| Home | Publications | Projects | News | CV | GitHub · Google Scholar · Email |
“Quantum-Inspired Trainable and Parameter-Efficient Tensor Networks for Image Inpainting” (with Konstantinos Slavakis) is now on arXiv: arXiv:2609.17298. It has been submitted to ICASSP 2027.
This is the follow-up to Fast Trainable Multilinear Bases for Image Compression: the same quantum Fourier transform circuits, trained this time for image inpainting rather than compression. Inpainting recovers an image from a subset of its pixels, and there the coherence of the transform with the pixel basis matters as much as its sparsity. The diagonal QFT relaxation keeps one fixed Hadamard per wire and frees only the controlled-phase gates, so its coherence is pinned at the minimum for every parameter value, by circuit topology alone. That removes the need for a coherence penalty and for Riemannian optimization: the phases are trained with plain Adam through an unrolled hard-thresholding recovery, and the transform stays exactly invertible with O(N2 log N) cost.
On DIV2K at 10% observed pixels, the diagonal model with 288 trainable parameters outperforms the DFT by 2.0 dB and the best fixed transform by 1.6 dB in PSNR, and comes within 0.2 dB of a learned butterfly factorization that carries 64 times as many parameters.
“Fast Trainable Multilinear Bases for Image Compression” (with Zhongyi Ni, Huanhai Zhou, and Jin-Guo Liu) is now on arXiv: arXiv:2608.00053.
We generalize the DFT, the DCT, and their block-wise variants to isometric multilinear bases that keep near-linear transform cost, exact invertibility, and a parameter count polylogarithmic in the image size. The basis is parameterized as an isometric tensor network inspired by quantum many-body theory and trained per dataset with Riemannian optimization on the manifold of unitary matrices. On Quick Draw line drawings, the trained basis stores images in roughly 20% fewer bytes than JPEG’s 8×8 block cosine transform at the same reconstruction quality.
The implementation ships as the JAX-based Python library zazabap/pdft, with the benchmark suite and reproduction scripts at zazabap/pdft-benchmarks.
“Riemannian Gauss–Newton Method for Open-System Projected Variational Quantum Dynamics” (with Konstantinos Slavakis) is accepted to the poster track of IEEE Quantum Week 2026 (QCE 2026), held in Toronto, Canada, on 13–18 September 2026. The two-page proceedings paper is available here.
Projected variational quantum dynamics (p-VQD) refits a fixed-depth circuit to each Trotter step by maximizing a pure-state fidelity. For an open system the target is a mixed reduced state, so that objective is undefined, and the obvious repair, fitting the reduced trace overlap, is provably capped below unity: its maximizer is the dominant eigenvector rather than the target. We purify the thermal bath with a thermofield ancilla so that the joint system–bath–ancilla state stays pure. The fidelity then applies exactly on the joint state, and tracing out the bath and the ancilla afterwards returns the physical reduced dynamics.
Each projection becomes a pure-state least-squares fit. A Riemannian Gauss–Newton inner loop with trust-region damping, whose normal matrix is the Fubini–Study metric, drives it to the ansatz ceiling in two or three iterations where first-order p-VQD stalls orders of magnitude short. On finite-temperature spin–boson and central-spin benchmarks, validated against an independent TDVP baseline, the ansatz sets the attainable fidelity and the optimizer sets the convergence rate. Where the ansatz is under-resourced, the second-order step attains the best tested fidelities.
The JAX reference implementation is at zazabap/RiemannianPVQD.
“Problem Reductions at Scale: Agentic Integration of Computationally Hard Problems” (with Xi-Wei Pan and Jin-Guo Liu) is now on arXiv: arXiv:2604.11535.
The companion Rust library and CLI (pred) is open-sourced at CodingThrust/problem-reductions, implementing 100+ NP-hard problem types and 200+ reduction rules, with automatic reduction-path search that lets a single registered solver be reused across the whole connected problem graph.
I will be exchanged at HKUST guangzhou campus from Jan.-Mar. 2026. Please check the work at Prof. Jinguo Liu’s group for Tensor Network related topics and research.
The official paper on LogosQ is published on the arXiv: arXiv:2512.23183. Please check it if you are interested in the implementation details.
The paper on graph encoding with variational quantum circuits, Tensor-Based Binary Graph Encoding for Variational Quantum Classifiers, is accepted by IEEE QCNC 2025 in Nara, Japan. The details are on the conference page.
“Search for Nearly Mass-Degenerate Higgsinos Using Low-Momentum Mildly Displaced Tracks in pp Collisions at √s = 13 TeV with the ATLAS Detector” is published in Physical Review Letters: Phys. Rev. Lett. 132, 221801 (2024). I am a co-author as a member of the ATLAS Collaboration. The preprint is arXiv:2401.14046.
