by up to half.
Cryptographic provenance standards such as C2PA and invisible watermarking are positioned as complementary defenses for content authentication, yet the two verification layers are technically independent: neither conditions on the output of the other. This work formalizes and empirically demonstrates the $\textit{Integrity Clash}$, a condition in which a digital asset carries a cryptographically valid C2PA manifest asserting human authorship while its pixels simultaneously carry a watermark identifying it as AI-generated, with both signals passing their respective verification checks in isolation. We construct metadata washing workflows that produce these authenticated fakes through standard editing pipelines, requiring no cryptographic compromise, only the semantic omission of a single assertion field permitted by the current C2PA specification. To close this gap, we propose a cross-layer audit protocol that jointly evaluates provenance metadata and watermark detection status, achieving 100% classification accuracy across 3,500 test images spanning four conflict-matrix states and three realistic perturbation conditions. Our results demonstrate that the gap between these verification layers is unnecessary and technically straightforward to close.
In recent years, LLM watermarking has emerged as an attractive safeguard against AI-generated content, with promising applications in many real-world domains. However, there are growing concerns that the current LLM watermarking schemes are vulnerable to expert adversaries wishing to reverse-engineer the watermarking mechanisms. Prior work in breaking or stealing LLM watermarks mainly focuses on the distribution-modifying algorithm of Kirchenbauer et al. (2023), which perturbs the logit vector before sampling. In this work, we focus on reverse-engineering the other prominent LLM watermarking scheme, distortion-free watermarking (Kuditipudi et al. 2024), which preserves the underlying token distribution by using a hidden watermarking key sequence. We demonstrate that, even under a more sophisticated watermarking scheme, it is possible to compromise the LLM and carry out a spoofing attack, i.e. generate a large number of (potentially harmful) texts that can be attributed to the original watermarked LLM. Specifically, we propose using adaptive prompting and a sorting-based algorithm to accurately recover the underlying secret key for watermarking the LLM. Our empirical findings on LLAMA-3.1-8B-Instruct, Mistral-7B-Instruct, Gemma-7b, and OPT-125M challenge the current theoretical claims on the robustness and usability of the distortion-free watermarking techniques.
In recent years, model collapse has become a critical issue in language model training, making it essential to understand the underlying mechanisms driving this phenomenon. In this paper, we investigate recursive parametric model training from a probabilistic perspective, aiming to characterize the conditions under which model collapse occurs and, crucially, how it can be mitigated. We conceptualize the recursive training process as a random walk of the model estimate, highlighting how the sample size influences the step size and how the estimation procedure determines the direction and potential bias of the random walk. Under mild conditions, we rigorously show that progressively increasing the sample size at each training step is necessary to prevent model collapse. In particular, when the estimation is unbiased, the required growth rate follows a superlinear pattern. This rate needs to be accelerated even further in the presence of substantial estimation bias. Building on this probabilistic framework, we also investigate the probability that recursive training on synthetic data yields models that outperform those trained solely on real data. Moreover, we extend these results to general parametric model family in an asymptotic regime. Finally, we validate our theoretical results through extensive simulations and a real-world dataset.
In this paper, we propose a novel statistical framework for watermarking generative categorical data. Our method systematically embeds pre-agreed secret signals by splitting the data distribution into two components and modifying one distribution based on a deterministic relationship with the other, ensuring the watermark is embedded at the distribution-level. To verify the watermark, we introduce an insertion inverse algorithm and detect its presence by measuring the total variation distance between the inverse-decoded data and the original distribution. Unlike previous categorical watermarking methods, which primarily focus on embedding watermarks into a given dataset, our approach operates at the distribution-level, allowing for verification from a statistical distributional perspective. This makes it particularly well-suited for the modern paradigm of synthetic data generation, where the underlying data distribution, rather than specific data points, is of primary importance. The effectiveness of our method is demonstrated through both theoretical analysis and empirical validation.
In this paper, we introduce a simple yet effective tabular data watermarking mechanism with statistical guarantees. We show theoretically that the proposed watermark can be effectively detected, while faithfully preserving the data fidelity, and also demonstrates appealing robustness against additive noise attack. The general idea is to achieve the watermarking through a strategic embedding based on simple data binning. Specifically, it divides the feature's value range into finely segmented intervals and embeds watermarks into selected ``green list" intervals. To detect the watermarks, we develop a principled statistical hypothesis-testing framework with minimal assumptions: it remains valid as long as the underlying data distribution has a continuous density function. The watermarking efficacy is demonstrated through rigorous theoretical analysis and empirical validation, highlighting its utility in enhancing the security of synthetic and real-world datasets.
We study the problem of dynamic matching in heterogeneous networks, where agents are subject to compatibility restrictions and stochastic arrival and departure times. In particular, we consider networks with one type of easy-to-match agents and multiple types of hard-to-match agents, each subject to its own compatibility constraints. Such a setting arises in many real-world applications, including kidney exchange programs and carpooling platforms. We introduce a novel approach to modeling dynamic matching by establishing the ordinary differential equation (ODE) model, which offers a new perspective for evaluating various matching algorithms. We study two algorithms, namely the Greedy and Patient Algorithms, where both algorithms prioritize matching compatible hard-to-match agents over easy-to-match agents in heterogeneous networks. Our results demonstrate the trade-off between the conflicting goals of matching agents quickly and optimally, offering insights into the design of real-world dynamic matching systems. We provide simulations and a real-world case study using data from the Organ Procurement and Transplantation Network to validate theoretical predictions.