A Computationally Efficient

Ground-Motion Selection Algorithm for

Matching a Target Response Spectrum
Mean and Variance
Nirmal Jayaram,a) Ting Lin,a) M.EERI, and Jack W. Baker,a) M.EERI

Dynamic structural analysis often requires the selection of input ground

motions with a target mean response spectrum. The variance of the target
response spectrum is usually ignored or accounted for in an ad hoc manner,
which can bias the structural response estimates. This manuscript proposes a
computationally efficient and theoretically consistent algorithm to select ground
motions that match the target response spectrum mean and variance. The selec-
tion algorithm probabilistically generates multiple response spectra from a target
distribution, and then selects recorded ground motions whose response spectra
individually match the simulated response spectra. A greedy optimization tech-
nique further improves the match between the target and the sample means and
variances. The proposed algorithm is used to select ground motions for the anal-
ysis of sample structures in order to assess the impact of considering ground-
motion variance on the structural response estimates. The implications for code-
based design and performance-based earthquake engineering are discussed.
[DOI: 10.1193/1.3608002]

INTRODUCTION
Dynamic structural analysis is commonly used in performance-based earthquake engi-
neering to predict the response of a structure subjected to earthquake ground motions. In the
past, a wide variety of techniques have been developed for selecting input ground motions
for the structural analysis (e.g., Haselton et al. 2009; Katsanos 2010). One commonly used
approach is to select recorded or simulated ground motions whose response spectra match a
target mean response spectrum (e.g., Beyer and Bommer 2007; Shantz 2006; Watson-
Lamprey and Abrahamson 2006). Commonly used target spectra include the uniform haz-
ard spectrum and more recently, the conditional mean spectrum (Baker 2011). Irrespective
of the procedure used to obtain a target response spectrum, there are several methods for
selecting input ground motions that match a desired spectrum. One approach is to select
ground motions (sometimes after scaling) that individually deviate the least from the target
response spectrum. The deviation can be measured using the sum of squared differences
between the response spectrum of the record and the target response spectrum (e.g., Youngs
et al. 2007). An alternate approach is to select a ground-motion set, rather than one record
at a time, by minimizing the mean spectrum of the selected records from the target response
spectrum. This is a more complicated subset selection optimization problem, and this opti-
mization has been carried out in the past using genetic algorithms (Naeim et al. 2004).

a)
Dept. of Civil and Environmental Engineering, Stanford University, Stanford, CA 94305-4020, USA

797
Earthquake Spectra, Volume 27, No. 3, pages 797–815, August 2011; V
C 2011, Earthquake Engineering Research Institute
798 N. JAYARAM, T. LIN, AND J. W. BAKER

In some situations, matching only a target mean response spectrum is not sufficient
since the approach ignores the inherent variance that may exist in the response spectrum.
For instance, the conditional mean spectrum is derived by conditioning on spectral accelera-
tion at only a single period, Sa ðT  Þ, so the response spectra at other periods have variance.
This is discussed in more detail in a subsequent section. Another approach might be to
choose ground motions associated with a scenario earthquake (having a specified magni-
tude, distance, etc.), but not conditioned on any spectral values (e.g., Jayaram and Baker
2010). In that case, the response spectrum will have variance at all periods. The commonly
used uniform hazard spectrum (UHS), on the other hand, is an envelope of the spectral
accelerations at all periods that are exceeded with a specified rate, as computed using proba-
bilistic seismic hazard analysis (PSHA). PSHA already accounts for variability in spectral
accelerations at each period being considered, and construction of the UHS is a conservative
method of combining these spectral values. No further variance needs to be applied to the
UHS, as varying the spectral values is equivalent to varying the associated exceedance rate
of the spectral accelerations from period to period (an operation which is unlikely to pro-
duce meaningful results). In other words, if the UHS is used as the target mean spectrum,
the ground motions should be chosen so that their spectra match the UHS and not have
additional variance about the UHS.
In cases where a response spectrum has a non-zero variance, estimates of structural
response obtained using the ground motions selected only based on the target mean spec-
trum will show smaller than “actual” dispersion (logarithmic standard deviation) and may
also have a biased mean value. Selecting ground motions based only on a target response
spectrum mean is computationally inexpensive since it can be done by choosing ground
motions whose response spectra individually deviate the least from the target, and so the
search time is proportional to the number of ground motions considered. When matching a
target response spectrum mean and variance, it does not suffice to treat ground motions indi-
vidually, but rather requires comparisons of the mean and variance of sets of ground
motions to the target values. That is, the suitability of a particular ground motion can only
be determined in the context of the complete ground-motion set in which it might be
included. There is generally an intractably large number of possible ground-motion sets,
and so identifying the best set is a computationally expensive combinatorial optimization
problem. No automated procedures are available in the literature, however, to select ground
motions that match the response spectrum mean and variance. One notable work in this
regard is that of Kottke and Rathje (2008), who proposed a semi-automated procedure that
first selects ground motions based on matching the mean spectrum, and subsequently
applies individual scale factors on the ground motions to achieve the target variance. This
technique, however, does not easily scale to work with large ground-motion datasets, and
also cannot be used for the selection of unscaled ground motions.
The current work proposes a new, computationally fast and theoretically consistent
ground-motion selection algorithm to match the target response spectrum mean and
variance. The selection algorithm first uses Monte Carlo simulation to probabilistically gen-
erate multiple response spectra from a distribution parameterized by the target means and
variances. For each simulated response spectrum, a ground motion with a similar response
spectrum is then selected. Since the Monte Carlo simulated response spectra have the
desired mean and variance, the response spectra of the selected recorded ground motions
A COMPUTATIONALLY EFFICIENT GROUND-MOTION SELECTION ALGORITHM 799

will also have the desired mean and variance. A greedy optimization technique then further
improves the match between the target and the sample means and variances. This step repla-
ces one previously selected ground motion at a time with a record from the ground-motion
database that causes the best improvement in the match between the target and the sample
means and variances.
The proposed algorithm is then used to select ground motions for estimating the seismic
response of sample single-degree-of-freedom (SDOF) and multiple-degree-of-freedom
(MDOF) structures, in order to assess the impact of considering ground-motion variance on
the structural response estimates. The implications for code-based design and performance-
based earthquake engineering are discussed.

