Corresponding author: P. Audet, Department of Earth Sciences, University of Ottawa, Ottawa, ON, Canada. (firstname.lastname@example.org)
 Identifying the sources of seismic anisotropy in subduction zone forearcs is key to understanding mantle deformation processes. Current models based on the interpretation of shear-wave splitting measurements favor the flow-induced alignment of olivine crystals around down-going slabs or the presence of foliated serpentine minerals due to subduction-related processes, thereby neglecting fossil slab fabric as a significant source of anisotropy. We use seismic receiver functions to show evidence for strong anisotropy within anhydrous uppermost mantle of subducting oceanic plates in the forearcs of Nankai, Cascadia, Mexico, and Costa Rica subduction zones. Seismic anisotropy models from the inversion of receiver function waveforms are consistent with fossil fabric generated at spreading ridges.
 Seismic anisotropy—the variations of seismic wave velocities as a function of the direction of propagation—is a powerful tool to study deformation in a wide range of tectonic environments, because it primarily reflects fabrics acquired by strain-induced alignment of anisotropic minerals within major rock constituents [Park and Levin, 2002]. In a subduction zone setting, seismic anisotropy arising from patterns of upper mantle flow can provide key information on factors affecting the rheology and deformation of mantle materials at high strain rates [Jadamec and Billen, 2010]; however, accurately locating the source of anisotropy remains a challenge [Long and Silver, 2009]. In the upper mantle the anisotropy is generally attributed to the crystal-preferred orientation of olivine, with fast wave polarization coinciding with the crystallographic a axis [Hess, 1964; Ribe, 1989]. The pervasive presence of serpentine minerals in regions where the upper mantle has been significantly hydrated may also impart strong seismic anisotropy with slow wave polarization along the crystallographic c axis [Mainprice and Ildefonse, 2009].
 The direction and magnitude of seismic anisotropy within subduction zone forearcs is generally estimated by measuring the splitting of shear waves along orthogonal slow and fast directions of propagation from either regional (direct S) or teleseismic (core-refracted SKS and SKKS) phases [Park and Levin, 2002; Long and Silver, 2009]. Global compilations of shear-wave splitting measurements in subduction zone forearcs show consistently trench-parallel fast directions of anisotropy, except in a few isolated cases (e.g., Cascadia and parts of South America) [Long and Silver, 2008]. Explanations of these results generally involve the mantle flow-induced alignment of olivine crystals around down-going slabs [Long and Silver, 2008, 2009]. Alternatively, some authors have suggested anisotropy arising from the presence of serpentine minerals within the hydrated mantle wedge [Katayama et al., 2009] or within hydrated through-going faults in the subducting plate created by bending of the plate at the outer rise before entering the trench [Faccenda et al., 2008]. None of these models consider primary (fossil) fabric within the down-going slab mantle as a significant source of anisotropy. In contrast, deformation experiments and seismic observations indicate that asthenospheric flow at mid-oceanic spreading ridges imparts a strong crystallographic fabric of dunites in the upper mantle section of oceanic lithosphere [Christensen, 2004] that may be preserved as fossil fabric in subducting slabs. The dismissal of fossil lithospheric fabric in studies of subduction zone anisotropy may be due to the lack of direct observational evidence for slab-induced anisotropy; indeed, shear-wave splitting measurements have poor depth resolution and cannot resolve thin (<20 km) anisotropic layers [Long and Silver, 2009].
