2.092/2.

093 Finite Element Analysis of Solids & Fluids I

Fall 09

Lecture 23 - Solution of K = M
Prof. K. J. Bathe Reading assignment: Chapters 10, 11 We have the solutions 0 < 1 2 . . . n . Recall that:
1 2 n

MIT OpenCourseWare

Ki = i M i

(1)

In summary, a necessary and sucient condition for i is that Eq. (1) is satised. The orthogonality conditions are not sucient, unless q = n. In other words, vectors exist which are K - and M -orthogonal, but are not eigenvectors of the problem. (2) = 1 . . . n zeros 1 T T . .. (3) M = I ; K = = zeros Assume we have an n q matrix P which gives us P T MP = I Is aii necessarily equal to i ? zeros a11 a22 .. . zeros
q q

; P T KP = A diagonal matrix
q q

If q = n, then A = , P = with some need for rearranging. If q < n, then P may contain eigenvectors (but not necessarily), and A may contain eigenvalues.

Rayleigh-Ritz Method
This method is used to calculate approximate eigenvalues and eigenvectors. (v ) = v T Kv vT M v

1 (v ) n 1 is the lowest eigenvalue, and n is the highest eigenvalue of the system. 1 is related to the least strain energy that can be stored with v T M v = 1: T 1 K1 = 1 (if T 1 M 1 = 1)

Lecture 23

Solution of K = M

2.092/2.093, Fall 09

Note that twice the strain energy is obtained when the system is subjected to 1 . If the second pick for v gives a smaller value of (v ), then the second pick is a better approximation to 1 .
q

Assume = i xi , and the Ritz vectors i are linearly independent. Also, = [ 1 . . . q ]. The xi will
i=1

be selected to minimize (). Hence, calculate

xi ()

= 0. (See Chapter 10.) The result is (4) (5)

x = M Kx = T K K = T M ; M

We solve Eq. (4) to obtain 1 , 2 , . . . , q and x1 , x2 , . . . , xq . Then our approximation to 1 , . . . , q is given by 1 , . . . , q . 1 1 ; 2 2 ; q q 1 1 where 1 . . . q = [ x1 . . . xq ].
nq nq q q

; 2 2

; etc.

If the q Ritz vectors span the subspace given by 1 , . . . , q , then we obtain (1 . . . q ) and (1 . . . q ). Pictorially, an example:

If 1 and 2 are in the x-y plane, then by the Rayleigh-Ritz analysis we get 1 , 2 . Major shortcoming: in general, we do not know the accuracy of (i , i ).

The Subspace Iteration Method

Pick X1 , then calculate for k = 1, 2, 3, . . .
nq

KX k+1 = M Xk This is inverse iteration with q vectors. Now perform the Rayleigh-Ritz solution: Kk+1 = X k+1 KX k+1
T

(a)

; Mk+1 = X k+1 M X k+1

(b) (c)

Kk+1 Qk+1 = Mk+1 Qk+1 k+1

Kk+1 , Mk+1 , and Qk+1 have dimensions q q . Recall that we have K = M from Eq. (1). We then have QT ; QT (d) k+1 Kk+1 Qk+1 = k+1 k+1 Mk+1 Qk+1 = I Finally, Xk+1 = X k+1 Qk+1 Equations (b), (c), and (e) correspond to the use of the Rayleigh-Ritz method. 2 (e)

Lecture 23

Solution of K = M

2.092/2.093, Fall 09

Then, provided the vectors in X1 are not M -orthogonal to the eigenvectors we seek, we have (with good ordering) that 1 .. k+1 . q Xk+1 1 . . . q In practice, we use q vectors to calculate the p lowest eigenvalues, with (say) q = 2p. In fact, the convergence i rate of the vectors is given by . q +1 If p = 2 and we have a multiplicity of 5 (or higher), q = 2p corresponds to not enough vectors. Ideally, we i want q+1 to be signicantly larger than p , so that is much less than 1 for i = 1, . . . , p. The quite q +1 conservative way is to use q = max(2p, p + 8) The textbook gives q = min(2p, p+8), which can also be used (apply the Sturm sequence check, see textbook); it will use less storage, but will generally need more iterations. For modern computers (specically with parallel processing), the above formula for q is frequently more eective.
T Notice that Xk +1 M Xk+1 = I because from (e),

QT k+1 X k+1 M X k+1 Qk+1 = I

Mk+1

MIT OpenCourseWare http://ocw.mit.edu

2.092 / 2.093 Finite Element Analysis of Solids and Fluids I

Fall 2009

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.