GEOPHYSICS,

VOL. 57, NO. 1 (JANUARY

1992): P. 15-26, 18 FIGS.

Wave-equation tomography

Marta Jo Woodward

ABSTRACT

The relation between ray-trace and diffraction tomography is usually obscured by formulation of the two methods in different domains: the former in space, the latter in wavenumber. Here diffraction tomography is reformulated in the space domain, under the title of wave-equation tomography. With this transformation, wave-equation tomography projects monochromatic, scattered wavefields back over sourcereceiver wavepaths, just as ray-trace tomography projects traveltime delays back over source-receiver raypaths. Derived under the Born approximation, these wavepaths are wave-theoretic backprojection patterns for reflected energy; derived under the Rytov approximation, they are wave-theoretic back-projection patterns for transmitted energy. Differences between ray-trace and wave-equation tomography are examined through comparison of wavepaths and raypaths, followed by their application to a transmission-geometry, synthetic data set. Rytov

wave-equation tomography proves superior to raytrace tomography in dealing with geometrical frequency dispersion and finite-aperture data, but inferior in robustness. Where ray-trace tomography assumes linear phase delay and inverts the arrival time of one well-understood event, wave-equation tomography accommodatesscattering and inverts all of the signal and noise on an infinite trace simultaneously. Interpreted through the uncertainty relation, these differences lead to a redefinition of Rytov wavepaths as monochromatic raypaths, and of raypaths as infinite-bandwidth wavepaths (Rytov wavepaths averaged over an infinite bandwidth). The infinite-bandwidth and infinite-time assumptions of ray-trace and Rytov, wave-equation tomography are reconciled through the introduction of handlimited raypaths (Rytov wavepaths averaged over a finite bandwidth). A compromise between rays and waves, bandlimited raypaths are broad backprojection patterns that account for the uncertainty inherent in picking traveltimes from bandlimited data.

INTRODUCTION

Seismic tomography encompassesa broad group of inversion schemes used for imaging seismic velocity fields. Defined as the reconstruction of a seismic held from integrals over paths through the field, the method is most familiar in its ray-theoretic form as traveltime inversion. In this application traveltimes are picked for source-geophone experiments, then compared to expected traveltimes calculated by tracing rays through an assumed background velocity field. The relation between traveltime and velocity perturbations is linearized with Fermat principle, and an updated velocs ity field is produced by projection of measured traveltime perturbations back through the medium over raypaths.

For evenly spacedsource-geophonepairs, the backprojt ( tion step may be performed analytically: in either the space domain as a generalized inverse Radon transform (Beylklr , 1982; Fawcett and Clayton, 1984) or in the wavenumbrr domain using the projection-slice theorem (Mersereau a 13 Oppenheim, 1974). However, given the irregular spati:tl coverage of the typical seismic experiment, most traveltime inversion is performed numerically in the space domalr , using iterative least squares (Dines and Lytle, 1979; Bishl)o et al.. 1985). Ray-theoretic tomography in this form is kno\si? as ray-trace tomography. Ray-trace tomography works well when two requiremer 1s are met (Wu and Toksoz, 1987). First, because the method relies on the high-frequency assumption of ray theory, t 1 velocity field being examined must vary slowly on the SC:I?

Presented at the 58th Annual International Meeting, Society of Exploration Geophysicists. Manuscript received by the Editor February : , I 1990;revised manuscript received April 25, 1991. *Formerly Stanford Exploration Project, Department of Geophysics, Stanford University; presently GECO-PRAKLA NSA, 1325South Da r7 Ashford, Houston, TX 77077. 0 1992Society of Exploration Geophysicists. All rights reserved. 15

16

Woodward
RAY-THEORETIC EQUATIONS

of the source wavelengths. In this case there is no scattering: phase delay is linear with frequency; the source wavelet is not distorted, and seismic events are completely characterized by traveltimes. Second, because without scattering rays sample very narrow regions in space, the source-geophone geometry must provide many view angles through the medium. When these requirements are not met. wave-theoretic tomography provides a better image. Wave-theoretic tomography accommodates scattering by replacing traveltime delays with scattered wavefields. Wavefields are recorded for source-geophone experiments, then compared to expected wavefields calculated by forward modeling through an assumed background velocity field. The wave equation is linearized with either the Born or Rytov approximation, and an updated velocity field is produced by propagation of measured wavefield perturbations back through the medium over wave-propagation paths. The backpropagation step of wave-theoretic tomography is usually formulated in the frequency-wavenumber domain, under the title of diffraction tomography (Mueller et al., 1979; Mueller, 1980; Devaney, 1982, 1984; Slaney et al., 1984; Wu and Toksoz, 1987). The method solves the problem analytically, for plane-wave scattering and independent monochromatic sources. The problem has also been solved in the time-space domain: as migration by inversion of a generalized Radon transform (Miller et al., 1987). and as nonlinear inversion (Tarantola. 1984, 1987; Mora. 1987). Here wave-theoretic tomography is reformulated for solution by iterative least squares in the frequency-space domain, under the title of wave-equation tomography (Woodward and Rocca, 1988; Woodward, 1989). This reformulation makes wave-equation tomography more flexible than diffraction tomography in dealing with irregularly sampled surveys and inhomogeneous background media. More importantly, it encourages physical understanding of the differences between ray and wave inversions through visual comparisons of their space-domain backprojection patterns. Where ray-trace tomography projects traveltime delays back over source-receiver raypaths, wave-equation tomography projects monochromatic, scattered wavefields back over source-receiver wavepaths. The first two sections of this paper develop the equations defining raypaths and wavepaths in a parallel fashion. The next two sections examine the resulting backprojection patterns in the space and wavenumber domains. Wavepaths derived under the Rytov approximation are linked to raypaths as wave-theoretic trajectories for transmitted energy; those derived under the Born approximation are linked to migration ellipses as wave-theoretic trajectories for reflected energy. The fifth section follows raypaths and Rytov wavepaths through the inversion of a transmission-geometry, synthetic data set. The final section summarizes the differences between ray-trace and Rytov wave-equation tomography, concluding that raypaths and wavepaths lie at two extremes of the uncertainty relation: the former assuming infinite bandwidth, the latter infinite time Rays and waves are reconciled through the uncertainty relation with the definition of bandlimited raypaths as frequency-averaged wavepaths. Bandlimited raypaths are broad backprojection patterns that account for the uncertainty inherent in picking traveltimes from bandlimited data.

