From the available data described above, we assembled records suitable for teleseismic tomography from 974 teleseismic earthquakes with origin times between January 2015 and July 2019 and moment magnitude 5.5 or higher. They encompass waveforms of all stations available in a 5∘ radius around a central position in the Alps located at 46∘ N and 11∘ E. Out of these we evaluated mantle phases for 765 events, all within distances between 35 and 135∘ relative to the central position, leading to a minimum event distance of 30∘ for the closest and a maximum event distance of 140∘ for the farthest station. Information on location and moment tensor was taken from the Global CMT catalogue distributed by the Lamont-Doherty Earth Observatory (LDEO) of Columbia University .

We produced a high-frequency dataset (db0.5) using a fourth-order Butterworth band-pass filter between 0.03 and 0.5 Hz, which turned out to be perfectly suited for combining automatic picking and cross-correlation of land station records. Because oceanic microseismic noise is rather strong in this frequency band, cross-correlation of OBS records was only possible for earthquakes with magnitudes above 6.2 to 6.5 depending on epicentral distance. For this reason, we assembled a second, low-frequency dataset (db0.1) with band-pass-filter upper-corner frequency of 0.1 Hz. In this way, most of the oceanic microseismic noise could be avoided, though at the expense of pick accuracy and resolution of teleseismic tomography.

The distribution of earthquakes of both datasets relative to the Alps strongly varies with azimuth and epicentral distance. Figure  shows the distribution of 370 events that were ultimately picked for the high-frequency dataset. The signal-to-noise ratio of the waveforms from the remaining events was too low owing to either an unfavourable magnitude-to-distance ratio or radiation pattern. The majority of the recorded waves reach the Alps from a sector between north and east (0 to 90∘) mainly originating from the Pacific Ring of Fire at epicentral distances between 80 and 90∘. A second concentration of sources in a sector between WSW and WNW with azimuths between 230 and 290∘ is produced by earthquakes in the subduction zones of North and South America. Epicentral distances in this sector are more broadly distributed than in the NE sector. There is a remarkable lack of events in a sector between about 100 and 230∘ as well as in the opposite direction between 290 and 340∘. To obtain at least a few usable records from the poorly covered sectors long recording periods are essential.

In the following, we will use the quantities absolute travel time at some station, τj, defined as the absolute arrival time minus source time; theoretical travel time, Tj, defined as the time relative to the earthquake source time predicted by a standard earth model using an available earthquake location; and averages of these quantities over the entire array, τ‾ and T‾, respectively. The travel-time residual is defined by 1rj=τj-Tj-(τ‾-T‾)=τj-τ‾-(Tj-T‾). We subtract array averages of observed and theoretical travel times to form residuals, because pure differences between the two quantities contain errors of source time and depend on the wave path through the entire earth. The difference between the array averages, τ‾-T‾, should absorb most of the heterogeneous earth structure remotely from the array, while the remaining residuals after average subtraction should rather reflect influences of heterogeneities below the array.

Highly accurate travel times and travel-time residuals are obtained as follows: first a very low-noise beam trace associated with some selected reference station is constructed by stacking appropriately shifted waveforms of all or selected stations on top of the reference station trace. Then, cross-correlation of the beam trace with all other traces is performed to determine highly accurate time lags relative to the beam. Finally, a travel time is read from the low-noise beam trace itself using automatic picking. Let us denote the beam travel time by τB and the time lag of station j relative to the beam by ΔτjB. Then, absolute travel time and its array average are given by τj=τB+ΔτjB,2τ‾=τB+Δτ‾, where Δτ‾ denotes the array average of the time lags relative to the beam. The travel-time residual is given by rj=τj-τ‾-(Tj-T‾)=τB+ΔτjB-τB-Δτ‾-(Tj-T‾)3=ΔτjB-Δτ‾-(Tj-T‾). Note that the travel-time residual is independent of the beam travel time and, hence, its accuracy is fully determined by the accuracy of the time lags ΔτjB.