GROUND-MOTION SELECTION ALGORITHM

The objective of the proposed algorithm is to select a suite of ground motions whose
response spectra have a specified mean and variance. This algorithm is based on the empiri-
cally verified observation that the set of logarithmic spectral accelerations (lnSa) at various
periods is a random vector that follows a multivariate normal distribution (Jayaram and
Baker 2008). The first step in this algorithm is to parameterize the multivariate normal dis-
tribution of lnSa’s at multiple periods. The parameters of the multivariate normal distribu-
tion are the means and the variances of the lnSa’s at all periods and the correlations between
the lnSa’s at all pairs of periods. Equivalently, the distribution can be parameterized using
the means of the lnSa’s and the covariances between the lnSa’s at all pairs of periods. In
order to achieve the desired properties in the selected ground motions, these parameters
should be set to their target values (i.e., target means and variances for the ground motions
to be selected). A subsequent section illustrates this parameterization.
Once the distribution means and covariances are set equal to the desired target values,
Monte Carlo simulation is used to probabilistically generate response spectra from the
above mentioned multivariate normal distribution. This can be performed using a standard
function in many programming languages. The number of response spectra to be simulated
equals the desired number of ground motions. For each simulated response spectrum, a
ground motion with a similar response spectrum is then selected. The similarity between a
ground-motion response spectrum and a Monte Carlo simulated response spectrum is eval-
uated using the sum of squared errors (SSE) described below:
P 
X 2
SSE ¼ ln Sa ðTj Þ  ln SaðsÞ ðTj Þ (1)
j¼1

where ln Sa ðTj Þ is the logarithmic spectral acceleration of the (optionally scaled) ground
motion in consideration at period Tj , ln SaðsÞ ðTj Þ is the target lnSa at period Tj from the simu-
lated response spectrum, p is the number of periods considered and SSE is the sum of
squared errors, which is a measure of dissimilarity. The measure of similarity defined by
Equation 1 is not unique, and discussion of other measures of similarity can be found in
Beyer and Bommer (2007) and Buratti et al. (2011). The selection is done by computing
SSE for each ground motion in the database, and then choosing the ground motion having
the smallest SSE. Other ground motion properties can be accounted for at this stage by, for
800 N. JAYARAM, T. LIN, AND J. W. BAKER

example, considering only ground motions falling within a specified range of magnitudes
and distances. Note that this is identical to comparison procedures in use today, except that
here we are comparing to simulated spectra rather than a target mean spectrum.
The mean and the variance of the simulated response spectra will approximately match
the corresponding target values because they were sampled from the desired distribution.
This match will be nearly exact if a large number of spectra are simulated and will be ap-
proximate otherwise. Since the simulated response spectra have approximately the desired
mean and variance, the response spectra selected using this approach will also have
approximately the desired mean and variance. Additionally, this ground-motion selection
approach also ensures that the selected set has the target correlation structure (i.e., correla-
tion between lnSa’s at pairs of periods) specified while parameterizing the distribution of
the response spectrum. This implies that, in the particular case where the logarithmic
response spectrum follows a multivariate normal distribution, the proposed algorithm
actually matches the entire response spectrum distribution. Another advantage of this
approach is that this algorithm allows the selection of unscaled ground motions (Jayaram
and Baker 2010).
As mentioned above, when ground motions are selected using the approach described
above, the sample means and variances may deviate slightly from the target values, partic-
ularly when the number of ground motions selected is small. Therefore, a “greedy” opti-
mization technique is used to further improve the match between the sample and the tar-
get means and variances. In this approach, each ground motion selected previously is
replaced one at a time with a ground motion from the database that causes the best
improvement in the match between the target and the sample means and variances. If
none of the potential replacements causes an improvement, the original ground motion is
retained. The mismatch is estimated as the sum of squared differences between the target
and the sample means and variances over the period range of interest. The deviation of
the set mean and variance from the target mean and variance (denoted SSEs ) is estimated
as follows:
X p  2  2 
ðtÞ ðtÞ
SSEs ¼ ^ ln Sa ðTj Þ  lln Sa ðTj Þ þw ^sln Sa ðTj Þ  rln Sa ðTj Þ
m (2)
j¼1

where SSEs is the sum of squared errors of the set, which is the parameter to be minimized,
ðtÞ
^ ln Sa ðTj Þ is the set mean lnSa at period Tj , lln Sa ðTj Þ is the target mean lnSa at period Tj ,
m
ðtÞ
^sln Sa ðTj Þ is the set standard deviation of the lnSa at period Tj , rln Sa ðTj Þ is the target standard
deviation of the lnSa at period Tj , w is a weighting factor indicating the relative importance
of the errors in the standard deviation and the mean (A possible value for w is 1, but it can
be chosen depending on the desired accuracy in the match between the sample and the tar-
get means and standard deviations), and p is the number of periods (Tj ) at which the error is
computed.
The set mean and standard deviation can be calculated as follows:

1X n
^ ln Sa ðTj Þ ¼
m ln Sai ðTj Þ (3)
n i¼1
A COMPUTATIONALLY EFFICIENT GROUND-MOTION SELECTION ALGORITHM 801

sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
n  2
1 X
^sln Sa ðTj Þ ¼ ln Sai ðTj Þ  m ^ ln Sa ðTj Þ (4)
n  1 i¼1

where ln Sai ðTj Þ denotes the lnSa of the ith record in the set at period Tj , and n denotes the
number of records in the set.
Note that the greedy optimization technique does not explicitly account for the correla-
tion structure of selected sets. This correction structure is captured in the initial selection
step, and is approximately retained after the greedy optimization as well.
The steps involved in the greedy optimization technique are summarized below.
• Step 1: Set j ¼ 1.
• Step 2: Set i ¼ 1. Denote the SSEs of the set as SSEs;old
• Step 3: If the ith database ground motion (Gi) is not already present in the set,
replace the jth ground motion in the set with Gi. Compute SSEs;i (i.e, the SSEs of
the set after the replacement is carried out).
• Step 4: Reverse the replacement carried out in Step 3. Increment i by 1.
• Step 5: If i is less than or equal to the size of the ground-motion database, go to
Step 3. Otherwise, identify the ground motion ~i that results in the minimum value
of SSEs;~i . If SSEs;~i < SSEs;old , replace the jth ground motion in the set with the ~i th
ground motion in the database.
• Step 6: Increment j by 1. If j is less than the size of the set, go to Step 2. Otherwise,
terminate the algorithm.
This is called a “greedy” optimization technique because it tries to improve the match
between the target and the sample properties in every iteration locally without necessarily
achieving a global optimum solution. In this application, the initial simulation and selection
steps result in a ground motion set that is already approximately optimal (for reasonably
large sets), so it has been observed that only this greedy technique is needed to find solu-
tions that are essentially globally optimal. Observational experience suggests that this algo-
rithm never produces sets of ground motions with poor matches between the sample and the
target means and variances (even for sets with as few as 10 ground motions, as illustrated in
a subsequent section).
The Appendix “An Alternate Ground-Motion Selection Algorithm” describes an alter-
nate selection algorithm that does not require knowledge of the response spectrum distribu-
tion or the correlation structure.

ILLUSTRATIVE GROUND-MOTION SELECTION

This section describes the application of the proposed algorithm for selecting structure-
specific ground motions that have a specified spectral acceleration at the structure’s funda-
mental period. In this example, the target response spectrum mean and covariance matrices
are obtained using the conditional mean spectrum (CMS) method (Baker 2011), which pro-
vides the mean and variance (and correlations) of the response spectrum conditioned on the
specified spectral acceleration. It is to be noted that while this example uses the targets from
802 N. JAYARAM, T. LIN, AND J. W. BAKER

the CMS method, the proposed algorithm can be used with any arbitrary target mean and
covariance (e.g., Jayaram and Baker 2010).
PARAMETERIZATION OF THE TARGET RESPONSE SPECTRUM DISTRIBUTION
As described in the previous section, the first step in the algorithm is to parameterize the
multivariate normal distribution of the lnSa’s using the means and the variances of the spec-
tral accelerations (chosen to equal the target mean and the target variance respectively) and
the correlations between the spectral accelerations at two different periods. The steps
involved in parameterizing the distribution using the CMS method are listed below.
• Step 1: Determine the target spectral acceleration (Sa) at a given period T* (e.g., the
fundamental period of the structure), and the associated magnitude (M), distance to
source (R) and e(T*), where e(T*) is the number of standard deviations by which a
given lnSa differs from the mean predicted (by a ground-motion model) lnSa at the
period of interest T*. In general,
ln Sa ðT Þ  lln Sa ðT Þ
eðT Þ ¼ (5)
rln Sa ðT Þ

where ln Sa ðT Þ is the ground motion’s logarithmic spectral acceleration at period T,

and lln Sa ðT Þ and rln Sa ðT Þ are the predicted mean and standard deviation, respec-
tively, of ln Sa ðT Þ given M, R, etc. (e.g., Campbell and Bozorgnia 2008). The val-
ues of M, R and e(T*) can be obtained from deaggregation (e.g., USGS 2008).
• Step 2: For all Tj of interest, compute the unconditional mean and the unconditional
standard deviation of the response spectrum, given M and R. In other words, com-
pute lln Sa ðT Þ and rln Sa ðT Þ .
• Step 3: Compute the mean of ðln Sa ðT1 Þ; ln Sa ðT2 Þ; :::; ln Sa ðTn ÞÞ conditioned on
e(T*). This mean matrix (denoted l) is computed as follows:
2 3
llnSa ðT1 Þ þ qðT1 ; T  ÞeðT  ÞrlnSa ðT1 Þ
6 llnSa ðT2 Þ þ qðT2 ; T  ÞeðT  ÞrlnSa ðT2 Þ 7
6 7
l¼6 6 : 7
7 (6)
4 : 5
llnSa ðTn Þ þ qðTn ; T  ÞeðT  ÞrlnSa ðTn Þ

where q(Tj,T*) is the correlation between e(Tj) and e(T*) provided by, for
instance, Baker and Jayaram (2008).
• Step 4: Compute the covariance of ðln Sa ðT1 Þ; ln Sa ðT2 Þ; :::; ln Sa ðTn ÞÞ condi-
tioned on e(T*). This covariance matrix (denoted R) is estimated as follows:
Let R0 denote the (unconditional) covariance matrix of the vector
ðln Sa ðT1 Þ; ln Sa ðT2 Þ; :::; ln Sa ðTn ÞÞ.
2 3
r2lnSa ðT1 Þ qðT1 ;T2 ÞrlnSa ðT1 Þ rlnSa ðT2 Þ  qðT1 ;Tn ÞrlnSa ðT1 Þ rlnSa ðTn Þ
6 7
6 qðT2 ;T1 ÞrlnSa ðT2 Þ rlnSa ðT1 Þ r2lnSa ðT2 Þ  qðT2 ;Tn ÞrlnSa ðT2 Þ rlnSa ðTn Þ 7
6 7
R0 ¼ 6 : :  : 7 (7)
6 7
4 : :  : 5
2
qðTn ;T1 ÞrlnSa ðTn Þ rlnSa ðT1 Þ qðTn ;T2 ÞrlnSa ðTn Þ rlnSa ðT2 Þ  rlnSa ðTn Þ
A COMPUTATIONALLY EFFICIENT GROUND-MOTION SELECTION ALGORITHM 803