Because ray-theoretic tomography relies on the hlk hfrequency approximation of ray theory, its traveltime pi,:ks must represent ray arrivals. Where traveltimes are pic.::d from bandlimited wavelets, this requires either that tie wavelet peaks are undistorted (that phase delay is linear v th frequency) or that Fermat-path first breaks (ray arrivals) .:In be determined. In practical bandlimited applications these requirements are often unmet. Where the velocity f tld varies rapidly on the scale of the source wavelength, wa\ elets are distorted by geometrical frequency dispersii n; where events overlap and signal level is low, first breaks r re difficult to pick. Given these limitations, the source-geophone travelt rle integrals of ray-theoretic tomography are:

r(gls1 =

n,(r)l[rls.

g,

w(r)1dr.

(1)

Here the source-geophone pair is indicated by s, g; r is t?e space-coordinate vector; u is inverse velocity or slownc~~s; and f. is the raypath from s to g through ~$3 along tie (1 raypath. 0 elsewhere). The expected source-geophone tr iveltime integrals, calculated by ray-tracing though the firstguess background velocity field, are:

Here MJ~ the background slowness field and Lo the rayf: th is through that field. Because L is a function of ~3, the acou,; ic tomography problem is nonlinear. To create a linear relat i n-t between Ar and AM,. ray theory invokes Fermat print [lle s and approximates L by L,, producing

At(&) =

Aw(r)Lo(rls,

g) dr.

(3)

The backprojection step of ray-theoretic tomography nvolves the solution of the system of linear equations reslllting from consideration of a number of source-geoph5)le pairs: I,Aw = At. The nonlinear part of the problenl is attacked iteratively, for successively updated backgrol I Id velocity fields.
WAVE-THEORETIC EQUATIONS

Because wave-theoretic tomography is full waveform nversion, it makes no assumptions about the characterizat i >n of an event by a single traveltime pick. However, it di:es assume that the source wavelet is sufficiently well known tor calculation of a background wavefield, and that the recorr I :d seismic traces are complete and free of noise. Given these assumptions, the wave-theoretic equivalent of equation (3) may be generated by linearizing the scalar wave-equal I )n with either the first-order Born or the first-order Ry. IV approximation. Born Where ray-trace tomography creates a linear relatur between velocity and traveltime perturbations, the B )rn

Wave-equationTomography
approximation creates a linear relation between velocity and wavefield-amplitude perturbations: AT(w) = q w) - q,,(w). ( These are complex amplitudes (V = Ae ), * measured in the temporal-frequency domain for source-geophone pairs. The Born approximation begins with the wave equation written as: Rytov

17

A s) = W

O(rFo[glr, vo(r)l x {To[r/s, vo(r)l + AWrls, O(r)l} dr

(4) or the

The first-order Rytov approximation creates a linear rel;ltion between velocity and wavefield-phase perturbation ;: A@.(o) = In [q(w)] - In [Ur,(w)]. As with the Born appro,;imation, these complex phases are measured in the tempor tIfrequency domain for source-geophone pairs. The Ryt IV approximation begins with the wave equation written as:

(Slaney et al., 1984). 0 is the object function, perturbed velocity field expressed as: O(r) = ki(r)[ 1 - vjj(r)/v r)] ( = 2ki(r)Av(r)/u(r).

x {[V(A@(rls,

O(r))]

+ O(r)} dr

(1:)

(5)

(Slaney et al., 1984). This equation is more difficult to interpret physically than the Born equivalent. Howevc,r, under the Rytov approximation [V(A@)I =C 0, and t 1: equation becomes

Go is the Green function or impulse response for the s background medium: for constant velocity and two or three dimensions, e w

Awls) =
This is just the Born [The Rytov formula weak-scattering limit, Following the Born rewritten as:

s
= /

O(r)

Go(glr) o(r/s)dr, J To (ids)

(1 I) oAl C P . in t I? 1981 I can ):

Go(r) =q

or

$Hi,(k0lrl),

(6)

where k. is the background wavenumber W/ZI~, and Hh is a zero-order Hankel function of the first kind. Equation (4) is usually interpreted by Huygens construcs tion. The anomalous wavefield at a specific geophone is generated by superposition: each point in the medium acts as a scatterer, emitting an impulse response scaled by the product of the wavefield and the object function at that point. By using reciprocity to replace G,(rlg) with G,(gjr), then regrouping terms and introducing the concept of a wavepath Y, Au(r) 7 r) ( %r/s,

equation with Aq replaced by reduces to the Born formula where *,A@ = A\V (Devaney, development, equation (I 1)

@(g/s)

AT!(r) 30(r/s, 4r)

g) dr (I !)

