You are on page 1of 34
E igenval ve Problems Prepared by : Hamid Reza Ronagh April- 194 Sydney ~ Australia Eigenvalves & Eyenvectors characteristic Functions Associated with each square matrix A oP order n is a Function Bu-k a _ an PO)S[A-AT] 2 OM ak 7 san lay ant Ona Onn-d called the characteristic Function of A. The equation (2) POAY=|A-dI] ae San be expressed in the pelynemial form cl ox 4a Ty... and js called the characteristic equation of A, # Cy dt Case @) Example 1 Find the characteristic equation of matrix r A, where \3 2 8 Ae 2 2 2 ° 2 3 sol”: The characteristic equa on of ATS tes 2 ° 2 a+ 2] =0 a 2 Bek thot is, cae) + oN =a 2 (I-A) A -SA4 2) — 2 (6-2A) a0 B62 434 lo = 0 In some instances the task of expressing the characteristic equation oP amatrix in Polynomial Form may be simpliPied by introducing the concept oF the trace of a matrix. The sum of the diagonal elemenis oP a matrix A is called the trace oF A and is denoted by trCA): For example, the trace oF matrix Ain 17 Example 1 ss Ie243 5 thatis, 6 Let te trCA2, ta«trCA),...., taete(A"). Tt can be shown that the ceePPicients of the characteristic equation are given ky the equations s Coe I Qs -t, Cae tA (Crt + 42), (4) Ca. ~ a (Catiecita+ts), ene Ya Centre Crate sere Cita + tr). Equations (4) make st possible to calculate the CoefPicients of the characteristic equation ok a matrix AL by summing the diggenal elements of the matrices of the Form AY This numerical process can be easily programmed on the computers, or for small values of 9 may be computed manually cithout cliPPiculty - Example 2 Find the characteristic equation of matrix hin Example 1 by using equations C4). sol”: The characteristic equation of A is of the form Cdr Xe Cada Cyeo. NOH, tretrCA)]= 144243 -6- 3 and since 7 2: & 13 30024 2 3 Aw} & 12 10 and Ax |3. 56 St 4 to 43 at 54 59 tas tr(M)s 54124I3=30, ta. teCA) = 174564 34s 152. Then using equations cy Coals Cie 6, Cas ~ “OO 430] 23, eae Hig [(326) + 8000) + 3a] = 10. DeFinition: The n roots of the characteristic equation oR a matrix A are called the eigenvalves of A. Example 3 Determine the eigenvalues oP the matrix A ohere 3 1 As . 7 2 Sol”: Bed 1 The characteristic equation oF A is 2 2-) In expanded Form this becomes \7_5) 44-0. The eygenvalues Met and hast are the roots of the characteristic equation. In general, eigenvalues oF a real matria may be negatives positive, identical and even comples seen for the matrices shawn below It is 1 3 ° As 3 -2 - Nye 4 dael Aad — a> o 7! | -ve “Ne 3 ° Aq died daed identicod 5 3 3 (6 A: dyei¥S ga iv Complew 2 + Some theorems: Theorem: The trace of a matrix A of order nis equal to the sum of the n eqgenvalues oF AS thatis trih)e ZN. The above theorem may be used as a good check pont on the ecgjenvalues * Theorem: IF A is a real symmetric matrix, then the eigenvalues oF A ore real. The above theorem is & good news Por structural engineers, as most oF the matrices weare dealing with are real and syrmetnic . Stipfness matrix ts a. geod example fir this. Eigen vector Roy renzere column vector, denoted by Xi, such that CA -A: T) Ki =o ) is called an ejgenvector of matrix A. Ibis guaranteed that at least one exgen vector exists far each hi Since equation (5) represents a system of 7 linear homogenous equations which has a nontrivial solution Xi #2 1P and only iP [a-d.T] 3 that is, iP ancl only iP Ai is an eigenvalue ok & Fuethermore, note that any nenzero scalar maltiple of an elgenvecter associated with an eigenvalue is also an eigenvector asse ciated with that eigenvalue Example. Determine a set of evgenvecters of the matrix win Example 3. ae sol”, Associated with hel are the eigenvectors (x x2) Por which (A-D) (x x2 oo EE JET-GL It Fo lows that ~2 my aXe IF %1 iS chosen as Some Convenient arbitrary Scalar, Sayt» a becomes -2 Hence, © -2)" is an