コーヒーミーティング過去ログ

2025-10-10

Gravitational redshift and wave packets in QFT

杉下 宗太郎 (理化学研究所 数理創造研究センター (iTHEMS) 数理基礎部門 研究員 / 京都大学 大学院理学研究科 助教)

Light emitted near a black hole appears redshifted to a distant observer; this is the gravitational redshift. I will explain how to describe it using wave packets in quantum field theory.

2025-10-03

Gelfand duality

ウアジーミャ・ソスニロ (理化学研究所 数理創造研究センター (iTHEMS) 数理基礎部門 研究員)

In a broad sense, geometry studies those mathematical objects that require geometric intuition, while algebra studies abstract structures that are common in all fields of mathematics. In various instances, it turns out that there is a way to describe geometric objects entirely using algebraic objects. As one such example, we discuss Gelfand duality between compact hausdorf topological spaces and commutative C*-algebras.

2025-09-26

Representation learning for astronomy

平島 敬也 (理化学研究所 数理創造研究センター (iTHEMS) 数理基礎部門 基礎科学特別研究員)

Stellar mass is a fundamental driver of stellar evolution, yet estimating it in star-forming regions is difficult due to heavy obscuration and strong inhomogeneity, which undermines simple dynamical models. Supervised ML is promising but limited by the cost of generating large, high-quality labeled datasets from high-resolution MHD simulations. In this talk, I will present a data-efficient alternative: pretraining a Vision Transformer with DINOv2 on one million synthetic fractal images, then transferring the frozen encoder to limited MHD maps. Synthetic pretraining improves frozen-feature stellar-mass predictions, slightly outperforming a supervised model trained on the same limited simulations. Principal component analysis of the embeddings reveals semantically meaningful structure (e.g., dense cores and inflows), enabling unsupervised segmentation without labels or lightweight fine-tuning.

2025-09-19

Scale transformations in physics and beyond

アルバロ・パストール グティエレス (理化学研究所 数理創造研究センター (iTHEMS) 数理基礎部門 基礎科学特別研究員)

The evolution of systems across different scales is a common theme in many areas of research, from fundamental physics to biology and machine learning. Understanding how macroscopic behaviour arises from microscopic dynamics is essential both for explaining emergent phenomena and for reconstructing a fundamental picture of nature. In this talk, I will introduce the renormalisation group in a non-technical way, highlighting its interpretation as a framework for performing scale transformations applicable to a broad range of contexts beyond physics. Beginning with the block-spinning idea, I will show how scale transformations lead to phase transitions and universal phenomena, and then sketch the modern Wilsonian perspective. Finally, I will show how this framework underlies our most fundamental picture of nature and offers a toolkit for addressing the open problems and puzzles in the Standard Model of particle physics.

2025-09-12

The coalescent history of humans

シュパイデル 玲雄 (理化学研究所 数理創造研究センター (iTHEMS) 数理基礎部門 数理遺伝学理研ECL研究ユニット 理研ECL研究ユニットリーダー / 理化学研究所 開拓研究所 (PRI) シュパイデル数理遺伝学理研ECL研究ユニット 理研ECL研究ユニットリーダー)

Our DNA differs from person to person, but how have these differences come about? I will introduce the fundamental statistical framework for understanding genetic variation, the standard coalescent model. I will demonstrate how the coalescent model captures key properties of genetic variation. The coalescent is central to statistical inference techniques used to reconstruct human evolution, and remarkably, allow us to learn about the earliest days of our species and up to events in a nation’s recent past.

2025-09-05

What is the definition of black holes in quantum theory?

横倉 祐貴 (理化学研究所 数理創造研究センター (iTHEMS) 数理基礎部門 上級研究員)

What is a black hole? No one knows the answer yet. Current observation data has not confirmed the existence of horizons yet, and there is no theoretical description consistent with quantum theory. Therefore, one natural approach would be to reconsider the definition of black holes in the context of quantum theory. In this talk, I will share my idea and provide one possibility for quantum black holes.

2025-08-29

Balancing Research and Parenthood: From Pregnancy to Parenting a School-Aged Child

辰馬 未沙子 (理化学研究所 数理創造研究センター (iTHEMS) 数理基礎部門 研究員)

The decision of when to have children is never simple for researchers, with each career stage posing different challenges. In this talk, I will share personal insights on balancing pregnancy, childbirth, and parenting with an academic career, even though I have not always been able to balance them successfully. Topics include the physical and mental problems of pregnancy, the pressure in a publish-or-perish environment, the difficulty of traveling with young children, and systemic issues such as daycare availability and school schedules. By sharing these experiences, I hope to foster a more open conversation about how academia can better support researchers with caregiving responsibilities.