%(rls, g) =

Go(r~g)*o(rls)
2k,?(r)

u'o(sls)

A*(&)

g,

u(r)1dr
+ AWrls, O(r)]}, (7)

and the scattered complex phases again viewed as integn Ii through the perturbed velocity field over wavepaths Zo. F )r a point source at s, To is the Green function for tl, s : background medium, and the Rytov wavepath is

Lf(r/s, g) = 2k,:(r)Go(rlg){~lo(rls)

2o(r/s, g) =

Xi(r)

Go(rIdGo(rls)

Go(&)

(I 1

this interpretation can be altered to resemble that of equation (2). Now the scattered complex amplitudes of wave-theoretic tomography are revealed as integrals through a perturbed velocity field over monochromatic wavepaths %-just as the traveltime delays of ray-theoretic tomography are integrals through the perturbed velocity field over raypaths L. As in the ray-trace application, Y (specifically A\u) is a function of AZ), and the problem is nonlinear. Under the Born approximation the equation is linearized by assuming the wavepath L!Z be independent of the velocity perturbato tion (Aul +Z Ur,), yielding the monochromatic analog of equation (3): Au(r) Y0(r/s, z!(r)

Where ray-theoretic tomography forms a system of equ Itions through consideration of a number of sources at (I geophones, wave-theoretic tomography forms a syste II of equations-LJ,,A,,, = AVr or LfoA,,iV= AQ, throul11 consideration of a number of sources, geophones, and fr :. quencies. The extra dimension of information in wav :. theoretic as compared to ray-theoretic tomography w II become apparent later. As with ray-theoretic tomograph .I the nonlinear part of the problem is attacked iteratively.
WAVEPATHS VERSUS RAYPATHS: SPACE DOMAIN

A s) W

g) dr

For a point source at s, TCj is the Green function for the s background medium, and the Born wavepath is %)(rls, g) = 2koZ(r)Go(rJg)Go(rls). (9)

Figures la and lb show Rytov and Born examples I 1 two-dimensional wavepaths for a homogeneous backgrour t velocity field. Their familiar raypath analog is shown r Figure 2. Unlike the raypath, the wavepaths are monochrc 1 matic, complex, and infinite in spatial extent. While tt t complex absolute values of the wavepaths decay with di tance from the source and receiver, the phase of the patterr ! oscillates from 7r to -7~. Because the Born and Rytcr

approximations are based on scattering theory, contours c1

18

Woodward

constant phase on the wavepaths yield confocal ellipses: curves of constant source-scatterer-receiver traveltimes, with the source and receiver located at the foci. The equation describing these ellipses is:
+-=I; a2-c2

a2

a > c,

(14)

where a is the semimajor axis (half the scattering path from source to receiver) and c is half the source-receiver offset. Beyond these observations, the fine structure of the wavepaths depends on whether they are Born- or Rytov-generated patterns.
Rytov wavepaths

Phase and amplitude separate naturally in the Rytov wavepath. When multiplied by AVIV, the imaginary part of the wavepath yields the time-delay-like phase delay A+(w), the real part the log of the amplitude ratio In [A(o)lA This natural separation underscoresthe parallelism between Rytov and transmission ray-trace tomography. Reinterpreted as IV In (AlA,)lk,, =C I and IV(A$)lk,I 5 I, the Rytov approximation is compatible with transmission ray theory in requiring both the amount of scattering per wavelength and the scattering angle to be small (Chemov, 1960).

In full waveform inversion, these restrictions are best met by forward-scattered energy in transmission-geometry implementations: applications where traveltime delays accumulate through a velocity perturbation. In fact, tomography under the Rytov and eikonal approximations are equivalent in the very-short-wavelength limit (Devaney, 1981). For the two-dimensional Rytov wavepaths, the imaginary parts of the patterns pass through zeros at the boundaries between the first, second, third, etc., Fresnel zones. Since a scatterer within the first Fresnel zone generates a wavefield reaching the geophonewithin a half wavelength of the source wavefield, a low-velocity scatterer in this zone produces a phase delay, a high-velocity scatterer a phase advance. A scatterer in the second Fresnel zone generates a wavefield reaching the geophone between a half and a full wavelength behind the source wavefield: a low velocity scatterer produces a phase advance, a high velocity scatterer a phase delay. Similar arguments about amplitudes explain the oscillations of the real parts of the Rytov wavepaths. For the imaginary, three-dimensional wavepath slice in Figure 3, there is an additional, paradoxical zero on the sourcereceiver Fermat path. Since a point on this curve scatters energy in-phase with the background wavefield, it yields an amplitude perturbation but no phase perturbation at the receiver.

a) Rytov

2-D

5 & 10 Hz navepaths

b) Born

2-D

5 & 10 Hz wavepaths

FIG. I. Frequency-space domain wavepaths for a constant-velocity background field. (a) Rytov 5 and IO Hz wavepaths. (b) Born 5 and IO Hz wavepaths. The velocity field is 2000 m/s; the source and receiver are separated by 2000 m. White is positive; black is negative; grey is zero.

