Manuel Tiglio Hearne Institute for Theoretical Physics

Slides:



Advertisements
Similar presentations
Example Project and Numerical Integration Computational Neuroscience 03 Lecture 11.
Advertisements

Point-wise Discretization Errors in Boundary Element Method for Elasticity Problem Bart F. Zalewski Case Western Reserve University Robert L. Mullen Case.
Numerical Relativity & Gravitational waves I.Introduction II.Status III.Latest results IV.Summary M. Shibata (U. Tokyo)
1 Iterative Solvers for Linear Systems of Equations Presented by: Kaveh Rahnema Supervisor: Dr. Stefan Zimmer
The Finite Element Method Defined
2. Numerical differentiation. Approximate a derivative of a given function. Approximate a derivative of a function defined by discrete data at the discrete.
MATH 685/ CSI 700/ OR 682 Lecture Notes
Visual Recognition Tutorial
Finite Element Method Introduction General Principle
8 TECHNIQUES OF INTEGRATION. In defining a definite integral, we dealt with a function f defined on a finite interval [a, b] and we assumed that f does.
Controller Tuning: A Motivational Example
Approximation on Finite Elements Bruce A. Finlayson Rehnberg Professor of Chemical Engineering.
Introduction to Numerical Methods I
Tutorial 10 Iterative Methods and Matrix Norms. 2 In an iterative process, the k+1 step is defined via: Iterative processes Eigenvector decomposition.
Chapter 4 Numerical Solutions to the Diffusion Equation.
Partial differential equations Function depends on two or more independent variables This is a very simple one - there are many more complicated ones.
PETE 603 Lecture Session #29 Thursday, 7/29/ Iterative Solution Methods Older methods, such as PSOR, and LSOR require user supplied iteration.
Ordinary Differential Equations (ODEs)
Boyce/DiPrima 10th ed, Ch 10.5: Separation of Variables; Heat Conduction in a Rod Elementary Differential Equations and Boundary Value Problems, 10th.
SOLUTION FOR THE BOUNDARY LAYER ON A FLAT PLATE
Pseudospectral Methods
Reaction-Diffusion Systems - Continued Reactive Random Walks.
Systems of Linear Equations Iterative Methods
Finite Differences Finite Difference Approximations  Simple geophysical partial differential equations  Finite differences - definitions  Finite-difference.
Finite Element Method.
Extrapolation Models for Convergence Acceleration and Function ’ s Extension David Levin Tel-Aviv University MAIA Erice 2013.
ECE 8443 – Pattern Recognition ECE 8423 – Adaptive Signal Processing Objectives: Deterministic vs. Random Maximum A Posteriori Maximum Likelihood Minimum.
1 Lecture 3: March 6, 2007 Topic: 1. Frequency-Sampling Methods (Part I)
Controller Design (to determine controller settings for P, PI or PID controllers) Based on Transient Response Criteria Chapter 12.
Progress in identification of damping: Energy-based method with incomplete and noisy data Marco Prandina University of Liverpool.
© Fluent Inc. 11/24/2015J1 Fluids Review TRN Overview of CFD Solution Methodologies.
CHAPTER 3 NUMERICAL METHODS
A Non-iterative Hyperbolic, First-order Conservation Law Approach to Divergence-free Solutions to Maxwell’s Equations Richard J. Thompson 1 and Trevor.
Black Hole Universe -BH in an expanding box- Yoo, Chulmoon ( YITP) Hiroyuki Abe (Osaka City Univ.) Ken-ichi Nakao (Osaka City Univ.) Yohsuke Takamori (Osaka.
Application: Multiresolution Curves Jyun-Ming Chen Spring 2001.
Carles Bona Tomas Ledvinka Carlos Palenzuela Miroslav Zacek Mexico, December 2003 Checking AwA tests with Z4 ( comenzando la revolucion rapida) ( comenzando.
Lecture 9: PID Controller.
Numerical Relativity in Cosmology - my personal perspective - Yoo, Chulmoon ( Nagoya U. ) with Hirotada Okawa ( Lisbon, IST ) New Perspectives on Cosmology.
NUMERICAL ANALYSIS I. Introduction Numerical analysis is concerned with the process by which mathematical problems are solved by the operations.
P2 Chapter 8 CIE Centre A-level Pure Maths © Adam Gibson.
5.3 Trigonometric Graphs.
Numerical Solutions to the Diffusion Equation
Lecture 11 Alessandra Nardi
Ch 11.6: Series of Orthogonal Functions: Mean Convergence
Segmentation COMP 755.
Lecture 4: Numerical Stability
A Signal Processing Approach to Vibration Control and Analysis with Applications in Financial Modeling By Danny Kovach.
1.1 A Preview of Calculus What is Calculus????????????????????
Introduction to Numerical Methods I
Solver & Optimization Problems
THE METHOD OF LINES ANALYSIS OF ASYMMETRIC OPTICAL WAVEGUIDES Ary Syahriar.
Geometric Integrators in the Study of Gravitational Collapse
21th Lecture - Advection - Diffusion
Controller Tuning: A Motivational Example
CHAPTER 29: Multiple Regression*
Tachyon vacuum in Schnabl gauge in level truncation
Hidden Markov Models Part 2: Algorithms
CSE 245: Computer Aided Circuit Simulation and Verification
The MESSM The Minimal Exceptional Supersymmetric Standard Model
Lecture #6 section pages pages72-77
Adaptive Perturbation Theory: QM and Field Theory
Large Time Scale Molecular Paths Using Least Action.
دانشگاه صنعتي اميركبير
TECHNIQUES OF INTEGRATION
Chapter 8: Estimating with Confidence
Topic 8 Pressure Correction
Programming assignment #1 Solving an elliptic PDE using finite differences Numerical Methods for PDEs Spring 2007 Jim E. Jones.
Software Development Techniques
EE, NCKU Tien-Hao Chang (Darby Chang)
Presentation transcript:

Dynamic control of discrete constraint violations in 3D black hole evolutions Manuel Tiglio Hearne Institute for Theoretical Physics Louisiana State University Mexico, Dec 2003

The technique (M. Tiglio, gr-qc/0304062) Given a set of evolution equations, add the constraints Ci = Ci (u, Du), to the right hand side These evolution equations determine how these constraints propagate Define an energy for these constraints Ec = (C,C). You can show that It satisfies the evolution equation Choose m to get the desired behavior, e.g. Do it on the fly, during evolution. In order to achieve symmetric hyperbolicity, do it for one resolution, interpolate, and keep that m(t) from there on. It is then a priori given. If you control the constraints (=they do not grow) for one resolution, and you are in the convergence regime, you are controlling them for higher resolutions as well The discretized constraints will not grow in time

Some general remarks Instead of asking the constraints to be exactly satisfied, fix a tolerance value T, and ask the constraints to decay to T after na=1 timesteps: That is, ask N(t+Dt)=T and solve for na=1 : A good tolerance value is the initial discretization value for the constraints Ideally, one would like na=1, in that way controlling the constraints at each timestep. Although a fully discrete approach could be followed, here all these calculations are semidiscrete: space is assumed to be discrete (with gridspacing not necessariliy small) and time continuous. Will come back to this later.

DBF! DBF! A proof of concept: a spherically symmetric black hole. Without constraint minimization, crashes at ~10M DBF! With constraint minimization, runs for at least 10,000M. Here na=1, and m is defined by the coarsest resolution run.

The 3D equations (O. Sarbach + M The 3D equations (O. Sarbach + M. Tiglio, PRD 66, 064023 (2002), gr-qc/0205086) Symmetric hyperbolic formulation with (a slight generalization of) the Bona-Masso slicing conditions. Two completely free functions of time (and space, if one wants) to use to minimize the constraints growth, here called g and h. It really trivial to add more 21 more free functions.

