Searcharxiv⌕ Search

arXiv subjects

Wei-Chun Chen

Publications and source records attributed to Wei-Chun Chen.

11 recordsLinked to original sources

Local regularity and p-harmonic approximation for weak inverse mean curvature flow

We prove local Lipschitz regularity for locally bounded $BV$ solutions of the calibration equations for inverse mean curvature flow. We also approximate every prescribed variational weak flow locally by Moser's logarithmic transforms of positive $p$-harmonic functions. On a fixed smooth domain, with the trace of a calibrated Lipschitz solution prescribed, both the uniform error and the absolute difference of total variations are $O((p-1)^{1/3})$ as $p\downarrow1$. In particular, the approximations converge strictly in $BV$.

math.AP↗

The Jacobi Spectrum and Global Second Variation of Pansu Spheres

For $n\ge2$, we establish second differentiability of horizontal perimeter for compact $C^2$ boundaries in $\mathbb H^n$ and the continuous extension of their affine second variation to $C^1$ vector fields along the boundary. For the unit Pansu sphere, the Hessian agrees with the closed Jacobi form across both characteristic poles, including variations which move them. We compute the full Friedrichs spectrum of its Jacobi operator. The operator has compact resolvent, Morse index one and nullity $2n+1$; its kernel consists exactly of normalized translation speeds. Consequently, the sphere is stable under volume-preserving ambient variations.

math.DG↗

Stable Rigidity in the Heisenberg Groups

For $n\ge2$, we prove that the $C^2$ boundary of a bounded nonempty open set in $\mathbb H^n$ is volume-preserving stable if and only if it is a Pansu sphere up to left translation and Heisenberg dilation. The proof combines centre reconstruction with a volume-constrained test field derived from Pansu dilation. Consequently, every isoperimetric region admitting an open representative with $C^2$ boundary is a Pansu ball up to these symmetries.

math.DG↗

Quantitative Centre Rigidity under Riemannian Approximation in Heisenberg Groups

For $n\ge2$, we prove a quantitative centre-rigidity estimate for bounded sets of finite Riemannian perimeter that are stationary and volume-constrained stable for normalized Riemannian perimeter with prescribed curvature $h_j$. Under confinement to a fixed compact set and cutoffs of vanishing full surface Dirichlet energy on the smooth regular boundary, the squared $L^2$ norms of the horizontal centre derivative, angle defect and weighted shape defect are bounded by $C(\varepsilon_j^2+\|h_j-2n\|_{C^2})|Ω_j|$ when the parenthesized quantity is sufficiently small. The excess of normalized Riemannian perimeter over horizontal perimeter satisfies the same bound.

math.DG↗

The Exact Comass Criterion for the Tsai--Wang Form

For a map $F:Ω\subset\mathbb{R}^n\to\mathbb{R}^m$, let $Θ(F)$ be the $n$-form on $Ω\times\mathbb{R}^m$ introduced by Tsai and Wang [arXiv:2604.04336]. If $λ_1,\ldots,λ_{r}$ are the nonzero singular values of $d F$ at any fixed point, we prove that $$ Θ(F) \text{ has comass one if and only if }~ \mathcal{S}(λ):=\sum_{i=1}^{r}\frac{λ_i^2}{1+λ_i^2}\leq1. $$ Note that if $\text{rank} d F\le1$, $\mathcal{S}$ is always less than $1$. We also prove a dichotomy that if $\mathcal{S}\leq 1$, then either $\mathcal{S} < 1$ or $\mathcal{S}\equiv1$. When $\mathcal{S}<1$, the graph tangent plane is the only calibrated plane, generalizing the corresponding result for hypersurfaces. When $\mathcal{S}\equiv1$, we prove that $F$ is affine or of rank $2$.

math.DG↗

CGU-ILALab at FoodBench-QA 2026: Comparing Traditional and LLM-based Approaches for Recipe Nutrient Estimation

