Numerical Methods for Partial Differential Equations

Slides:



Advertisements
Similar presentations
Courant and all that Consistency, Convergence Stability Numerical Dispersion Computational grids and numerical anisotropy The goal of this lecture is to.
Advertisements

Finite Difference Discretization of Hyperbolic Equations: Linear Problems Lectures 8, 9 and 10.
Example Project and Numerical Integration Computational Neuroscience 03 Lecture 11.
Formal Computational Skills
Numerical Methods for Partial Differential Equations CAAM 452 Spring 2005 Lecture 9 Instructor: Tim Warburton.
Roundoff and truncation errors
Chapter 8 Elliptic Equation.
MATH 685/ CSI 700/ OR 682 Lecture Notes
1 Predator-Prey Oscillations in Space (again) Sandi Merchant D-dudes meeting November 21, 2005.
Numerical Uncertainty Assessment Reference: P.J. Roache, Verification and Validation in Computational Science and Engineering, Hermosa Press, Albuquerque.
Numerical Integration of Partial Differential Equations (PDEs)
1cs542g-term Notes  Notes for last part of Oct 11 and all of Oct 12 lecture online now  Another extra class this Friday 1-2pm.
Total Recall Math, Part 2 Ordinary diff. equations First order ODE, one boundary/initial condition: Second order ODE.
Lecture 2: Numerical Differentiation. Derivative as a gradient
Atms 4320 Lab 2 Anthony R. Lupo. Lab 2 -Methodologies for evaluating the total derviatives in the fundamental equations of hydrodynamics  Recall that.
Introduction to Numerical Methods I
Chapter 4 Numerical Solutions to the Diffusion Equation.
Numerical Methods for Partial Differential Equations CAAM 452 Spring 2005 Lecture 11 Instructor: Tim Warburton.
Numerical Methods for Partial Differential Equations CAAM 452 Spring 2005 Instructor: Tim Warburton.
Numerical Methods for Partial Differential Equations CAAM 452 Spring 2005 Lecture 9 Instructor: Tim Warburton.
Numerical Methods for Partial Differential Equations CAAM 452 Spring 2005 Lecture 2 Instructor: Tim Warburton.
Numerical Methods for Partial Differential Equations CAAM 452 Spring 2005 Lecture 6 Various Finite Difference Discretizations for the Advection Equations.
Numerical Methods for Partial Differential Equations CAAM 452 Spring 2005 Lecture 8 Instructor: Tim Warburton.
ECE 8443 – Pattern Recognition EE 3512 – Signals: Continuous and Discrete Objectives: Stability and the s-Plane Stability of an RC Circuit 1 st and 2 nd.
Ch 8.1 Numerical Methods: The Euler or Tangent Line Method
Finite Differences Finite Difference Approximations  Simple geophysical partial differential equations  Finite differences - definitions  Finite-difference.
Sistem Kontrol I Kuliah II : Transformasi Laplace Imron Rosyadi, ST 1.
Simulating Electron Dynamics in 1D
Lecture 22 MA471 Fall Advection Equation Recall the 2D advection equation: We will use a Runge-Kutta time integrator and spectral representation.
Variants of the 1D Wave Equation Jason Batchelder 6/28/07.
1 Spring 2003 Prof. Tim Warburton MA557/MA578/CS557 Lecture 5a.
Numerical Methods for Partial Differential Equations CAAM 452 Spring 2005 Lecture 3 AB2,AB3, Stability, Accuracy Instructor: Tim Warburton.
Lecture 3.
MA/CS 375 Fall MA/CS 375 Fall 2002 Lecture 31.
Integration of 3-body encounter. Figure taken from
Numerical Methods for Partial Differential Equations CAAM 452 Spring 2005 Lecture 5 Summary of convergence checks for LMM and one-step time stepping Instructor:
1 Spring 2003 Prof. Tim Warburton MA557/MA578/CS557 Lecture 8.
6. Introduction to Spectral method. Finite difference method – approximate a function locally using lower order interpolating polynomials. Spectral method.
7. Introduction to the numerical integration of PDE. As an example, we consider the following PDE with one variable; Finite difference method is one of.
Physics 430: Lecture 25 Coupled Oscillations
MA/CS 375 Fall MA/CS 375 Fall 2002 Lecture 12.
Engineering Analysis – Computational Fluid Dynamics –
16.920J/SMA 5212 Numerical Methods for PDEs Lecture 5 Finite Differences: Parabolic Problems B. C. Khoo Thanks to Franklin Tan SMA-HPC ©2002 NUS.
1 Spring 2003 Prof. Tim Warburton MA557/MA578/CS557 Lecture 24.
Numerical Methods for Partial Differential Equations CAAM 452 Spring 2005 Lecture 12 Instructor: Tim Warburton.
5. Integration method for Hamiltonian system. In many of formulas (e.g. the classical RK4), the errors in conserved quantities (energy, angular momentum)
ECE 576 – Power System Dynamics and Stability Prof. Tom Overbye Dept. of Electrical and Computer Engineering University of Illinois at Urbana-Champaign.
Topics 1 Specific topics to be covered are: Discrete-time signals Z-transforms Sampling and reconstruction Aliasing and anti-aliasing filters Sampled-data.
Lecture 39 Numerical Analysis. Chapter 7 Ordinary Differential Equations.
The story so far… ATM 562 Fovell Fall, Convergence We will deploy finite difference (FD) approximations to our model partial differential equations.
Lecture 2. Finite Difference (FD) Methods Conservation Law: 1D Linear Differential Equation:
Copyright © The McGraw-Hill Companies, Inc. Permission required for reproduction or display. 1 Part 6 - Chapters 22 and 23.
1 Spring 2003 Prof. Tim Warburton MA557/MA578/CS557 Lecture 25.
1 Spring 2003 Prof. Tim Warburton MA557/MA578/CS557 Lecture 32.
1 Spring 2003 Prof. Tim Warburton MA557/MA578/CS557 Lecture 28.
1 EEE 431 Computational Methods in Electrodynamics Lecture 13 By Dr. Rasime Uyguroglu
Numerical Solutions to the Diffusion Equation
Lecture 11 Alessandra Nardi
Chapter 4 Dynamical Behavior of Processes Homework 6 Construct an s-Function model of the interacting tank-in-series system and compare its simulation.
PDE Methods for Image Restoration
Numerical Methods for Partial Differential Equations
EEE 431 Computational Methods in Electrodynamics
Chapter 4 Dynamical Behavior of Processes Homework 6 Construct an s-Function model of the interacting tank-in-series system and compare its simulation.
Lecture 4: Numerical Stability
Introduction to Numerical Methods I
Convergence in Computational Science
Class Notes 18: Numerical Methods (1/2)
Homework 3 Q1) Build a finite-difference solver for: with initial condition: Q1a) use Heun’s RK3 time integrator Q1b) use the 4th order central difference.
6th Lecture : Numerical Methods
Presentation transcript:

