Documente Academic
Documente Profesional
Documente Cultură
org/
Short Note
Centered and staggered Fourier derivatives and
Hilbert transforms
Gustavo J. P. Correa
, Marc Spiegelman , Suzanne Carbotte ,
and John C. Mutter
INTRODUCTION
When ringing derivatives are used in seismic modeling, migration, and other wave propagation problems in heterogeneous media, they interact with boundaries of blocks or layers
of different physical properties, generating noncausal and nonphysical waves. These spurious waves can be as strong as reflected or converted waves and other legitimate physical events,
leading to errors and misinterpretations. Therefore, compact
(i.e., short and nonringing) derivatives should be used.
Figure 1 shows snapshots of the vertical component of a particle displacement field, modeled with the elastic wave equation, using the Fourier pseudospectral method (Kosloff et al.,
1984; Witte, 1989; Witte and Richards, 1990). The model is
composed of two half-spaces separated by a horizontal interface at 1.6 km depth. The P-wave and S-wave velocities and
densities of the upper and lower half-spaces are V P = 2.0 km/s,
VS = 0.8 km/s, = 2.4 g/cm3 , and V P = 4.0 km/s, VS = 2.3 km/s,
= 2.6 g/cm3 , respectively. There are 256 256 points in the
grid, spaced by 12.5 m in both horizontal x- and depth
z-directions. An explosive source is triggered at time t = 0 at
the point of coordinates x = 0 and z = 1.2 km. The snapshots
were taken at time t = 0.6 s. We used centered space derivatives to model the left panel, and staggered space derivatives
to model the right panel. Besides the derivatives, everything
else is the same in the algorithms used to produce both panels.
A number of noncausal and nonphysical waves (and other
artifacts) appear on the left panel of Figure 1 because of the
interaction of the ringing tail of the centered derivatives with
the horizontal interface at the center of the model and with the
grid boundaries. These artifacts are not present on the right
panel of Figure 1, which was modeled using staggered derivatives. The artifacts produced by the centered derivatives are
about 20 dB below the maximum amplitude of the wavefield.
Published on Geophysics Online May 23, 2002. Manuscript received by the Editor February 11, 2000; revised manuscript received January 14, 2002.
Formerly Lamont-Doherty Earth Observatory, Columbia University, P.O. Box 1000, Palisades, New York 10964-8000; presently University of
Houston, Department of Geosciences, 312 Science & Research Building 1, Houston, Texas 77204-5007. E-mail: gus@uh.edu.
Lamont-Doherty Earth Observatory, Columbia University, P.O. Box 1000, Palisades, New York 10964-8000. E-mail: mspieg@ldeo.columbia.edu;
carbotte@ldeo.columbia.edu.
Fourier Derivatives
(1)
S () =
+
1t
ei 2
(2)
d n() p(t)
= DFT 1 ((i)n S ()DFT ( p))(t).
dt n
(3)
Here the term staggered denotes the use of half-gridspacing phase-shift operators S in equation (3). However,
equation (3) is equivalent to the application of the centered
derivatives of equation (1) to p(t) discretized over the staggered grid t = ( j + 1/2)1t, j = 0, . . . , N 1. In other words,
staggered derivatives of grid-centered functions are the
same as centered derivatives of grid-staggered functions. We
call Fourier derivatives any derivatives calculated using
equations (1) or (3).
-1.6
0
Figure 2 shows the result of applying the centered and staggered first- and second-order Fourier derivatives to a spike at
the center of grids of sizes N = 125 and N = 128 (impulse response). Observe in Figure 2 that first-order derivatives are
ringing when centered on the grid and compact when staggered, whereas second-order derivatives are compact when
centered on the grid and ringing when staggered. Figure 2 also
suggests that this behavior of derivatives does not depend on
the parity of the grid size (N ). To prove these conjectures, and
to generalize them to derivatives of higher order, we next analyze the spectra of centered and staggered derivatives of any
order.
In frequency domain, the centered first-order derivative of
equation (1) is obtained through multiplication by d (c) /dt =
i, which can be factored as the product d (c)/dt = ||Q(),
where
Q() = i sgn()
Q () = S ()Q().
1.6
-1.6
0
Depth Z (km)
2.4
1.6
1.6
2.4
Centered
(5)
The spectra of the staggered first-order derivatives of equation (3) are given by d ()/dt = ||Q ().
Figure 3 shows the amplitude and phase spectra of Q(),
S (), and Q (), and their Fourier transforms in time domain. Analogous results apply to S + () and Q + (). These
functions are discrete and periodic beyond the interval
[ N , + N ]; hence, they wrap around the Nyquist frequency
N (Oppenheim and Schafer, 1975). Figure 3 also shows that
Q() has phase jumps of at = 0 and = N , S () has
a phase jump of at = N , and Q () has a phase jump
of at = 0. Functions that have large phase jumps at N
present ripples in the time domain due to Gibbs phenomenon
(Bracewell, 1986). By contrast, functions that have small phase
jumps at N do not show ripples in the time domain.