Wave-equation Born wavepaths

Tomography

19

Phaseand amplitude fail to separatein the Born wavepath: when multiplied by the object function, the real part yields a[V(o) - qO(o)], the imaginary part 9[q(w) - qO(o)]. Without the Rytov wavepath normalization factor of s Go(gls), Born-wavepath zero crossings shift in a complex way with source-receiver separation and frequency. Born wavepaths are more naturally examined in the time-space domain of prestack migration. Figure 4 shows the constantvelocity wavepaths of Figure lb, Fourier transformed into this domain and sliced at two times. The patterns are recognizable as the ellipses over which time arrivals are smeared in prestack migration (Schneider, 1971). While Born and Rytov wavepaths look alike in this domain (except for a time shift), the Born data parameters are much more suited to time than the Rytov parameters. Since the Fourier transform of A*(o) is the difference between the measured and expected seismic traces, Born tomography implemented in time-space corresponds to the projection of unexpected time arrivals back over isochronal scattering ellipses (Miller et al., 1987). Rytov tomography implemented in time-space corresponds to the projection of the Fourier transform of complex-phaseperturbations back over these same patterns, a much less intuitively meaningful operation. The suitability of the Born approximation for backscattered energy and reflection geometries arises from its weak-scattering as-

sumption: A* < Vr,, (or equivalently, IAAIAo[ * 1 and iA+\ < 1). By requiring both the total amount of scattering and the total change in phase to be small, the method trades the ability to handle large cumulative phase delays for the ability to handle large scattering angles (Chernov, 1960).
Inhomogeneous background velocity field

Figure 5 shows examples of the imaginary parts of Rytov wavepaths in inhomogeneous media. The two-dimensional, 15 Hz wavepaths encompass their expected semicircular, refracted, and direct-plus-reflected-plus-refracted raypath equivalents. Because the direct-plus-reflected-plus-refracted wavepath contains three overlapping events, its zeros no longer fall on Fresnel-zone boundaries. The interior regions of the overlapping ellipses change sign from frequency to frequency, as determined by interference between the three events. This geometry does not fit the transmission geometry applications for which Rytov tomography is best suited. Where multiple waves are considered simultaneously, the linear relation between cumulative phase delay and velocity perturbation breaks down (Keller, 1969). For this application, the different wavefields would have to be separated before Rytov inversion.
WAVEPATHS VERSUS RAYPATHS: WAVENUMBER DOMAIN Propagating energy

Ltl

FIG. 2. Raypath analog of the wavepaths in Figure 1.

Figure 6 shows the two-dimensional wavenumber-amplitude spectra of the 10 Hz wavepath and raypath of Figures 1 and 2. The gross structure of the spectra is easily related to the space-domain patterns: both ray-theoretic and wavetheoretic tomography are most sensitive to spatial frequencies representing velocity variations paralleling the sourcegeophone axis and least sensitive to those perpendicular to the axis. The fine structure of the wavepath spectra requires more scrutiny. Devaney (1984) and Wu and Toksijz (1987) show how to construct spectra similar to those in Figure 6 from planewave scatteringarguments. Adapted from their work, Figure 7 illustrates two plane waves: emanating from source s, scattering at point a or b, and arriving at geophone g. For source and geophone plane waves described by k, and k,,

FIG. 3. Slice through a 3-D, 5 Hz Rytov wavepath in the constant-velocity background field of Figure I.

FIG. 4. Time-space domain wavepaths (S{L!$}) for the constant-velocity background field of Figure 1.

20

Woodward

the scatterer wavenumber is k, - k, . In plane rectangular s coordinates the wavenumber becomes: k,=2zsin (y) sin(T) (15)

where oS and og describe the angles k, and k, make with the x-axis, respectively. For a single source-geophone experiment, the possible combinations of plane-wave angles are constrained by the appropriate scattering ellipses. Parameterizing CY~ equation (15) in terms of equation (14) u and in s c produces Figure 8. The left and right panels relate the scattering ellipses of Figure 1 to the wavepath spectrum of Figure 6. The wavepath region on the source-receiver axis correspondsto a flat ellipse of eccentricity l(a = 1) and is sensitive to scatterers with wavenumbers outlining the circular holes. The wavepath region surrounding the sourcereceiver axis correspondsto ellipses of decreasingeccentricity (increasing a) and is sensitive to higher wavenumbers as labeled. The wavepath has a maximum spatial frequency twice that of the source wavefield (2ko), approached as a limit by the spacing of the most distant confocal ellipses.
Evanescent energy

The characteristic holes in the spatial-amplitude spectra of the monochromatic wavepaths arise from a causality (source/sink) condition placed on the source and geophone Green functions in equation (6). Replacing the source s Green function with its complex conjugate yields the s wavepaths and spatial-amplitude spectra shown in Figure 9. The familiar ellipses (the loci of points the sum of whose distances from source and geophone is constant) have been

FIG. 5. Imaginary parts of Rytov, frequency-space domain wavepaths for three inhomogeneous background-velocity fields, along with their raypath analogs. Top panels: refraction through a vertical gradient. Middle panels: transmission through three layers. Bottom panels: reflection above an interface (with direct, reflected, and refracted arrivals). FIG. 7. Plane-wave scattering: propagating energy.

FIG. 6. Two-dimensional wavenumber-amplitude spectra, origins in the centers. Left panel: 10 Hz wavepath. Right panel: raypath. (k,, = 2ke, k, = 2n * IO/Us.)

FIG. 8. Diagram relating wavepath scattering ellipses to wavepath spatial-amplitude spectra through the semimajor axis parameter a; a > 1. (The foci half-offset c equals 1.)

Wave-equation Tomography replaced by hyperbolas (the loci of points the difference of whose distances from source and geophone is constant): x2 -++-_=
a2-c2

21

z2
a2

1;

a < c.