Let R1 denote the covariance between ðln Sa ðT1 Þ; ln Sa ðT2 Þ; :::; ln Sa ðTn ÞÞ and ln Sa ðT  Þ,
defined as follows:
2 3
qðT1 ; T  ÞrlnSa ðT1 Þ rlnSa ðT  Þ
6 qðT2 ; T  ÞrlnSa ðT2 Þ rlnSa ðT  Þ 7
6 7
R1 ¼ 66 : 7
7 (8)
4 : 5
qðTn ; T  ÞrlnSa ðTn Þ rlnSa ðT  Þ

The covariance matrix of ðln Sa ðT1 Þ; ln Sa ðT2 Þ; :::; ln Sa ðTn ÞÞ conditioned on ln Sa ðT  Þ can
be computed as follows (e.g., Johnson and Wichern 2007):
1
R ¼ R0  R1 R01 (9)
r2lnSa ðT  Þ

where R01 denotes the transpose of R1 . The conditional standard deviation of the lnSa’s is
the square root of the diagonals of R, also given by Equation 10.
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
rlnSa ðT ÞjlnSa ðT  Þ ¼ rlnSa ðT Þ 1  qðT ; T  Þ2 (10)

Figure 1 shows the target conditional response spectrum mean and standard deviation
obtained corresponding to magnitude ¼ 7, distance to the rupture ¼ 10km, T* ¼ 2.63s and
e(T*) ¼ 2. These values have been chosen to be compatible with ground-motion studies car-
ried out by Haselton et al. (2009). The unconditional lnSa means and standard deviations
corresponding to this scenario, llnSa ðTj Þ and rlnSa ðTj Þ , are obtained from the Campbell and
Bozorgnia (2008) ground-motion model. (Since lnSa’s at multiple periods follow a multi-
variate normal distribution, the exponential of the mean lnSa equals the median spectral
acceleration. This is why the axis of Figure 1 is labeled as “Median Sa.”)

RESPONSE SPECTRUM SIMULATION

Forty response spectra are simulated (using Monte Carlo simulation) by sampling from
a multivariate normal distribution with the mean and covariance matrices defined by Equa-
tions 6 and 9 for the target scenario described above. The response spectra are simulated at
20 periods logarithmically spaced between 0.05s and 10.0s, and are shown in Figure 2a. A
large period range is used to ensure a good match in the entire response spectrum that cov-
ers regions of higher modes and nonlinearity. Because individual spectra may vary while
still achieving a target mean and variance of the overall set, there is often little penalty in
considering a broad period range in this step.
Figure 1a compares the mean of the Monte Carlo simulated response spectra to the tar-
get mean. It can be seen that the mean values agree reasonably well. Figure 1b shows a rea-
sonable agreement between the standard deviation of the simulated lnSa values and the tar-
get standard deviation. The small deviation seen in these figures is because the sample
mean and standard deviation for moderately small sample sizes do not necessarily match
the target mean and standard deviation.
804 N. JAYARAM, T. LIN, AND J. W. BAKER

Figure 1. (a) Response spectrum mean and (b) response spectrum standard deviation.

SELECTION OF GROUND MOTIONS TO MATCH SIMULATED SPECTRA

Forty ground motions are selected from the Next Generation Attenuation (NGA) data-
base (Chiou et al. 2008) that individually match the forty response spectra simulated in the
previous step. For two-dimensional structural models, a single ground motion component is
required as an input for every time history analysis. (For three-dimensional structural mod-
els, two ground motion components can be selected by considering their geometric mean
response spectrum, as described in Jayaram and Baker 2010.) Here, each horizontal compo-
nent of a recording from the same station in the NGA database is treated separately as an
individual ground motion. No constraints on, for example, the magnitudes and distances of
the selected recordings are used, but such constraints are easily accommodated by simply
restricting the set of ground motions considered for selection. Prior to selection, each of the
available 7102 ground motions in the NGA database is scaled so that its Sa ðT  Þ matches the
target Sa ðT  Þ from the target mean spectrum (seen in Figure 1a) when T* is equal to 2.63s.
Figure 2b shows the response spectra of the selected ground motions. The sample and the
target means and standard deviations are shown in Figure 1. The figure shows that the sam-
ple and the target response spectrum mean and variance match reasonably well. Addition-
ally, the selected ground motion spectra also match the specified target correlation structure
(specified by the nondiagonal terms of the covariance matrix in Equation 9) reasonably
well, as indicated by a mean absolute error between the sample and the target correlations
of 0.12.
The computational time required for selecting the set of 40 ground motions is 10 sec-
onds using a MATLAB implementation on an 8GB RAM 2.33GHz quad core processor.
This computational efficiency allows for the algorithm to be optionally applied multiple
times if one wants several candidate sets to choose from. While selecting the ground
motions shown in Figure 2, we applied the algorithm multiple times (twenty times, in par-
ticular) to obtain multiple candidate ground-motion sets and chose the set with the minimum
value of SSE. This approach is also beneficial to work around situations where recorded
ground motion spectra that adequately match one or more of the simulated spectra are not
available.
A COMPUTATIONALLY EFFICIENT GROUND-MOTION SELECTION ALGORITHM 805

Figure 2. (a) Simulated response spectra, (b) response spectra of ground motions selected before
greedy optimization, and (c) response spectra of ground motions selected after greedy optimization.

GREEDY OPTIMIZATION TECHNIQUE

The greedy optimization technique is used to modify the ground-motion suite selected
in the previous step. The spectra of the selected ground motions are shown in Figure 2c.
The means and the standard deviations of the set are shown in Figure 1, and have a near per-
fect match with the target means and standard deviations. The mean absolute error between
the sample and the target correlations is 0.15.
In total, the computational time required to select the set of 40 ground motions from the
7102 available ground motions is about 180 seconds using a MATLAB implementation on
an 8GB RAM 2.33GHz quad core processor. A MATLAB implementation of the proposed
ground-motion selection algorithm can be downloaded from http://www.stanford.edu/
~bakerjw/gm_selection.html.