eigenvector associated with the eigenvalue 1. Similarly, associated vith Aae4 are the egenvectors Ge xa For which CAW4IDCx) xa 2 8 tet is = 1 x oO [ 2. a] [2] -[:] Hence, Xiexa, and (17 is an eigenvector asscciatedl with the ejegnvalue a a It should be noted that CK -2k) and (kK), where Kis any nenzers scalar, represent the general Forms ef the eigenvectors oF & The eigenvalues oF a matrix are also called the Proper valves, ‘the latent values, omd the characteristic yalues of the matrix and similarly the eigen vectors are called te prooer vectors and Se on Eigenvalue Finding techniques Se Far we have seen one method for calca lating the eigenvalues, thatis, Finding the roots of the charauc- tertstic equation of the matrix. In general when the order of the matrix is greater than 3, the roots of te polynomial may be only Found by iteration Fechniques- There are numerovs amouat of other techniques available which are sometimes better than Calculating the roots of the characteristic equation Bathe ['] has categorized the solution methocls in the Pallowing Four groups, depending en the basic property that i useel as the basis of the solution algorithm. 4. Vector iteration methods 2. Transformation metheds ti}- K.g. Bathe, “Finite element procedures in Eqgineering analysis’, 1825 Prentice-Hall, New Jersey 3. pelynemial iteration techniques 4. Methods based on the Sturm Sequence preperty Some of the most notable methods in each greupar, ta. Inverse iteration 12. Forward iteration [3- Rayleigh quotient iteration 24+ Jacobi methed 2.2. Generalized gJacebi method 2.3. Heuseholder- GR- Inverse iteration 24 GR iteration 34. Explicit Poly nomial iteration 3.2. Impkcit polynomial iteration P Poly AL Sturm sequence 4.2. A combined sturm sequence, and inverse iteratia 40 In general, none of these methods are superior teancther and a single algo rithm which aloes gives a very ePPi cient solution of the eigenvalue problem dees nof exist. However among these, the sturm sequence method is probably the easiest one te be understood and te be pragrammed. Sturm sequence methed This method is applicable te ay real symmetric matrix A. IF we choose atrial value Por A, say At, te calculate the matrix BeA-deT, and use a Gaussian eliminahon Technique without piveting te transform this matrix te an upper triangular matrix @°, then the number & of negatives on the diagonal ef B is equal to the number of roots of the characteristic pelynomial that are less than At. Example 5S Prove that the eigenvalues oF matrix K are between =Siand +4. { 3 ° Sol": we calculate the number of negatives on the diagonal of 8°, choosing 8 as A-AsI and de equal to -5 ond +4 6 30 with Ar=-5 Be] 3 3 ol ° -1 6 Using Gaussian elimination 6 3 ° a B=] ° HS -l ° ° 5.33 12 As is seen, there are no negatives on the diagonal. This Means that there is no eigenvalue less than -~3 a 3 oO with Ape+t By} 3-6 -t ° a 4d -3 3 ° a a - p.| ¢ 3 | ° ° -2-64 There are 3 negatives on the diagenal, which means that all & eigenvalues are less than +4. As is shown on Page 3, the eigenvalues of & are =4, 1 ond 3. This property allows Faro trial routine te be developed to Converge ta any eigenvalue required: The trial At values will be chosen successively untill convergence to the required umber of negatives en the diagenal occurs. Example 6 Calculate the smallest eigenvalue oF A, Knowing that the eigenvalues are between -Sand +t. Sel": we have to cheese At values, agood value is the average of the bounds of the ejgenvalue range Here the Pirst At is — = 0:3. Calculaticg B, aind 8° and Counting the number oF negatives on the diagenal oF B> are the next steps. For 25, there is one negative onthe diggonal of BP, showing thet there is one. eigenvalue less than -05. We also Kno+ that alll eigenvalues are gresiber than -5. So the range in which we have