 Recent observations of converted teleseismic waves (or receiver functions) in southern Mexico reveal depth-localized anisotropy within the uppermost mantle of the subducting Cocos plate [Song and Kim, 2012]. Receiver functions are particularly well suited for studies of layered anisotropy because they are sensitive to jumps in velocity structure with scale lengths of 1–10 km. These observations indicate the presence of a 2–6 km subcrustal layer with ~7% anisotropy, interpreted as fossil fabric from Pacific-Cocos ridge spreading [Song and Kim, 2012]; however, whether or not such anisotropy is widespread in the uppermost mantle of subducting slabs is currently unresolved. In this paper we use observations of receiver functions recorded over broadband stations in the forearc of shallow subduction zones in southwest Japan, southern Vancouver Island (Canada), southern Mexico, and the Nicoya Peninsula (Costa Rica) to detect the presence of the anisotropic mantle lid (AML) previously found in Mexico. The rationale for selecting these regions is based on three criteria for the accurate extraction of structural parameters: (1) the low dip of young subducting slabs where the plate interface is shallow, which is an important condition for untangling the trade-off between depth and velocity estimates using free-surface reverberations; (2) the subduction of normal oceanic crust with well resolved ocean basin structure; and (3) the availability of broadband stations with long recording time periods to ensure proper sampling of back-azimuth and incidence angles. The last condition is particularly important in resolving structural parameters for dipping, anisotropic media.
2 Data and Method
 Data used in this study come from several permanent and portable networks of broadband seismic stations (see Table S1 in the auxiliary material). At each station we compile all available P-wave seismograms with high signal-to-noise ratio (> 7.5 dB) on the vertical component from M > 5.8 earthquakes at teleseismic distances that occurred between January 1990 and December 2011. The number of useful events at each station depends on earthquake occurrence relative to station location, as well as total deployment time period (Table S1). Three-component seismograms are decomposed into P and S up-going wave modes and deconvolved using Wiener spectral deconvolution to obtain radial and transverse P receiver functions. After deconvolution, receiver functions are filtered at corner frequencies of 0.05 to 0.5 Hz using a two-pass Butterworth filter and stacked into 7.5° back-azimuth and 0.002 s km–1 slowness bins. Examples of binned receiver functions for station PGC (northern Cascadia) are shown in Figure 1. The resulting receiver functions show coherent signals representing direct P-to-S (Ps) conversions and free-surface P-to-S (Pps) and S-to-S (Pss) reverberations from velocity contrasts within the underlying column.
 To first order these data are dominated by the signature of a dipping low-velocity layer (LVL) [Audet et al., 2009]. This signature comprises double polarity pulses of converted P-to-S phases (Ps at ~5 s at station PGC) and back-scattered (Pps and Pss at ~15 and ~20 s, respectively at station PGC) waves seen in the radial component of band-pass filtered receiver functions (Figures 1A–1C). Arriving 1–2 s after LVL-related Ps signals we find a converted phase with strong azimuthal variations in amplitude and/or polarity in both radial and transverse components. At station PGC, this phase is evident as a strong negative pulse in the radial component at back-azimuths corresponding to events arriving from the southeast, and absent at other azimuths (Figures 1C and 1D). For the former case, the strong negative arrivals indicate conversions from normal isotropic oceanic mantle velocity below to directionally fast velocity within a localized layer that imparts comparably strong positive conversions from the top of the layer and bottom of the LVL. The absence of such signal at other azimuths indicates directionally normal velocity contrasts from both the bottom and top of this layer. These signals represent layering within the subducting slab mantle whose seismic signature does not correspond to dipping low-velocity material and therefore represent the signature of a localized anisotropic layer immediately below the subducting oceanic crust, corresponding to the AML previously found in Mexico [Song and Kim, 2012].