Numerical Methods for Partial Differential Equations CAAM 452 Spring 2005 Lecture 7 Instructor: Tim Warburton

Summary of Small Theta Analysis The dominant remainder term in this analysis relates to a commonly used, physically motivated description of the shortfall of the method: Dissipative Unstable Dispersive CAAM 452 Spring 2005

Dissipation in Action Consider the right difference version: We are going to experimentally determine how much dissipation the solution experiences. CAAM 452 Spring 2005

Testing Methodology Reduce the time-stepping error to a secondary effect by choosing a 4th order Runge-Kutta (JST) time-integrator and a small dt. Fix the method (choose one of the difference operators for the spatial derivative) Do a parameter study in M, i.e. we ask the questions: how does increasing the number of data points change the error. We need to understand what questions we are asking: Is the computer code stable (as predicted by the theory)? Does the computer code converge (as predicted by the theory)? What order of accuracy in M do we achieve? We hypothesize that if the theory holds then we should achieve: What is the actual approximate p achieved? CAAM 452 Spring 2005

Initial Condition Because we do not wish to introduce uncertainty over the source of errors in the computation we use an initial condition which is infinitely smooth. CAAM 452 Spring 2005

Computing Approximate Order of Accuracy We compute the error by: We will compute the approximate order of accuracy by assuming: where C is independent of M If we measure the error for two different M then we can eliminate the constant: In the case when M2=2M1 CAAM 452 Spring 2005

After 4 Periods The numerical pulses are in the right place but have severely diminished amplitude. CAAM 452 Spring 2005

