Finite Elements in Analysis and Design 36 (2000) 261}278

A new load-dependent Ritz vector method for structural

dynamics analyses: quasi-static Ritz vectors
Jianmin Gu , Zheng-Dong Ma
, Gregory M. Hulbert
*
Vehicle NVH Department, Ford Motor Company, Dearborn, MI 48124, USA

Department of Mechanical Engineering and Applied Mechanics, 2250 G. G. Brown, The University of Michigan,
Ann Arbor, MI 48109-2125, USA

Abstract

Existing load-dependent Ritz vector (LDRV) methods employ static recurrence procedures to generate the
Ritz vectors. As such, these vector methods are best suited for low-frequency problems. For higher-frequency
problems, the existing methods may engender large sets of Ritz vectors, which signi"cantly reduces the
methods' e$ciency. A new algorithm is presented for LDRV generation using a quasi-static recurrence
procedure, denoted as the quasi-static Ritz vector (QSRV) method. A tuning parameter, designated as the
centering frequency, controls the behavior of the QSRV approach, enabling the new method to improve
upon existing LDRV methods for particular frequency ranges of interest. Compared with existing LDRV
methods, the QSRV method is more e$cient (in terms of the number of Ritz vectors), more accurate (in terms
of response errors), and more stable (in terms of orthogonality). Numerical examples are provided to
illustrate the accuracy, e$ciency and generality of the proposed method.  2000 Elsevier Science B.V. All
rights reserved.

Keywords: Structural dynamics; Load-dependent Ritz vector method; Model reduction; Quasi-static recurrence

1. Introduction

Ritz methods have long been used to approximate the dynamic response of structures and as
model reduction techniques to reduce the size of large-scale models. Provided that the Ritz vectors
span the same con"guration space, they are an attractive alternative to standard eigenvectors
(normal modes), since Ritz vectors can be computed with signi"cantly less computational expense.
Using appropriately chosen Ritz vectors can improve accuracy in cases where eigenvectors are not
the best choice of basis, e.g., eigenvectors that are orthogonal to the loading are not excited, even if

* Corresponding author. Tel.: 1-734-763-4456; fax: 1-734-647-3170.

0168-874X/00/$ - see front matter  2000 Elsevier Science B.V. All rights reserved.
PII: S 0 1 6 8 - 8 7 4 X ( 0 0 ) 0 0 0 3 6 - 6
262 J. Gu et al. / Finite Elements in Analysis and Design 36 (2000) 261}278

