Numerical Analysis Seminar

Seminar information archive ~03/28Next seminarFuture seminars 03/29~

Date, time & place Tuesday 16:30 - 18:00 002Room #002 (Graduate School of Math. Sci. Bldg.)
Organizer(s) Norikazu Saito, Takahito Kashiwabara

Seminar information archive

2024/03/13

16:30-17:30   Online
David Sommer (Weierstrass Institute for Applied Analysis and Stochastics)
Approximating Langevin Monte Carlo with ResNet-like neural network architectures (English)
[ Abstract ]
We analyse a method to sample from a given target distribution by constructing a neural network which maps samples from a simple reference distribution, e.g. the standard normal, to samples from the target distribution. For this, we propose using a neural network architecture inspired by the Langevin Monte Carlo (LMC) algorithm. Based on LMC perturbation results, approximation rates of the proposed architecture for smooth, log-concave target distributions measured in the Wasserstein-2 distance are shown. The analysis heavily relies on the notion of sub-Gaussianity of the intermediate measures of the perturbed LMC process. In particular, we derive bounds on the growth of the intermediate variance proxies under different assumptions on the perturbations. Moreover, we propose an architecture similar to deep residual neural networks (ResNets) and derive expressivity results for approximating the sample to target distribution map.
[ Reference URL ]
https://sites.google.com/g.ecc.u-tokyo.ac.jp/utnas-bulletin-board/

2024/03/13

17:30-18:30   Online
Andreas Rathsfeld (Weierstrass Institute for Applied Analysis and Stochastics)
Analysis of the Scattering Matrix Algorithm (RCWA) for Diffraction by Periodic Surface Structures (English)
[ Abstract ]
The scattering matrix algorithm is a popular numerical method for the diffraction of optical waves by periodic surfaces. The computational domain is divided into horizontal slices and, by a clever recursion, an approximated operator, mapping incoming into outgoing waves, is obtained. Combining this with numerical schemes inside the slices, methods like RCWA and FMM have been designed.
The key for the analysis is the scattering problem with special radiation conditions for inhomogeneous cover materials. If the numerical scheme inside the slices is the FEM, then the scattering matrix algorithm is nothing else than a clever version of a domain decomposition method.
[ Reference URL ]
https://sites.google.com/g.ecc.u-tokyo.ac.jp/utnas-bulletin-board/

2024/01/09

16:30-18:00   Room #002 (Graduate School of Math. Sci. Bldg.)
Takashi Matsubara (Osaka University)
Deep learning that learns from, becomes part of, or replaces numerical methods for differential equations (Japanese)
[ Reference URL ]
https://sites.google.com/g.ecc.u-tokyo.ac.jp/utnas-bulletin-board/

2023/11/14

16:30-18:00   Room #002 (Graduate School of Math. Sci. Bldg.)
Ken Furukawa (RIKEN)
On some dynamical systems and their prediction using data assimilation (Japanese)
[ Reference URL ]
ハイブリッド開催です。参加の詳細は参考URLをご覧ください。

2023/10/24

16:30-18:00   Room #002 (Graduate School of Math. Sci. Bldg.)
Kazuaki Tanaka (Waseda University)
Neural Network-based Enclosure of Solutions to Differential Equations and Reconsideration of the Sub- and Super-solution Method (Japanese)
[ Reference URL ]
https://sites.google.com/g.ecc.u-tokyo.ac.jp/utnas-bulletin-board/

2023/10/17

16:30-18:00   Room #002 (Graduate School of Math. Sci. Bldg.)
Makoto Okumura (Konan University)
Structure-preserving schemes for the Cahn-Hilliard equation with dynamic boundary conditions in two spatial dimensions (Japanese)
[ Reference URL ]
https://sites.google.com/g.ecc.u-tokyo.ac.jp/utnas-bulletin-board/

2023/06/27

16:30-18:00   Room #002 (Graduate School of Math. Sci. Bldg.)
Toshihiro Yamada (Hitotsubashi University)
Solving high-dimensional partial differential equations via deep learning and probabilistic methods (Japanese)
[ Reference URL ]
https://sites.google.com/g.ecc.u-tokyo.ac.jp/utnas-bulletin-board/

2023/06/06

16:30-18:00   Room #002 (Graduate School of Math. Sci. Bldg.)
Hideyuki Azegami (Nagoya Industrial Science Research Institute)
Relation between regularity and numerical solutions of shape optimization problems
(Japanese)
[ Reference URL ]
https://sites.google.com/g.ecc.u-tokyo.ac.jp/utnas-bulletin-board/