After 4 Periods M Maximum error at t=8 ~order 20 0.69149169495179 40 0.69137534363447 80 0.68329408124856 0.01 160 0.63163960492185 0.11 320 0.52742560650408 0.26 After 4 periods the solution is totally flattened in all but the last 2 results. If we bothered to keep increasing M we would eventually see the error decline as 1/M CAAM 452 Spring 2005

The Unstable Left Stencil I repeated this with everything the same, except the choice of instead of We clearly see that there is initial growth near the pulses, but eventually the dominant feature is the highly oscillatory and explosive growth (large m in the above red term). CAAM 452 Spring 2005

Cont (snapshot) CAAM 452 Spring 2005

Dispersion in Action Consider the central difference version: With the same time integrator before, M=100,… We note that the remainder terms are dispersive corrections – i.e. they indicate that modes of different frequency will travel with different speed. Furthermore, to leading order accuracy the higher order modes will travel more and more slowly as m increases. CAAM 452 Spring 2005

2nd Order Central Difference After 4 Periods We notice that the humps start to shed rear oscillations as the higher frequency Fourier components lag behind the lower frequency Fourier components. CAAM 452 Spring 2005

Convergence Study (t=8) What should we use as an error Indicator ?? CAAM 452 Spring 2005

2nd Order Central Difference M Maximum error at t=8 ~order 20 0.94240801474506 40 0.35029899850733 1.42 80 0.38298512776859 0.13 160 0.19118127466878 1.00 320 0.05471730524482 1.80 We do not see a convincing 2nd order accuracy I computed this by log2(error_M/error{M-1}) CAAM 452 Spring 2005

4th Order Central Differencing CAAM 452 Spring 2005

4th Order Central Difference M Maximum error at t=8 ~order 20 0.44353878542277 40 0.12852952972116 1.79 80 0.02040190886767 2.66 160 0.00134076478259 3.93 320 0.00008319438414 4.01 We see pretty convincing 4th order accuracy I computed this by log2(error_M/error{M-1}) CAAM 452 Spring 2005

6th Order Central Difference M Maximum error at t=8 ~order 20 0.19409575615520 40 0.04953652526983 1.97 80 0.00169299155603 4.87 160 0.00002891724105 5.87 320 0.00000046041415 5.97 We see pretty convincing 6th order accuracy I computed this by log2(error_M/error{M-1}) CAAM 452 Spring 2005

Summary of Testing Procedure Understand what you want to test Keep as many parameters fixed as possible If possible, perform an analysis before hand Run parameter sweeps to determine if code agrees with analysis. Estimate error scaling with a single parameter if possible. It should be apparent here that the errors computed in the simulations only approximate the tidy power law (dx^p) in the limit of small dx (large M). It is quite common to refer to the range of M before the error scaling applied as the “preasymptotic convergence range”. The preasymptotic range is due to the other factors in the error estimate which are much larger than the dx^p term (i.e. the p’th derivatives may be very large, or the constant may be large…) In this case we heuristically set dt small, using a high-order Runge-Kutta time integrator, and then performed a parameter sweep on M. We could have been unlucky and the time-stepping errors may have been dominant which would mask the dx behavior. To be careful we would try this experiment with even smaller dt and check if the scalings still hold. CAAM 452 Spring 2005

Summary of our Approach to Designing Finite Difference Methods We have systematically created finite difference methods by separating the treatment of space and time derivatives. Then designing a solver consists of choosing/testing: a time integrator (so far we covered Euler-Forward, LeapFrog, Adams-Bashford(2,3,4), Runge-Kutta) a discretization for spatial derivatives a discrete differential operator which has all eigenvalues in the left half of the complex plane (assuming the PDE only admits non-growing solutions). Possibly using Gerschgorin’s theorem to localize the eigenvalues of the discrete differentiation operator dt so that dt*largest eigenvalue (by magnitude) of the derivative operator sits inside the stability region  stability. (small dx) spatial truncation analysis  consistency and accuracy. Fourier analysis for classification of the differential operator. Writing code and testing. CAAM 452 Spring 2005

Some Classical Finite Difference Methods However, there are a number of classical methods which we have not discussed and do not quite fit into this framework. We include these for completeness.. CAAM 452 Spring 2005