their frequencies are contained in the loading frequency bandwidth [1]. In addition, Ritz vectors
are employed as starting vectors within some eigenvalue solvers because a properly chosen set of
Ritz vectors can signi"cantly accelerate the iteration process of an eigenvalue solver.
Load-dependent Ritz vectors (LDRVs) are a particular class of Ritz vectors in which information
about the loading on the structure is used to generate the vectors. Wilson et al. [1] "rst developed
the WYD algorithm, in which the LDRVs were derived by a Krylov sequence [2]. These
mass-orthogonalized Ritz vectors are much less expensive to generate than normal modes.
Moreover, static completeness is automatically satis"ed, since the computed static displacement is
used as the initial Ritz vector. The method, further developed by Nour-Omid and Clough [3,4] and
Leger et al. [5] (Ref. [6] provides perhaps a more accessible version of this original reference),
retains the good convergence characteristics of the mode}acceleration method but with much less
computational cost. Nour-Omid and Clough employed the Lanczos algorithm [7], which has been
used extensively to solve large eigenvalue problems, to generate the LDRVs. Their method is
essentially identical to the WYD algorithm when applied with the same initial vector and partial
orthogonalization. Therefore, the WYD generated vectors may be regarded as a special class of
Lanczos vectors. Leger et al. later proposed a modi"ed version of the WYD algorithm, called the
LWYD algorithm, to eliminate the loss of orthogonality associated with the WYD algorithm.
Existing Ritz vector methods employ static recurrence procedures to generate the Ritz vectors,
which satis"es the static completeness condition. As such, the LDRV methods have similar
advantages and disadvantages as the mode}acceleration method. Consequently, the existing
LDRV methods are best suited for relatively low-frequency problems. For higher-frequency, or
banded-frequency problems, they may engender large sets of Ritz vectors to span the con"guration
space associated with the high frequencies. This decreases the LDRV's e$ciency, compared to the
standard modal superposition technique. It also can result in the loss of orthogonality of the Ritz
vectors, which causes numerical problems in solving the reduced system. For multiple loading
cases, these disadvantages become more pronounced.
A good set of Ritz vectors should satisfy two basic conditions. First, the Ritz basis should be
complete with respect to possible loading patterns associated with the problem, at least for
a desired frequency of interest. While the normal mode set (mode}displacement method) does not
satisfy this condition, existing LDRV and mode}acceleration methods do meet this condition at
zero frequency. Second, the Ritz vectors should span any desired frequency range. A normal mode
set, comprising all modes in the frequency range of interest, does satisfy this frequency condition,
but the LDRV and mode-acceleration methods do not meet this second condition.
In this paper, a new LDRV method is proposed that employs a quasi-static recurrence procedure
based upon a new modal superposition technique. The introduction of a tuning parameter,
designated as the centering frequency, into the quasi-static recurrence procedure, enables the new
quasi-static Ritz vector (QSRV) method to satisfy both conditions needed to achieve a good set of
Ritz vectors. Compared with other LDRV methods, the QSRV method is more e$cient (in terms of
the number of recurrence steps and LDRVs employed), more accurate (in terms of response errors)
and more stable (in terms of the orthogonality of the Ritz vectors), even for low-frequency
problems. The QSRV method can include multiple centering frequencies and accommodate
multiple loading cases, further improving the method's performance. Numerical examples in both
frequency response analysis and transient response analysis are presented, which illustrate the
accuracy, e$ciency, and applicability obtained by the proposed method.
 J. Gu et al. / Finite Elements in Analysis and Design 36 (2000) 261}278 263

2. Load-dependent Ritz vector methods

The matrix equation of linear structural dynamics is

MX$ #CXQ #KX"F, (1)
where M, C and K are the mass, damping, and sti!ness matrices, respectively; F is the vector of
applied loads (a given function of time), X is the vector of displacement unknowns (of length N), and
superposed dots indicate di!erentiation with respect to time. Furthermore, the external load vector
F often is represented as a superposition of the spatial vectors G (loading patterns) with the
G
time-dependent factors h (t) (i"1, 2,2, k), i.e.,
G
I
F" G h (t)"GH, (2)
G G
G
where G"[G , G , 2, G ], H"[h (t), h (t), 2, h (t)]2 and k denotes the total number of
  I   I
loading patterns. For ease of exposition, we "rst discuss the single loading pattern case. Therefore,
in the following G is a vector and H is a scalar function.
A standard model reduction process for the system (1) can be described as a simple coordinate
transformation, i.e.,
X"WZ, (3)
where Z is a vector of the n-generalized coordinates of the reduced system (for e$ciency), n;N
and W is the corresponding transformation matrix. Substituting Eq. (3) into Eq. (1) and premultip-
lying the resultant equation by W2 yields
MZ#CZQ #KZ"W2F, (4)
where the reduced matrices are
M"W2MW, (5)
C"W2CW, (6)
K"W2KW. (7)
Transformation matrix W could be a modal matrix containing n normal modes of the system.
The model reduction process is then just the standard modal superposition technique (the
mode}displacement method), and Eq. (4) is a set of uncoupled linear ordinary di!erential equations
provided the system damping model uncouples properly. More generally, W contains n general Ritz
vectors or the so-called `assumed modesa [8]. Obviously, the selection of the Ritz vectors to form
W is critical. In the following, we focus on the traditional LDRV methods.