Accurate nutrient estimation from unstructured recipe text is an important yet challenging problem in dietary monitoring, due to ambiguous ingredient terminology and highly variable quantity expressions. We systematically evaluate models spanning a wide range of representational capacity, from lexical matching methods (TF-IDF with Ridge Regression), to deep semantic encoders (DeBERTa-v3), to generative reasoning with large language models (LLMs). Under the strict tolerance criteria defined by EU Regulation 1169/2011, our empirical results reveal a clear trade-off between predictive accuracy and computational efficiency. The TF-IDF baseline achieves moderate nutrient estimation performance with near-instantaneous inference, whereas the DeBERTa-v3 encoder performs poorly under task-specific data scarcity. In contrast, few-shot LLM inference (e.g., Gemini 2.5 Flash) and a hybrid LLM refinement pipeline (TF-IDF combined with Gemini 2.5 Flash) deliver the highest validation accuracy across all nutrient categories. These improvements likely arise from the ability of LLMs to leverage pre-trained world knowledge to resolve ambiguous terminology and normalize non-standard units, which remain difficult for purely lexical approaches. However, these gains come at the cost of substantially higher inference latency, highlighting a practical deployment trade-off between real-time efficiency and nutritional precision in dietary monitoring systems.

cs.CL↗

Bounds on two-distance sets in Euclidean space and Unit Sphere

We establish upper bounds for the size of two-distance sets in Euclidean space and spherical two-distance sets. The main recipe for obtaining upper bounds is the spectral method. We construct Seidel matrices to encode the distance relations and apply eigenvalue analysis to obtain explicit bounds. For Euclidean space, we have the upper bounds for the cardinality $n$ of a two-distance set. \[ n \le \dfrac{(d+1)\left(\left(\frac{1+δ^2}{1-δ^2}\right)^2 - 1\right)}{\left(\frac{1+δ^2}{1-δ^2}\right)^2-(d+1)}+1. \] if the two distances are $1$ and $δ$ in $\mathbb{R}^d$. For spherical two-distance sets with $n$ points and inner products $a, b$ on $\mathbb{S}^{d-1}$, we will have the following: \[ \begin{cases} n \le \dfrac{d\left(\left(\dfrac{a+b-2}{b-a}\right)^2-1\right)}{\left(\dfrac{a+b-2}{b-a}\right)^2-d}, &a+b \ge 0; n \le \dfrac{(d+1)\left(\left(\dfrac{a+b-2}{b-a}\right)^2-1\right)}{\left(\dfrac{a+b-2}{b-a}\right)^2-(d+1)}, &a+b < 0. \end{cases} \] Notice that the second bound (for $a+b < 0$) is the same as the relative bound for the equiangular lines in one higher dimension.

math.CO↗

LLM Pruning and Distillation in Practice: The Minitron Approach

We present a comprehensive report on compressing the Llama 3.1 8B and Mistral NeMo 12B models to 4B and 8B parameters, respectively, using pruning and distillation. We explore two distinct pruning strategies: (1) depth pruning and (2) joint hidden/attention/MLP (width) pruning, and evaluate the results on common benchmarks from the LM Evaluation Harness. The models are then aligned with NeMo Aligner and tested in instruct-tuned versions. This approach produces a compelling 4B model from Llama 3.1 8B and a state-of-the-art Mistral-NeMo-Minitron-8B (MN-Minitron-8B for brevity) model from Mistral NeMo 12B. We found that with no access to the original data, it is beneficial to slightly fine-tune teacher models on the distillation dataset. We open-source our base model weights on Hugging Face with a permissive license.

cs.CL↗

Knowledge Distillation with Feature Maps for Image Classification

The model reduction problem that eases the computation costs and latency of complex deep learning architectures has received an increasing number of investigations owing to its importance in model deployment. One promising method is knowledge distillation (KD), which creates a fast-to-execute student model to mimic a large teacher network. In this paper, we propose a method, called KDFM (Knowledge Distillation with Feature Maps), which improves the effectiveness of KD by learning the feature maps from the teacher network. Two major techniques used in KDFM are shared classifier and generative adversarial network. Experimental results show that KDFM can use a four layers CNN to mimic DenseNet-40 and use MobileNet to mimic DenseNet-100. Both student networks have less than 1\% accuracy loss comparing to their teacher models for CIFAR-100 datasets. The student networks are 2-6 times faster than their teacher models for inference, and the model size of MobileNet is less than half of DenseNet-100's.

cs.LG↗