2025-08-22

Predicting and Explaining Animal Behavior with a Neural Network

藤原 輝史 (理化学研究所 開拓研究所 (PRI) 藤原適応運動制御理研白眉研究チーム 理研白眉研究チームリーダー)

Explaining behavior with brain activity is a central goal of systems neuroscience. An insect, such as a fruit fly, exhibits sophisticated movements despite its compact brain, suggesting that low-dimensional neural computation underlies its high-dimensional behavior. Nevertheless, the technical challenge in measuring the entire neural activity, even for a small insect, limits our understanding of how the brain coordinates behavior. Therefore, we adopt a neural network simulating animal behavior and utilize its encoding feature to gain insight into low-dimensional computation. We train a neural network to predict a walking fly’s future step location (output) based on its previous step history (input). We then analyze the compressed representation at the computational bottleneck, paired with its input, to explain how the fly decides on the next step location. We have recently been particularly interested in a Mixture Density Network. Here, the fly’s “mind”, represented as a mixture of Gaussian distributions, can select a particular step location not only from a single Gaussian lobe but also from well-separated, multiple Gaussian lobes, providing greater flexibility, as observed in the actual animal. We hypothesize that analyzing the structure of the Gaussian mixture and corresponding input-output step patterns elucidates how a small animal can take complex action with limited neural resources.

2025-08-08

Mapping the Universe with DESI

Andrei Cuceu (NASA Einstein Fellow, Lawrence Berkeley National Laboratory (LBNL), USA)

The Dark Energy Spectroscopic Instrument (DESI) is in the process of creating the largest three-dimensional map of the Universe. I will explain how we are creating this map by measuring precise distances to more than 40 million galaxies, and also how this map is being used to study the evolution of our Universe.

2025-08-01

GWAS: unravelling the links between the human genome and measurable traits

リュカ・ソール (理化学研究所 数理創造研究センター (iTHEMS) 数理基礎部門 数理遺伝学理研ECL研究ユニット 特別研究員)

Understanding how the genome defines lifeforms has long been a central question in biology. In an effort to address this, a new class of studies, Genome-Wide Association Studies (GWAS), emerged and proliferated from the mid-2000s, largely driven by advances in DNA sequencing technologies. In this talk, I will introduce how GWAS are conducted to find associations between the genome and specific observable traits, ranging from disease susceptibility to morphological characteristics.

2025-07-18

Machine Learning for Holographic Entanglement Inequalities

大栗 博司 (Fred Kavli Professor and Director, Walter Burke Institute for Theoretical Physics, California Institute of Technology, USA)

2025-07-04

Nonperturbative Quantum Gravity in a Closed Universe

野村 泰紀 (理化学研究所 数理創造研究センター (iTHEMS) 国内外連携・人材育成部門 理研バークレーセンター 上級研究員 / Professor/Director, Berkeley Center for Theoretical Physics, University of California, Berkeley, USA)

I explain why the nonperturbative quantum gravitational Hilbert space in a closed universe is one-dimensional and real-valued (for each alpha-sector). This result is derived from an analysis involving spacetime wormholes. I also discuss how meaningful physical predictions can arise in this case as a consequence of partial observability: physical observers can access only a subsystem of the universe.

2025-06-27

Describing Quantum Phenomena with Classical Gravity

竹田 大地 (理化学研究所 数理創造研究センター (iTHEMS) 数理基礎部門 基礎科学特別研究員)

The holographic principle asserts an equivalence between gravitational theories and quantum theories without gravity. Through the holographic principle, it has become possible to perform non-perturbative analyses of systems such as QCD and condensed matter by using classical field theories in curved spacetime, or to approach unknown quantum gravity theories through the language of quantum field theories. In this talk, we focus on the former and introduce methods to handle linear response and phase transition in quantum theories using classical gravity.

2025-06-20

Maximum Entropy Reinforcement Learning

田中 章詞 (理化学研究所 数理創造研究センター (iTHEMS) 数理基礎部門 上級研究員 / 理化学研究所 革新知能統合研究センター (AIP) 上級研究員)

Maximum Entropy Reinforcement Learning (MaxEnt RL) augments standard RL by encouraging policies to remain stochastic through entropy maximization. This leads to better exploration, more robust behavior, and improved performance in complex environments. I'd like to review the idea shortly, and comment some connections to thermodynamics/statistical systems.