3D black hole simulations (M. Tiglio, L. Lehner and D 3D black hole simulations (M. Tiglio, L. Lehner and D. Neilsen, gr-qc/0312001). Very generous help from P. Diener and O. Sarbach. Without constraint minimization (g=0=h). The run with coarsest resolution crashes around 20M

Constraint minimization: how good is the semidiscrete picture? Although fully discrete calculations could be done, for simplicity we use semidiscrete ones. How accurately do they represent the fully discrete simulation? Black hole simulation of the previous slide Gauge wave Time derivative of the constraints energy, using: a)the semidiscrete energy calculations.b) a posteriori computing the time derivative of the actual energy through finite differencing. The agreement is remarkable. Actually not so surprising…

Dynamic minimization: preliminaries Let’s warm up by keeping one of the free constraint-functions fixed, say h=0, and dynamically defining g by setting a tolerance value, as described in the second slide. Here the boundaries are at 5M and he tolerance value for the energy here is equal to the initial discretization one. One notices that large (but not too large) values of na perform better. Why? na=1 would be the ideal choice. However, small values of na make the free parameters vary too much in order to keep the constraints under control. Not possible with a fixed Courant factor. On the other hand, large values of na allow the constraints to grow toomuch between two iterations.

Resulting functions for two runs of the previous slide. The titles are wrong. The minimization is done at every timestep, in one case with na =1, and in the with other na =1000. Small values of na unfortunately cause the free constraint-function h to vary too much. What can we do about it? One could fine-tune na, but this is against the whole spirit of minimizing the needed numerical experimentation.

Bi-parametric minimization Having found the limitation of adjusting the equations to suppress instabilies, it is clear how to proceed: Namely, use more than one free constraint-function, in order to avoid large or quick variations in them. Two unkowns for one equation, can satisfy one more equation. Ask the code to choose the pair (h and g) that not only achieves the desired constraint behaviour, but also but minimizes the change in this pair between two timesteps.

Not only quite stable, but also very accurate. Apparent horizon mass, using Jonathan Thornburg’s new fast apparent horizon finder, for the runs of the previous slide. The run with higher resolution ran out of computing time. With the coarsest (very coarse) resolution the errors are initially less than one percent, and later less than one part in 103. For the higher (not that high at all) resolution the errors after a while are of the order of one part in 105 !

Moving out the boundaries: 10M and 15M The dependence of the liftetime on the position of the outer boundaries is not monotonic: Boundaries at 5M runs for ~800M. Boundaries at 10M runs for ~700M. Boundaries at 15M runs for ~1000M. No fine tuning here on the tolerance value or na (here na typically is 100, still not 1 !!!). All these runs are fully 3D ones, no octant, bitant or any other symmetry is imposed. The run with boundaries at 15M is single resolution only because of lack of computing time.

Understanding the mechanism in more detail The resulting, dynamically defined constraint-functions h and g After some time they settle down to some values, and at late times the change very quickly. These late time variations are consequence and not cause of the code crashing. Comparing a run with parametrs from this figure to another run where those parameters are kept fixed after t=750M. The code still crashes, therefore the late time oscillations in the constraint-functions are consequence and not cause of the code crash.

Fixed, fine-tuned constraint-functions After some time the constraint-functions remain fairly constant until the code crashes. What happens if one runs with fixed constraint-functions, with their values given the those asymptotic, quite constant values?. Not big difference. However, presumably this would change in the case of a dynamical solution

How much do the constraint-functions depend on the position of the outer boundaries? a) Boundaries at 5M: they settle down to h=-1.9, g=-1.0 b) Boundaries at 10M: they settle down to h=-3.0 10-1, g=-2.5 c) Boundaries at 15M: they settle down to h=-1.3 10-1 , g=-3.4 Not yet converging to anything when the domain is enlarged. Constraint functions obtained with certain boundary and used for other computational domain is not optimal, but much better than doing nothing. This plot shows a run with boundaries at 10M and fixed constraint-functions given by (a) and (b).

Don’t Be Fooled (TM)!, ask for the convergence tests! DBF! Definition 16 (IDIOT) Anyone who publishes a calculation without checking it against an identical computation with smaller N OR without evaluating the residual of the … approximations via finite differences is an IDIOT John P. Boyd, in "Chebyshev and Fourier Spectral Methods", Second Edition.