COMMON MIDPOINT VERSUS COMMON OFFSET ACQUISITION GEOMETRY IN SEISMIC IMAGING

. We compare and contrast the qualitative nature of backprojected images obtained in seismic imaging when common oﬀset data are used versus when common midpoint data are used. Our results show that the image obtained using common midpoint data contains artifacts which are not present with common oﬀset data. Although there are situations where one would still want to use common midpoint data, this result points out a shortcoming that should be kept in mind when interpreting the images.


1.
Introduction. In seismic imaging, pressure waves generated on the surface of the earth travel into the subsurface, where they are scattered by heterogeneities. The scattered pressure waves returning to the surface are picked up by receivers and the inverse problem involves imaging the subsurface from these scattered waves. This problem is non-linear and as such is very difficult to solve. In this article, we model the heterogeneities in the subsurface as singular perturbations about a constant velocity field. We consider the corresponding linearized map F which maps the perturbation in velocity to the corresponding perturbation in the pressure waves. The goal is to recover, to the extent possible, the perturbation in the background wave speed from the perturbations in the pressure waves measured on the surface. In other words, we would like to study the invertibility of the linear operator F .
We compare two standard (see [33]) ways of collecting data in seismic experiments; the common midpoint acquisition geometry and the common offset acquisition geometry to determine which geometry shows more features of the subsurface and adds fewer artifacts. To do this, we will use microlocal analysis, which provides a precise characterization of how operators like F and its adjoint, F * , map singularities.
The nature of F and F * F depends on the set of sources and receivers. In the case of a single source and receivers occupying a relatively open subset of the surface, Beylkin [2] proved that if caustics do not occur then F * F is a pseudodifferential operator (ΨDO). Rakesh then showed that F is a Fourier integral operator (FIO) [32] under the assumption of no caustics in the source ray field, but caustics can occur on the receiver side. Later, Nolan and Symes [26] studied a variety of acquisition geometries and gave sufficient conditions for F * F to be a ΨDO. Ten Kroode, Smit, and Verdel [39] considered relatively open sets of sources and receivers and investigated the microlocal properties of F in relation to the so-called travel time injectivity condition; see [26,Assumption 3,pp. 929] and [39]. Stolk [35] studied the microlocal analysis of F and F * F even when the travel time injectivity condition fails.
To describe the data acquisition geometries, we introduce some notation. We denote the subsurface to be imaged by the half space, (1)

COMMON MIDPOINT VERSUS COMMON OFFSET 3
Seismic sources and receivers will be parametrized by points in a bounded open set Ω ⊂ R 2 (where we identify R 2 with the plane x 3 = 0). For each point s = (s 1 , s 2 ) ∈ Ω we specify a source S(s) and a receiver R(s) on the surface of the earth, x 3 = 0. Data are taken for different times, t > 0, so the data space is In the common midpoint data acquisition geometry, the source and receiver move away from a common midpoint. For this case, Ω ⊂ (s 1 , s 2 ) ∈ R 2 : s 2 > 0 and for each s = (s 1 , s 2 ) ∈ Ω, Hence the sources and the detectors move over the plane and are symmetric with respect to the x 1 axis. In the common offset data acquisition geometry, the source and receiver collecting data on the surface of the earth are offset by a constant vector. Choose α > 0 and for each s = (s 1 , s 2 ) ∈ Ω let So the sources and the detectors move so they are always (0, 2α, 0) apart. These data acquisition geometries are realistic because seismic data can be synthesized to provide both common midpoint and common offset data [33]. Although we have made simplifying assumptions of constant background velocity, if one data acquisition method is shown to be better in this case, it is likely to be better at least for small perturbations away from it. This is because bicharacteristics depend smoothly on the background velocity as they involve solving an ODE and one can invoke smooth dependence of ODE solutions on parameters. This means that the wavefront relation also depends smoothly on the parameters. Therefore if artifacts appear for a constant background velocity, they would still be present when we make small perturbations away from the constant background velocity.
