multistep methods Last modified: November 28, 2017 Recall that we are interested in the numerical solution of the initial value problem (IVP):

Size: px

Start display at page:

Download "multistep methods Last modified: November 28, 2017 Recall that we are interested in the numerical solution of the initial value problem (IVP):"

Sheila Jacobs
5 years ago
Views:

1 MATH 351 Fall 217 multistep methods rozman/courses/m351_17f/ Last modified: November 28, 217 Recall that we are interested in the numerical solution of the initial value problem (IVP): dy dt = f (t,y), y() = y. (1) We denote the time at the nth time step by, the computed solution at the nth time step by y n, y n = y( ), and the value of the right hand side of Eq. (1) at the nth time step by f n, f n = f (,y n ). The step size h (assumed to be constant for the sake of simplicity) is given by h = 1. So far all of the numerical methods that we have developed for solving initial value problems are one-step methods they only use the information about the solution at time to approximate the solution at time +1 (possibly taking intermediate steps but then discarding this additional information). Multistep methods attempt to gain efficiency by keeping and using the information from the previous steps. A method is called linear multistep method if a linear combination of the values of the computed solution and possibly its derivative in the previous points are used. An example of multistep method is the following expression y n+1 = y n + h s β i f n+1 i, (2) where s is the number of steps in the method. If β =, the multistep method is said to be explicit, because then y n+1 can be described using an explicit formula. If β, the method i= Page 1 of 6

2 MATH 351 Multistep methods Fall 217 is said to be implicit, because then an equation, generally nonlinear, must be solved to compute y n+1. The constants β i satisfy the constraint s β i = 1 (3) i= originating from the requirement that if f (t,y) = α = const, the correct solution y(t) = αt+c is produced by the numerical scheme. A broad category of multistep methods that are called Adams methods involve the integral form of Eq. (1): y(+1 ) = y( ) + f (τ,y(τ))dτ. (4) The general idea behind Adams methods is to approximate the above integral using polynomial interpolation of f at the points +1 s, +2 s,..., if the method is explicit, and +1 as well if the method is implicit. Explicit Adams methods are called Adams-Bashforth methods. Implicit Adams methods are known as Adams-Moulton methods. 1 Adams-Bashforth methods To derive the integration formula for Adams-Bashforth method, we interpolate f at the points +1 s, +2 s,..., with a polynomial of the degree s 1. We then integrate this polynomial exactly. Let s derive the three-step Adams-Bashforth method with the local truncation error O(h 3 ). The interpolating polynomial for f (τ,y(τ)) passing through three points ( 2,f n 2 ), ( 1,f n 1 ), and (,f n ) is as following: P 3 (τ) = f n 2 L n 2 (τ) + f n 1 L n 1 (τ) + f n L n (τ), (5) Page 2 of 6

3 MATH 351 Multistep methods Fall 217 where the Lagrange polynomials are L n 2 (τ) = L n 1 (τ) = (τ 1 )(τ ) ( 2 1 )( 2 ) = 1 2h 2 (τ 1)(τ ), (τ 2 )(τ ) ( 1 2 )( 1 ) = 1 h 2 (τ 2)(τ ), (6) L n (τ) = (τ 2)(τ 1 ) ( 2 )( 1 ) = 1 2h 2 (τ 2)(τ 1 ). Substituting Eq. (5) into Eq. (4), we obtain: y n+1 = y n + = y n + f n 2 P 3 (τ)dτ = L n 2 (τ)dτ + f n 1 L n 1 (τ)dτ + f n L n (τ)dτ. The integrals in Eq. (7) can be evaluated by introducing a new integration variable such that u = τ, u 1, (7) h τ = + hu, dτ = hdu, (8) τ +1 = h + hu τ = hu, τ 1 = h + hu, τ 2 = 2h + hu. (9) Lagrange polynomials as the functions of u are, L n 2 (u) = 1 u(u + 1), 2 L n 1 (u) = u(u + 2), (1) L n (u) = 1 (u + 1)(u + 2). 2 Page 3 of 6