2.1. WYD algorithm

Wilson et al. [1] employed a special Krylov sequence to generate the LDRVs, having the
recurrence form
WH"K\MW , i"2, 3, 2, n, (8)
G G\
264 J. Gu et al. / Finite Elements in Analysis and Design 36 (2000) 261}278

where the starting vector is

WH"K\G (9)

and mass-orthonormalization is employed at each step
G\
WM "WH! (W2MWH)W , (10)
G G H G H
H
WM
W" G . (11)
G (WM 2MWM )
G G
The WYD algorithm has a physical interpretation. The "rst Ritz vector, W , represents the static

response mode to the given loading pattern G. The inertial term, which is neglected when
calculating W in Eq. (9), is approximated by MW and applied as a static load in the next step to
 
generate a new Ritz vector in order to approximate inertial e!ects. This process is repeated until
enough Ritz vectors have been generated or no more independent Ritz vector can be generated by
the process. As noted in Section 1, the static completeness condition for the generated Ritz vector
set is satis"ed automatically.

2.2. Lanczos algorithm

Based on the standard Lanczos algorithm, Nour-Omid and Clough [3] proposed another set of
Ritz vectors, known as Lanczos vectors. This algorithm uses the same Krylov sequence as used in
the WYD algorithm (Eq. (8)). In general, the Lanczos algorithm does not require a typical starting
vector as required by the WYD algorithm. However, Nour-Omid and Clough suggested that it has
particular advantages for structural dynamic problems if the static response mode as in Eq. (9) is
chosen as the "rst Lanczos vector.
The major di!erence between the Lanczos algorithm and WYD algorithm is the orthonormaliz-
ation procedure. Instead of full orthogonalization used by the WYD algorithm (Eqs. (10) and (11)),
the orthogonalization procedure employed by the Lanczos algorithm is applied only to the
previous two vectors
WM "WH!a W !b W , (12)
G G G G\ G G\
where
a "W2 MWH, (13)
G G\ G
b "(WM 2 MWM ). (14)
G G\ G\
The same normalization condition, Eq. (11) is employed by the Lanczos algorithm.
Nour-Omid and Clough showed that the orthogonalization procedure employed in the Lanczos
algorithm is equivalent to the full orthogonalization employed in the WYD algorithm. However, in
practice, due to roundo! error, the Ritz vectors generated by the Lanczos algorithm tend to lose
their orthogonality. If such errors are not corrected when they reach a critical size, the vectors may
become linearly dependent. Therefore, a full reorthogonalization against all previous vectors or an
equivalent action is still necessary when certain linear dependence conditions are reached [3].
 J. Gu et al. / Finite Elements in Analysis and Design 36 (2000) 261}278 265

2.3. LWYD algorithm

Both the WYD and Lanczos algorithms lose orthogonality due to the numerical roundo! errors
when the number of the Ritz vectors becomes large. This produces numerical problems when
solving the reduced system (4). To obtain a more stable Ritz vector generation algorithm, Leger
et al. [5] introduced an additional set of temporary vectors and orthogonalization procedure. In
the LWYD algorithm, the Ritz vectors are generated using

WH"K\Mu , (15)
G G
where u is the static mode shape due to the loading pattern G, i.e.,

u "K\G. (16)

For i*2, the vector u is mass-orthogonalized before employing Eq. (15), i.e.,
G
u "u !(u2 MW )W , i"2, 3, 2, n. (17)
G G\ G\ G\ G\
The standard Ritz vector orthonormalization, (10) and (11) are used for each recurrence step.
The above-reviewed LDRV methods employ static recurrence procedures to generate the Ritz
vectors. Therefore, these LDRV methods are best suited only for low-frequency problems. For
higher-frequency or banded-frequency problems, these methods can lead to large sets of Ritz
vectors to span the higher-frequency con"guration space and thus increase computational expense.
When the Ritz vector set becomes too large, the Ritz vectors become nearly linearly dependent,
resulting in loss of accuracy and if severe enough, prohibiting the model reduction process from
producing any useful results.