0.8
1.6
(4)
0.8
Depth Z (km)
Downloaded 03/10/15 to 200.20.187.13. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
d n(c) p(t)
= DFT 1 ((i)n DFT ( p))(t),
dt n
1559
Staggered
FIG. 1. Snapshots of the vertical component of a particle displacement field modeled using the elastic wave
equation. The space derivatives were calculated using Fourier transforms. The left panel was modeled with
centered derivatives, and presents visible artifacts due to the interaction of the ringing tail of the centered
derivative with a horizontal interface at the center of the model and with the grid boundaries. The compact
staggered derivatives on the right panel do not produce such artifacts. The amplitudes in both panels were
clipped at the same level. The small white square marks the source position.
1560
Correa
et al.
1
-1
-1
-60
-40
-20
20
40
60
80
Non-dimensional values
-40
-20
20
40
60
80
-80
0
-2
-60
-40
-20
20
40
60
80
-80
2
-2
-2
-80
-40
-20
20
40
60
80
-1
-60
-2
-80
-60
-1
-80
-80
1
-60
-40
-20
0
20
Sample index
40
60
80
-80
-60
-40
-20
20
40
60
-60
-40
-20
20
40
60
-40
-20
80
-60
80
N = 125
0
20
Sample index
40
60
80
N = 128
-5 N
N 5
10
-1
-5 N
N 5
10
1
0
-1
-2
-10
-5 N
N 5
N 5
10
-20
20
0.5
0
-0.5
-5 N
N 5
10
1
0
-1
/2
-2
-10
10
2
/2
Phase of Q ()
(radians)
-2
-10
-5 N
-1
/2
-2
-10
0.5
-0.5
2
/2
1
-1
/2
-2
-10
Phase of S ()
(radians)
Amplitude of S ()
(arbitrary units)
-2
-10
Fourier Tr. of S ()
(arbitrary units)
-1
2
/2
1
-20
Fourier Tr. of Q ()
(arbitrary units)
Phase of Q()
(radians)
Amplitude of Q()
(arbitrary units)
FIG. 2. Fourier derivative impulse responses on grids of odd (left) and even (right) sizes (N ). The staggered
derivatives are time advanced (d + /dt), but similar results apply to time delayed derivatives (d /dt). First-order
derivatives are ringing when centered and compact when staggered, whereas second-order derivatives are compact when centered and ringing when staggered, regardless of the parity of N .
Amplitude of Q ()
(arbitrary units)
Downloaded 03/10/15 to 200.20.187.13. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
-80
20
0.5
0
-0.5
-5 N
N 5
10
-20
20
FIG. 3. Amplitude spectra (left), phase spectra (center), and the Fourier transform to time domain (right) of (ac) the centered
quadrature function Q(), (df) the time-delay phase-shift function S (), and (gi) the time-delayed quadrature function Q ().
All spectra are periodic functions of the angular frequency, and the graphs wrap around the Nyquist frequency ( N ). The functions
Q() and S () have a phase jump of and across N , respectively, whereas Q () does not have a noticeable phase jump there.
Spectra with large phase jumps across the Nyquist frequency correspond to ringing time-domain functions (Gibbs phenomenon),
whereas spectra with no apparent phase jumps correspond to compact time functions.
Downloaded 03/10/15 to 200.20.187.13. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
Fourier Derivatives
1561
Table 1. The complex spectra of nth-order Fourier derivatives on centered and staggered grids of odd and even sizes (N). Any
spectrum can be factored using only the following terms: (1) the powers of the absolute value of the angular frequency ||n , (2) the
(centered) quadrature function Q(), (3) the half-grid-point phase-shift operator S (), and (4) the staggered quadrature function
Q (). Any compact derivative spectrum has a negligible phase jump at the Nyquist frequency; hence, it either contains the factor
Q () or it is simply ||n . Derivative spectra that contain either the factor Q() or S () have large phase jumps across the
Nyquist frequency; therefore, they present Gibbs ripples and are ringing. C = centered, S = staggered.
Grid Size N
Derivative Order n
Centered or Staggered
any
1 (mod 4)
any
1 (mod 4)
any
2 (mod 4)
any
2 (mod 4)
any
3 (mod 4)
any
3 (mod 4)
any
0 (mod 4)
any
0 (mod 4)
Derivative Spectrum
d n(c)
= +||n Q()
dt n
d n()
= +||n Q ()
dt n
d n(c)
= ||n
dt n
d n()
= ||n S ()
dt n
n(c)
d
= ||n Q()
dt n
d n()
= ||n Q ()
dt n
d n(c)
= +||n
dt n
d n()
= +||n S ()
dt n
Ringing?
ringing
compact
compact
ringing
ringing
compact
compact
ringing
Downloaded 03/10/15 to 200.20.187.13. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
1562
Correa
et al.
even N , an inspection of Table 1 shows that centered derivatives of even order and staggered derivatives of odd order are
Hermitian, whereas centered derivatives of odd order and staggered derivatives of even order are not Hermitian. These results are summarized in Table 2. A comparison of Tables 1
and 2 shows that the parity of N does not control the ringing or
compact nature of derivatives, that N alone guarantees Hermitian derivatives only if it is odd, and that Hermitian derivatives
are not necessarily compact.
