Stability and Convergence theorems for Newmarks method

Fumihiro Chiba February 7, 2012

Abstract For the second order time evolution equation with a general dissipation term, we introduce a recurrence relation of Newmarks method. Deriving an energy inequality from this relation, we obtain the stability and the convergence theories of Newmarks method. Detail of proofs for theorems and their application are stated in Chiba-Kako[2].

keywords convergence, stability, newmark, dissipation term, recurrence relation, energy inequality, second order time evolution equation

Introduction of Newmarks method

We recall the basic ideas of Newmarks method[7],[11] for the second order time evolution equation in the d-dimensional complex Euclidean space Cd . Let M , C and K be dd Hermitian matrices on Cd which are constant in time, and let f (t) be a given Cd -valued function on [0, ). We use the usual Hermitian scalar product (, ) and the corresponding norm in Cd . Throughout this chapter, we assume that M is positive denite: (M x, x) > 0 for x Cd with x = 0, which is equivalent to the condition: (M x, x) m(x, x) for x Cd , (2) (1)

with a positive constant m. The identity matrix on Cd is denoted by I. We consider the approximation method for the following initial value problem of the second order time evolution equation for an Cd -valued function u(t): M
d2 u (t) dt2

+ C du (t) + Ku(t) = f (t), dt

u(0) = u0 ,

du (0) dt

= v0 .

(3)

Let an , vn and un be approximations of ( d2 / dt2 )u(t), ( d/ dt)u(t) and u(t) respectively at t = n with a positive time step and an integer n, and let fn = f ( n). Then Newmarks method for (3) is dened through the following relations: M an + Cvn + Kun = fn , un+1 = un + vn + 1 2 an + 2 (an+1 an ), (4) 2 vn+1 = vn + an + (an+1 an ).

chibaf@mac.com, http://math.digi2.jp/

The rst relation corresponds to the equation (3), and the second and the third relations correspond to the Taylor expansions of u(t + ) and v(t + ) at t = n respectively. Here, and are positive tuning parameters of the method. An interpretation of the parameter related to the acceleration ( d2 / dt2 )u(t) can be seen in [7]: The case = 1/6 corresponds to the approximation where the acceleration is a linear function of t in each time interval; the case = 1/4 corresponds to the approximation of the acceleration to be a constant function during the time interval which is equal to the mean value of the initial and nal values of acceleration; the case = 1/8 corresponds to the case where the acceleration is a step function with the initial value for the rst half of each time interval and the nal value for the second half of the interval. It is often the case that is equal to 1/2.

Iteration scheme of Newmarks method

Based on the formulas in (4), Newmarks method generates the approximation sequence un , n = 0, 1, 2, . . . , N by the following iteration scheme: Step 1. For n = 0, compute a0 from the initial data u0 and v0 : a0 = M 1 (f0 Cv0 Ku0 ). Step 2. Compute an+1 from fn+1 , un , vn and an solving a linear equation: an+1 = (M + C + 2 K)1 { } 1 2 Kun (C + K)vn + {( 1) C + ( ) K}an + fn+1 . 2 Step 3. Compute un+1 from un , vn , an and an+1 : un+1 = un + vn + ( 1 ) 2 an + 2 an+1 . 2 Step 4. Compute vn+1 from vn , an and an+1 : vn+1 = vn + (1 ) an + an+1 . Step 5. Replace n by n + 1, and return to Step 2. Here, Step 1 is nothing but the rst relation of (4) with n = 0. The expression of an+1 in Step 2 is obtained by eliminating un+1 and vn+1 from the second and the third equalities in (4) together with the rst equality in (4) with n replaced by n + 1: M an+1 + Cvn+1 + Kun+1 = fn+1 . Step 3 and Step 4 are from the second and the third relations of (4).

Recurrence relation of Newmarks method

Newmarks method (4) for the second order equation (3) is reformulated as follows in the recurrence relation([1] and [10]). Eliminating vn and an from (4) by a series of lengthy computations, we have (M + 2 K)D un + CD un + {(1 )C + ( 1 )K}D un + Kun 2 1 2 = {I + ( 2 )D + D }fn , 2 (5)

where

D un = (un+1 un )/, D un = (un un1 )/, D un = (D un D un )/.