To obtain the beam itself, we first select a reference station and consider travel-time differences to all other stations, τj-τR, which are again determined by cross-correlation. The reference station should be close to the centre of the array to minimise waveform discrepancies to other stations and exhibit a high data availability and low noise. We then use these time differences to shift the station traces and stack them on top of the reference trace to form the beam. Stacking is restricted to traces with sufficiently high correlation with the reference trace. To perform these initial cross-correlations efficiently, we take advantage of automatic readings at the stations based on higher-order statistics and the Akaike information criterion . The complete workflow is illustrated in Fig. .

To get initial P wave onsets as reference times for cross-correlation in records of teleseismic earthquakes we use the HOS/AIC algorithm by , which was originally designed for precise local to regional earthquake detection, location and focal mechanism estimation but not for teleseismic phase reading. Therefore, all wavelength-dependent parameters were adapted to our needs.

We choose kurtosis, the central moment of order 4, as the characteristic function, which is calculated on a mean removed seismogram in a moving window of N time samples at index j as 4m^4(j)=1N∑l=0N-1xj-l4. The AIC, which estimates the information loss of a function, is applied to the kurtosis function in the following way : AIC(k)=(k-1)lg⁡1k∑j=1km4(j)25+(L-k+1)lg⁡1L-k+1∑j=kLm4(j)2, with L being the length of the kurtosis function and k ranging from 0 to L.

As initial guess, we use theoretical onsets of the phase estimated for a spherically symmetric earth model and calculate characteristic functions in a properly chosen time window around those onsets. The most probable pick (mpp) is defined as the minimum of the AIC of the phase in this window. We select the moving time window a full order of magnitude larger than those typically used for local earthquake onset determination and calculate the most probable onset. Subsequently, an automatic quality is assigned to the onset based on the signal-to-noise ratio and the difference between the latest and earliest possible pick . This quality determines whether the pick is used for further processing. The earliest possible pick, tepp, is calculated as half the signal period before the most probable pick, tmpp, accounting for a possibly missed first oscillation before the most probable pick. The signal period for this step is estimated by the mean time differences of zero-crossings within a characteristic time window after the most probable pick. The latest possible pick, tlpp, is set to the time where the signal amplitude exceeds the noise level which is calculated as the root mean square of the noise in a window preceding the most probable pick. A symmetrised pick error (SPE) is then calculated as a weighted average of both pick uncertainties with double weight on the uncertainty derived from the latest possible pick: SPE=Δtearliest+2Δtlatest3=(tmpp-tepp)+2(tlpp-tmpp)36=2tlpp-tepp-tmpp3.

By definition, using a maximum frequency of 0.5 Hz, we obtain a minimum uncertainty from the earliest possible pick of a full second. Assuming Δtlatest=0, the minimum possible SPE will be 0.33 s. However, more realistic uncertainties will likely range of the order of 1 to 2 s close to the maximum travel-time residuals expected from mantle heterogeneities below the Alpine orogen. We will show later (Fig. a) that in many cases pick uncertainties even exceed typical travel-time residuals of interest. To resolve the fine-scale mantle structure below the Alps, it is crucial to reduce the uncertainties of the onsets using additional constraints provided by the high station density of the AlpArray network.

By visual inspection of selected examples, we validated that the large uncertainties result from difficulties of the characteristic function algorithm to find that part of the first P wave onset which is similar in all traces. The reason for this is the relatively low amplitude of the P onset which is often hidden in site-specific noise. The resulting most probable onsets therefore strongly scatter, confirming estimated uncertainties of about one half of the signal period. Another limitation of the characteristic function approach is the false picking of either later-arriving phases due to the first motion being completely masked by noise or of other signals produced in the vicinity of the station leading to a severe number of outliers and to a time-intensive manual postprocessing.

Although too uncertain to be used for tomographic inversion, the AIC onsets turned out to be more precise than onsets predicted with standard 1D earth models and are therefore better suited as reference times for signal cross-correlation. We found that theoretical phase onsets can differ from actual arrivals by up to some tens of seconds, most probably owing to differences of the true physical properties in the global earth from those of the spherically symmetric earth model, uncertainties in origin time, dispersion processes along the travel path and the use of centroid times of the gCMT catalogue as earthquake origin times (which we want to use for the FWI). The resulting need for large cross-correlation shifts to catch all overlapping phases would involve a high risk of cycle skipping.

