From June 12 to 14, 2026, the 2026 Interdisciplinary Conference on Statistics and Optimization, organized by the School of Mathematical Sciences at NanjingNormal University, was successfully held. The conference focused on cutting-edge interdisciplinary directions including statistics, optimization theory, and financial technology. Experts and scholars from Renmin University of China, Beijing Jiaotong University, the University of California, Riverside, and Nanjing Audit University were invited to deliver keynote presentations and exchange the latest researchfindings.On the morning of June 12, Professor Zhu Liping from Renmin University of China opened the conference with a talk titled “Nonsmoothness in Support Vector Machines and Overfitting in Deep Learning”. Centered on regression andclassification problems in supervised learning, the report systematically reviewed the development of methods such as linear discriminant analysis, probit regression, logistic regression, support vector machines, and deep learning. Professor Zhu focused on the computational challenges caused by nonsmooth optimization in support vector machines and explained the role of convolution smoothing methods in improving computational efficiency and adapting to big-data environments.Meanwhile, the report analyzed the strong predictive power of deep learning in processing high-dimensional complex data, as well as its potential risk of overfitting.It emphasized the importance of evaluating model generalization ability throughappropriate assessment frameworks. This was followed by Professor Kong Lingchen from Beijing Jiaotong University, who presented "Fast Algorithms for Sparse
Decentralized Federated Learning." The report introduced the background of federated learning’s transition from centralized to decentralized paradigms and highlighted the significance of decentralized federated learning in scenarios where data are distributed, privacy protection is increasingly important, and communication resources are limited. Professor Kong analyzed the major challenges faced by sparse decentralized federated learning in terms of communication cost, privacy protection, and theoretical convergence, and introduced two fast algorithms, CEPS and PaME. These methods improve communication efficiency through compressed transmission, partial message exchange, and adaptive aggregation, while also taking privacy protection and convergence guarantees into account. The report further demonstrated the effectiveness of the proposed algorithms through theoretical analysis andnumerical experiments, offering new ideas for the design of large-scale distributed machine learning methods.On the afternoon of June 12, Professor Yao Weixin from the University of California, Riverside, delivered a talk titled "Online Kernel-Based Mode Learning."Professor Yao addressed Modern big data streams routinely present dual challenges:massive volume that overwhelms single-machine memory, and severe data contamination that undermines traditional mean-based methods like Ordinary Least Squares (OLS). To address these issues simultaneously, they propose a novel online kernel-based mode learning framework designed for robust and memory-efficient parameter estimation over continuous data streams. Processing data sequentially in blocks, the approach leverages a specialized Modal Expectation-Maximization (MEM) algorithm to localize conditional density peaks within each chunk. By extracting and aggregating local empirical Hessians as directional precision weights, the global parameter vector is updated recursively through a single-step algebraic
combination, enabling the immediate deletion of historical raw records. Crucially, they establish the asymptotic normality of the resulting online estimator and prove its statistical equivalence to the global batch estimator, confirming that computationalacceleration introduces no efficiency penalty. Extensive simulation studies and an empirical application involving over 7.2 million American Airline flight recordsdemonstrate that our method achieves superior resistance to asymmetriccontamination, matches the predictive accuracy of batch workflows, and delivers substantial speedups over standard benchmarks. Following this, Professor ZhouXingcai from Nanjing Audit University presented "FedFask: Fast Sketching Distributed PCA for Large-Scale Federated Data." His team pointed out that Principal component analysis (PCA) is a fundamental tool for linear dimensionalityreduction, but its application to large-scale federated data faces significant challengesin communication cost, computational complexity, and statistical efficiency. They propose FedFask, a fast sketching distributed PCA algorithm that integrates multiple random sketches, orthogonal Procrustes alignment, and a Kolmogorov–Nagumo-type average on the Stiefel manifold to optimally integrate local eigenspaces. The methoddynamically determines the number of sketches via an automatic stopping criterion. Theoretical analysis shows that FedFask achieves the same statistical error rate as centralized PCA under sub-Gaussian assumptions, while substantially reducing communication and computation overhead. Extensive numerical experiments on synthetic data and the 1000 Genomes dataset validate the efficiency and accuracy of the proposed method.The third afternoon presentation, "Single-index Measurement Error Jump Regression Model in Alzheimer's Disease Studies," was given by Professor ZhaoYanyong from Nanjing Audit University. His team proposed a single-index measurement error jump regression model (SMEJRM) for analyzing the relationship
between neurocognitive scores and various predictors in Alzheimer's disease research. The model integrates single-index structure, measurement error correction, and jump discontinuities in the unknown link function. The study employed a two-stageapproach: Stage 1 used the SIMEX method to correct measurement errors, and Stage 2 applied clustering-based jump point estimation. Theoretical analysis establishedasymptotic properties and confidence regions. Simulation studies and real dataanalysis from ADNI demonstrated that the method outperforms existing approaches,identifies clinically meaningful subgroups, and provides new insights into AD progression risk factors. The fourth presentation, "Tail Risk in Dependent Default Systems with Systematic Risk and Sector-Specific Common Shocks," was given by Professor Yang Yang from Nanjing Audit University. They investigated tail risk in two interdependent default systems driven jointly by systematic risk and sector-specific common shocks. The model considered multi-level influences—commonshocks, systematic risk, and idiosyncratic risk—under low default probability scenarios. Using multivariate regular variation and extreme value theory, they derivedasymptotic expressions for marginal exceedance probabilities of sector default losses and conditional cross-sector exceedance probabilities, further analyzing the asymptotic behavior of VaR, ES, and conditional expected loss measures. Numerical simulations validated the theoretical results and revealed the significant impact of different risk-factor dominance patterns on tail risk.With rich content and clearthemes, the conference not only addressed data-driven problems in real-world economic and financial markets, but also explored fundamental theories andmethodological innovations in statistical learning, optimization algorithms, andartificial intelligence models. Participants engaged in active discussions on the relatedresearch topics, creating a lively academic atmosphere. The successful organization of this conference further promoted the interdisciplinary integration of statistics, optimization, machine learning, and financial technology, while also providing a valuable platform for academic collaboration and talent development in related fields.
On the morning of June 13, participants engaged in discussions on the application of statistical and optimization methods in statistical modeling competitions and market research competitions. Teams led by Professor Lin Jinguanand Associate Professor Huang Xingfang from Nanjing Audit University systematically presented a comprehensive technical framework for statistical modeling, covering core aspects including topic selection strategies, data cleaningand transformation, variable selection, parameter estimation, model evaluation, anddecision-making applications. The presentation closely aligned with the 2026 theme of "Serving National Strategy, Empowering Innovation through Statistics," using the modern industrial system and the real economy as typical case studies to elaborate on the application pathways of statistical modeling in major national strategic needs. A dedicated session on "Statistical Modeling Research on Industrial Chain Resilience" constructed a four-dimensional evaluation index system encompassing supply, demand, technology, and policy environment, comprehensively applyingcomplex network analysis and econometric methods, and proposing two-way policy recommendations at both enterprise and government levels. This report provided a systematic theoretical framework and practical guide for applying statistical methods to high-quality economic and social development. The conference featured rich and diverse content with clear thematic focus, addressing both data-driven issues in real economic and financial markets, and delving into fundamental theories, methodological innovations, and applications in statistical learning, optimization algorithms, and AI models. Faculty and students actively exchanged ideas and engaged in discussions on related research topics, fostering a strong academic atmosphere. The successful organization of this conference further promoted interdisciplinaryintegration among statistics, optimization, machine learning, and financial technology, while also establishing a valuable platform for academic collaboration and talent development in these fields.（By Xiuli Du）