3. The quasi-static Ritz vector method

The quasi-static Ritz vector (QSRV) method extends the previous LDRV methods by employing
a quasi-static recurrence procedure, based on the concept of the quasi-static completeness of the
Ritz basis. The basic idea is to let the Ritz vectors span the con"guration space at a desired
frequency or frequencies, e$ciently capturing possible dynamic deformations for the desired
frequency range. The "rst QSRV is chosen as a quasi-static mode corresponding to the loading
pattern G by solving the following quasi-static equilibrium equation:

WH"(K!uM)\G, (18)
 A
where u is denoted as the centering frequency in Ref. [8], since it usually is chosen at the midpoint
A
of the frequency range of interest. Applying normalization, the "rst QSRV is given by

W*
W "  . (19)
 (W*2MW*)
 
266 J. Gu et al. / Finite Elements in Analysis and Design 36 (2000) 261}278

For i"2, 3, 2, n, the quasi-static recurrence procedure is

WH"(K!uM)\MW , (20)
G  G\
G\
WM "WH! (W2MWH)W , (21)
G G H G H
H
WM
W" G . (22)
G (WM 2MWM )
G G
Physically, the starting QSRV, W , represents a normalized frequency response deformation

mode of the undamped system under the loading pattern G at the frequency u . That is, by using

a quasi-static solution, the dynamic e!ect of the loading or the inertial term neglected in the static
solution of Eq. (9), is included. If u is the only frequency content of the external load, then

W describes the exact steady state deformation mode of the structure. In general, assuming u is
 
a major frequency component in the frequency domain of interest, then W gives the most likely

deformation shape for the dynamic response of the structure. The next Ritz vector, WH, represents

the frequency response deformation mode shape due to the inertial force MW . After orthonormal-

ization, the new Ritz vector set, including W and W , spans a wider con"guration space for the
 
dynamic response of the structure. This quasi-static recurrence is continued until a given criterion
is satis"ed. Note that if u "0, then the QSRV algorithm reduces to the WYD algorithm. Thus, the

tuning parameter, u , extends the #exibility and generality of the LDRV method, which allows the
A
Ritz vectors to best represent both the deformation shape and the frequency content of the dynamic
response.
In the above procedure, only one centering frequency is employed. However, more than one
centering frequency can be employed in the QSRV algorithm. Using multiple centering frequencies
makes the technique more robust for general cases. For example, if two centering frequencies
u and u (two recurrence groups) are used, the quasi-static recurrence procedure proposed
 
above can be extended, viz., for i"1,

WH"(K!u M)\G. (23)

 A
For i"2, 3, 2, m,

WH"(K!u M)\MW . (24)

G A G\
For i"m#1,

WH "(K!u M)\G. (25)

K> A
For i"m#2, m#3,2, n,

WH"(K!u M)\MW , (26)

G A G\

 Damping is ignored for the sake of simplifying our discussions. However, all discussions can be extended easily to the
damped system.
 J. Gu et al. / Finite Elements in Analysis and Design 36 (2000) 261}278 267

where m denotes the number of Ritz vectors in the recurrence group corresponding to u ; n!m is
A
the number of Ritz vectors in the recurrence group corresponding to u ; mass orthonormaliz-
A
ation, Eqs. (10) and (11) is employed after each recurrence step.
The QSRV method with multiple centering frequencies is particularly well suited for problems
with widely distributed loading frequency components or with wide frequency ranges of interest.
Additionally, by deriving each new group of LDRVs with a di!erent centering frequency and
independent recurrence procedure, the method is able to span a wider con"guration space with
much less chance of orthogonality loss.

4. Ritz vector termination procedures

To determine how many LDRVs are necessary for a given problem, a participation factor, p ,
G
was de"ned by Wilson et al. [1] and Nour-Omid and Clough [3] to measure the signi"cance of one
particular Ritz vector, W , in the total response
G
p "W2G. (27)
G G
This participation factor is computed for each Ritz vector, and is used to terminate the vector
generation process. However, since it does not take dynamic e!ects into account, this is a purely
static measure and only suitable for low-frequency problems.
In this paper, a new measure for the participation factor is proposed, having the form
"W2s"
p" G , (28)
G [(W2W )(s2s)]
G G
where
s"(K!XM)\G (29)
is the frequency response due to the loading pattern, G, and X is a speci"ed frequency. At the
maximum value of one, the Ritz vector exactly matches the frequency response deformation shape,
while the minimum value of zero denotes that the Ritz vector is orthogonal to the frequency
response deformation and hence not likely to contribute signi"cantly to the reduced model
response. The frequency parameter, X, is chosen to represent a dominant frequency of the loading
pattern, G.
If the frequency content of loading is broad (multiple loading frequencies), the following
participation factor is suggested
"W2s "
p " G H , (30)
GH [(W2W )(s2s )]
G G H H
where
s "(K!XM)\G (31)
H H
in which X is a chosen excitation frequency of interest. This de"nition can be extended to the case
H
where there are multiple loading patterns, as discussed below.
268 J. Gu et al. / Finite Elements in Analysis and Design 36 (2000) 261}278