An analysis of the necessary shift in travel times predicted by the standard earth model AK135  for the final picks of 370 events in a frequency band between 0.03 and 0.5 Hz yielded an average value of -3.71 s, implying that the average travel time in the area of study is less than that predicted by the AK135 earth model. The standard deviation is σ=5.84 s. We found an absolute time shift of over 10 s for 22 events with a peak value of -53 s. Hence, it is not reasonable to directly use 1D theoretical onsets as starting points for a signal cross-correlation.

Especially for lower-magnitude events and high-noise OBS records it may happen for some stations that useful automatic picks are not available. Provided that there are sufficient records left with a reliable automatic pick, we go back to theoretical travel times as correlation reference times which have been corrected by the median time difference between the available automatic picks and the corresponding theoretical travel times. In this way, we still obtain good time references for cross-correlation and avoid omitting all records with unreliable automatic picks. This approach can greatly increase the number of picks obtained with the cross-correlation technique.

Applying a cross-correlation method to improve first arrivals on a large regional array like the AlpArray seismic network is based on the hypothesis of a high similarity of the waveforms of the selected phase across the array. We found this requirement to be satisfied especially well for teleseismic P waves travelling through the mantle but not for PKP phases that penetrate the core. In contrast to mantle P phases, PKP phases are composed of several arrivals which modify the shape of the waveform across the array owing to the different epicentral distances, making signal correlation challenging.

We start by searching for a reference station which represents the waveforms of the entire array best for each single event. The most important criterion for such a station is a continuous operation with high data quality. Therefore, we only consider permanent stations with low noise that were ideally running for the entire time span of events in our database. Also, we want this station to be in a central position in the Alps within the shortest possible distance to all other stations to minimise possible changes in waveform related to large-scale heterogeneities in the global earth (see Sect. ). For each event we start with a small pool of stations meeting those criteria and correlate the signals of all other stations in small time windows around the reference times we get from the AIC picks and the corrected theoretical travel times. The reference station with the highest mean correlation is then chosen to be representative of the full set of stations for this event. Combining each station selected as reference for an event with all other available stations leads to 187 000 correlation pairs for the 370 events in our database. The average cross-correlation coefficient for those signal pairs is 0.78.

After correlating all stations with the reference station, we align the waveforms according to the time of the maximum of the cross-correlation function. For each event we then form a beam representing the onset of the first P wave phase by stacking the vertical component traces onto the reference station if the maximum cross-correlation coefficient is ≥0.8 (Fig. ). To find the exact time of highest correlation independent from sampling, a parabola is fitted around the concave part of the cross-correlation function and analytically evaluated for its apex .

The resulting beam (Fig. ) is of very high quality with an increase in signal-to-noise ratio (SNR) by a factor of about 30. On the beam, first motion becomes clearly visible and can be determined precisely either automatically or manually. In our case, we applied the automatic picking procedure of Sect.  to determine the onset on the beam trace. Alternatively, the reference onset can be determined by hand once for each event. After determination of the absolute pick on the beam, vertical components of all stations are correlated with it. The travel time at each station is then calculated as the beam travel time plus the time difference obtained from the lag time associated with the maximum correlation between the beam waveform and the waveform at the station. After onset determination our algorithm also searches for outliers within a time window around an expectation value we calculate for each station to further assess possible cycle-skipping issues. Finally, to assure the consistency of our travel-time dataset, wave fronts are constructed and visually inspected. Outliers can be easily recognised as they create strong distortions of the travel-time isochrones.

The different role of theoretical, AIC and correlation-corrected travel time is illustrated in Fig. . If the traces are aligned according to theoretical travel time (Fig. a), the alignment with the beam trace is the worst. Evidently, this must be due to lateral heterogeneities below the array not contained in the standard earth model. If the traces are aligned according to their AIC automatic pick (Fig. b), overlap with the beam trace improves, but there are still significant deviations, for example for stations Z3.A013A and Z3.A286A. The agreement with the stacked trace is best when the traces are aligned according to their correlation-corrected onsets. Fig. c demonstrates the scatter of the AIC picks which makes them insufficient for teleseismic tomography.