Higgsinos with masses near the electroweak scale could solve the hierarchy problem and provide a dark matter candidate, but they are hard to detect at the LHC when the charged and neutral states differ in mass by about 1 GeV. The search selects events with an energetic jet, missing transverse momentum, and a low-momentum track with a significant transverse impact parameter, in 140 fb−1 of proton-proton collisions at 13 TeV. For the first time since LEP, it excludes mass splittings between the lightest charged and neutral Higgsinos from 0.3 to 0.9 GeV at 95% confidence level, reaching Higgsino masses of about 170 GeV.
“Combination of searches for heavy spin-1 resonances using 139 fb−1 of proton-proton collision data at √s = 13 TeV with the ATLAS detector” is published in the Journal of High Energy Physics: JHEP 04 (2024) 118. I am a co-author as a member of the ATLAS Collaboration. The preprint is arXiv:2402.10607.
The paper combines searches for new heavy spin-1 resonances that decay into pairs of W, Z, or Higgs bosons, or directly into leptons or quarks, using the full Run 2 dataset of 139 fb−1 collected in 2015–2018. Searches in quark pairs and third-generation leptons enter such a combination for the first time. In a heavy vector-boson triplet model, the combined data exclude masses below 5.8 TeV in a weakly coupled scenario and below 4.4 TeV in a strongly coupled one, constraints stronger than those from any individual analysis.
I completed my M.Sc. in experimental high energy physics at KEK. From April 2021 to September 2023 I was a junior physicist in the Inner Tracking Detector group of the ATLAS experiment at CERN. I worked on Higgs self-coupling physics in the HH → bbττ channel with custom CNN and GNN pipelines on ATLAS pixel-detector data, and with CERN on detector testing and data interpretation.
In September 2023 I started the Ph.D. in Information and Communication Engineering at Tokyo Institute of Technology (now Institute of Science Tokyo), supervised by Prof. Konstantinos Slavakis.
I presented “Background Estimation for Non-resonant Di-Higgs Production Search in the bbtautau Final State on the ATLAS experiment” (with Tulin Varol, Yanlin Liu, Manabu Togawa, Yu Nakahama, and the ATLAS Japan Group) at the 2023 Spring Meeting of the Physical Society of Japan, held online on 22–25 March 2023 (talk 24aT1-6). The abstract is on J-STAGE.
The Higgs boson trilinear self-coupling λHHH sets the shape of the Higgs potential and can be measured through Higgs boson pair production. The bbττ final state is one of the most sensitive channels, but when the taus decay hadronically it is contaminated by multijet background. The talk presents a data-driven fake-factor estimate of that background, derived in control regions defined by b-tagging and the charge sign of the two taus, for use in new signal regions with better sensitivity to λHHH.
“Overnight Stock Movement Prediction with Contextualized Embedding Incorporating News and Stock” (with Junjie H. Xu, Liangcheng Lyu, and Yiming Bai) appears in the proceedings of the 2022 IEEE 11th Global Conference on Consumer Electronics (GCCE), held in Osaka, Japan, on 18–21 October 2022, pp. 476–478.
Given a piece of news and a stock code, the model predicts the overnight movement of that stock from a combined news–stock embedding. One part is a contextualized embedding of the news, produced by a language model trained on news paired with the stock movements that followed it. The other is a stock embedding, produced by a pre-trained language model trained on information about the stocks. The paper also discusses what the stock embedding adds.
I presented “Performance Evaluation of Quality Control Data Acquisition System for Silicon Pixel Detector Production in ATLAS Experiment High-Luminosity LHC Upgrade” (with Manabu Togawa, Koji Nakamura, Yoichi Ikegami, Kazunori Hanagaki, and the ATLAS Japan Silicon Group) at the 77th Annual Meeting of the Physical Society of Japan, held at Okayama University, Okayama University of Science, and online on 15–19 March 2022 (talk 16pA562-12). The abstract is on J-STAGE.
For the High-Luminosity LHC, ATLAS will replace its inner tracking detector, and every new silicon pixel module has to be threshold-tuned for the tracker to reconstruct tracks and calibrate energies well. The talk evaluates and improves the data acquisition system used for quality control in module production: the tuning algorithm, the charge-injection setup, and the choice of scans. The final evaluation of each newly assembled module combines its built-in charge-injection circuits, an X-ray source, and statistical tools in the local database.
At the same meeting I co-authored two more talks from the ATLAS Japan Silicon Group: S. Arakuta et al. on the readout test system for the first prototype ATLAS pixel modules for the HL-LHC (16aA562-12), and K. Miyamoto et al. on assembling the module cooling parts of the new silicon pixel detector (17aA561-5).
Last updated November 1, 2026.