CMM-2003 – Computer Methods in Mechanics June 3-6, 2003, Gliwice, Poland

A block Lanczos method with spectral transformations for natural vibrations and seismic
analysis of large structures in SCAD software
[Link]. Fialko, E.Z. Kriksunov and [Link]
Software company SCAD Soft
13, Chokolovsky bld., room 508
Kiev, 252680 GSP, Ukraine
E-mail: fialko@[Link]

Abstract
A powerful eigenvalue extraction method for natural vibration analysis of large-scale structures is presented. This method is based on
a block Lanczos algorithm with shift implementation [4] and has four modes of operation: modal mode, interval mode, seismic mode
and verification mode. The modal mode is developed to extract the first n required eigenpairs. The interval mode means the
extracting of all eigenpairs whose frequencies fall into an interval [ a, b] . The seismic mode is intended for extracting eigenpairs
while ensuring that a required sum of modal masses be provided for all seismic input directions. The verification mode is used to
detect hard-to-find errors of a finite element model, such as a local and global dimensional instability, lack of supports and so on. The
shift technique is applied to split the frequency interval into small subintervals and restricts the growth of the computation effort for
large-scale dynamic problems. The sparse direct solver, based on multifrontal technique, is used to factor the stiffness matrix. It
provides not only a faster factoring procedure, but fast forward-back substitutions during repeated Lanczos steps. The performance of
this method is compared here with that of different solvers. Its robustness and efficiency is confirmed by numerous examples.

Keywords: eigenvalue, shift, trust interval, block, multifrontal

helps detect these hard-to-find errors. This approach has proved

1. Introduction to be better than an analysis based on singularities detected
during the matrix factoring.
Recently, large-scale problems with large number of Numerical examples presented here illustrate the
eigenpairs arise more and more often. Therefore it requires capabilities of the proposed method.
robust and fast methods for eigenvalue analysis to be developed.
The great number of eigenpairs is needed often for seismic 2. Block Lanczos method with spectral transformations
analysis when a lot of low vibration modes are local and/or do
not make a representative contribution to the seismic response The Lanczos method [5], [6], [7] is recognized as a most
of a structure in question. It is very difficult to ensure the powerful tool for extraction of large number of eigenpairs in
sufficient percentage of the sum of modal masses [1], [3], [9] large-scale problems of structural mechanics. The stability of its
along all or some principal seismic directions. computational process is ensured by a selective or/and partial
The conventional procedure of extracting low vibration re-orthogonalization [5], [6], [7]. The experience of authors [3]
modes (a modal mode) consists of following: the user assigns indicates that the selective re-orthogonalization based on
the number of desired modes, runs the analysis and checks the Paige’s theorem, is efficient only for relatively short Lanczos
sum of modal masses in each direction. If the sum of modal processes when the number of generated Lanczos vectors is not
masses is insufficient, the user increases the required number of too great. For long Lanczos processes the selective re-
modes and runs the calculation again. And so on. For “hard” orthogonalization fails to succeed. On the contrary, the partial
problems this procedure is usually repeated multiple times. It is reorthogonalization [7], [8] keeps Lanczos vectors highly
not an efficient way to solve seismic problems. The seismic orthogonal and ensures a highly stable computation even for
mode (regime) of a block Lanczos method proposed here is long Lanczos processes which generate 300-700 and more
based on shift implementation and a posteriori bounding of Lanczos vectors. Therefore we employ the block version of
residual [2], [4], and it allows us to obtain all modes, provide partial re-orthogonalization [4].
the required sum of modal masses during one run of algorithm Conventional algorithms of Lanczos method possess the
and avoid time-consuming attempts which are typical for following disadvantage when the large-scale problems
conventional approaches. containing 60 000 - 500 000 and more equations are to be
The design of nuclear power plants and other industrial solved: at each step of Lanczos vector generation it is necessary
buildings often requires to extract all natural vibration to make the forward-back substitutions only for a single right-
frequencies belonging to a given frequency interval [ a, b] . The side vector (r. s. v.). Due to this a lot of computation time is
interval mode of the proposed block Lanczos method, based on spent for input-output (I/O) operations because the upper part of
Sturm check sequences and a shift technique, implements this the factored matrix U , where K = LU and K is a stiffness
approach. matrix, must be read twice block-by-block from a secondary
A considerable effort and design time are spent to detect storage (disk) per each Lanczos vector.
errors in large-scale finite element models, like local and global
dimensional instability, lack of supports and so on. The
verification mode proposed here, one based on a shift technique,
 Fig.1 Typical distribution of eigenvalues, Ritz approximations and coarse approximations. • - eigenpairs; ο - Ritz pairs; - coarse
approximations