Rows 1 and 2 of Table 2 tell us that centered first-order
Fourier derivatives are Hermitian only when N is odd. Hence,
they can be calculated using RFFT algorithms only if N is odd,
but not if N is even. By contrast, rows 1 and 3 of Table 2 tell
us that staggered first-order Fourier derivatives are Hermitian
for all values of N . Hence, they can be calculated using the
efficient RFFT algorithms regardless of the parity of N (Witte
and Richards, 1990).
An alternative efficient technique to calculate first-order
Fourier derivatives was presented in the literature (Zhou,
1992). This technique also relies on the Hermitian symmetry
of the DFT of real functions. Therefore, its application to centered derivatives is restricted to odd N , but it can be applied
to staggered derivatives for any value of N .
The most efficient FFT and RFFT algorithms are still those
where N is a power of two (Press et al., 1994; Frigo, 1999),
though many algorithms have been developed in recent years
for other prime and composite numbers. We did a benchmark
test where derivatives were calculated 4000 times using the
RFFT routines of the prize-winning FFTW package (Frigo,
1999) in a 850-MHz Pentium III processor under the Linux
operational system. Centered derivatives calculated with
N = 125 = 53 and N = 127 (a prime number) took total CPU
times of 2.93 s and 43.24 s, respectively. Staggered derivatives
calculated with N = 128 = 27 took a total CPU time of 1.97 s,
which was about 1.5 times faster than the centered derivatives
with N = 125 and about 22 times faster than the centered
derivatives with N = 127. These tests suggest that, using
RFFT algorithms and choosing N a power of two, staggered
derivatives may be calculated more efficiently than centered
derivatives.
CENTERED AND STAGGERED HILBERT TRANSFORMS
Table 2. Hermitian symmetry of nth-order Fourier derivatives on centered and staggered grids of odd and even sizes
(N). C = centered, S = staggered.
Grid Size N
odd
even
even
even
even
Derivative
Order n
Centered or
Staggered
Hermitian
Symmetry?
any
odd
odd
even
even
both
C
S
C
S
yes
no
yes
yes
no
We thank John Diebold and Peter Buhl for useful discussions, and an anonymous reviewer for helpful suggestions
about the Hilbert transform. This research has been supported
Fourier Derivatives
1563
0.8
0.4
nondimensional value
Downloaded 03/10/15 to 200.20.187.13. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
0.6
0.2
0.0
-0.2
-0.4
-0.6
-0.8
-20
-15
-10
-5
10
15
20
sample index
FIG. 4. The continuous Hilbert transform kernel sampled at the centered and staggered grids (circles, dotted
line, 64 samples shown), the staggered Hilbert transform kernel for N = 32 (squares, continuous line) and the
centered Hilbert transform kernel for N = 32 (diamonds, dashed line). Notice that the staggered kernel fits
almost perfectly to the continuous one, whereas the centered kernel does not.
in part by NSF grants OCE 98-11483 and 94-02172, Department of Earth and Environmental Sciences of Columbia University, and Petrobras.
L.D.E.O. contribution 6287.
The following free software was used in the preparation of
this paper: SU, FFTW, Xgraphic, Linux, Gnu Gcc, Gnu Emacs,
and Latex.
REFERENCES
Bracewell, R. N., 1986, The Fourier transform and its applications, 2nd
ed.: McGraw-Hill, Inc.
Claerbout, J. F., 1976, Fundamentals of geophysical data processing:
McGraw-Hill, Inc.
Fornberg, B., 1990, High order finite difference and pseudo-spectral
method on staggered grids: SIAM J. Numer. Analysis, 27, 904918.
Frigo, M., 1999, A Fast Fourier Transform compiler: Proc. 1999 ACM
SIGPLAN, 169180.
Kosloff, D., Reshef, M., and Loewenthal, D., 1984, Elastic wave calculations by the Fourier method: Bull., Seis. Soc. Am., 74, 875891.
Ozdenvar,
T., and McMechan, G. A., 1996, Causes and reduction of
numerical artefacts in pseudo-spectral wavefield extrapolation: Geophys. J. Internat., 126, 819828.
Press, W. H., Teukolsky, S. A., Vetterling, W. T., and Flannery, B. P.,
1994, Numerical recipes in C: The art of scientific computing, 2nd
ed.: Cambridge Univ. Press.
Saito, M., 1974, Hilbert transforms for sampled data: J. Phys. Earth, 22,
313324.
Taner, M. T., Koehler, F., and Sheriff, R. E., 1979, Complex seismic
trace analysis: Geophysics, 44, 10411063.
Witte, D. C., 1989, The pseudospectral method for simualting wave
propagation: Ph.D. thesis, Columbia Univ.
Witte, D. C., and Richards, P. G., 1990, The pseudospectral method
for simualting wave propagation, in Lee, D., Cakmak, A., and
Vichnevetsky, R., Eds., Computational acoustics: Elsevier Science
Publ., 118.
Zhou, B., 1992, Discussion on The use of Hartley transform in geophysical applications (R. Saatcilar, S. Ergintav, and N. Canitez):
Geophysics, 57, 196197.