We conrmed this result by a formula manipulation software NCAlgebra [5] on Mathematica[9]. Especially, in the case = 1/2, we have: (M + 2 K)D un + 1 C(D + D )un + Kun = (I + 2 D )fn . 2 (6)

These recurrence relations are useful for the stability and error analyses of Newmarks method. See [6] and [8] for the case with C 0.

Derivation of an energy inequality

Taking a scalar product of (5) and (D + D )un , we can derive an energy inequality for New marks method. From this inequality, we obtain the stability conditions for Newmarks method. From now on, let T be a xed positive number and N be a positive integer, and dene := T /N and n = 0, 1, 2, . . . , N . Let C be nonnegative: (Cx, x) 0, K be positive denite: (Kx, x) k(x, x) with k > 0, and m > 0 be the smallest eigenvalue of M . We also assume that 1 . We have the following theorem for the energy inequality. 2 Theorem 4.1 Let {un }N be a sequence generated by the scheme in Section 2. Then for n=0 a positive constant 0 determined as below, there exists a positive constant C0 such that the inequality: M 1/2 D un 2 + 2 { 1 ( 1 ) 1 }K 1/2 D un 2 + (1 2 2 2 , 2
1 )K 1/2 un 2 2

C0 ,

(7)

holds for all 0 and an arbitrary positive constant , where 0 is determined either as 0 > 0 when or as 0 < 0 < (1 2 (8)

m when 0 < . 2 )K

(9)

Remark 4.1 The constant C0 is dened as follows: )1 [ ( { }2 ] T 1 w0 + 2 (4 + 2) sup0tT f (t) C0 := 1 2 0 ( ( )1 ) 1 T , exp 1 0 2 where { ( )} w0 = M 1/2 D u0 2 + 2 1 1 K 1/2 D u0 2 2 2 1 +Re{ (KD u0 , u0 )} + K 1/2 u0 2 + ( 2 )C 1/2 D u0 2 , 0 < m, in the case (8), or
2 0 < m 0

(10)

(11)

and in (11) is dened as: (12) (13) (1

2

) K, in the case (9). 3

Remark 4.2 When C = 0, = 1/2 and f (t) 0, the following equality holds. ((M + 2 K)D un , D un ) + (Kun+1 , un ) = ((M + 2 K)D u0 , D u0 ) + (Ku1 , u0 ). Remark 4.3 The constant C0 in the theorem depends primarily on the coecient matrices M , C and K, and secondarily on the tuning parameters and and it also depends on the initial values u0 and v0 and the inhomogeneous term f , and nally on 0 in the way described above. Remark 4.4 In the case C is negative or K is not positive, we consider v(t) := et u(t) instead of u(t). By choice of appropriate > 0, we may obtain new matrices satisfying the condition in Theorem 4.1 as coecients of equation of v(t).

Stability and convergence conditions for Newmarks method

In this section, using the energy inequality (7) we derive stability conditions for Newmarks method. With respect to a parameter , we divide the stability condition into two cases and lead to the next theorem.

5.1

Stability theorem

Theorem 5.1 Let M and K be positive denite and C be nonnegative. Let m and k be the smallest eigenvalues of M and K. Assume 1/2. Then, Newmarks method for (3) in a time interval [0, T ] is stable in the following two cases with respect to : Case 1: If 1 < , then with 0 > 0 and given in (12) and (13) we have, for < 0 and 2 N = T / , (14) un (42+1) C0 , n = 0, 1, 2, . . . , N. (42)k Case 2: If 0 1 , then for 0 satisfying 2 0 < 0 <

m , ( 1 )K 2

(15)

we have for 0 and N = T / un u0 + m 2 ( 1C0 n, )K

0 2

n = 0, 1, 2, . . . , N.

(16)

Here, in both cases, C0 is given by (10) together with (11). Remark 5.1 H. Fujii investigated in [3] and [4] the stability condition for the Rayleigh damping case with C = aK + bM, a, b R.