The second problem is related to a drastic increase of 4. If in > 1 , modify σ = λ l + ∆λ , where the choice of ∆λ is
amount of operations when the dimension of Krylov subspace based on estimation of Ritz pairs (see below). If σ ≠ 0 ,
(number of generated Lanczos vectors) is big enough (usually
exceeds ~100). perform the factoring of K σ = K − σM = L σ U σ .
The block version of Lanczos algorithm is intended to 5. Internal loop: j = 1,2,...,
reduce the I/O effort. 6. Calculate:
Spectral transformations are implemented to split the long
Lanczos process into a few relatively short ones. It reduces K σ U j = MQ j ⇒ U j
essentially the computation time when a large number of
eigenpairs (100 - 1000 and more) are required. U j := U j − Q j −1B Tj
(2)
The original natural vibration problem is as follows: A j = U Tj MQ j

Kϕ − λ M ϕ = 0 (1) R j +1 = U j − Q j A j

where K is a positive definite symmetrical stiffness matrix and 7. Q j +1B j +1 = R j +1 , where QTj +1MQ j +1 = I
M is a positive definite or semi-definite lumped or consistent 8. Update the block triangle matrix
mass matrix, λ, ϕ is an eigenpair.
The idea of the shifted block Lanczos algorithm [4] is A1 B2 
adopted as a basis for the algorithm presented here, although  T 
many details have been modified and adjusted to peculiarities of B 2 A2 B3 
 
structural mechanics problems. Tj =  B T3 A3 B4  (3)
The basic algorithm of the block Lanczos method with  ... ... ... ... ... 
spectral transformations is presented below.  
 B Tj A j 
1. λ l = λ r = σ = a , where λ l , λ r are left and right
boundaries of the trust interval, a is the leftmost boundary 9. Solve the reduced eigenvalue problem
of the interval [ a, b] , σ is a shift value. The trust interval
contains only converged Ritz pairs (we refer to them as
eigenpairs) and does not have any missed ones, that is, all
Tj S j = Θ j S j (4)
eigenpairs from trust interval [λ l , λ r ] are thought to have
10. Compute the precision of Ritz pairs using a posteriori
been extracted. Set the block size p (usually p = 3 ) and
boundary of residual [2],[4]:
in = 1 .
2. External loop: continue solution till all required eigenpairs MK σ−1My i − My i θi = B j +1ETj s i = β ij , i = 1,2,..., j (5)
are extracted with a given accuracy. M −1 2

3. { }
Set the start block Q 0 = q10 , q 02 , , , q 0p = 0 and
{ 0 0 0
}
R 1 = r , r ,..., r , where all vectors in block R 1 are
{ }
where Y j = y 1j , y 2j ,..., y jj = Q j S j . The Ritz pair is taken
1 2 p
β ij β ij
linearly independent. Then: Q1B1 = R1 , where B 1 is an as converged, if ≤ tol , where λ i − ν ij ≤ and
upper triangle matrix p × p and Q MQ1 = I . It is a
T (θ ) i
j 2
(θ )
i
j 2

matrix form of the block Gram-Schmidt orthogonalization λ i is an exact eigenvalue, v ij is an approximation at the
algorithm, which in this article differs from that presented Lanczos step j , tol is an adopted tolerance (usually
in [4]. The size of blocks Q, R is N × p where N is the 1.0×10-4 ÷ 1.0×10-8 . We provide the following
number of equations of (1).