The new vector participation factor is a more realistic indication if the Ritz vector will participate
in the dynamic solution. Therefore, it is used to terminate the recurrence sequence when the
participation factor magnitude drops below a speci"ed tolerance value. The sequence of quasi-
static recurrence steps also terminates when either the Ritz vector generated is no longer numer-
ically orthogonal to the vectors previously obtained, or the number of the Ritz vectors found
exceeds the speci"ed limit.

5. Multiple loading case

For the case of multiple loading patterns (k'1, in Eq. (2)), Nour-Omid and Clough [4]
introduced a block Lanczos algorithm, based on the Lanczos algorithm discussed in a previous
section. The block Lanczos algorithm creates a block of (k) Lanczos vectors for k di!erent loading
patterns at each recurrence step. As such, orthogonalization is imposed not only among the blocks,
but also among di!erent Ritz vectors in each block. Therefore, the block Lanczos method is
essentially an extension of the simple Lanczos method, where each set of LDRVs is replaced by
a block of LDRVs. The block Ritz vector algorithm has also been implemented into the WYD and
LWYD algorithms. However, due to the size of the Ritz vector set required by the multiple
loadings, and the use of the same static equilibrium condition, these algorithms induced much more
linear dependence than the single loading case.
The blocking idea can be employed to extend the QSRV algorithm for multiple loading cases.
Generally, there will be a block of LDRVs for multiple loading patterns at each recurrence step.
For each loading pattern, the LDRVs generated by the quasi-static recurrence procedure around
the centering frequency of that loading are in a recurrence group. For exposition, assume there are
two loading patterns, G and G , with their major frequency components of X and X ,
   
respectively. Choosing u "X and u "X , a block QSRV algorithm can be described as: For
A  A 
block number i"1,

WH "(K!u M)\G , k"1, 2, (32)

) AI I
WH
W "  , (33)
  (WH2 MWH )
 
W "WH !(W2 MWH )W , (34)
    
WM
W "  . (35)
  (WM 2 MWM )
 
For block number i"2, 3,2, n,

WH "(K!u M)\MW , k"1, 2, (36)

GI AI G\ I
G\ G\
W "WH ! (W2 MWH )W ! (W2 MWH )W , (37)
G G H G H H G H
H H
 J. Gu et al. / Finite Elements in Analysis and Design 36 (2000) 261}278 269

WM
W " G , (38)
G  (WM 2 MWM )
G G
G G\
WM "WH ! (W2 MWH )W ! (W2 MWH )W , (39)
G G H G H H G H
H H
WM
W " G . (40)
G  (WM 2 MWM )
G G
The number of LDRVs in each block or recurrence group can be determined with the help of vector
participation factor introduced in Eq. (30).
The QSRV method gives more #exibility than the previous block methods for multiple loading
problems, and the number of vectors for di!erent loading patterns need not be equal. In addition,
since the LDRVs in di!erent groups are generated using di!erent centering frequencies, the Ritz
vectors have better orthogonality properties, which make the solution process more robust and
stable than other block Ritz vector methods.