(16)

This transformation physically results from the replacement of an exploding, causal source by an imploding, anticausal source. Just as energy from an exploding source that is scatteredby points along a single ellipse reaches a geophone in-phase, energy from an imploding source scattered by points along a single hyperbola reachesa geophonein-phase. Alternatively, Figure 6 represents propagating energy and Figure 9 evanescent energy. Anticausal scattering and the relation between space-domain hyperbolas and wavenumber-domain holes are diagramed in Figures 10 and 11.
INVERSION

showsthe velocity field used for the two-dimensional experiment: an anomalous circular region 500 m in diameter and 5-percent slower than a 2000 m/s background field. A single shot was positioned on the surface directly above the anomaly; multiple geophones were positioned every 40 m at a depth of 2000 m, up to a maximum offset of 1480 m. Rytov data As shown elsewhere (Woodward, 1989), the Rytov approximation is accurate for this example, thereby separating linearization problems from experimental-geometry effects. The data for the Rytov inversion was calculated from the shot profiles shown in the lower panels of Figure 12; the resulting phase delays and log amplitude ratios for offsets of 0,360, and 1060m appear in Figure 13. The phase delays are plotted as time delays, having been normalized by frequency. Because the phase delays are small, phase unwrapping was not the problem that it sometimes becomes in more complicated applications (Tribolet, 1977; Kaveh et al., 1984; Soumekh, 1988). Figure 14 illustrates the physical meaning of imaginary Rytov wavepaths as forward-modeling tools. The leftmost panels in Figure 14 superimpose5, 10, 15,20, 25, and 30 Hz first-Fresnel zones over the circular anomaly, for the three different offsets of Figure 13. Given that Fresnel-zone boundaries are equivalent to imaginary Rytov-wavepath zero crossings, the absolute maxima in Figure 13 can be

This section follows the differences between raypaths and Rytov wavepaths through inversion of a synthetic data set. The example is limited to wave-equation tomography under the Rytov approximation, as Rytov phase delays are both s more similar than Born scattered amplitudes to the travels time delays of ray theory and more appropriate for the transmission geometry of the test experiment. (See Oristaglio, 1985; Beydoun and Tarantola, 1988; and Lo et al., 1988, for Born-Rytov comparisons.) The upper panel in Figure 12

FIG. 10. Plane-wave scattering: evanescent energy.

FIG. 9. Evanescent, two-dimensional, frequency-space domain wavepath for a constant-velocity background field, along with its spatial-amplitude spectrum. The wavepath is formed by replacement of the source Green function in s equation (13) by iH,$2)(kolr1)/4. (Hh2 is a zero-order Hankel function of the second kind.) The velocity and axes are the same as those of Figures 1 and 6.

FIG. 11. Diagram relating evanescent wavepath scattering ellipses to evanescent wavepath spatial-amplitude spectra through the semimajor axis parameter a; a < 1. (The foci /half-offset c equals 1.)

22

Woodward trate relative minima and maxima at 1I and 17 Hz for the 360 m offset trace, and at 8 Hz for the 1060 m offset trace. Ray data Since weak scattering is apparent in the finite-difference data set of Figure 12, the linear phase-delay assumption of ray theory is slightly violated by this experiment. To prevent this linearization problem from interfering with the experimental geometry effects, ray traveltimes were not picked from the data set. Instead, traveltimes were calculated by tracing straight rays through the model. (The issue of event picking in the presence of geometrical frequency dispersion-nonlinear phase delay-is discussed below, and in Wielandt, 1987.) Figure 15 shows these traveltime delays, along with a raypath diagram for a geophone at 1000 m offset. The time delays agree with the high-frequency delays in the Rytov data.
Results

predicted by inspection of these diagrams: for each offset, they occur at that frequency for which the first Fresnel zone just encompassesthe anomaly. When the anomaly protrudes into the second Fresnel zone, it underlies a negative portion of the wavepath and contributes to a phase perturbation of opposite sign. In a similar manner, several relative maxima and minima in Figure 13 can be predicted by inspection of the rightmost panels in Figure 14. Here the boundaries of the first five Fresnel zones for several frequencies are shown superimposed on the anomaly, the sign of each wavepath alternating from positive to negative from the inside out. When the edge of the anomaly just grazes the inside boundary of a negative oscillation, it produces a relative minimum; when the edge grazes the inside boundary of a positive oscillation, it produces a relative maximum. The plots illus-

-1000

b-4

1 Ix-m

-0 El0 -0

N-

Velocity -1000

field
1000

Figure 16 showsthe results for: ray; monochromatic 5 Hz; monochromatic I5 Hz; monochromatic 25 Hz, and multifrequency 5, 10, 15,20, and 25 Hz inversions. All the examples were computed using LSQR, the conjugate gradient, linear system solver of Paige and Saunders (1982). The elongation of the images in the z-direction arises from the familiar finite-aperture problem of tomography (Menke, 1984): without horizontally placed shots and geophones, the spectral patterns of Figure 6 never sweep across the k, axis. The deterioration in the monochromatic images from the lowfrequency 5 Hz to the high-frequency 25 Hz-and on through the asymptotically high-frequency ray imagearises from the decreasing match between the central resolution vectors of the respective inversions and the anomaly s size and shape. (See Woodward, 1989, for a singular-value decomposition analysis of the example.) More simply ex-

;y

b
N-

Z-4 O-

36 \
2 Background
-1000 rl,

\
..

-A---_--___

wavefield
1000

zio-..-. ______________.._ \:. _.---

cu-

, 5 10

15

20

25

30

35

Full

wavefield

Frequency(Hz)
FIG. 13. Rytov data: complex-phase delay versus frequency, imaginary part normalized by frequency. The solid, dashed and dotted lines indicate offsets of 0, 360, and 1060 m, respectively.

