The effect of epsilon on response spectra, and the

conditional mean spectrum as a potential target

spectrum

Jack W. Baker
Stanford University
 The goal addressed here
• Estimate (the median and distribution of) structural
response for a given spectral acceleration value at the first-
mode period of a structure: Sa(T1)

• Why only Sa(T1)?

– We can model its occurrence using PSHA
– This allows us to link ground motion occurrence with structural
response for performance-based engineering

• Questions
– What is a “typical” or “average” response spectrum for a ground
motion having an extreme Sa(T1)?
– How does this relate to code spectra or uniform hazard spectra?
 A typical code uniform hazard spectrum

A Los Angeles area UHS (10% in 50)

USGS firm soil hazard with IBC 2000
1.5
Spectral Acceleration [g]

Assume 1s as our structure’s

first-mode period
0.5

0
0 0.5 1 1.5 2 2.5 3

Period [s]
Hazard disaggregation for the UHS at Sa(1s)

Mbar ≈ 7
Rbar ≈ 10 km

(from USGS)
The UHS and the median prediction for the M ≈ 7,
R ≈ 10 km event from deaggregation

1.5
UHS Spectrum
Median Spectrum for M, R
Spectral Acceleration [g]

1 The gap between the median spectrum

and the code spectrum means that we
also have a positive “ε”

0.5

0
0 0.5 1 1.5 2 2.5 3

Period [s]
 A goal from the PEER GMSM program

Compute the median maximum interstory drift ratio from a given

earthquake event and site, additionally conditioned on Sa(1s)

2.5
Median spectrum
Median plus two σ spectrum
2 M=7, R=10 km
Spectral acceleration [g]

1.5
Target Sa(T1) value

0.5

0
0 0.5 1 1.5 2 2.5 3
Period [s]
Response spectra of 40 real records with M ≈ 7 and R ≈ 10 km

To think about how a uniform hazard spectrum (or, alternatively, a “+2 σ” spectrum,
relates to real records, let’s plot the spectrum of 40 real records with magnitudes and
distances similar to the target M and R (as seen in green below)

2.5

2
Spectral acceleration [g]

1.5

0.5

0
0 0.5 1 1.5 2 2.5 3
Period [s]
Response spectra of 40 real records with M ≈ 7 and R ≈ 10 km
Note that one record is as large as the “+2 σ” spectrum at the 1 second period we are
interested in, but it’s not equally extreme at all periods. This raises teh question, what is
the “average” spectrum from an M=7, R=20 km ground motion with Sa(1s) = 1g?

2.5

2
Spectral acceleration [g]

1.5

0.5

0
0 0.5 1 1.5 2 2.5 3
Period [s]
Response spectrum of a single record

ln( observed Sa(T) ) - ln( predicted median Sa(T) )

ε(T) =
predicted standard deviation of ln( Sa(T) )

2.5
Record spectrum
At 1s, this record’s
Predicted median spectrum
spectrum is about 2
2 Median +/1 one standard deviation
standard deviations
Spectral acceleration [g]

larger than predicted by

ε(0.5s) ≈ 1 an attenuation equation
1.5 (or “+2 ε”), but at 0.5s, it
is only about 1 standard
ε(1s) ≈ 2 deviation larger than
1 predicted (“+1 ε”).
We can compute ε’s for
many records in this
0.5 manner, and plot them
on a scatter plot, as seen
in the next slide.
0
0 0.5 1 1.5 2 2.5 3
Period [s]
 A scatter-plot of these ε’s for many records illustrates
their joint distribution

4
We see that the ε’s at

Sa(0.5s)
0.5s and 1s are related to

residuals(ε)forforSa(0.5s)
each other, but they are
not equal to each other. 2
Predictionresiduals

0
The record from
the previous slide
Prediction

-2

-4
-4 -2 0 2 4
Predictionresiduals
Prediction residuals(ε)for
forSa(1s)
Sa(1s)
 Joint distribution of the ε’s

4
Since we have

Sa(0.5s)
residuals(ε)forforSa(0.5s)
specified a target
Sa(T1) value (and can
find the associated M 2
and R from deagg),
we can also find the ε
value at T1 (which is 2
Predictionresiduals

in this case). 0
The question is then,
what are the
Prediction

associated ε values at
-2
other periods, given
this ε at T1?
Sa(1s)
We see from this target
scatter plot that for -4
-4 -2 0 2 4
this extreme positive ε
Predictionresiduals
Prediction residuals(ε)for
forSa(1s)
Sa(1s)
at 1s, the epsilons at
0.5s tend to be
relatively lower.
 Joint distribution of the ε’s

4
As a first guess for a

Sa(0.5s)
residuals(ε)forforSa(0.5s)
target spectrum, we ρ = 0.75
might want to use the
average (or expected) 2
value of ε at other
periods, given ε at T1. Conditional mean of
ε(0.5s) given ε(1s)
Predictionresiduals