SELECTION OF A SMALLER NUMBER OF GROUND MOTIONS

To test the effectiveness of the algorithm in sampling smaller ground motion sets, it is
repeated to select a set of 10 ground motions for the scenario described earlier
(magnitude ¼ 7, distance to rupture ¼ 10km, T* ¼ 2.63s and e(T*) ¼ 2). The response
806 N. JAYARAM, T. LIN, AND J. W. BAKER

spectra of the selected records are shown in Figure 3a. The set means and standard devia-
tions are compared to the target means and standard deviations in Figures 3b and 3c. It can
be seen that the matches are good, illustrating the effectiveness of the algorithm in selecting
small sets of ground motions. The mean absolute error between the sample and the target
correlations is 0.17. The computational time required to select the set of 10 ground motions
is about 25 seconds using a MATLAB implementation on an 8GB RAM 2.33GHz quad
core processor. The computational time required for selecting the set of 10 ground motions
without using the greedy optimization technique is 4 seconds.

IMPACT OF MATCHING SPECTRUM VARIANCE ON

STRUCTURAL RESPONSE
Code-based structural design and performance-based earthquake engineering applica-
tions require statistics such as the mean (e.g., ASCE 2005) or the median and the dispersion
(e.g., ATC 2009) of the structural response. It is of interest in this section to evaluate the
impact of ground-motion selection considering a target response spectrum mean and var-
iance (as compared to considering only a target mean) on these statistics.

Figure 3. (a) Response spectra of 10 selected ground motions, (b) response spectrum mean,
and (c) response spectrum standard deviation.
A COMPUTATIONALLY EFFICIENT GROUND-MOTION SELECTION ALGORITHM 807

Figure 4. Response spectra of 40 selected ground motions for e ¼ 2 and T* ¼ 2.63s. (a) Using
Method 1: Match target response spectrum mean and (b) using Method 2: Match target response
spectrum mean and variance.

GROUND MOTION SELECTION

The ground motions used for evaluating structural response are selected using the
method described in the previous section for a target scenario with magnitude ¼ 7, distance
to rupture ¼ 10km, Vs30 ¼ 400m=s, and a strike-slip mechanism. The Campbell and
Bozorgnia (2008) ground-motion model is used to estimate the mean and variance of the
response spectrum. The values of e and period T* are varied to obtain multiple test scenar-
ios. Three typical e values of 0, 1 and 2 are considered. The structures considered in this
work have periods (T*) ranging between 0.5s and 2.63s.
In order to investigate the impact of matching response spectrum variance (Equation 9)
on the structural response statistics, sets of forty ground motions are selected using two meth-
ods: “Method 1” in which only the target mean is matched (a common approach in current
practice, e.g., Baker and Cornell 2006 and Method 300 in Haselton et al. 2009) and “Method
2” in which both the target mean and the target variance are matched using the approach pro-
posed here. The target response spectrum mean and covariance matrices are evaluated using
Equations 6 and 9 for each combination of e and T*. Figure 4 shows example response spec-
tra of ground motions selected using these two methods (for e ¼ 2 and T* ¼ 2.63s).

STRUCTURAL RESPONSE
This section describes the response of sample nonlinear single-degree-of-freedom
(SDOF) structures and multiple-degree-of-freedom (MDOF) buildings designed according
to modern building codes. In this work, we consider only maximum displacement for the
SDOF structures and maximum interstory drift ratio (MIDR) for the MDOF structures.

Description of Structural Systems

The SDOF structures considered in this work follow a non-deteriorating, bilinear force-
displacement relationship (Chopra 2007). They have T* ¼ 0.5s, 5% damping and post-
808 N. JAYARAM, T. LIN, AND J. W. BAKER

yielding stiffness equal to 10% of elastic stiffness. SDOF structures with R factors (the ratio
of the target spectral acceleration at the period of the structure, Sa ðT  Þ, to the yield spectral
acceleration ¼ x2 * yield displacement, where x is the structure’s fundamental circular fre-
quency) of 1, 4 and 8 are considered to study varying levels of non-linear behavior. The R
factor is controlled by varying the yield displacements of the SDOF structures relative to
the Sa ðT  Þ value obtained from the target spectrum. The SDOF structures are non-deterio-
rating systems, so structural collapse is not considered.
The MDOF structures used in this study were designed per modern building codes and
modeled utilizing the Open System for Earthquake Engineering Simulation (OpenSEES;
2007) by Haselton and Deierlein (2007). The structural models consider strength and stiff-
ness deterioration (Ibarra et al. 2005) unlike in the SDOF case. The designs for these build-
ings have been checked by practicing engineers as part of the Applied Technology Council
Project ATC-63 (FEMA 2009). They have also been used for previous extensive ground-
motion studies (Haselton et al. 2009). The two buildings used in the current study are a
four-story reinforced concrete moment frame structure with T* ¼ 0.94s, and a 20-story rein-
forced concrete moment frame structure with T* ¼ 2.63s. The buildings show deterioration,
and collapse is said to occur if dynamic instability (large increases in the drift for
small increases in the ground-motion intensity) is reached in the model (Haselton and
Deierlein 2007).