FIG. 12. Two-dimensional transmission seismic experiment: velocity model with 5-percent slow circular velocity anomaly; shot profile through the background-velocity model; shot profile through the full model.

Wave-equation

Tomography

2. B

plained in terms of wavepath and raypath pictures, it is easier to construct the localized anomaly from a linear combination of broad ellipsoids than from a linear combination of narrow lines-given the narrow range of viewing angles available. All the monochromatic and ray inversions suffer from strong amplitudes at the source, the most highly illuminated area in the model. By using scattering information (the nonlinearity in the phase delays), the multifrequency inversion offers the sharpestimage and finally moves energy away from the source, truly localizing the anomaly in the center of the model.
BANDLIMITED RAYPATHS

The preceding sections have shown that multifrequency Rytov wave-equation tomography is superior to ray-trace tomography in dealing with geometrical frequency dispersion and inverting finite-aperture data. The method achieves this superiority because it is monochromatic, representing one extreme of the uncertainty relation: Athw 2 112. Since Ao is infinitely small, A.t is infinitely long and the entire seismic coda is utilized along with the first arrivals. The method accounts for scattered energy arriving at any time from any distance and its wavepath backprojection patterns cover the entire x, z plane. Unfortunately, wave-equation tomography superiority is gained at the expense of robusts ness. Because it is monochromatic, modeled events cannot simply be windowed from unmodeled events in time all of the signal and noise on a trace must be dealt with simultaneously. For any seismic experiment, the validity of the

-0

fl0 -0

l-4

monochromatic assumption depends on the knowledge 11 the source, the quality of the data, and whether modeht events can be separated from unmodeled events befolf: inversion. Ray-trace tomography is more robust than wave-equatic IL tomography because it selectsas signal the time delay of om: well-understood event. While in theory the fundament 11 assumption of ray-trace tomography is that of high fr.:. quency, in bandlimited practice it is more often that of ro geometrical frequency dispersion: that phase delay is a I proximately linear with frequency and described by : II average time delay picked from a wavelet peak. Instead 4I inverting amplitude changes and phase delays for a singIt: high frequency, ray-trace tomography actually inverts die slope of a phase-delay versus frequency curve that is a.;. sumed linear over an infinite bandwidth. Behavior over it broad range of frequencies is characterized by a singIf: measurement; scattering is forbidden and energy travel,, along the fastest source-receiver path. Ray-trace tomogr I. phy representsthe other extreme of the uncertainty relatio I. where At is infinitely short and Aw infinitely large. Raypatl, are compact because ray-theoretic tomography is essential ! infinite bandwidth. [The broadband nature of ray theory hi:,, also been pointed out by Foreman (1989).] In real seismic applications phase delay is rarely linear ar (I bandwidth is never infinite. The uncertainty inherent II picking a peak from a distorted, bandlimited wavelet I; usually incorporated in ray-trace tomography only inc i. rectly. Algorithms acknowledge that raypaths are far highs . in wavenumber than the model being inverted by smoothirlj: the inversion in several ways. The final cell representaticsih can be bandpassed;the model can be parameterized in tern, ; of smooth basis functions instead of cells, reducing ttic, high-wavenumber indeterminacy and expense of the pro ). lem at the outset (Dziewonski et al., 1977; Van Trier, 1988); the indeterminacy can be removed by imposing smoothne ; I constraints on the model during the inversion with damps(1 least squares(Menke, 1984a; Sword, 1988); the result can lil, smoothed by broadening the backprojection raypaths ther I.

-0 00 -0 Nd

-0 a0 -0
Nd

-1000

Offset (m)

1000

-1000

Offeet (m)

1000

a0 -0 *-0

FIG. 14. Imaginary Rytov wavepaths superimposed on a contoured outline of the circular velocity anomaly. Left panels: first-Fresnel zones for 5, 10, 15, 20, and 25 Hz at offsets0,360, and 1060 m. The ~-HZ first-Fresnel zone is the widest; the 25Hz zone the narrowest. Right panels: the boundaries of the first five Fresnel zones for several frequencies and offsets.

r.zl
1000
m

A
0
Offset (m)

-1000

1000

FIG. 15. Ray-data traveltime delays for all offsets, and die raypath for a geophone at 1000 m offset.

24
selves with convolutional quelling Rytov (a specialized et al., provide information form

Woodward of Three observations can be made illustrate about these bandlimited

weighted,

damped

least squares;

Meyerholtz wavepaths

1989). a way into reag(A@)

raypaths. First, they graphically in full waveform by a single domain when time smooths the medium because pick. how the extra information an event in the domain: time only highinto at inversion the event is lost when an event is specified Windowing in the frequency

Transmission-geometry of incorporating ray-trace sonable fashion. tomography and %(A@) respectively, tion normalized frequency: bandwidth tomography under

and uncertainty

in a more direct

and physically with

Ray tomography a nondispersive to be linear delay picked

can be redefined constraint: in frequency Under for an event

as Rytov and zero, this defini-

is nondispersive redundant. take they

is the discarded spreading

assumed

wavenumber Second, account, the source

information

over a specified average

bandwidth.

geometrical

the traveltime

becomes

the bandlimited and receiver

raypaths

are high in amplitude

of imaginary

phase perturbations

over

and low in amplitude of wavefront ray-trace to traditional

elsewhere. tomography, will a signal than

This weighting ~mrr .a[A@(w)] At =& I Wmln 0 pattern becomes d w . (17) a velocity

is an expression

healing (Claer-

bout, 1985). In contrast perturbation larger have a much Third,

close to the source or receiver on the recorded magnitude. the breadth of similar inversely Rytov

impact

more distant perturbation a similar wavepaths r)] over frequency: ited raypaths depends

The appropriate normalized

backprojection

and most importantly,

of the bandlimof the freas patband-

average of imaginary I
LA,(r) =

on the width

quency band summed over and not on the central frequency wrndr J+[Y(w, I Wrnl A0
-

dw.