te search Por the eigenvalue Js limited to (-5s -5) and the next trial At can be chesen as C5#C05), 2.75 . As this process continues a we can ame up ith atable like: de range no. of negatives 05 -5 to 44 1 -2-75 -5 to -05 1 -3.895 -S te 2.75 1 HH 4B 7S -5 to -3-875 8 =4:15625 a 44395 to - 3.875 ° - toon -4.ovce to - 3.199% oft Tt is seen that with continuing the Procedure, th difference between the lower and upper bouncls of the range reduces and we can Converge te the required eigenvalue te any desired accuracy: For this problem it is seen thot the procedures converges to _+. In stability Problems, the buckling equation is different Prom the standard equation (K-41) Xe. There we have (K-AS)4er, where Kis the stiFfness matrix, Sis the stability matrix, Ais the Pactor ky which the present loads on the structures should be scaled te cause buckling ana 4 is the buckling mode corresponding to Xs This type of equation can be transPormed te the standard Format if it is premaltiplied by 3 Cs'k -AS's) Gee (4) or (A -ATD4Gae where ASK, For sturm sequence method however, we Can work with the non standard Form by introducing B as Bs K-AtS w and continve using Gaussion elimination and counting the negatives oF 8. This is an advantage as mest of the other metheds require the transformati- ». 16 envector Finding techniques se Mode Finding is the procedure that gives the eigen- vectors oF a matrix corresponding to its eigenvalues, and is needed to complete the eigen selution with those. methods that denst calculate the value and Vector simultaneously. Some oF the methods given on pages 4-10, Solve for the eigenvalue and eigenvector simultaneously » while others solve for the eigenvalue first with the eigen- mode calculated afterwards. Sturm sequence methed can only provide the eigenvalues of a matria and a complement to this method is required for the calculation of eigenvectors. To the Knowledge of the author, the best complement is the Hopper and williams [2] technique which is called the *Random Force Vector’ method. [2] ¢.T Hopper ane Fw. williams "Mode Finding in nonlinear structural eigenvalue calculations” J. struct. Mech, § 03), Q25D-2FB, 1977 Random Force Vector method. Ibis easier te explain this methed fer the eigenvalve problems that are encountered in structures. The basic concept which was the basis far the develop ment oP the Random Force Vector method was thot + when a structure is loaded upto buckling, a small perturbing Force can easily Porce it to the buckled shape. To make the basic concept clearer, lets consider the mechanism shown below. It is interesting te think of mechanisms as stractares with zero buckling load: -? t a/ This structure is now at buckling without any, external load. IF a small Purturbing Force Pis applied to the assembly, as shown, then the only possible normalized mode is the one with the dashed line and is of Sel. Structures at buckling are also mechanisms with the mechanism mode being the buckled shape corresponding to the applied buckling load. So the idea is te load the structure upto buckling and then force it ky a purturbing Force, the shape that it adopts is the buckling shape- rp the exact buckling load is applied, the sti PPness becomes singular: This singularity can be handled numerically by consiclering loads close te, but nef ot the bucKling load. Assume Fs, the mode carresponding fo Ac» has to be calculated Prom At, which 1s only approximately equal toi. The Hopper and williams methed te Pind 4s can be represented as calculatia (which Is only app P! mo Ly ap reximotely equal to 4) from _ex random Forces (K- Aes) Jer Gt Example F- Consider A as the 202 matrix given ky 3 8 he [‘ ae Now Knowing that one of the eigenvalues of A is 13 calculate the correspending ejgenvecter using the dePinition for eigenvectors and then