Response of SDOF Systems

Table 1 shows the mean, median and dispersion (dispersion refers to logarithmic stand-
ard deviation) of ductility ratios (spectral displacement divided by the yield displacement)
of the SDOF structures under the different ground-motion scenarios described earlier. The
ductility statistics are estimated using the two sets of 40 ground motions selected using
Method 1 (ground motions selected by matching only the target response spectrum mean)
and Method 2 (ground motions selected by matching the target response spectrum mean
and variance). It can be seen from Table 1 that the median ductilities are similar across the

Table 1. Ductility ratio of SDOF structure

Median Ductility Dispersion of Ductility Mean Ductility

e R Method 1 Method 2 Method 1 Method 2 Method 1 Method 2

0 1 1.00 1.00 0 0 1.00 1.00

4 3.93 3.76 0.24 0.31 4.21 4.18
8 10.76 9.97 0.28 0.42 10.82 10.74
1 1 1.00 1.00 0 0 1.00 1.00
4 3.55 3.35 0.22 0.33 3.79 3.93
8 8.04 8.16 0.28 0.46 8.57 9.46
2 1 1.00 1.00 0 0 1.00 1.00
4 3.27 3.04 0.19 0.28 3.39 3.34
8 6.90 7.44 0.24 0.41 7.34 7.98
A COMPUTATIONALLY EFFICIENT GROUND-MOTION SELECTION ALGORITHM 809

Figure 5. Distribution of the structural response of the SDOF structure corresponding to R ¼ 8

and e(T*) ¼ 1: (a) linear scale and (b) logarithmic scale.

two ground-motion selection methods, while the mean and the dispersion of the response
are higher in Method 2, when the ground-motion variance is considered. The higher disper-
sion of the response seen while using Method 2 is a result of considering the uncertainty in
the response spectra, which is ignored in Method 1. As expected, the increase in dispersion
is particularly significant at large R values when the structure behaves in a non-linear man-
ner. Note that there are no differences between the methods when R ¼ 1, because the
response is dependent only on Sa ðT  Þ, which is identical in both cases.
Figure 5 shows the fraction of response analyses that result in a ductility less than a
specified value for the SDOF structure with R ¼ 8 in the e ¼ 1 scenario, estimated using
Methods 1 and 2. This type of plot is referred to as an empirical cumulative distribution
function, or CDF. The CDFs intersect at a value of approximately 0.5 due to the similarity
in the median response in both cases. The CDF obtained using Method 2 is flatter with
heavier tails as a result of the larger dispersion observed in this case. As seen from Figure
5a, the upper tails of the CDFs are heavier than the lower tails. Since the mean response is
the area above the CDF (the mean of a random variable is the area under the complementary
CDF, which equals 1 - CDF), it can be visually observed that the difference in the heaviness
of the upper tails results in a larger mean value of the response in case of Method 2 as com-
pared to Method 1. This is a graphical evidence of the larger mean values reported earlier in
Table 1. Analytically, if the responses were to follow a lognormal distribution (a common
assumption in performance-based earthquake engineering), the properties of the lognormal
distribution will imply that a larger dispersion results in a larger mean for a fixed median,
which also explains the larger means observed in Method 2.

Response of MDOF Systems

Table 2 summarizes the maximum interstory drift ratio (MIDR) estimates for the MDOF
structures considered in this study under various ground-motion scenarios, estimated using
Methods 1 and 2. The distributions of responses are summarized using the probability of
collapse (i.e., counted fraction of responses indicating collapse) and the median and the dis-
persion of the non-collapse responses.
810 N. JAYARAM, T. LIN, AND J. W. BAKER

Table 2. Maximum interstory drift ratio (MIDR) of 20-story and 4-story moment frames

Median MIDR Dispersion of MIDR Collapse Probability

Building e Method 1 Method 2 Method 1 Method 2 Method 1 Method 2

20-story 0 0.0044 0.0043 0.18 0.32 0 0

moment 1 0.0096 0.0086 0.24 0.29 0 0
frame 2 0.0186 0.0196 0.25 0.43 0 0.05
4-story 0 0.0072 0.0072 0.09 0.09 0 0
moment 1 0.0137 0.0139 0.26 0.29 0 0
frame 2 0.0279 0.0237 0.28 0.46 0.10 0.20

From Table 2, it can be seen that, as observed in the SDOF case, the medians are similar
across Methods 1 and 2 in all the considered scenarios. The dispersions are larger, however,
when the ground-motion variance is considered in Method 2. The increase in the dispersion
also results in an increased probability of observing large values of structural response. This
can result in an increased probability of structural collapse while using Method 2, as evi-
denced, for example, when e ¼ 2 in Table 2.
Figure 6 shows the empirical CDF of the MIDR of the 20-story frame corresponding to
the e ¼ 2 ground-motion scenario. As seen in the SDOF case, the CDF obtained using
Method 2 is flatter and has heavier tails on account of larger dispersion. The maximum plot-
ted values of the CDFs differ from one, and the difference equals the probability of
collapse.
In summary, the response estimates for the SDOF and the MDOF structures across sev-
eral ground-motion scenarios show that the consideration of the response spectrum variance
while selecting ground motions does not significantly impact the median structural
response, but tends to increase the mean response and the dispersion in the response. The

Figure 6. Distribution of the structural response of the 20-story moment frame building corre-
sponding to e(T*) ¼ 2: (a) linear scale and (b) logarithmic scale.
A COMPUTATIONALLY EFFICIENT GROUND-MOTION SELECTION ALGORITHM 811

increased dispersion can result in more extreme responses, which can lead to a larger proba-
bility of structural collapse.
These example analysis cases serve to illustrate the potential importance of matching
response spectrum variance. More detailed investigations regarding the impact are impor-
tant, and will be carried out in the future.