Under our assumptions (and with our data acquisition geometries), the forward operator can be described (see [26] for example) as an FIO of order m, where f models the singular velocity perturbations in the subsurface from the known constant background velocity: with phase function and a smooth cutoff function ϕ(s) that is supported in Ω and identically 1 in an arbitrary large compact proper subset of the support of ϕ. The canonical relation of F is given by where ∂ s is the differential in the s variables, etc. [40]. The formal L 2 adjoint operator, F * , is an FIO associated with C t (which is C with T * (X) and T * (Y ) coordinates flipped, if C = {x, ξ; y, η} then C t = {y, η; x, ξ}).
One standard reconstruction method in imaging is to apply F * to the data. The result is the normal operator F * F applied to f . Analyzing what this normal operator does to singularities of f will help determine whether it is a good reconstruction method and, if so, whether one data acquisition method is better than another. Remark 1.1. In order to have a heuristic understanding of whether one acquisition geometry can be better than the other, consider the following special case. Assume the smooth component of the velocity in the subsurface is layered, i.e., only depends on x 3 . Reflectors in the subsurface can be detected in the data when there is an ellipsoidal surface (with foci at the source and receiver) which is tangent to the reflector. That is, the "dip" or normal to the reflector is also normal to the ellipsoid at the point of tangency. With this in mind, we see that the common midpoint data can only detect reflectors which are located on the plane x 2 = 0 and which also have a vertical dip. In contrast, the common offset geometry can detect reflectors which have a range of locations (including away from x 2 = 0) and also a range of dips.
Heuristically, this remark shows why imaging using a common offset geometry could be expected to be better than imaging using a common midpoint geometry. The authors thank one of the referees for pointing out this special case. However, it is also worth pointing out that the results of this paper (regarding the nature of the normal operator F * F ) do not directly follow from such heuristic arguments.
This pattern, that common offset can be better in some ways than common midpoint, also occurs in synthetic aperture radar (SAR), in which the two-dimensional topography of the earth is imaged. In the common offset SAR geometry it was shown that the normal operator introduces one artifact singularity into the reconstructed image in addition to each true singularity, whereas in the common midpoint case, three artifact singularities are introduced into the reconstructed image in addition to each true one. A more precise description of the normal operators in these cases are given in [23,1].
We now present the main results of this article and then interpret them.
We will define I p,l classes in section 2 but this theorem has the following practical implication. Microlocalized away fromC, F * cm F cm is a standard ΨDO of order 2m. So singularities of f can be visible in the image F * cm F cm (f ). Microlocalized away from ∆, F * cm F cm is a FIO associated toC of order 2m. So a singularity of f at (x, ξ) can produce an artifact in the image which is of the same strength at On the other hand, the common offset normal operator has the following properties.

Theorem 1.3 (Microlocal Properties for Common Offset).
Let F co be the common offset forward operator defined in (3), (4), and (5), of order m then F * co F co is a standard ΨDO of order 2m.
This theorem means that singularities of f can be visible in F * co F co and artifacts are not added to the reconstruction. Our theorems show that common offset can be better than common midpoint data acquisition since singularities are not added to the reconstruction using the normal operator. We remark that despite this advantage of common offset data acquisition compared to common midpoint data, different (and perhaps more desirable) singularities might be visible with common midpoint acquisition.
Our proof rests on a fundamental assumption of Guillemin and Sternberg [20,17]: the Bolker Assumption,i.e., π L is an injective immersion (see section 3.2). We show that F co satisfies this Assumption and this implies that F * co F co is a ΨDO, thus the singularities of f can be reconstructed from the singularities of F co . We would like to point out that [3] provides another way to obtain the reconstruction of the singularities. It uses the Beylkin Determinant for common offset, which is nonzero (see pp. 247-249,(5.2.30))and an explicit inversion formula for the ground reflectivity function (5.1.56). Using microlocal techniques we obtain the same nonzero determinant of dπ L (see (22) and subsequent calculations in section 3.2) which implies that this projection is an immersion. This is a part of the Bolker Assumption.
The proofs in Section 3 use similar techniques as the ones done in [8,1] and the pertinent details will be given.