(18)
raypaths integrating of Figures

of the band. Imaginary monochromatic imagined terns for infinite bandlimited-raypath as collapsing

wavepaths bandlimited The inverse

can be defined raypaths relation raypath

raypaths; bandwidth. width

can be between

to traditionally and

narrow

Figure

I7 shows examples (l8), monochromatic

of the bandlimited by numerically wavepaths Rytov

defined by equation the imaginary, integrations wavepaths (The wavepath between

formed

temporal-frequency

1 and 5 from 5 to 25 Hz, 15 to 2.5 Hz, and 20 to 25 Hz. The

cancel the rapidly oscillating while adding the smooth, of the equivalent outer regions of the first-Fresnel a linear zones. relation reflection-geometry

bandlimited in Figure

5 is omitted: and frequency transmission

since

phase delay

does not hold for this cannot be defined.)

case, a bandlimited

raypath

FIG. 16. Inversion

results.

FIG. 17. Two-dimensional bandlimited stant, vertical-gradient and layered fields of Figures I and 5.

raypaths for the conbackground-velocity

Wave-equation width is shown in more detail in Figure 18, with plots of half cross-sections through 3-D bandlimited raypaths for different combinations of o,,,, and wmax. While the relation is demonstrated more mathematically in Appendix A, it can be derived most intuitively from the uncertainty relation. In the uncertainty relation Ar corresponds to the time window of the trace examined (i.e., for traveltime picks. the sample rate). Since it dictates the averaging of frequency information in the Fourier domain, it is inversely proportional to the bandwidth of the applicable bandlimited raypath. Because it also limits the distance detectable scatterers can stray from the Fermat path, it is directly proportional to the bandlimited raypath width. s
CONCLUSIONS

Tomography

2!$

(Woodward, 1989 p. 55). This frequency dependence supgests the simultaneous inversion of separate bandregions b, a modified ray-trace technique.
ACKNOWLEDGMENTS

I thank the sponsors of the Stanford Exploration Projec 1 and Jon Claerbout for their generous support. I also than-: Fabio Rocca, who first intuited the existence of some lini; between my early fat-ray ideas and Rytov diffraction tomo& raphy. His guidance and encouragement were indispensabli: in the formalization and completion of this work. Finally, I must acknowledge the inspiration of this research by J. C Hagedoorn seismic-beam discussion, in his classic papt I s on seismic reflection interpretation (1954).
REFERENCES Beydoun, W. B., and Tardntola, A.. 1988, First Born and Rytcr approximations: Modeling and inversion conditions in a canonic I example: J. Acoust. Sot. Am., 83, 1045-1055. Beylkin. G.. 1982, Generalized radon transform and its application Ph.D. thesis, New York Univ. Bishop. T. N., Bube, K. P.. Cutler. R. T., Langan, R. T.. Lovt P. L.. Resnick, J. R., Shuey, R. T., Spindler. D. A., and Wyll H. W.. 1985, Tomographic determination of velocity and depth r laterally varying media: Geophysics, 50, 903-923. Chernov. L. A., 1960, Wave propagation in a random mediurl McGraw-Hill Inc. Claerbout, J. F., 1985, Imaging the earth s interior: Blackwcl Scientific Publications. Devaney, A. J.. 1981. Inverse-scattering theory within the Rytcj approximation: Optics Lett., 6, 374-376. ~ 1982, A filtered backpropagation algorithm for diffractic I tomography: Ultrasonic Imaging, 4, 336-350. 1984, Geophysical diffraction tomography: IEEE Tran I Geosci. Remote Sensing, GE-22. 3-13. Dines. K. A.. and Lytle, R. J., 1979, Computerized geophysic I tomography: Proc. IEEE. 67, 1065-1073. Dziewonski. A. M.. Hager, B. H., and O onnell. C R. J., 197 Large-scale heterogeneities in the lower mantle: J. Geophy ) Res.. 82. 239-255. Fawcett. J. A.. and Clayton, R. W., 1984. Tomographic reconstru : tion of velocity anomalies: Bull. Seis. Sot. Am.. 74, 2201-2215. Foreman, T. L.. 1989, An exact ray theoretical formulation of tl ( Helmholtz equation: J. Acoust. Sot. Am.. 86, 23&246. Hagedoorn. J. G., 1954, A process of seismic reflection interpret I. tion: Geophys. Prosp.. 2, 85-127. Kaveh. M.. Soumekh, M.. and Greet-deaf, J. F., 1984, Sign.ll processing for diffraction tomography: IEEE Trans. Son. Ultr I\on.. SU-31. 23C-238. Keller. J. B.. 1969. Accuracy and validity of the Born and Rytc I approximations: J. Opt. Sot. Am.. 59, 1003-1004. Lo. T.. Toksiiz. M. N.. Xu, S.. and Wu, R.. 1988. Ultrasor 1: laboratory tests of geophysical tomographic reconstruction: Ge 1. physics. 53. 947-956. Menke. W.. 1984. The resolving power of cross-borehole tomogr.lphy: Geophys. Res. Lett., 11, 105-108. Mersereau. R. M., and Oppenheim, A. V.. 1974. Digital reconstru,:tion of multidimensional signals from their projections: Pro,:. IEEE. 62. 1319-1338. Meyerholtz. K. A.. Pavlis, G. L., and Szpakowski. S. A., 19e . J Convolutional quelling in seismic tomography: Geophysics, 5.l. 57g-580. Miller D., Oristaglio, M., and Beylkin, G.. 1987, A new slant tII seismic imaging: Migration and integral geometry: Geophysic q, 52. 943-964. Mora. P., 1987. Elastic wavefield inversion: Ph.D. thesis, Stanfo -11 University. Mueller. R. K.. 1980. Diffraction tomoeraohv I: The wave eauatioll: Ultrasonic Imaging,2, 213-222. - Mueller. R. K., Kaveh, M., and Wade, G., 1979, Reconstructi , : tomography and applications to ultrasonics: Proc. IEEE, 67,
567LFK7