6. Numerical examples

Fig. 1 depicts a cantilever plane truss structure with 25 repeated structures excited by two
separate external loadings. The excitation f indicates the driving point location for the calculation

of the system's driving point response, while the second excitation f is set to zero for the single

loading cases. Table 1 lists some of the structure's system frequencies of interest. Both frequency
response analyses and transient response analyses were performed to compare the e$cacy of the
quasi-static Ritz vectors, the exact eigenvectors (normal modes) and the standard Ritz vectors in
structural dynamics analyses.
For all cases, the reference solution was obtained by direct solution of the underlying problem.
The relative response error is de"ned by
"R !R "
E "   ;100% (41)
 max("R ", R )
 
where R is the solution computed using one of the model reduction methods, and R is the
 
reference solution. To prevent exaggeration of the error when the reference solution value is small,

Fig. 1. Cantilever truss structure.

270 J. Gu et al. / Finite Elements in Analysis and Design 36 (2000) 261}278

Table 1
System frequencies, in Hz, with values around the frequency bands of interest

Mode no. 1 2 3 4 5 * 10 11 12 13 14
Frequency 2.03 12.5 34.0 50.5 64.1 * 238.8 251.1 289.5 341.0 349.7

Mode no. 15 * 32 33 34 35 36 37 38 39 40
Frequency 392.7 * 847.5 860.6 868.5 875.0 875.9 889.4 910.3 937.2 941.1

R is equal to 0.01 times the static response for frequency response analyses or the time-averaged

response for transient response analyses.

6.1. Frequency response analysis

For purpose of exposition, two frequency bands with single loading f are studied. The "rst

frequency range of interest is a low-pass band from 0 to 50 Hz. All 5 system normal modes
under a cut-o! frequency of 75 Hz were included in the calculations for the mode}displacement
(M}D) method. Recurrence termination leads to 6 vectors for all Ritz vector methods. The
centering frequency (CF) is selected at the center of the frequency range, CF"25 Hz, for the QSRV
method with one CF (QSRV 1CF). For the two CF method (QSRV 2CF), CF "10 Hz, and

CF "40 Hz.

Fig. 2 shows the driving point response errors for di!erent methods. It is clear that the
WYD and Lanczos algorithms are more accurate in the low-frequency range near 0 Hz,
while the QSRV methods always provide much more accurate results in the frequency ranges
around the CF. This demonstrates that the QSRV method is dependent on the choice of the CF. If
the response around a particular frequency is of major concern, the WYD and Lanczos algorithms
can always be improved using the QSRV method by choosing a CF close to the frequency value of
interest.
The second frequency band was chosen as a mid-frequency band from 850 to 900 Hz. All 40
system normal modes between 0 and 950 Hz are retained in the M}D computation. For the QSRV
1CF, CF "875 Hz; for the QSRV 3CF, CF "860 Hz, CF "875 Hz, CF "890 Hz. Based on
  
the terminating conditions, there are 8 (QSRV 1CF) and 9 (QSRV 3CF) Ritz vectors in the QSRV
methods, and 35 vectors in the WYD, Lanczos and LWYD methods.
From Fig. 3, the improvement of QSRV methods around the CF is clear. For the WYD, Lanczos
and LWYD algorithms, accuracy is relatively low although they employ more Ritz vectors than the
QSRV methods. In contrast, if the number of vectors in these standard methods is set to that used
by the QSRV method, the accuracy will be much worse due to the insu$cient Ritz basis.
Fig. 4 shows the frequency response errors obtained by employing 50 Ritz vectors for the WYD,
Lanczos and LWYD methods, obtained by suspending the other recurrence terminating condi-
tions. These methods are still inferior to the QSRV method (8 Ritz vectors). The WYD algorithm is
quite unstable in this frequency band because of the linear dependence of its Ritz vectors.
Therefore, the WYD, Lanczos and LWYD methods are far less robust than the QSRV method for
mid-frequency or high-frequency problems.
 J. Gu et al. / Finite Elements in Analysis and Design 36 (2000) 261}278 271

Fig. 2. Frequency response errors for di!erent methods in 0}50 Hz band.

Fig. 3. Frequency response errors for di!erent methods in 850}900 Hz band, 35 vectors for standard methods.
272 J. Gu et al. / Finite Elements in Analysis and Design 36 (2000) 261}278

Fig. 4. Frequency response errors for di!erent methods in 850}900 Hz band, 50 vectors for standard methods.