compare ft. results with the results oP the Random Force Vector me thod- sel": vsigg the deBinition, we have te find a vecter [x i) which satisPies 2 val] % ° ve ot [pm ° TRis gives xa2-YE x1, and with choesing xis!) the eigenvector which is normalized with respect te x, Weuld be ti -i2] . Now we will selve the problem using the Rand-m Force Yector method. we hove te choose a value for At which ig close te ALC=, and then solve Por CA- 1) fp-Ge. As the First trial, lets cheese At 20.94, then 3B-0-49, Ya dn Qs, m 2-02-99] | Fra On we also need some random numbers for Gi, and ree I open the white Pages and Pickup a ¥ digit phone number Prom the top righthand side of the opened pages: It happens te be 663 1200 FSFIai2 Now we solve for 4s in [T ETP4n] f eeaizes ve etl lg) f eerceur then 4, ler 2 77663 1200 \ Jy Geeen-2 [VE atl faeziaig| NB Sate “| ave io6 6a ' K iv and when normalized gives wage (ert This result may be compared with the normalized veetor Pound from the definition of eigenvectors, ie C1 T+ The First clement is the same number in beth cases, so the relative diPPerence oP the secend elements canbe chosen as the relative error in the eigenvector be 510FF = AINE! igo 2 2.7502/ [ae BEBE error = Now we continue the method with afher values of At that are closer to Au Similar random Forces of 66352 . . °°) 3s chosen. Fellewing is the result. 387d. At eigenvector error /. 1 0. 494, oat a1 180514 1 0 4994 9.024! -1 414 5968 1 2 49994 0. a02F A F514, As is seen the differences in the results of the ths when Mt Systems of evaluation are dis appeari becomes closer to Ai - The other interesting point is that : if the emors in the eigenvectors are draun Dei] in a legrithmic scale, the points will i approximately fellow a line of uni versus t slope. This is shown in The Figure below a log (error) (554-59) 4 (4,357) 2 (39257) 2 (4 b.56) Similar geed results can be achieved if a new set oF random forces is chesen everytime rather than continuing with the same random farce vector Following are the results with the same dt values but with diPPerent random Porce vectors chosen every time From white Pages. 23 Me Force Vector eigenvector error, 4699895 ' 0. 444 0.4233 580 82 67 abols 68 6703252 ' 0 4494 01232 5341405 =1-41595549 4532148 1 0.44944 0.010 2213340 nist be 3676 Now if the errors are drawn versus as, the points x are seen to be scattered about a line of ynit slope. Te scatter js due to different randem Forces = leg (error) (5,392) (Ce nal (352-03) a5 Summary #! #2 #3 #4 characteristic equation is [A-dalee eigenvalues are the roots of #I. the matrices that We ore dealing are real symmetric matrices. eigenvalves of a real symmetric mateix are real. several methods exists but we are going te work with a. Sturm Sequence method (eigenvalve) b- Random Force vecter methad Ceigenvecter) intre duct 25 Computer program On the Fe Ilowing Pages, a computer program 1s presented eshich solves for the generalized eigenvalue problem of (K- As) 4-9. The program isa fortran ## subroatine which has a call Format as; call eigen (Rg, evctr, evalue, mocleno, dof, maxdof, Kminusg)) where 5 k, gp medens, doP and maxdeP are the input variables. Kminusg is an intermediate Variable which is only required for necessary calculations and i's not of ony interest to the user. evetr and evalue are the output variables. rather than those thet are obvious by Variables used their names have Pollowin meanings : doF s degrees of Preedom (dimension of the matrices nxn then doPon. Korg; iP Kis Psy maxdek : modenos the maximum number of degrees of freedom which can be aay integer number olepencing on the size of the computer memory or the maximum size of the problem thot the ver intends to solve the number corresponding to the i gen value/vector required; modens =! is corresponding to the smallest eigenvalve