RESEARCH INTERESTS

Learning, Stochastic Dynamics, and Theoretical Physics

My research is organized around a recurring question: how do complex systems move, fluctuate, learn, and respond? I have worked on stochastic optimization, nonequilibrium physics, quantum information, and gravitational dynamics. Across these areas, I tend to look for a compact dynamical description that exposes the mechanism rather than only the final observable.

I am currently extending this viewpoint toward learning and reasoning systems, especially problems where stochastic dynamics, optimization, and scientific inference meet.

• Learning and stochastic dynamics
  • For finite-learning-rate stochastic gradient descent (SGD), we derived analytical formulae for the noise and model fluctuations, treating the discrete update scale as part of the dynamics rather than an infinitesimal limit.[ICML 2021]
  • We analyzed the fundamental properties of minibatch noise in discrete-time SGD.[ICLR 2022 Spotlight]
  • We derived the stationary distribution near a local minimum and found a power-law escape rate for minibatch SGD.[ICML 2022 Spotlight]
  • In generic nonequilibrium Markov processes, we showed that dynamical activity universally constrains static response, giving the response kinetic uncertainty relation (R-KUR).[Commun. Phys. 8, 62 (2025)]
  • Earlier, we derived thermodynamic uncertainty relations for arbitrary initial states in continuous- and discrete-time Markov processes and applied them to measurement and feedback control.[Phys. Rev. Lett. 125, 140602 (2020)]
• Quantum information, response, and transport
  • We extended the quantum information engine to many fermions and showed that the leftmost-particle position is the minimal feedback record. A particle-number-independent measurement-energy bound keeps the record cost compact, enabling a collective work-per-information advantage that reaches 44% at a calibrated operating point before Pauli blocking suppresses the gain.[arXiv:2609.24558]
  • We quantified how a finite-support state-preparation constraint limits ballistic transport in a hard-wall half-line continuous-time quantum walk. The optimal mean drift reduces exactly to a principal-eigenvalue problem, and its deficit from the maximal group velocity decays as π²J/(4M²), with the coefficient set by a continuum Fourier-window variational problem.[Phys. Rev. A 114, 032434 (2026)]
  • We showed that the conjectured square-root bound for quantum random access codes (QRACs) fails in the standard density-operator model because classical random access codes with private randomness form an embedded diagonal-state subclass. The resulting counterexamples fill the full asymptotic interval between the square-root curve and Nayak’s entropy bound at every fixed compression rate, and achieve order-optimal logarithmic qubit scaling at vanishing recovery bias.[arXiv:2607.15617]
  • We derived finite-frequency fluctuation-response bounds for Markovian open quantum systems in an input-output setting, showing that any detector’s lock-in response precision is limited by the emitted-field quantum Fisher information and, for dissipative coupling modulation, by calibrated signal-channel activity.[Quantum Science and Technology (2026)]
  • We derived a quantum response kinetic uncertainty relation (QR-KUR) for continuously monitored Markovian open quantum systems, showing that the static response precision of a trajectory observable is bounded by the conventional quantum dynamical activity together with a perturbation-induced inter-subspace transition term that disappears in the classical limit and is further constrained by symmetry-sector selection rules.[Phys. Rev. A 113, 062443 (2026)]
  • We constructed a new type of quantum information engine that can store useful work cumulatively and transport a quantum particle unidirectionally by harnessing purely quantum fluctuations with the aid of Maxwell’s demon, whose maximum power and maximum transport velocity are well-defined and the optimal operation time is specified. We proposed an improved definition of the efficiency by including all possible energy flow involved in the engine cycle. We discussed possible experimental implementations with existing techniques, especially those for cold atom systems and optical lattices.[Phys. Rev. A 113, 022436 (2026)]
  • We experimentally realized Maxwell’s demon using a 62-qubit superconducting quantum processor, being the first experiment using an isolated quantum many-body system. [Phys. Rev. A 109, 062614 (2024)]
• Complex dynamics and gravitation
  • We identified horizon–orbit scale competition as the mechanism controlling chaos-bound violation for charged spinning particles in black-bounce–Kerr–Newman spacetime. Background deformation can suppress surface gravity faster than orbital instability, while at fixed background particle angular momentum, spin, and charge shift the instability threshold.[arXiv:2609.18143]
  • We studied the classical chaos bound for charged spinning particles in Kerr–Newman–AdS spacetime. Violation is mainly controlled by the interplay of black-hole rotation, charge, AdS curvature, and particle angular momentum, with counter-rotation generally favoring violation and intrinsic spin producing smaller threshold shifts.[arXiv:2607.00432]
  • We investigated Lyapunov-exponent bounds for spinning charged particles in Kerr–Newman black-hole spacetimes, showing how black-hole rotation, particle spin, charge, and angular momentum control whether the exponent exceeds the surface-gravity/MSS scale in both non-extremal and extremal regimes.[Eur. Phys. J. C 86, 677 (2026)]