4 MATH 351 Multistep methods Fall 217 Integrating, we obtain L n 2 (τ)dτ = h L n 1 (τ)dτ = h L n (τ)dτ = h L n 2 (u)du = h 2 L n 1 (u)du = h L n (u)du = h 2 We conclude that the three-step Adams-Bashforth method is u(u + 1)du = 5 12 h, u(u + 2)du = 4 h, (11) 3 (u + 1)(u + 2)du = h. y n+1 = y n + h 12 (5f n 2 16f n f n ). (12) Note that the coefficients in Eq. (18) satisfy the constraint Eq. (3). 2 Adams-Moulton methods The same approach can be used to derive the integration formulas for implicit Adams- Moulton methods. The resulting interpolating polynomial is of degree one greater than in the explicit case, so the error in an s-step Adams-Moulton method is O(h s+1 ), as opposed to O(h s ) for an s-step Adams-Bashforth method. The interpolating polynomial through ( 2,f n 2 ), ( 1,f n 1 ), (,f n ), and (+1,f n+1 ) is as following: P 4 (τ) = f n 2 L n 2 (τ) + f n 1 L n 1 (τ) + f n L n (τ) + f n+1 L n+1 (τ), (13) where the Lagrange polynomials are (τ t L n 2 (τ) = n 1 )(τ )(τ +1 ) ( 2 1 )( 2 )( 2 +1 ) = 1 6h 3 (τ 1)(τ )(τ +1 ), (τ t L n 1 (τ) = n 2 )(τ )(τ +1 ) ( 1 2 )( 1 )( 1 +1 ) = 1 2h 3 (τ 2)(τ )(τ +1 ), (14) L n (τ) = (τ 2)(τ 1 )(τ +1 ) = 1 ( 2 )( 1 )( +1 ) 2h 3 (τ 2)(τ 1 )(τ +1 ), (τ t L n+1 (τ) = n 2 )(τ 1 )(τ ) (+1 2 )(+1 1 )(+1 ) = 1 6h 3 (τ 2)(τ 1 )(τ ). Page 4 of 6

5 MATH 351 Multistep methods Fall 217 y n+1 = y n + = y n + f n 2 P 4 (τ)dτ = (15) L n 2 (τ)dτ + f n 1 L n 1 (τ)dτ + f n L n (τ)dτ + f n+1 Lagrange polynomials as the functions of the variable u Eq. (7) are, Integrating, we obtain L n 2 (τ)dτ = h L n 1 (τ)dτ = h L n (τ)dτ. = h L n+1 (τ)dτ. = h L n 2 (u) = 1 6 u(1 u2 ), L n+1 (τ)dτ. L n 1 (u) = 1 u(u + 2)(1 u), 2 (16) L n (u) = 1 2 (u + 2)(1 u2 ), L n+1 (u) = 1 u(u + 1)(u + 2). 6 L n 2 (u)du = h 6 L n 1 (u)du = h 2 L n (u)du = h 2 L n+1 (u)du = h 6 We conclude that the three-step Adams-Moulton method is u(1 u 2 )du = 1 24 h, u(u + 2)(1 u)du = 5 h, (17) 24 (u + 2)(1 u 2 )du = h. u(u + 1)(u + 2)du = 3 8 h. y n+1 = y n + h 24 (f n 2 5f n f n + 9f n+1 ). (18) Note that the coefficients in Eq. (18) satisfy the constraint Eq. (3). Page 5 of 6

6 MATH 351 Multistep methods Fall Predictor-corrector method An Adams-Moulton method can be impractical because, being implicit, it requires solving nonlinear equations during every time step. An alternative is to pair an Adams-Bashforth method with an Adams-Moulton method to obtain an Adams-Moulton predictor-corrector method. Such a method proceeds as follows: Predict: use the Adams-Bashforth method to compute a first approximation to y n+1, which we denote by ŷ n+1. Evaluate: evaluate f (+1,ŷ n+1 ). Correct: use the Adams-Moulton method to compute y n+1, but instead of solving an equation, use f (+1,ŷ n+1 ) in place of f (+1,y n+1 ) so that the Adams-Moulton method can be used as if it was an explicit method. Evaluate: evaluate f (+1,y n+1 ) to use during the next time step. 4 Conclusions A drawback of multistep methods is that because they rely on values of the solution from previous time steps, they cannot be used during the first time steps. Therefore, it is necessary to use a one-step method, with the same order of accuracy, to compute enough starting values of the solution to be able to use the multistep method. For example, to use the three-step Adams-Bashforth method, it is necessary to first use a one-step method such as the fourth-order Runge-Kutta method to compute y 1 and y 2, and then the Adams-Bashforth method can be used to compute y 3 using y 2, y 1 and y. Page 6 of 6

Jim Lambers MAT 772 Fall Semester Lecture 21 Notes

Jim Lambers MAT 772 Fall Semester 21-11 Lecture 21 Notes These notes correspond to Sections 12.6, 12.7 and 12.8 in the text. Multistep Methods All of the numerical methods that we have developed for solving