2023/05/23

16:30-18:00   Room #002 (Graduate School of Math. Sci. Bldg.)
Masaaki Imaizumi (The University of Tokyo)
Theory of Deep Learning and Over-Parameterization (Japanese)
[ Reference URL ]
https://sites.google.com/g.ecc.u-tokyo.ac.jp/utnas-bulletin-board/

2023/05/16

16:30-18:00   Room #002 (Graduate School of Math. Sci. Bldg.)
Yuuki Shimizu (The University of Tokyo)
Numerical analysis of the Plateau problem by the method of fundamental solutions (Japanese)
[ Reference URL ]
https://sites.google.com/g.ecc.u-tokyo.ac.jp/utnas-bulletin-board/

2023/04/25

16:30-18:00   Room #002 (Graduate School of Math. Sci. Bldg.)
Taihei Oki (The University of Tokyo)
Combinatorial preprocessing methods for differential-algebraic equations (Japanese)
[ Reference URL ]
https://sites.google.com/g.ecc.u-tokyo.ac.jp/utnas-bulletin-board/

2021/07/06

16:30-18:00   Online
Ken Hayami (National Institute of Informatics (Professor Emeritus))
Iterative solution methods for least squares problems and their applications
(Japanese)
[ Reference URL ]
https://forms.gle/B5Hwxa7o8F36hZKr7

2021/06/22

17:00-18:30   Online
Taiji Suzuki (The University of Tokyo)
On approximation ability and adaptivity of deep neural network (Japanese)
[ Reference URL ]
https://forms.gle/HwetNGXCzbCyMC7B7

2021/06/08

16:30-18:00   Online
Kohei Soga (Keio University)
Action minimizing random walks and numerical analysis of Hamilton-Jacobi equations (Japanese)
[ Reference URL ]
https://forms.gle/kjhqne4nV6fqEFWB8

2021/05/11

16:30-18:00   Online
Takuya Tsuchiya (Ehime University )
Topics on finite element error analysis on anisotropic meshes (Japanese)
[ Reference URL ]
https://forms.gle/CoaM4vSE1GvDRuDR6

2021/04/27

16:30-18:00   Online
Akitoshi Takayasu (University of Tsukuba)
Rigorous numerics for nonlinear heat equations in the complex plane of time (Japanese)
[ Reference URL ]
https://forms.gle/qW5ktphBB6dsh8Np7

2021/01/12

16:30-18:00   Online
Takaharu Yaguchi (Kobe University)
DGNet: Deep Energy-Based Modeling of Discrete-Time Physics and Related Topics (Japanese)
[ Reference URL ]
https://forms.gle/DpuhGupZ7NYbot5d7

2020/12/15

16:30-18:00   Online
Ming-Cheng Shiue (National Chiao Tung University)
Iterated pressure-correction projection methods for the 2d Navier-Stokes equations based on the scalar auxiliary variable approach (English)
[ Abstract ]
In this talk, the first-order iterated pressure-correction projection methods based on the scalar auxiliary variable approach is proposed and studied for the 2d Navier-Stokes equations and Boussinesq equations.
In the literature, enormous amounts of work have contributed to the study of numerical schemes for computing the Navier-Stokes equations. In general, two of the main numerical difficulties for solving Navier-Stokes equations are the incompressible condition and the nonlinear term. One of the approaches to deal with the incompressible condition is the so-called projection. The typical projection method only needs to solve the Poisson type of equations depending on the nonlinear term's treatment, which is efficient. However, the pressure-correction projection methods suffer from the splitting error, leading to spurious numerical boundary layers and the limitation of accuracy in time. In the literature, an iterated pressure-correction projection method has been proposed to overcome the difficulty.
As for the nonlinear term treatment, it is better to treat the nonlinear term explicitly so that one only requires to solve the corresponding linear system with constant coefficients at each time step. However, such treatment often results in a restricted time step due to the stable issue. Recently, the scalar auxiliary variable approach has been constructed to have an unconditional energy stable numerical scheme.
In this work, a new iterated pressure-correction projection method based on the scalar auxiliary variable's simple choice is proposed. We find that this new scheme can enjoy two properties, including reducing the splitting errors and having unconditional energy stability. The proofs of the energy stability and error convergence are provided and analyzed. Finally, numerical examples are provided to illustrate the theoretical work. This is joint work with Tony Chang.
[ Reference URL ]
https://forms.gle/y7w2nmaYtHNeoDSn8