This paper clarifies the differences between ray and wave tomography through pictorial comparisons of their respective backprojection patterns. These comparisons underscore the parallelism between ray-trace and Rytov, wave-equation tomography, and suggest ways of modifying rays to reflect the bandlimited nature of most seismic data. In contemplating any ray inversion, thought must be given to what picked traveltimes represent: asymptotically high-frequency ray arrivals, or group velocities characteristic of the source bandwidth. Very preliminary results suggest traveltimes picked from wavelet peaks may be best modeled with bandlimited raypaths, formed by narrowly averaging monochromatic wavepaths around the dominant source frequency

.-l

6i)O
Half-width

1600
x (m)

i,

600

1600
x (m)

Half-width

Lh:

55-75

Hz

-3

GO Half-width

id00 x (m)

500 Half-width

lciO0
x (m)

FIG. 18. Bandlimited-raypath width. Half cross-sections through 3-D constant-velocity bandlimited raypaths (dotted lines) and their envelopes (solid lines) for four different combinations of mm, and mrnax.The cross-section is taken midway between the source and receiver: the source-receiver separation is 2000 m; and the background velocity is 2000 m/s.

Oristaglio. M. L.. 1985, Accuracy of the Born and Rytov appro. Imations for reflection and refraction at a plane interface: J. 01 . 1 Sot. Am.. A. 2. 1987-1992.

26

Woodward 1987. Inversion of travel times and seismic waveforms in Nolet. G., Ed., Seismic tomography: D. Reidel PublishingCI In-

Paige, C., and Saunders,M., 1982, LSQR: An algorithm for sparse linear equations and sparse least squares: ACM Trans. Math. Software, 8, 43-71. Schneider, W. A., 1971, Developments in seismic data processing and analysis (1968-1970): Geophysics, 36, 1043-1073. Slaney, M., Kak. A. C., and Larsen, L., 1984. Limitations of imaging with first-order diffraction tomography: IEEE Trans. Microwave Th. Tech., MTT-32, 860-873. Soumekh, M., 1988, Phase reconstruction/unwrapping from amplitude for diffracted waves usinga perturbationsolutionof the wave equation: IEEE Trans. Acoust. Speech Signal Proc., ASSP-36, 107&1085. Sword, C. H., 1987, Tomographicdetermination of interval velocities from reflection seismic data: the method of controlled directional reception: Ph.D. thesis, Stanford Univ. Tarantola, A., 1984, Inversion of seismic reflection data in the
acoustic approximation: Geophysics, 49, 1X9-1266.

pany, 135-157. Tribolet, J., 1977,A new phaseunwrappingalgorithm: IEEE Tn I s. Acoust. Speech, Signal Proc., ASSP-25, 17&177. Van Trier, J., 1988, Migration-velocity analysis using geologic al constraints: 58th Ann. Internat. Mtg., Sot. Explor. Geoph,,;., Expanded Abstracts, 897-900. Wielandt, E., 1987, On the validity of the ray approximation lor interpreting delay times, in Nolet, G., Ed., Seismic tomograp1y: D. Reidel PublishingCompany, 85-98. Woodward, M. J., 1989,Wave-equation tomography: Ph.D. theiIs, Stanford Univ. Woodward, M. J., and Rocca, F., 1988, Wave-equation tomo: a phy: 58th Ann. Intemat. Mtg., Sot. Explor. Geophys., ExpanIIed Abstracts. 1232-1235. (Also available as SEG slide set No. 10.) Wu, R. S.. and Toks(iz, M. N., 1987, Diffraction tomography 3nd
multisource holography applied to seismic imaging: Geophy: i :s, 52. 11-25.

APPENDIXA
BANDLIMITED-RAY WIDTH The equation describing the envelope (A) of the bandlimited raypaths show in Figure I8 is: The width of the bandlimited raypath is determined by the term dominating A behavior for small x. In equation (1 .I) s ~ this is the factor multiplying 2/At4: 1 - cos (Athw). I he first zero of this term marks the boundary of the first n,;iin lobe in A, echoing the uncertainty relation discussion H i.h: At& = 27~. The simple inverse relation between At and \w yields a more complicated inverse relation when solved r o the half-width X:

[I -

cos

(AtA@)] - 2

ho sin (AhrAw)

cos

(AtAm) + 1,x + Wiin]

(A-1)

Here b is the product of the Green function geometricals spreading terms in equations (6) and (13): 4rrd b= x2 f d214 (A-2) x = v : (E+ d). (1lL.4)

where d is the source-receiver separation and .Y is the offset from the source-receiver axis. At is the time delay between the direct and x-scattered arrivals:

For large Aw (or d >> ~zJ/Aw) this simplifies to:

At=;

(2+2+;-d).

(A-3)

x=

ml6 -. A0

(. i-5)