2025-06-13

Quantum data and quantum machine learning

松浦 俊司 (理化学研究所 数理創造研究センター (iTHEMS) 数理基礎部門 上級研究員)

While classical machine learning has had a great impact on our daily life, the potential impact of quantum machine learning is still unclear. I will discuss this from an information-theoretic perspective. Information-theoretic bounds don’t always provide direct, practical advantages, but they help clarify the fundamental limits and opportunities for quantum machine learning in scientific discovery.

2025-06-06

The Price equation: from breeding to Bayes

トーマス・ヒッチコック (理化学研究所 数理創造研究センター (iTHEMS) 数理基礎部門 基礎科学特別研究員)

In the mid 1960’s an American chemist – George Price - moved to London to understand the origins of altruism. In just the few short years he worked on this problem, he made several key contributions to our understanding of evolution, the most important of which is the equation that now bears his name. Commonly claimed to be the most fundamental in evolution, the Price equation provides an abstractness that allows many other classic results to be derived from it when additional assumptions are made. In this talk will explain how this equation is derived, how other classic expressions emerge as special cases, and finally some of its links beyond evolutionary biology.

2025-05-30

Null hypothesis: Hypotheses in evolutionary biology are not falsifiable

ホセ サイード・グティエレス オルテガ (理化学研究所 数理創造研究センター (iTHEMS) 数理基礎部門 研究員)

Evolutionary biology provides a multidisciplinary framework to understand past events that originated biodiversity. It relies on observational data, and based on observed patterns, researchers can propose theories that clarify the underlying dynamics ruling evolutionary phenomena. However, a common criticism is that hypotheses in evolutionary biology are not falsifiable, since it is impossible to predict an evolutionary phenomenon and set up an experiment to test its certainty (i.e., we don’t have time machines). I would like to mention that falsifiability, through the reasoning of induction, cannot be used to prove historical events, but this is not necessary. Making sense of observational data and proposing plausible hypotheses based on abductive reasoning are the ways of thinking in evolution, which, however, bring some epistemological concerns. For example, phylogenetic analyses based on different genes can produce trees with differing topologies, thereby describing contradictory evolutionary events. Statistical and probability methods are often used to decide which tree topology is most likely or probable to be true. But are not statistics and probability solutions for inductive and deductive (not abductive) problems? Also, how valid is the method of imposing a substitution rate model to a DNA sequence data “after” observing the data? At most, we can tell that evolutionary hypotheses are plausible explanations that can be rejected in the presence of new and stronger data. But in this respect, hypotheses in evolutionary biology are similar to those from other fields: they welcome scrutiny, and their explanatory power increases or decreases as scientific knowledge advances.

2025-05-23

Toy models in data assimilation studies for weather prediction

大塚 成徳 (理化学研究所 数理創造研究センター (iTHEMS) 数理展開部門 予測科学研究チーム 上級研究員 / 理化学研究所 計算科学研究センター (R-CCS) データ同化研究チーム 上級研究員)

In numerical weather prediction systems, we typically use a huge model based on complex physics of the atmosphere. However, it is inconvenient to use such a model for testing new ideas. Therefore, we have a hierarchy of models, ranging from so-called toy models to high-end, expensive ones. Today I will introduce one of the most used toy models to study data assimilation algorithms, the Lorenz96 40-variable model. I will also introduce the DA-ML study group, in which we use this model to learn about the basics of data assimilation and machine learning.

2025-05-16

Shocks: from fluid discontinuities to cosmic particle accelerators

カミリア・デミデム (理化学研究所 数理創造研究センター (iTHEMS) 数理基礎部門 研究員)

In my previous coffee meeting talk, I attempted to offer a gentle introduction to turbulence, a key ingredient of particle acceleration in cosmic environments. This presentation will borrow the same general approach to discuss another important phenomenon of high-energy astrophysics: shock waves. My goal is to provide an intuitive grasp of shocks as understood in fluid dynamics, while also highlighting the distinct complexities inherent to astrophysical collisionless plasmas.

2025-05-09

Statistical Signatures of Quantum Chaos in Non-Hermitian Systems

プラティック・ナンディ (理化学研究所 数理創造研究センター (iTHEMS) 数理基礎部門 特別研究員)

Recent advances have sparked growing interest in the chaotic behavior of open quantum systems, particularly those governed by non-Hermitian Hamiltonians. In this talk, I will present a framework for analyzing the statistical correlations between the complex eigenvalues and the singular values of such Hamiltonians. The methods I will discuss draw from and extend tools in random matrix theory and spectral analysis, with potential applications that reach beyond physics, such as in statistical data analysis, machine learning, and complex network theory.