Leap Frog (space and time) We use centered differencing for both space and time. We know that the leap frog time stepping method is only stable for operators with eigenvalues in the range: However, we also know that the centered difference derivative matrix is a skew symmetric matrix with eigenvalues: So we are left with a condition: i.e. CAAM 452 Spring 2005

cont We can perform a full truncation analysis (in space and time): We know that dt <= dx so CAAM 452 Spring 2005

Lax-Friedrichs Method We immediately conclude that the following method is not stable: Because the Euler-Forward time integrator only admits one stable point (the origin) on the imaginary axis, but the central differencing matrix has all eigenvalues on the imaginary axis. However, we can stabilize this formula by replacing the second term in the time-stepping formula: CAAM 452 Spring 2005

cont This formula does not quite fit into our constructive process (method of lines approach). We have admitted spatial averaging into the discretization of the time derivative. We can rearrange this: CAAM 452 Spring 2005

cont We can immediately determine that this is a stable method as long as c*dt/dx <=1 Given, this condition we observe that the time updated solution is always bounded between the values of the left and right neighbor at the previous time because this is an interpolation formula. CAAM 452 Spring 2005

cont A formal stability analysis would involve: Which gives stability for each mode if: CAAM 452 Spring 2005

cont Thus considering all the possible modes We note that the middle mode requires: This condition gives a sufficient condition for all modes to be bounded. By the invertibility and boundedness of the Fourier transform we conclude that the original equation is stable. CAAM 452 Spring 2005

cont We can recast the Lax-Friedrichs method again This method consists of Euler Forward, central differencing for the space derivative and an extra dissipative term (i.e. a grid dependent advection diffusion equation: ) CAAM 452 Spring 2005

CFL Number The ratio appears repeatedly, in particular in the estimates for the maximum possible time step. We refer to this quantity as the CFL (Courant-Friedrichs-Lewy) number. Bounds of the form: which result from a stability analysis are frequently referred to as CFL conditions. CAAM 452 Spring 2005

Lax-Wendroff Method We are fairly free to choose the parameter in the stabilizing term: The artificial viscosity term acts to shift the eigenvalues of the spatial operator into the left half-plane. Recall – the Euler-Forward stability region is the unit circle centered at -1 in the complex plane. So pushing the eigenvalues off the imaginary axis allows us to choose a dt small enough to push the eigenvalues of the discrete spatial operator into the stability region… CAAM 452 Spring 2005

Class Exercise Please provide a CFL condition for: In terms of: What is the truncation error? CAAM 452 Spring 2005

Von Neumann Analysis By stealth I have introduced the classical Von Neumann analysis. The first step is to Fourier transform the finite difference equation in space. A short cut is to make the following substitutions: CAAM 452 Spring 2005

Analysis cont We can also make the substitutions for the difference operators: CAAM 452 Spring 2005

Example The Lax-Wendroff scheme becomes CAAM 452 Spring 2005

Example cont So the Lax-Wendroff scheme becomes a neat one step method for each Fourier coefficient: We have neatly decoupled the Fourier coefficients so we are back to solving a recurrence relation in n for each m: Hence we need to examine the roots of the stability polynomial for the above CAAM 452 Spring 2005

Example cont This case is trivial as we just need to bound the single root by 1 Plot it this as a function of theta… CAAM 452 Spring 2005

Final Von Neuman Shortcut We can skip lots of steps by making the direct substitution: Here g is referred to as the amplification factor and i*m*theta is our previous theta_m CAAM 452 Spring 2005

Next Time Definition of consistency Lax-Richtmyer equivalence theory Treating higher order derivatives CAAM 452 Spring 2005

Homework 3 Q1) Build a finite-difference solver for Q1a) use your Cash-Karp Runge-Kutta time integrator from HW2 for time stepping Q1b) use the 4th order central difference in space (periodic domain) Q1c) perform a stability analysis for the time-stepping (based on visual inspection of the CK R-K stability region containing the imaginary axis) Q1d) bound the spectral radius of the spatial operator Q1e) choose a dt well in the stability region Q1f) perform four runs with initial condition (use M=20,40,80,160) and compute maximum error at t=8 Q1g) estimate the accuracy order of the solution. Q1h) extra credit: perform adaptive time-stepping to keep the local truncation error from time stepping bounded by a tolerance. CAAM 452 Spring 2005