5.2

Convergence theorem

Using the recurrence relation (5) and the stability theorem, we can show in this section the convergence of Newmarks method together with its convergence order. Let u(t) be the solution of (3) and un (n = 0, 1, 2, . . . , N ) be the solution of Newmarks method for (3). We assume that M and K are positive denite and C is nonnegative. The discretization error en is dened as en := u( n) un . Then we have the following theorem. 4

Theorem 5.2 We assume that f C 2 ([0, T ]), then we have the estimate en = O( l ), where l = 2 for = 1 , 2 l = 1 for > 1 . 2

5.3

Energy inequality and discrete energy

Related to an energy inequality, we dene the following quantity.: ( ) wn := (M D un , D un ) + 1 (CD un , D un ) 2 { ( )} 1 1 2 +Re{ (KD un , D un )} + Re{(Kun+1 , un )} 2 2 ( ) ( ( )) 1 1 1 2 = ((M + C + K)D un , D un ) 2 2 2 +Re{(Kun+1 , un )} = (MD un , D un ) + Re{(Kun+1 , un )}, where ( ) ( )) ( 1 M := M + 1 C + 2 1 2 K M. 2 2
wn := (M D un , D un ) + (Kun , un ),

(17)

We call wn a discrete energy. We also dene other quantities wn and wn :

(18)

and
wn := (M vn , vn ) + (Kun , un ),

(19)

where vn is a velocity given in Newmarks method together with un . These are discrete analogues of the continuous energy: ) ( d d (20) E(t) := M dt u(t), dt u(t) + (Ku(t), u(t)). The dierence quotient D un approximates the derivative of u at t = n + /2 with the second order in . So the average of (MD un , D un ) and (MD un1 , D un1 ) is a second d d order approximation of (M dt u, dt u) at t = n. The average of (Kun+1 , un ) and (Kun , un1 ) is represented as follows:
1 {(Kun+1 , un ) 2

+ (Kun , un1 )} = (Kun , 1 {un+1 + un1 }), 2

1 where we use the assumption that K is Hermitian. Since 2 {un+1 + un1 } is a second order 1 approximation in of un at t = n, 2 {(Kun+1 , un ) + (Kun , un1 )} is a second order approximation of (Ku, u) at t = n. Now we dene the following quantity Wn as a second order approximation in of (20) at t = n: { w0 for n = 0, Wn := 1 (wn + wn1 ) for n 1. 2

5.4

Convergence theorem of Wn

Theorem 5.3 For a xed T > 0, the convergence order of E(T )WN with T = N is estimated as follows: { 1 O( 2 ) for = 2 , |E(T ) WN | = 1 O( ) for > 2 . 5

References
[1] Chiba, F., and Kako, T. On the stability of newmarks method. Research Institute of Mathematical Sciences, Kyoto University 1040 (1998), 3944. [2] Chiba, F., and Kako, T. Newmarks method and discrete energy applied to resistive mhd equation. Vietnam Journal of Mathematics 30, SI (2002), 501520. [3] Fujii, H. Finite-Element Galerkin Method for Mixed Initial-Boundary Value Problems in Elasticity Theory. PhD thesis, Center for Numerical Analysis, The University of Texas, Austin, 1971. [4] Fujii, H. A note on nite element approximation of evolution equations. RIMS Kkyroku o u 202 (1974), 96117. [5] Helton, J. W., and Miller, R. L. NCAlgebra. Department of Mathematics, University of California San Diego, http://math.ucsd.edu/ncalg/, 2005. [6] Matsuki, M., and Ushijima, T. Error estimation of newmark method for conservative second order linear evolution. Proc. Japan Acad Ser. A. 69 (1993), 219223. [7] Newmark, N. M. A method of computation for structural dynamics. Proceedings of the American Society of Civil Engineers, Journal of the Engineering Mechanics Division 85, EM 3 (1959), 6794. [8] Raviart, P. A., and Thomas, J. M. Introduction ` lAnalyse Numrique des Equations a e aux Drives Partielles. Masson, Paris, 1983. e e [9] Wolfram, S. The Mathematica Book, fth ed. Wolfram Media Inc, Illinois, 2003. [10] Wood, W. L. A further look at newmark, houbolt, etc., time-stepping formulate. International Journal for Numerical Methods in Engineering 20 (1984), 10091017. [11] Wood, W. L. Practical Time-Stepping Schemes. Clarendon Press, Oxford, 1990.