6.2. Transient response analysis

Three cases of transient response analysis are studied, single loading for the "rst two cases and
multiple loading for the third case. The "rst case is low-frequency excitation in which f is

a harmonic force of 25 Hz frequency; the driving point responses from 0 to 2.0 s are calculated by
di!erent methods. For this case, the mode}displacement method includes all 4 system normal
modes below 50 Hz. To provide a fair comparison of accuracy at a "xed cost, the same number of
(4) Ritz vectors are employed for all Ritz vector methods. Due to the assumption of no damping,
the transient response is not suppressed, thus both transient part and steady-state part exist in the
total response. As such, two centering frequencies with CF "0 Hz and CF "25 Hz are em-
 
ployed for the QSRV method with two sets of Ritz vectors for each CF, respectively.
Figs. 5 and 6 show the computed driving point displacement histories, between 1.5 and 2.0 s,
using sine and cosine functions for the loading, respectively. Compared with the M}D method,
both LWYD and QSRV 2CF methods are much closer to the reference solution. Therefore, for
low-frequency transient response problems, the WYD or Lanczos method can always be improved
either by using the LWYD method (however, more computational cost is required since additional
sets of temporary vectors need to be computed and orthogonalized), or by using the QSRV method
and by choosing the "rst CF at 0 Hz and the other CF near the excitation frequency.
Mid-frequency excitation (300 Hz for the harmonic force f ) was employed in the second case.

All system normal modes in the range from 0 Hz to a high cut-o! frequency are included in
 J. Gu et al. / Finite Elements in Analysis and Design 36 (2000) 261}278 273

Fig. 5. Driving point responses, between 1.5 and 2.0 s, for di!erent methods under 25 Hz excitation (sine function).

Fig. 6. Driving point responses, between 1.5 and 2.0 s, for di!erent methods under 25 Hz excitation (cosine function).
274 J. Gu et al. / Finite Elements in Analysis and Design 36 (2000) 261}278

Fig. 7. Maximum driving point response errors for di!erent methods using di!erent number of vectors under 300 Hz
excitation.

calculations for the M}D method. To compare the convergence characteristics of di!erent
methods, the transient response analyses are performed from 0 to 0.75 s using various numbers of
Ritz vectors or eigenvectors. For QSRV 2CF, CF "0 Hz and CF "300 Hz.
 
In Figs. 7 and 8, direct comparisons among the maximum and average driving point response
errors of those methods are performed. These two "gures clearly illustrate the e!ectiveness of the
quasi-static Ritz vectors in the convergence and reduction of errors compared with the standard
Ritz vectors. For this case, the results using 8 quasi-static Ritz vectors are comparable to the results
obtained using more than 14 standard Ritz vectors. It is clear that the response of the structure can
be obtained quite accurately with as few as 13 quasi-static Ritz vectors (7 vectors for CF , 6 vectors

for CF ). The convergence behavior of the new approach is much faster than convergence of the

standard methods due to the fact that, the QSRV 2CF method satis"es both static completeness
and quasi-static completeness requirements of the Ritz vector space by using two recurrence
groups of LDRVs with two CF.
For mid- or high-frequency problems, the standard Ritz vector methods are unstable since they
may provide entirely inadequate results with certain number of Ritz vectors. For example,
Fig. 9 shows part of the driving point response for di!erent Ritz vector methods using 13 vectors. It
can be seen that the response magnitude computed by the LWYD method is in error.
In the third case, the structure was subject to multiple loadings, in which f and f are equal
 
magnitude harmonic forces of 350 and 500 Hz, respectively. The driving point responses from 0 to
 J. Gu et al. / Finite Elements in Analysis and Design 36 (2000) 261}278 275

Fig. 8. Average driving point response errors for di!erent methods using di!erent number of vectors under 300 Hz excitation.

Fig. 9. Driving point responses, between 0.42 and 0.50 s, for di!erent methods under 300 Hz excitation.
276 J. Gu et al. / Finite Elements in Analysis and Design 36 (2000) 261}278

Fig. 10. Maximum driving point response errors for di!erent methods using di!erent number of vectors under 350 and
500 Hz excitations.