IMPLICATIONS
Code-based design is often concerned with the average response of the structure (e.g.,
ASCE 2005). The average response is typically interpreted as the mean response, although
sometimes it is interpreted as the median. If median structural response is of interest, the
consideration of the response spectrum variance while selecting ground motions does not
have a significant impact in the limited investigation performed here. On the other hand, if
mean structural response is of interest, the consideration of the response spectrum variance
appears to increase the mean structural response and may thus impact code-based design
calculations.
Performance-based earthquake engineering (PBEE), in contrast, often requires knowl-
edge about the full distribution of structural response (ATC 2009). Matching target response
spectrum variance increases the dispersion of structural response, thereby affecting the dis-
tribution of structural response and consequently the damage state and loss estimation com-
putations in PBEE. The increase in the dispersion leads to higher and lower extremes of
structural response and the associated damage states and losses. The increased dispersion
can also lead to a larger probability of structural collapse. PBEE calculations will thus
almost certainly be affected by this issue.
In summary, the example analyses presented above and engineering intuition suggest
that the target response spectrum variance used when selecting ground motions has an
impact on the distribution of structural responses obtained from resulting dynamic analysis.
It appears that this is true for both code-based design checks and performance-based earth-
quake engineering analysis. Further study is needed to quantify the magnitude of these
impacts, and this new algorithm will facilitate such studies.

CONCLUSIONS
A computationally efficient, theoretically consistent ground-motion selection algorithm
was proposed to enable selection of a suite of ground motions whose response spectra have
a target mean and a target variance. The algorithm first uses Monte Carlo simulation to
probabilistically generate multiple realizations of response spectra from a target distribution,
and then selects recorded ground motions whose response spectra individually match the
simulated response spectra. A greedy optimization technique then further improves the
match between the target and the sample means and variances by replacing one previously
selected ground motion at a time with a record from the ground-motion database that causes
the best improvement in the match. It was shown empirically that this selection algorithm
selects ground motions whose response spectra have the target mean and variance.
The proposed algorithm was then used to select ground motions for estimating the seis-
mic response of sample single-degree-of-freedom (SDOF) and multiple-degree-of-freedom
812 N. JAYARAM, T. LIN, AND J. W. BAKER

(MDOF) structures, in order to assess the impact of considering response spectrum variance
on the structural response estimates. SDOF structures with different levels of nonlinearity
(as indicated by their R factors) were analyzed using the selected ground motions. It was
seen that considering the response spectrum variance does not significantly affect the result-
ing median response, but slightly increases the mean response and considerably increases
the dispersion (logarithmic standard deviation) of the response. The increase in the mean
and the dispersion is larger for more nonlinear SDOF structures. Two code-compliant
MDOF structures with heights of 4 and 20 stories were also analyzed using the selected
ground motions. As with the SDOF structures, it was seen that considering the response
spectrum variance does not significantly affect the median response but increases the disper-
sion of the response and the probability of observing collapse. These observations have
implications for applications where the dispersion of the response is an important considera-
tion, such as in many performance-based engineering evaluations. A MATLAB implemen-
tation of the proposed ground-motion selection algorithm can be downloaded from http://
www.stanford.edu/~bakerjw/gm_selection.html.

ACKNOWLEDGMENTS
The authors thank Curt Haselton, Tom Shantz, Nilesh Shome, Peter Stafford, and an
anonymous reviewer for their helpful reviews of the manuscript. Also, thanks to Curt
Haselton for providing the structural models used for the example analyses. This work was
supported by the State of California through the Transportation Systems Research Program
of the Pacific Earthquake Engineering Research Center (PEER), and by Cooperative Agree-
ment Number 08HQAG0115 from the United States Geological Survey. Any opinions,
findings, conclusions or recommendations expressed in this material are those of the authors
and do not necessarily reflect those of the funding agencies.

APPENDIX A: A GREEDY GROUND-MOTION SELECTION TECHNIQUE

The ground-motion selection algorithm described in the body of this manuscript selects
an initial set of ground motions whose response spectra match a set of simulated response
spectra. These simulations are obtained from a multivariate normal distribution parameter-
ized by the target mean and covariance matrices. A greedy optimization technique then fur-
ther improves the match between the target and the sample means and variances and obtains
the final set of ground motions.
Sometimes, it may not be possible to completely parameterize the distribution of the
response spectra using the mean and covariance information. This includes situations where
ground motions are selected to match the UHS (where only the mean spectrum needs to be
considered) or where the mean and the variance information, but not the correlation informa-
tion, are available. There may also be situations where the response spectrum does not follow
a multivariate normal distribution. For such situations, the authors propose the following tech-
nique for selecting the initial ground-motion set that can be subsequently improved by the
greedy optimization technique. The steps involved in the technique are summarized below.

• Step 1: Initialize the algorithm with an empty ground-motion set.

• Step 2: Set i ¼ 1.
A COMPUTATIONALLY EFFICIENT GROUND-MOTION SELECTION ALGORITHM 813

Figure A1. (a) Response spectra of 40 ground motions selected using the greedy selection and
optimization techniques. (b) Response spectrum mean and (c) response spectrum standard
deviation.

• Step 3: If the ith database ground motion (Gi) is not already present in the ground-
motion set, include it in the set and compute SSEs;i (i.e, the SSEs of the set after Gi
is included, where SSEs is defined in Equation 2).
• Step 4: Delete Gi from the set, if included in Step 3. Increment i by 1.
• Step 5: If i is less than or equal to the size of the ground-motion database, go
to Step 3. Otherwise, identify the ground motion ~i that results in the minimum
value of SSEs;~i . Add the ~i th ground motion in the database to the ground-
motion set.
• Step 6: If the size of the set equals the desired number of ground motions,
terminate the algorithm. Otherwise, go to Step 2.
This selection technique will provide a reasonable starting set of ground motions
that can be subsequently improved using the greedy optimization technique described ear-
lier in the manuscript. This selection technique does not take advantage of the knowledge of
the response spectrum distribution or the correlation structure, but is therefore more general
in its application. It is also empirically seen to produce sets of ground motions with
response spectrum mean and variance closely matching the corresponding target values.
814 N. JAYARAM, T. LIN, AND J. W. BAKER

To test the effectiveness of the technique, it is used to select a set of 40 ground motions
for the scenario described earlier (magnitude ¼ 7, distance to rupture ¼ 10km, T* ¼ 2.63s
and e(T*) ¼ 2). The response spectra of the selected records are shown in Figure A1a. The
ground-motion set means and standard deviations are compared to the target means and
standard deviations in Figures A1b and A1c. It can be seen that the matches are good, illus-
trating the effectiveness of the technique. Incidentally, despite the fact that the technique
does not use the correlation information, it is seen that the mean absolute error between the
sample and the target correlations (Equation 7) is only 0.15.