Estimating an error for automatically determined as well as for manually assigned travel times is a difficult task and can be rather subjective. The concept of earliest and latest possible pick for error estimation uses information of a single trace only and is not suited for travel-time residuals determined by cross-correlation, as the credibility of a time difference to a reference trace associated with a high cross-correlation coefficient is by far higher. This argument also applies for uncertainties of the absolute onsets, if the reference trace is a low-noise beam where the concept of estimating the earliest possible pick as half the signal wavelength is questionable, as the first onset may be clearly identifiable without any risk of missing the first oscillation.

As the beam represents the waveform of the majority of stations, we consider the maximum cross-correlation between station and reference trace as the most important indicator for the relative accuracy of a travel-time difference. However, this assumption only holds if the stations forming the beam trace are evenly distributed in the array and not just representing a part of the array (for example stations close to the reference station). This is vital for the consistency of the full dataset.

Moreover, using the cross-correlation coefficient as a measure of accuracy might lead to a down-weighting of traces of stations influenced by strong local heterogeneities whose waveform does not fit the shape of the reference trace. Fortunately, this matter can be easily identified by looking at spatial distributions of maximum correlation. Affected stations should stand out in comparison to adjacent stations when looking at correlation coefficients averaged over many events (Sect. ). We tried to find such regional dependencies by creating spatial plots of the cross-correlation coefficient for randomly selected events but could find neither evidence of a decrease in the correlation coefficient with distance to a reference station nor regional clusters of high or low cross-correlation coefficients.

A second criterion for a good match of station and reference trace is the shape of the cross-correlation function itself. Hence, we also evaluate the full width at half-maximum (FWHM) of the cross-correlation function. If the FWHM increases, the cross-correlation maximum loses sharpness and the accuracy of a travel-time difference decreases. This approach implies a frequency dependency of travel-time uncertainty, leading to a higher uncertainty for longer periods (and hence wavelengths).

For a parabola fitted to the maximum of the cross-correlation function of the form 7f(x)=ax2+bx+c, the full width at half maximum (FWHM) can be calculated as 8FWHM=2b2a2+Cmax⁡-2c2a, where Cmax⁡ denotes the maximum correlation. To combine both criteria, we chose to calculate the travel-time difference uncertainty as follows: 9σ=(1-Cmax⁡)FWHM.

The influence of a bad fit owing to signal coda on the cross-correlation coefficient and hence travel-time residual uncertainty is illustrated in Fig. c. The contribution of the width of the cross-correlation function, depending on signal period, is practically the same for all traces of this event. However, the maximum correlation decreases for stations with additional complexity in the signal (ZS.D005 and Z3.A010A).

An evaluation of the regional distribution of the median of travel-time uncertainty per station in the db0.5 dataset (Fig. a) exhibits lower values north and east of the Alpine arc, in central and southern Germany, as well as in the Czech Republic, eastern Austria and Slovenia. We hypothesise that this effect originates in the spatial segregation of those areas from the Alpine orogen, as the subsurface structure of the surrounding area of the Alps is simple in comparison to that beneath the Alps. In contrast to that, travel-time uncertainty increases above the highly complex structures in the Alpine arc where the P wave fronts are significantly altered by the strongly heterogeneous subsurface. This decreases their correlation with the waveforms on other stations of the array and to the stacked reference trace. It is also likely that uncertainty increases due to local site effects which can be significant in narrow valleys where anthropogenic activities such as traffic are harder to evade. These influences should be visible on single stations which show a high daytime noise level. We expect those effects to be present equally in both the high- and the low-frequency dataset. However, most of the station outliers we see in one dataset are not present in the other.

The travel-time uncertainty distribution pattern of the lower-frequency dataset db0.1 (Fig. b) shows a shape comparable to the high-frequency one with the lowest uncertainty in the northeastern parts of the array. However, overall uncertainty is higher, and the contrast between regions of high and low uncertainty is decreased. We assume that this is an effect of signals of larger wavelengths being less sensitive to small-scale anomalies due to their lower-resolution capability (finite frequency effect) and hence waveform fit with the reference trace being easier to achieve. The only area where we see a totally opposite behaviour is the Ligurian Sea, where the positive impact on pick uncertainty using lower frequencies is salient. Here, not only the number of total picks greatly increased but also average pick quality is raised by a full class for nearly all OBS whilst for the remaining stations quality tends to decrease by almost one class in comparison to the high-frequency dataset. We also note that there are only small changes in uncertainty for the SWATH-D stations. They even show slightly the counterintuitive behaviour of having higher uncertainties than average in db0.5 but lower uncertainties than average in db0.1.