0.75 s are calculated by di!erent methods with various number of vectors. For the block WYD,
Lanczos and LWYD methods, there are two sets of LDRVs corresponding to the two loading
patterns in each block. As such, the total number of vectors for these standard block methods is
always even. The QSRV method is applied in four recurrence groups as described by Eqs. (32)}(40)
where CF "0 Hz, CF "350 Hz and CF "500 Hz.
  
In Figs. 10 and 11, the respective maximum and average driving point response errors of the
methods using various number of vectors are shown. Again, the QSRV method achieves the best
performance among the LDRV methods. It is clear that the standard block methods become
unstable as the total number of vectors increases, as evidenced by the oscillations in error of the
block WYD and Lanczos methods as a function of the number of vectors. The block LWYD
method fails to give any solution due to heavy loss of orthogonality when there are more than 12
LDRV included. While the convergence characteristics of the quasi-static Ritz vectors are similar
to those of the exact eigenvectors, the cost of computing eigenvectors is more expensive than that
for the quasi-static Ritz vectors.

7. Conclusions

A quasi-static recurrence procedure, denoted as the quasi-static Ritz vector (QSRV) algorithm
was presented, which provides a more energetic basis than existing load-dependent Ritz vector
 J. Gu et al. / Finite Elements in Analysis and Design 36 (2000) 261}278 277

Fig. 11. Average driving point response errors for di!erent methods using di!erent number of vectors under 350 and
500 Hz excitations.

methods, for structural dynamics analysis. A centering frequency parameter controls the dynamic
range and behavior of the new approach. The QSRV method can employ more than one centering
frequency and more than one set of loading patterns, using a blocking strategy. These techniques
further improve the accuracy and robustness of the QSRV method. The e!ectiveness of the QSRV
method was demonstrated using frequency response analysis and transient response analysis exam-
ples. Thus, the QSRV provides an attractive alternative to standard load-dependent Ritz vector and
mode}displacement model reduction procedures for e$cient structural dynamics analysis.

Acknowledgements

The authors gratefully acknowledge support of this research by the US Army Tank-Automotive
Research, Development and Engineering Center, through the Automotive Research Center, a US
Army Center of Excellence. The authors also thank Prof. K.K. Choi for suggesting we take a look
at Ref. [9] and the application of Ritz vectors, which provided the impetus for this work.

References

[1] E.L. Wilson, M.W. Yuan, J.M. Dickens, Dynamic analysis by direct superposition of Ritz vectors, Earthquake Eng.
Struct. Dyn. 10 (1982) 813}821.
278 J. Gu et al. / Finite Elements in Analysis and Design 36 (2000) 261}278

[2] B.N. Parlett, The Symmetric Eigenvalue Problem, Prentice-Hall, Englewood Cli!s, NJ, 1980.
[3] B. Nour-Omid, R.W. Clough, Dynamic analysis of structures using Lanczos co-ordinates, Earthquake Eng. Struct.
Dyn. 12 (1984) 565}577.
[4] B. Nour-Omid, R.W. Clough, Block Lanczos method for dynamic analysis of structures, Earthquake Eng. Struct.
Dyn. 13 (1985) 271}275.
[5] P. Leger, E.L. Wilson, R.W. Clough, The use of load dependent vectors for dynamic and earthquake analyses,
Earthquake Engineering Research Center Report, University of California, Berkeley, UCB/EERC-86/04, 1986.
[6] P. Leger, Load-dependent subspace reduction methods for structural dynamic computations, Comput. Struct. 29
(1988) 993}999.
[7] C. Lanczos, An iteration method for the solution of the eigenvalue problem of linear di!erential and integral
operators, J. Res. Natl. Bur. Stand. 45 (1950) 255}282.
[8] Z.-D. Ma, I. Hagiwara, Improved mode-superposition technique for modal frequency response analysis of coupled
acoustic-structural systems, AIAA J. 29 (1991) 1720}1726.
[9] S. Wang, K.K. Choi, Continuum design sensitivity analysis of eigenvectors using Ritz vectors, J. Aircraft 31 (1994)
457}459.

View publication stats