2
 β ij This algorithm ensures that the skipped eigenpairs do not
classification: if ≤ tol then it is a converged fall into the interval [a, σ last ] where σ last is the last shift value.
(θ ) i
j 2

The algorithm presented above implements the modal and

β ij interval modes. In addition, when the interval mode is used, we
eigenpair; if tol < ≤ 1.0 × 10 − 2 then it is a Ritz pair; if
(θ ) j 2 define the number of eigenvalues enclosed in the given interval
i
[a, b] :
β ij n def = Z (σ = b ) − Z (σ = a ) (7)
> 1.0 × 10 − 2 then it is a coarse approximation (see
(θ )
i
j 2

Fig.1). The initial shift value is taken as σ = a .

In the seismic mode the computation stops as soon as the
11. Break the internal iteration loop at j , if: required sum of modal masses is reached.
• (in = 1) ∧ ( j ≥ NL ) , where NL ≅ 150 − 200 is the In the verification mode, a small negative shift is
maximal number of Lanczos vectors implemented to avoid a singularity caused by potential errors in
• [ ( )]
(in > 1) ∧ ncin ≥ Z (σ) − n def , where n cin is the number the computation model.

of converged modes at the given step in of the 3. Numerical results

external loop, n def is the number of eigenpairs already
found, Z (σ ) is the number of negative signs detected 3.1. The modal mode
on the diagonal of U σ .
The modal mode of the proposed method is illustrated by
These conditions being satisfied ensures that the
the following example. The model of a multi-storey building
skipped eigenpairs do not fall into the interval [λ l , σ] . contains 19 409 nodes, 19 456 finite elements and 115 362
If the internal loop is interrupted, go to step 14. equations (see Fig. 2).
12. If the orthogonality between Lanczos vectors is less than The comparison of computation time vs. the number of
required, perform a partial re-orthogonalization according required modes (modal mode) for different methods is presented
with [4] – corrected matrices Q j +1 , R j +1 , B j +1 are the result in Table 1. The subspace iteration (SI), block subspace iteration
of it. (BSI), a conventional Lanczos method (LM) and the block
13. j + + , go to 5 Lanczos method with spectral transformations (BLST) proposed
here are presented. A Pentium-III computer (Intel-1000MHz
14. Evaluate the eigenvectors for all eigenvalues enclosed in
[ ]
~ ~
the interval λ l , λ (see Fig.1), where λ is a largest
processor, 512 MB RAM) is used. The precision of all obtained
eigenpairs estimated by (6) is not worse than 1.0×10-04 .
eigenvalue from the continuous part of the spectrum. In the The block subspace iteration method has been developed
seismic mode, calculate the sum of modal masses in each according to [9] and is presented in [3]. The conventional
principal direction [1], [9]. Lanczos method has been developed according to [5], [7] and is
15. Evaluate a new approximation of ∆λ so that the number presented in [3]. This example illustrates a typical performance
of converged eigenvalues, Ritz and coarse approximations, of these methods. It can be clearly seen that the BSI method is
~ preferable to SI. If the factored stiffness matrix can fit in RAM,
located to the right of λ , be equal to the expected number
then the Lanczos method is essentially faster than BSI. The
of eigenpairs which we want to obtain from the next trust
more the number of required modes, the bigger the advantage of
subinterval (see Fig.1). Usually this value is adopted to be
the Lanczos method.
15-20 that corresponds to about 100-150 maximal number
Otherwise, if the size of the factored stiffness matrix is too
of Lanczos vectors. Evaluate λ r and put λ l = λ r to the big for it to be allocated in RAM, while the number of required
next shift interval. modes is not so big (about 10 - 50), the BSI method usually
16. In the modal, interval and verification modes: if happens to be more efficient than that of Lanczos due to
n def ≥ n req , where n req is the required number of blocking of I/O operations during forward-back substitutions
eigenpairs enclosed in the interval [a, b] and n def is the (see examples from [3]).
The BLST method is essentially faster than other considered
number of eigenpairs obtained and enclosed in the interval methods, especially when the number of required modes is
[a, λ r ] , break the external loop and go to step 17. In the large. The size of Krylov subspace (number of Lanczos vectors)
seismic mode: break the external loop and go to step 17, if remains relatively small during spectral transformations. An
the sum of modal masses for all converged modes reaches appropriately chosen shift divides the required interval into a
the accepted values (default: sequence of non-overlapped trust subintervals. We pass from a
∑ m x = ∑ m y = 90%, ∑ m z = 75% ). Otherwise (no preceding subinterval to the following one till the required
number of modes are converged or the proper sum of modal
finish criterion is satisfied), in + + and go to step 2. masses is achieved. The required interval is a trust interval too,
17. Evaluate the precision for each eigenpair: because it is a sum of adjacent non-overlapped trust
subintervals. Inasmuch as each modification of shift leads to the
y i − λ i K −1My i
preci = 2
, i = 1,2,..., n def (6) factoring of matrix K σ = K − σM , the fast multi-frontal solver
yi 2 with sparse reordering plays an important role.
The last row of Table 1 illustrates great capabilities of the
BLST method presented here.