REFERENCES
American Society of Civil Engineers (ASCE), 2005. Minimum design loads for buildings and
other structures. ASCE 7-05, American Society of Civil Engineers=Structural Engineering
Institute, Reston, VA.
Applied Technology Council (ATC), 2009. Guidelines for Seismic Performance Assessment of
Buildings ATC-58 50% Draft, The Applied Technology Council, Redwood City, CA.
Baker, J. W., and Cornell, C. A., 2006. Spectral shape, epsilon and record selection, Earthquake
Engineering & Structural Dynamics 35, 1077–1095.
Baker, J. W., 2011. The conditional mean spectrum: A tool for ground motion selection, ASCE
Journal of Structural Engineering 137, 322–331.
Baker, J. W., and Jayaram, N., 2008. Correlation of spectral acceleration values from NGA
ground motion models, Earthquake Spectra 24, 299–317.
Beyer, K., and Bommer, J. J., 2007. Selection and scaling of real accelerograms for
bi-directional loading: A review of current practice and code provisions, Journal of Earth-
quake Engineering 11, 13–45.
Buratti, N., Stafford, P. J., and Bommer, J. J., 2011. Earthquake accelerogram selection and
scaling procedures for estimating the distribution of drift response, ASCE Journal of Struc-
tural Engineering 137, 345–357.
Campbell, K. W., and Bozorgnia, Y., 2008. NGA ground motion model for the geometric mean
horizontal component of PGA, PGV, PGD and 5% damped linear elastic response spectra for
periods ranging from 0.01 to 10s, Earthquake Spectra 24, 139–171.
Chiou, B. S. J., Darragh, R. B., Gregor, N. J., and Silva, W. J., 2008. NGA project strong-
motion database, Earthquake Spectra 24, 23–44.
Chopra, A. K., 2007. Dynamics of structures, Prentice Hall, Upper Saddle River, NJ.
Federal Emergency Management Agency (FEMA), 2009. Recommended Methodology for
Quantification of Building System Performance and Response Parameters, FEMA P695A,
Prepared for the Federal Emergency Management Agency, The Applied Technology Coun-
cil, Redwood City, CA.
Haselton, C. B., Baker, J. W., Bozorgnia, Y., Goulet, C. A., Kalkan, E., Luco, N., Shantz, T. J.,
Shome, N., Stewart, J. P., Tothong, P., Watson-Lamprey, J. A., and Zareian, F., 2009. Evalu-
ation of ground motion selection and modification methods: Predicting median interstory
drift response of buildings, Technical report, Report 2009=01, Pacific Earthquake Engineer-
ing Research Center, University of California, Berkeley, CA.
Haselton, C. B., and Deierlein, G. G., 2007. Assessing seismic collapse safety of modern
reinforced concrete frame buildings, Technical report, Report 2007=08, Pacific Earthquake
Engineering Research Center, University of California, Berkeley, CA.
A COMPUTATIONALLY EFFICIENT GROUND-MOTION SELECTION ALGORITHM 815

Ibarra, L. F., Medina, R. A., and Krawinkler, H., 2005. Hysteretic models that incorporate
strength and stiffness deterioration, Earthquake Engineering and Structural Dynamics 34,
1489–1511.
Jayaram, N., and Baker, J. W., 2010. Ground-motion selection for PEER Transportation Sys-
tems Research Program, in Proceedings, 7th CUEE and 5th ICEE Joint Conference, Tokyo,
Japan.
Jayaram, N., and Baker, J. W., 2008. Statistical tests of the joint distribution of spectral accelera-
tion values, Bulletin of the Seismological Society of America 98, 2231–2243.
Johnson, R. A., and Wichern, D. W., 2007. Applied Multivariate Statistical Analysis, Prentice
Hall, Upper Saddle River, NJ.
Katsanos, E. I., 2010. Selection of earthquake ground motion records: A state-of-the-art review
from a structural engineering perspective, Soil Dynamics and Earthquake Engineering 30,
157–169.
Kottke, A. R., and Rathje, E. M., 2008. A semi-automated procedure for selecting and scaling
recorded earthquake motions for dynamic analysis, Earthquake Spectra 24, 911–932.
Naeim, F., Alimoradi, A., and Pezeshk, S., 2004. Selection and scaling of ground motion time
histories for structural design using genetic algorithms, Earthquake Spectra 20, 413–426.
Open System for Earthquake Engineering Simulation (OpenSEES), 2007. Pacific Earthquake
Engineering Research Center, University of California, Berkeley, http://opensees.
berkeley.edu/ (last verified 03=29=2010).
Shantz, T. J., 2006. Selection and scaling of earthquake records for nonlinear dynamic analysis
of first model dominated bridge structures, in Proceedings of the 8th U.S. National Confer-
ence on Earthquake Engineering, San Francisco, CA.
U.S. Geological Survey (USGS), 2008. 2008 Interactive Deaggregations (Beta). http://
eqint.cr.usgs.gov/deaggint/2008/ (last verified 03=29=2010).
Watson-Lamprey, J. A., and Abrahamson, N. A., 2006. Selection of ground motion time series
and limits on scaling, Soil Dynamics and Earthquake Engineering 26, 477–482.
Youngs, R. R., Power, M. S., Wang, G., Makdisi, F. I., and Chin, C. C., 2007. Design ground
motion library (DGML) - Tool for selecting time history records for specific engineering
applications, in SMIP Seminar on Utilization of Strong-Motion Data.
(Received 21 April 2010; accepted 16 July 2010)