 To characterize subduction zone seismic structure, we invert receiver function waveforms using a fast ray-based forward algorithm for waves in dipping, anisotropic media [Frederiksen and Bostock, 2000]. The Monte Carlo inversion is carried out using a Neighborhood Algorithm [Frederiksen et al., 2003]. We use a correlation-based misfit since receiver functions are more sensitive to phase than absolute amplitudes [Frederiksen et al., 2003]. Cumulative variance within each receiver function bin is used as inverse weight in the misfit calculation. We first obtain an isotropic background velocity model for each subduction zone by modeling radial and transverse component waveforms (Ps, Pps, and Pss phases converted from the top and bottom of the LVL) using a planar dipping oceanic crust overlain by continental forearc crust and underlain by a mantle half-space. Some stations further resolve structure within the oceanic crust consisting of an upper LVL and a lower layer with normal oceanic velocities referred to as the lower oceanic crust (LOC) [Hansen et al., 2012]. The presence or absence of this feature has no impact on the robustness of the inversion for the AML (below). From this model we estimate the strike and dip of the LVL (with or without the LOC), the compressional-to-shear velocity ratio (Vp/Vs) and thickness of the dipping LVL (and LOC) and the overlying forearc crust. Example results for station PGC are shown in Figure S2.
 In a second step, we use the isotropic background velocity model and search for the presence of a dipping AML by modeling forward converted Ps phases from the top and bottom of the AML. The AML is modeled using background oceanic mantle P and S velocities and the inversion searches for four parameters: thickness, percent anisotropy, trend and plunge of a fast axis of hexagonal symmetry. Models that use hexagonal symmetry with a slow axis of anisotropy are unable to reproduce receiver function signatures (i.e., anisotropy is strictly positive in Figures S3 and S4). In the inversion upper mantle P and S velocities are fixed to 8 km s–1 and 4.5 km s–1 for all subduction zones to simplify comparisons, although absolute velocities at uppermost mantle conditions will depend on thermal structure that may vary between different subduction zones [Hacker et al., 2003]. Only the forward conversions are used in the inversion, which hampers our ability to simultaneously resolve the background S-velocity and the thickness of the AML. Example results for stations PGC (northern Cascadia) and PNCB (Costa Rica) are shown in Figures S3 and S4. Simple tests indicate that the orientation of the fast axis is well resolved, whereas the thickness and strength of anisotropy trade-off with the selected background velocity model and imply that our results represent upper bounds (auxiliary material). In particular, a 5% difference in background mantle velocities can result in errors in the estimated thickness of the AML of ~15% (Figure S5).
3 AML and Fossil Spreading
 Our results show consistent structural parameters of the AML within each subduction zone but significant variations across them (Figure 2). The thickness of the AML averages 5–25 km with lowest values (~5 km) found in Cascadia and highest values (>20 km) found beneath the Nicoya Peninsula. Calculation of the total incoming oceanic lithosphere thickness from thermal modeling [Conrad and Lithgow-Bertelloni, 2006] indicates that the AML can make up >50% of the lithospheric column; however, the AML thickness does not scale with the age of the subducting plate (Figure 3A). The magnitude of anisotropy is variable and ranges between 10 and 25% (Figure 3B). These values allow us to rule out dense crustal metamorphic facies (e.g., eclogite) as candidates to explain the anisotropic layer for several reasons. First, the AML thickness represents a significant fraction of the total oceanic lithosphere column. Second, high-grade metamorphic crustal rocks exhibit only weak anisotropy [Christensen, 2004] and cannot explain the high values found here. Instead, our measurements are more likely to represent highly sheared mantle rocks such as dunites or serpentinites that are abundant in the uppermost mantle section of ophiolite outcrops [Christensen, 2004].
 Bearing in mind that our estimates likely represent upper bounds, the large anisotropies found here are close to single-crystal values for olivine obtained from simple shear experiments conducted at large (150%) strains [Zhang and Karato, 1995], but could also represent significant contributions from pyroxenes with measured anisotropy levels of about 15% that are compatible with those found here [e.g., Pera et al., 2003]. Possible alternative lithologies are serpentinites (antigorite being the stable variety at uppermost mantle conditions), which exhibit large anisotropies [Mainprice and Ildefonse, 2009]. However, serpentinites are characterized by much lower seismic velocities (comparable to those of oceanic crust) as well as a slow axis of hexagonal symmetry [Bezacier et al., 2010; Kern et al., 1997; Mainprice and Ildefonse, 2009].