3
 Fig. 2 A multi-storey building

Table 1: Comparison of computation time for different methods. A multi-storey building (Fig.2)

Number of Subspace iterations Block subspace Lanczos method [3] Block Lanczos method with
required modes iterations [3] spectral transformations
25 2 h 28 min 31s 1 h 49 min 38 s 54 min 24 s 38 min 14 s
50 5 h 18 min 33 s 3 h 06 min 16 s 1 h 22 min 37 s 55 min 56 s
100 — — 2 h 22 min 14 s 1 h 52 min 14 s
1 000 — — — 11 h 25 min 02 s

4
 Fig. 3 An industrial building. The frequency band of equipment is 8 to 20 Hz.

The precision of the extracted eigenpairs estimated by (6) is

3.2. The interval mode
not worse than 1.0e − 08 . It is a very high accuracy, and we
The example shown in Fig. 3 illustrates the interval mode of believe that the proposed method has given nearly exact
the block Lanczos method with spectral transformations. eigenpairs of the presented finite-element model.
This model contains 10 439 nodes, 11 431 finite elements The computation time is 45 min 45 s on a Pentium-III
and 60 760 equations. The frequency band of equipment is 8 to computer (Intel-1000 MHz processor, 512 MB RAM).
20 Hz.
First of all, the proposed algorithm evaluates the number of 3.3. The seismic mode
eigenpairs enclosed in the interval [8, 20] Hz by (7). The first Big difficulties often arise in the seismic analysis when a lot
Sturm sequence check (SSCH) is performed for the shift of eigenpairs are required to achieve a good sum of modal
corresponding to the top boundary of the frequency (20 Hz). masses. The following examples illustrate this problem (see
The second SSCH is performed for the bottom boundary (8 Hz). Tables 2,3) and capabilities of the proposed method.
The required number of eigenpairs is 97.
Three subintervals are required to extract 97 eigenpairs. The Table 2: Sum of modal masses versus number of extracted
first subinterval is [8, 13.837] Hz ( σ = 8 Hz, 46 eigenpairs have eigenmodes. The model of church (Fig.4)
converged), the second one is [13.837, 18.578] Hz ( σ = 16.343
Hz, 41 eigenmodes have converged) and the third one is Number of ∑ mX ,% ∑ mY ,% ∑ mZ ,%
[18.578, 22.624] Hz ( σ = 20.807 Hz, 39 eigenpairs have modes
converged). In fact, the algorithm evaluates 126 eigenpairs from 10 50.0 51.9 0.0
the interval [8, 22.624] Hz that is wider than required because 50 58.2 77.3 5.1
the left boundary of each trust subinterval must match the right 100 68.4 84.5 13.9
boundary of its preceding subinterval. Otherwise, the continuity 300 95.4 95.2 74.3
of the spectrum is not safely guaranteed. 638 97.9 97.3 90.0
Four factorings of matrix K σ = K − σM are required.