2.
Microlocal analysis and I p,l classes. We present some basic definitions in microlocal analysis.
Now we define the types of geometric singularities exhibited by the canonical relation of our operators.
(iii) f * Ω vanishes exactly to first order on Σ, then we say that f has a fold singularity along Σ. (b) If for all σ ∈ Σ, we have (i) the rank of f is constant; let us call this constant k, (ii) ker(df σ ) ⊂ T σ Σ, (iii) f * Ω vanishes exactly to order n − k on Σ, then we say that f has a blowdown singularity along Σ.
The fundamental mathematics for I p,l classes is in [21,25] and the techniques we use appeared initially in [15] and we follow [9]. They were used in the context of radar imaging in [29,8,9]. First, consider the following example.
The class of product-type symbols S p,l (m, n, k) is defined as follows.
Since any two sets of cleanly intersecting Lagrangians are equivalent via a canonical transformation [21], we first define I p,l classes for the case in Example 2.3.
v i is locally finite and v i = Aw i where A is a zero order FIO associated to χ −1 , the canonical transformation from above, and w i ∈ I p,l (Λ 0 ,Λ 1 ).
If u is the Schwartz kernel of the linear operator F , then we say F ∈ I p,l (Λ 0 , Λ 1 ).
One way to show that a distribution belongs to an I p,l class is by using the iterated regularity property: if there is an s 0 ∈ R such that for all first order pseudodifferential operators P i with principal symbols vanishing on Λ 0 ∪ Λ 1 , we have P 1 P 2 . . . P r u ∈ H s0 loc .
3. Proofs. We first calculate the canonical relation C for each operator. Then we analyze the geometry of the right projection π R : C → T * (X) \ {0} and the left projection π L : For the common midpoint case, we show that π L and π R both drop rank by 1 on the submanifold Σ of C cm that is above x 2 = 0. We show π L has a fold singularity on Σ and π R has a blowdown singularity (similar to common midpoint SAR). To find the microlocal properties of F * cm F cm , we prove C t cm • C cm ⊂ ∆ ∪C whereC is the graph of the reflection We then show that F * cm F cm ∈ I 2m,0 (∆, G) using arguments similar to those in [1] for common midpoint SAR that use the machinery in [15,9]. The normal operator F * cm F cm is simpler than in common midpoint SAR [1] because there is only one geometric symmetry (across x 2 = 0), whereas in common midpoint SAR [1], there is symmetry about the x 1 = 0 and x 2 = 0 axes.
Common offset is easier and, we show that π L : C co → T * (Y ) \ {0} is an injective immersion so the Bolker condition holds and F * co F co is a standard pseudodifferential operator.
3.1. Proof of Theorem 1.2. In this case we consider S cm (s 1 , s 2 ) = (s 1 , s 2 , 0) and We use the notation: Notice that the canonical coordinates on C cm are (s 1 , s 2 , ω, x 1 , x 2 , x 3 ). We assume that x 3 = 0, that is, that scatterers are contained in the open subsurface x 3 > 0. Next, we consider the projections π L and π R . We have Note that, in the description of π L , we reordered the components to obtain the identity in the first variables. This ordering will help to find the determinant of dπ L . We get dπ L to be Hence π L drops rank by 1 on Σ = {(s, ω, x) ∈ C cm |x 2 = 0} since we assume x 3 , s 2 = 0 and since 1 Acm  velocity, we also make the assumption that no direct scattering takes place. In practice, such direct data are usually filtered out. Therefore this is a reasonable and valid assumption.
Next we classify the singularities of π L on Σ. Notice that on Σ, A cm = B cm and the second column of the matrix C above is 0. We have that v ∈ Ker dπ L , if v = (0, 0, 0, δx 1 , δx 2 , δx 3 from the first and the second rows of the above matrix. Notice that the last row does not give us new information about the kernel. The 2 × 2 system we considered has only the zero solution. Thus Ker dπ L =span ∂ ∂x2 . We have that Ker dπ L T Σ, hence π L has a fold singularity.
