学术成果
[Google Scholar] [Research Gate]
已发表(经同行评审)
- Kangqiao Liu, Zongping Gong, and Masahito Ueda
Thermodynamic Uncertainty Relation for Arbitrary Initial States
Physical Review Letters 125, 140602 (2020) (2020.09.29)
(IF = 9.185,中科院1区Top)
[arXiv:1912.11797]. PDF Supplements
本文提出适用于任意初态的热力学不确定性关系,以末态瞬时流取代稳态假设中的平均流。该关系在传统 TUR 可能失效的瞬态弛豫过程中仍成立,并可自然推广到测量反馈协议。 - Kangqiao Liu*, Liu Ziyin*, and Masahito Ueda (*equal contribution)
Noise and Fluctuation of Finite Learning Rate Stochastic Gradient Descent
Proceedings of the 38th International Conference on Machine Learning, PMLR 139:7045-7056, 2021 (ICML 2021) (2021.07.01)
(接收率 = 21.5%,CCF A类,机器学习顶会)
[arXiv:2012.03636]. PDF Supplements
有限学习率使 SGD 本质上成为离散动力学,其稳态协方差、稳定边界与逃逸效率均不同于连续时间 Langevin 近似。离散理论在白噪声和 minibatch 噪声下都能一直准确到失稳边缘。 - Liu Ziyin*, Kangqiao Liu*, Takashi Mori, and Masahito Ueda (*equal contribution)
Strength of Minibatch Noise in SGD
The 10th International Conference on Learning Representations (ICLR 2022) (2022.01.29)
受选为Spotlight (5% of all submissions)
(接收率 = 32.26%, Spotlight = top 5%,清华A类,机器学习顶会)
[arXiv:2102.05375]. PDF
本文建立有限学习率下 minibatch 噪声强度与参数涨落的离散时间理论,并给出方差发散阈值。该框架解释了大步长下连续时间近似的失效,并预测批量大小、模型宽度与正则化共同决定的稳定性。 - Takashi Mori, Liu Ziyin, Kangqiao Liu, and Masahito Ueda
Power-law escape rate of SGD
Proceedings of the 39th International Conference on Machine Learning, PMLR 162:15959-15975, 2022 (ICML 2022) (2022.07.15)
受选为Spotlight
(接收率 = 21.9%,CCF A类,机器学习顶会)
[arXiv:2105.09557]. PDF
当 SGD 噪声强度随损失函数变化时,稳态分布呈现重尾幂律,逃离局部极小值的速率也由幂律而非 Arrhenius 指数律控制。理论预测与神经网络数值实验给出了相符的首达时间标度。 - Jiale Yu*, Shiyu Wang*, Kangqiao Liu, Chen Zha, Yulin Wu, Fusheng Chen, Yangsen Ye, Shaowei Li, Qingling Zhu, Shaojun Guo, Haoran Qian, He-Liang Huang, Youwei Zhao, Chong Ying, Daojin Fan, Dachao Wu, Hong Su, Hui Deng, Hao Rong, Kaili Zhang, Sirui Cao, Jin Lin, Yu Xu, Cheng Guo, Na Li, Futian Liang, Yong-Heng Huo, Chao-Yang Lu, Cheng-Zhi Peng, Kae Nemoto, W. J. Munro, Xiaobo Zhu, Jian-Wei Pan, and Ming Gong
Experimental demonstration of a Maxwell’s demon quantum battery in a superconducting NISQ processor
Physical Review A 109, 062614 (2024) (2024.06.20)
(IF = 2.9,中科院2区). PDF
超导 NISQ 处理器通过基于测量结果的 Maxwell 妖反馈,把激发定向输运到目标量子比特子系统。实验展示了量子电池的定向充电,并测量了相应的熵与遍历性动力学。 - Kangqiao Liu and Jie Gu
Dynamical activity universally bounds precision of response in Markovian nonequilibrium systems
Communications Physics 8, 62 (2025) (2025.02.11)
(IF = 5.4,中科院1区Top)
栏目撰稿Behind the Paper
学院新闻
[arXiv:2410.20800]. PDF
对 Markov 非平衡过程,响应精度受到动力学活性(系统状态改变的总速率)的普适限制,而不仅由耗散决定。该界同样适用于弛豫过程,并可比响应热力学不确定性关系更紧。 - Kangqiao Liu, Masaya Nakagawa, and Masahito Ueda
Maxwell’s demon for quantum transport
Physical Review A 113, 022436 (2026) (2026.02.23)
(IF = 2.9,中科院2区),
学院新闻
[arXiv:2303.08326]. PDF
重复位置测量与反馈控制把倾斜晶格中的相干跃迁转化为定向上坡输运,从而在无热偏置条件下储存势能。本文系统分析了该量子信息引擎的速度、功率、效率与鲁棒性。 - Kangqiao Liu and Jie Gu
Response kinetic uncertainty relation for Markovian open quantum systems
Physical Review A 113, 062443 (2026) (2026.06.16)
(IF = 2.9,中科院2区),
学院新闻
[arXiv:2501.04895]. PDF
响应动力学不确定性关系把响应精度的代价分解为跃迁活性项与跨跃迁量子项。两部分在不同扰动下分别占主导,从而为受监测的 Markov 量子动力学给出统一上界。 - Chuang Yang, Chuanhong Gao, Deyou Chen, and Kangqiao Liu
Bound on Lyapunov exponents with spinning particles in Kerr–Newman spacetimes
The European Physical Journal C 86, 677 (2026) (2026.06.22)
(IF = 4.9,中科院2区). PDF
当旋转、粒子自旋、电荷或总角动量跨越依赖取向的阈值时,粒子的 Lyapunov 指数可以超过表面引力尺度。反向旋转构型与极端黑洞极限表现出更显著的违背。
未经同行评审
- Jie Gu and Kangqiao Liu
Finite-frequency fluctuation-response bounds for open quantum systems
[arXiv:2605.03340] (2026.05.06). PDF
本文建立了频率分辨的输入—输出涨落响应不等式:任意探测方案测得的锁相响应都受输出场涨落限制,其探测器无关的上限由量子 Fisher 信息与通道活性给出。驱动二能级系统和猫态谐振腔的算例验证了这一频域约束。 - Deyou Chen, Chuang Yang, and Kangqiao Liu
Chaos bound for spinning particles in Kerr-Newman-AdS black holes
[arXiv:2607.00432] (2026.07.01). PDF
数值相图刻画了 Kerr-Newman-AdS 时空中粒子 Lyapunov 指数超过黑洞表面引力的参数区域。旋转方向、电荷与宇宙学常数共同重塑混沌界的违背窗口。 - Kangqiao Liu
Classical codes violate the conjectured square-root bound for quantum random access codes
[arXiv:2607.15617] (2026.07.20). PDF
带私有随机性的经典随机访问码本身就是一般 QRAC 模型的子类,并可突破猜想中的平方根界。一般模型的渐近最坏情形边界实际由 Nayak 熵界决定。