5
 Fig. 4 The model of a church (4 525 nodes, 6 520 finite elements, 27 138 equations)

Table 3: Sum of modal masses vs. number of extracted The required sum of modal masses is taken as 90% in each
eigenmodes. A multi-storey building (Fig.5) principal direction (in compliance with UBC-97, PS-92 and
Eurocode-8 seismic codes).
Number of ∑ mX ,% ∑ mY ,% ∑ mZ ,% The Church problem requires 638 eigenmodes to satisfy the
modes seismic codes mentioned above. The second problem (a multi-
10 79.6 80.7 7.5 storey building) requires 924 eigenmodes. In fact, 1051
50 83.5 84.5 47.6 eigenpairs have been extracted for the second problem and 20
100 84.9 88.5 63.8 trust subintervals have been created. The precision of
300 91.6 92.3 82.1 eigenvectors is not worse than 1.0e − 08 . The computation time
924 94.8 95.6 90.0 is 13 h 52 min on a Pentium-III computer (Intel-1000MHz
processor, 512 MB RAM).

6
 Fig. 5 A multi-storey building (20 285 nodes, 29 932 finite elements, 119 874 equations)

3.4. The verification mode Table 4: Low frequencies of a multi-storey building

This mode of the method allows us to detect hard-to-find errors
of computational models like local or global dimensional Number Frequency, Hz Comments
instability, lack of supports and so on. 1 -8.228e-006 Free rigid body motion of
A small negative shift is implemented to avoid the 2 -2.104e-006 unconstrained part of structure.
singularity caused by potential errors in computation model. An 3 2.62e-006 Mode 1 is presented in Fig. 7
example taken from real engineering practice is presented in 4 3.172e-006
Fig. 6. First frequencies are shown in Table 4. 5 4.257e-006
The modes 1-6 have nearly zero frequencies. This indicates 6 5.623e-006
a dimensional instability. Indeed, the analysis of eigenmodes 7 0.5181 The bottoms of columns are not
detects an unconstrained part in the structure which has six 8 0.6726 constrained - see Fig. 8
modes of free rigid body motion. The first mode is presented in 9 0.986
Fig. 7. 10 1.042

7
 Fig. 6. A multi-storey building (24 434 nodes, 26 273 finite elements, 127 165 equations)

The multi-frontal solver informs us of such instability which large-scale structures. Proceedings of the second European
leads to very small pivots during the Gauss elimination. But the conference on computational mechanics, Poland, June 26-
modes of instability (mechanism modes) have been detected 29, 2001, 18 p.
only in the verification mode of the BLST method. Moreover,
the verification mode allows us to detect unconstrained bottoms [4] Grimes, R.G. Lewis, J.G., Simon, H.D., A shifted block
of columns too (Fig. 8). Lanczos algorithm for solving sparse symmetric generalized
eigenproblems, SIAM J. Matrix Anal. Appl, V.15, 1: pp. 1-
References 45, 1994.

[1] Clough, R. and Penzien, J., Dynamics of structures, [5] Hughes, T.J.R., The finite element method, Prentice Hall,
McGraw-Hill, Inc., 1975. Englewood Cliffs, NJ, 1987

[2] Ericsson, T. Ruhe, A., The spectral transformation Lanczos [6] Parlett, B., The symmetric eigenvalue problem, Prentice-
method for the numerical solution of large sparse generalize Hall, Englewood Cliffs, NJ, 1980.
symmetric eigenvalue problem, Math. Comput., 35, pp.
1251-1268, 1980. [7] Papadrakakis, M., Solving large –scale problems in
mechanics, John Wiley & Sons Ltd., 1993.
[3] Fialko, [Link]., High-performance iterative and sparse direct
solvers in Robot software for static and dynamic analysis of

8
 Fig.7 An unconstrained part of structure is detected

Fig.8 Unconstrained bottoms of columns are detected

[8] Simon, H., The Lanczos algorithm with partial [9] Wilson, E.L., Three dimensional dynamic analysis of
reorthogonalization, Math. Comp., 42, pp. 115-136, 1984. structures, Computers and Structures, Inc., Berkeley,
California, USA, 1996.