We now show that F * cm F cm ∈ I 2m,0 (∆, C) where F is given by (4) with the phase function φ = φ cm given by (6). Our strategy of proof is to use the iterated regularity property characterizing I p,l classes due to Melrose and Greenleaf-Uhlmann (Proposition 2.7) and is similar to the ones given in [8,1]. We have that the Schwartz kernel of F * cm F cm is given by K(x, y) = e iΦ(x,y,s1,s2,ω) a(x, y, s 1 , s 2 , ω)ds 1 ds 2 dω, where Φ(x, y, s 1 , s 2 , ω) = ω (y 1 − s 1 ) 2 + (y 2 − s 2 ) 2 + y 2 3 + (y 1 − s 1 ) 2 + (y 2 + s 2 ) 2 + y 2 The ideal of functions vanishing on ∆ ∪ C is generated by . We will now write in terms of derivatives of Φ and smooth functions, the generators x 1 − y 1 and x 3 − y 3 . Such expressions for the other generators are similar.
3.1.1. Expression for x 1 − y 1 . Let us write x 1 − y 1 in the form where f, g, h are smooth functions and Φ is the phase function (9) for the Schwartz kernel (7). Let us use the following prolate spheroidal coordinate system: x 3 = s 2 sinh ρ sin φ sin θ.
For y = (y 1 , y 2 , y 3 ), we use the above coordinates but with (ρ, φ, θ) replaced by (ρ ′ , φ ′ , θ ′ ). In this coordinate system, we have A ′ cm and B ′ cm are similarly defined with (ρ, φ) replaced by (ρ ′ , φ ′ ). Furthermore, Now Adding and subtracting cosh ρ in the first term on the right hand side of Equation (15) (the equation above), and rearranging, we get
Simplifying, we obtain, Hence We now write this expression in Cartesian coordinates using the notation from (11).
We have the following expressions: Thus x 1 − y 1 becomes 3.1.2. Expression for x 3 − y 3 . Next we consider x 3 − y 3 . We have We already have an expression for (x1 −y1) from (19), the second expression in the right hand side of (20) can be written in terms of ∂ωΦ (see (12)) and for the third expression in (20), we use (17). Now since x3 − y3 = x 3 +y 3 and due to the assumption that the support of the function lies above the x1-x2 plane, x3 + y3 will never be zero; see Remark 3.1. Therefore, we have an expression in terms of the derivatives (12), (13) and (14) for x3 − y3.
We can find expressions similar to the one for (x1 − y1) in (19) for the remaining terms in (10) by using (12)- (14) and (17) as in the above two calculations. Finally we proceed as in [8,1] to show that F * cm Fcm ∈ I 2m,0 (∆,C).
From (24), knowing ρ and φ we can uniquely and smoothly determine θ. Thus π L is an embedding, and the Bolker Assumption holds, so by [20,17], we have F * co F co ∈ I 2m (∆) which means that F * co F co is a pseudodifferential operator and hence the normal operator does not introduce additional singularities.
Remark 3.2. Now that we have proven that the normal operator for common offset geometry is a ΨDO but the normal operator for common midpoint is not, we would like to make some geometric observations about the associated canonical relations.
For the common midpoint wavefront relation, C cm , we see that pairs of wavefront set elements (x 1 , x 2 , x 3 ; ξ 1 , ξ 2 , ξ 3 ) and (x 1 , −x 2 , x 3 ; ξ 1 , −ξ 2 , ξ 3 ) are associated to the same wavefront element of the data. Note that this symmetry in the wavefront relation also essentially comes from an obvious reciprocity between the source and the receiver. This shows that the canonical relation C cm is (at least) a 2:1 relation. Hence we intuitively expect artifacts and that F * cm F cm is not a ΨDO; see Remark 1.1 as well. No obvious symmetry is present in the common offset geometry and it is not surprising that F * co F co is a ΨDO. Our theorems make precise these intuitive observations. Additionally the fact that F * cm F cm is in an I 2m,0 class shows that artifacts are added by our reconstruction method and they are of the same strength as the original singularities that generate them.