 Our anisotropy estimates are about twice as large as previous estimates obtained from Pn measurements sampling the top 5–10 km of oceanic mantle in ocean basins [Hess, 1964; Gaherty et al., 2003; Shinohara et al., 2008]. However, Shimamura and Asada  used long-offset profiles and found 13% P-wave anisotropy at depths of 40–50 km below seafloor beneath the western Pacific. Quantitatively, our estimates should be closer to those of Song and Kim  (7% anisotropy within the topmost 2–6 km of subducted oceanic mantle), who used receiver functions to show evidence of the AML beneath southern Mexico. We posit that differences between our estimates and those of Song and Kim  are due to different data resolution, processing, and modeling strategies (see auxiliary material).
 The orientation of the fast symmetry axis is largely coherent within each subduction zone, trending at a high angle to the direction of relative plate motion in Shikoku (Nankai, southwest Japan) and Cascadia and at a low angle in Mexico and the Nicoya Peninsula (Figure 2). The trends of the fast symmetry axis and fast shear-wave splitting directions from SKS phases are consistent in Japan [Long and van der Hilst, 2005] but not in Cascadia [Currie et al., 2004]. Variability in the orientation of anisotropy with respect to relative plate motion across different subduction zones is difficult to reconcile with flow-induced fabric due to subduction processes. In addition, it is unlikely that the uppermost ~20 km of the subducting slab mantle can be affected by subslab asthenospheric flow. Other explanations such as the presence of pervasive serpentinite dykes from bending-related faults in the upper mantle are again not favored as they predict consistent trench-parallel fast wave propagation [Faccenda et al., 2008].
 Alternatively, our results are more likely to represent fossil fabric acquired at mid-ocean ridges from the alignment of olivine crystals (crystallographic a axis) in the spreading direction [Blackman et al., 2002; Shinohara et al., 2008; Tono et al., 2009]. Our results agree with this model and show the trend of the fast symmetry axis oriented perpendicular to the fossil spreading ridge axis determined from magnetic lineations offshore (Figures 3C and S8). We thus posit that the AML is a ubiquitous feature of the oceanic upper mantle that retains its characteristic signature upon entering the subduction zone. These results also provide modern examples of subducting slab fabric that can be used to identify fossil slab fabric and paleo-spreading directions from Proterozoic suture zones [Song and Kim, 2012; Mercier et al., 2008].
 Our results suggest that the AML, if considered in isolation, can in theory produce up to 0.7 s of delay time of shear-wave splitting for the most extreme and favorable case; however, it is unlikely that splitting from SKS waves can resolve this signal mainly due to wavefield complications and interference patterns associated with scattering, nonplanar geometry as well as nonhorizontal symmetry of the fast axis of propagation [Levin et al., 2007]. Nevertheless, the correlation of SKS fast directions with the trend of the fast axis of the AML in Japan suggests that vertical deformation may be coherent throughout the oceanic lithosphere. Finally, the effects caused by fossil fabric combined with those due to subduction-related dynamic processes and anisotropic lithologies within the mantle wedge are likely to produce complicated interference patterns of shear-wave splitting. Interpretation of seismic anisotropy in terms of a single mechanism may therefore need to be reevaluated in light of these results.
 We use receiver functions to show the ubiquity of the uppermost AML previously found in Mexico [Song and Kim, 2012]. The AML is <20 km thick and is characterized by large magnitude anisotropy with a fast axis of hexagonal symmetry. The trend and plunge of the fast axis are consistent with fabric generated at oceanic spreading ridges and indicate that fossil slab fabric persists upon entering the subduction zone.
 Data from this study come from the Japan Meteorological Agency, Canadian National Seismograph Network, the Nicoya Seismogenic Zone Network, and the Tectonics Observatory Network at Caltech. Thoughtful reviews by YoungHee Kim and an anonymous reviewer are gratefully acknowledged.