2020/12/01

16:30-18:00   Online
Hiroyuki Sato (Kyoto University)
Conjugate gradient methods for optimization problems on manifolds (Japanese)
[ Reference URL ]
https://forms.gle/Ubeccm8neLkacjbk8

2020/10/27

16:30-18:00   Online
Buyang Li (The Hong Kong Polytechnic University)
Convergent evolving finite element algorithms for mean curvature flow and Willmore flow of closed surfaces (English)
[ Abstract ]
We construct evolving surface finite element methods for the mean curvature and Willmore flow through equivalently reformulating the original equations into coupled systems governing the evolution of surface position, velocity, normal vector and mean curvature. Then we prove $H^1$-norm convergence of the proposed evolving surface finite element methods for the reformulated systems, by combining stability estimates and consistency estimates. The stability analysis is based on the matrix–vector formulation of the finite element method and does not use geometric arguments. The geometry enters only into the consistency estimates. Numerical experiments illustrate and complement the theoretical results.
[1] https://doi.org/10.1007/s00211-019-01074-2
[2] https://arxiv.org/abs/2007.15257
[ Reference URL ]
https://forms.gle/HeuUxWLGa696KPvz8

2020/07/21

16:30-18:00   Online
Tomoya Kemmochi (Nagoya University)
Structure-preserving numerical schemes for constrained gradient flows of planar curves (Japanese)
[ Reference URL ]
https://forms.gle/3JiNEjWnrWLW8cFA9

2020/06/30

16:30-18:00   Online
Koya Sakakibara (Okayama University of Science)
Structure-preserving numerical methods for interface problems (Japanese)
[ Reference URL ]
https://forms.gle/ztK741vNdBT7hfGSA

2020/06/23

16:30-18:00   Online
Shun Sato (The University of Tokyo)
Linearly implicit and high-order conservative schemes for ordinary differential equations with a quadratic invariant (Japanese)
[ Reference URL ]
https://forms.gle/hvvvFLAhH1314UQK8

2020/01/20

16:50-18:20   Room #056 (Graduate School of Math. Sci. Bldg.)
Yves A. B. C. Barbosa (Politecnico di Milano)
Isogeometric Hierarchical Model Reduction: from analysis to patient-specific simulations (English)
[ Abstract ]
In the field of hemodynamics, numerical models have evolved to account for the demands in speed and accuracy of modern diagnostic medicine. In this context, we studied in detail Hierarchical Model Reduction technique combined with Isogeometric Analysis (HigaMOD), a technique recently developed in [Perotto, Reali, Rusconi and Veneziani (2017)]. HigaMod is a reduction procedure used to downscale models when the phenomenon at hand presents a preferential direction of flow, e.g., when modelling the blood flow in arteries or the water flow in a channel network. The method showed a significant improvement in reducing the computational power and simulation time, while giving enough information to analyze the problem at hand.

Recently, we focused our work in solving the ADR problem and the Stokes problem in a patient-specific framework. Specifically, we evaluate the computational efficiency of HigaMod in simulating the blood flow in coronary arteries and cerebral arteries. The main goal is to assess the
mprovement that 1D enriched models can provide, with respect to traditional full models, when dealing with demanding 3D CFD simulations. The results obtained, even though preliminary, are promising [Brandes, Barbosa and Perotto (2019); Brandes, Barbosa, Perotto and Suito (2020)].

2019/12/16

16:50-18:20   Room #117 (Graduate School of Math. Sci. Bldg.)
Yuki Ueda (The Hong Kong Polytechnic University)
A second-order stabilization method for linearizing and decoupling nonlinear parabolic systems (Japanese)
[ Abstract ]
We present a new time discretization method for strongly nonlinear parabolic systems. Our method is based on backward finite difference for the first derivative with second-order accuracy and the first-order linear discrete-time scheme for nonlinear systems which has been introduced by H. Murakawa. We propose a second-order stabilization method by combining these schemes.
Our error estimate requires testing the error equation by two test functions and showing $W^{1,\infty}$-boundedness which is proved by ($H^2$ or) $H^3$ energy estimate. We overcome the difficulty for establishing energy estimate by using the generating function technique which is popular in studying ordinary differential equations. Several numerical examples are provided to support the theoretical result.

12345 Next >