and modencs dof is corresponding to the biggest eigenvalue: at ° ° subroutine eigen (k,g,evctr,evalue, modeno, dof, maxdof,kmirus parameter (accuracy = 0.00001) real*8 k(maxdof,maxdof) ,g(maxdof,maxdof) real*8 kminusg (maxdof,maxdof) , evctr (maxdof) real*8 range, maxlambda,minlambda, ranout real*8 lambda, upperlambda, lowerlambda, evalue integer count, dof, ranin logical zerodiagonal A random generator will be used in the subroutine. It need to be started somewhere beforehand. Following will be the starter. call randomize (1,ranin, ranout) the bounds within which the eigenvalues are assumed to be are strored here minlambda = -le5 maxlambda = +1e5 now the eigenvalue is started to be calculated using the bisection technique between the bounds and if a zero on diagonal occured, by a random divison between the bounds. upperlambda = maxlambda lowerlambda = minlambda vange = upperlambda - lowerlambda ie =0 28 eo if (range.gt.accuracy) then ic = ictl range = upperlambda - lowerlambda lambda = (lowerlambda + upperlambda) /2. call makekminusg (x,g,kminusg, lambda, dof,maxdof call check (kminusg,dof,maxdof, zerodiagonal) if (zerodiagonal.eq./true’) then call randomize (0,ranin, ranout) lambda = lowerlambda + (upperlambda-lowerlambda) *ranout call makekminusg (k,g,kminusg, lambda, dof, maxdof) end if call factorize ( kminusg,dof,maxdof ) call countnegatives (kminusg, count, dof,maxdof) if (count.ge.modeno) upperlambda = lambda if (count.1t.modeno) lowerlambda = lambda go to 3 end if evalue = lambda at this stage the eigenvectors are calculated using a random disturbance call makekminusg (k,g, kminusg, lambda, dof,maxdo£) call factorize ( kminusg,dof,maxdof } call randomvetr (evetr,dof, maxdof, ranin) call postfactorize (evctr, kminusg, maxdof, dof) call normalize (evetr,maxdof, dof) return end 2a subroutine factorize (kminusg,dof,maxdof) xeal*8 kminusg(maxdof,maxdof) real*8 ratio integer dof do i=1,dof do j=it1,dof ratio = kminusg(j,i) /kminusg(i,i) do k=1,dof kminusg({j,k) = kminusg(j,k) - ratiot*kminusg (i,k) end do end do end do do i=1,dof do j=itl,dof kminusg(i,3) = kminusg(i, 3) /kminusg(i, 4) end do end do return end subroutine countnegatives (kminusg, count, dof,maxdof) real*8 kminusg(maxdof, maxdof) integer count,dof count = 0 do i=1,dof i£(kminusg(i,i).1t.0) count = count+1 end do return end 30 subroutine makekminusg(k,g,kminusg, lambda, dof, maxdof) real*8 (maxdof, maxdof) , g(maxdof,maxdof) real*8 kminusg(maxdof,maxdof) real*§ lambda integer dof call clearmatrix (kminusg, maxdof, maxdof) do i=l,dof do j=1,dof kminusg (i,j) = k(i,3) - lambda*g(i, 3) end do end do return end subroutine check (kminusg,dof,maxdof, zerodiagonal) real*8 kminusg(maxdof,maxdof) integer dof logical zerodiagonal do i=1,dof if (abs (kminusg(i,i)).1t.le-20) zerodiagonal = .true. end do return end BI subroutine randomize (when, ranin, ranout) real*8 ranout integer when,a,c,ranin if (when.eq.0) go to 1 ranin = 673 continue a=17 m = 1048576 c=l k = int ((a*ranintc) /m) ranin = a*tranin+c-m*k ranout = float (ranin)/float (m) if (ranovt.eg-1.or-ranout.eq.0) go to 1 return end subroutine randomvetr (evctr, dof, maxdof, ranin) real*8 evctr(maxdof) real*8 ranin,ranout integer dof do i=1,dof call randomize (0,ranin, ranovt) evetr (i) = ranout end do return end 32 subroutine postfactorize (rhs, kminusg,maxdof, dof) real*8 rhs (maxdof) , kminusg (maxdof,maxdof) real*8 sum integer dof forward reduction of the right~hand side do i=l,dof sum = 0 do j=i-1,1,-1 sum = sum + kminusg(3,i)*rhs (3) end do vhs(i) = rhs (i) -sum end do diagonal scaling of the right-hand side do i=1,dof vhs(i) = rhs (i) /kminusg (i, i) end do back substitution of the right-hand side do i= dof,1,-1 sum = 0 do j = i+i,dof sum = sum + kminusg(i, 3) *rhs (3) end do rhs(i) = rhs(i) - sum end do return end 33

You might also like