Some probability
0
calculations can show
that the expected ε
ε(0.5s) = ρ · ε(1s)
value at other periods
Prediction

is equal to the original

-2
ε multiplied by the
correlation coefficient
between the two
Sa(1s)
epsilons. In this case, target
-4
the correlation -4 -2 0 2 4
coefficient is 0.75, so Predictionresiduals
Prediction residuals(ε)for
forSa(1s)
Sa(1s)
the average ε at 0.5s is
75% of the ε at 1s.
 Joint distribution of the ε’s

4
ε(0.5s) = 0.75 · ε(1s)

Sa(0.5s)
residuals(ε)forforSa(0.5s)
2 ε(0.5s) = 1.5
Predictionresiduals

0
Conditional mean
at ε(0.5s) given our
target ε(1s)
Prediction

-2

ε(1s) = 2

-4
-4 -2 0 2 4
Predictionresiduals
Prediction residuals(ε)for
forSa(1s)
Sa(1s)
 The conditional mean spectrum

We can use this 1.5 ε value 2.5

at 0.5s to compute the Median spectrum
expected spectral Median plus two σ spectrum
acceleration value at that Conditional mean spectrum, given Sa(1s)
2
Spectral acceleration [g]
period, and can repeat the
process at all periods. To
build a full response Conditional mean Sa(0.5s)
spectrum. value, given target Sa(1s)
1.5
We see on the right that the Target Sa(1s) value
spectrum has a peak near 1s

ε(0.5s)=1.5
(because the ε’s are highly
correlated at closely spaced 1
periods), and decays towards

ε(1s)=2
the median spectrum at large
and small periods (as the
correlation of ε’s decreases). 0.5

0
0 0.5 1 1.5 2 2.5 3
Period [s]
 Computing the Conditional Mean Spectrum
Some comments to go along with the following slide:

• To do this computation, you need to start with an Sa at a given

period, along with a magnitude and distance. Those target values
could come either from a PSHA and deaggregation, or from a
scenario analysis of a single event.

• Once you know Sa, magnitude and distance, finding the associated ε
is simply a matter of back-calculating using an attenuation model.
Or you can get ε directly from a deaggregation calculation.

• The calculation requires correlation coefficients between epsilons at

many period pairs. I’ve gone and calculated those correlations using
a large set of motions, and fit the results to a simple predictive
equation, so that others can use it to get their correlations instead of
computing correlations themselves.

• Once you know your expected (or mean) ε’s at all periods, you can
simply plug them back into the attenuation equation to compute
target spectral accelerations at all periods, giving you the conditional
mean spectrum.
 Computing the Conditional Mean Spectrum
1. Compute the mean and standard deviation of logarithmic spectral
acceleration at all periods for a target magnitude and distance

ln Sa ( M , R, T )
(Mean and σ of log Sa from an attenuation function)
σ ln Sa (T )
2. Compute the target ε at T1 (from disaggregation or back-calculation)

ln Sa (T1 ) * −ln Sa ( M , R, T1 )
ε (T1 ) = (lnSa(T1)* is the target Sa at T1)
σ ln Sa (T1 )
3. Compute the conditional mean ε at other periods, given ε(T1)
⎛π ⎛ T ⎞ T ⎞
ρ = 1 − cos ⎜
ε (Ti ) = ρ ⋅ ε (T1 ) − ⎜ 0.359 + 0.163I (Tmin<0.189 ) ln min ⎟ ln max ⎟
⎝2 ⎝ 0.189 ⎠ Tmin ⎠
(Baker and Cornell, 2006, Bull. Seism. Soc. Am.)

4. Compute the spectral acceleration at all periods, using this

information
ln Sa (Ti )* = ln Sa ( M , R, Ti ) + σ ln Sa (Ti ) ⋅ ε (Ti ) (target Sa’s at other periods)
 Conclusions
• A conditional mean target spectrum can be constructed by
– Selecting a target Sa(T1) value
– Computing conditional mean ε values at other periods, given ε(T1)
– Computing Sa values at other periods, given the target M, R and ε’s

• Using “positive epsilon” spectral values at all periods (e.g.,

a UHS) is more conservative than many users might
believe
 References
The importance of “ε”:

Baker J.W. and Cornell C.A., 2005. A vector-valued ground motion intensity measure
consisting of spectral acceleration and epsilon, Earthquake Engineering & Structural
Dynamics, 34 (10), 1193-1217.

The conditional mean spectrum:

Baker J.W. and Cornell C.A., 2006. Spectral shape, epsilon and record selection,
Earthquake Engineering & Structural Dynamics, 35 (9), 1077–1095.

The correlation model for ε‘s:

Baker J.W. and Cornell C.A., 2006. Correlation of response spectral values for multi-
component ground motions, Bulletin of the Seismological Society of America, 96 (1),
215-227.