The total number of picks per station is highest on permanent station networks which are distributed most densely in the central Alps and Apennines. Temporal coverage decreases slightly in the western part of the Array due to a delayed start of deployment of temporary stations in this area.

We start with teleseismic P wave fronts constructed as isolines from the estimated travel times. To further demonstrate the improvement of correlation-corrected travel times over AIC travel times, we show interpolated P wave fronts constructed from both kinds of travel times. As an example, we take the M6.5 earthquake that happened on 17 November 2017 in the Eastern Xizang–India Border Region (Fig. ). In both cases, one can identify the P wave travelling across the array of about 700 stations from northeast to southwest. However, when constructed from the AIC onsets (Fig. a), the wave fronts are strongly irregular and several outliers are apparent, leading to distorted isolines which cannot be explained by mantle heterogeneities. After application of the cross-correlation correction, the resulting wave fronts no longer show the scatter inherent to the AIC onsets and become smooth except for some weak undulations (Fig. b). These seem to be produced by several adjacent stations and should be attributed to subsurface structures.

To illustrate the varying shapes of the wave fronts crossing the AlpArray network from different azimuths and epicentral distances, we have selected four different earthquakes as representative examples: two with nearly equal back-azimuth (75∘) but very different epicentral distances (104 and 45∘) and two others covering western (288∘) and southern (218∘) back-azimuths with differing epicentral distances (89 and 52∘) (Fig. ). In addition to the P wave fronts, we show the demeaned travel-time residuals associated with each particular event as defined in Eq. (). They should correlate with the wave fronts, as deformations of the wave front should lead to travel-time residuals and vice versa. To compensate for the influences of different station elevations, we apply a constant travel-time correction on all residuals shown, assuming vertical propagation and a surface P wave velocity of 5.5 kms-1.

Comparing Fig. a and c reveals a notable difference of the 1 s travel-time isoline spacing which is much greater for the distant event. This reflects the different horizontal apparent velocity of the two wave fields, which is controlled by epicentral distance and is much higher for the more distant event. While the wave fronts are generally regular and smooth, strong distortions become visible in some places. For example, in Fig. a, the spacing of the isolines locally broadens in northern Italy north and east of the Ligurian Sea. This increased spacing can be associated with very large negative residuals beginning at about 7.5∘ E and 45∘ N and continuing further to the northeast. The broadening can be explained by the transition from normal to large negative residuals further to the southwest. A second one occurs in the Apennines to the south where the wave front has a strong lag near the western coast of Italy compared to the areas north of it but takes up again while propagating over the areas with negative residuals in the western and central Apennines. In Fig. c, a very similar behaviour is visible.

A closer examination of travel-time residuals shown in Fig. b and d reveals that there is a general agreement between the patterns but also significant differences, for example in southeastern France, where we observe normal to negative residuals for the distant event but positive residuals for the close event. The opposite is the case in most of Switzerland where we observe negative residuals for the close event and rather normal residuals for the distant one. Apparently, the steeply upwards-propagating waves from the distant event see different subsurface structures than the more slanted waves of the close event do.

A comparable behaviour is observed for events arriving from other back-azimuths. Isoline spacing is again much larger for the more distant event whose waves arrive from a WNW direction. In Fig. g, there are again notable distortions of the wave fronts around 7.5∘ E and 45∘ N. These distortions are shifted to the NE for the waves arriving from the SSW in Fig. e. The associated residuals exhibit large-scale coherent patterns of negative and positive residuals but are again different in various regions. For example, residuals are generally positive in southeastern France for the waves arriving from SSW while they are normal to negative for the event from WNW. This is again an indication of heterogeneous mantle structure to be resolved by tomography later.

Although travel-time residuals differ with epicentral distance and event back-azimuth as waves move through high- or low-velocity zones from different angles before reaching the surface, there are certain features which tend to occur for a large number of events. The most prominent ones are the negative residuals along the Apenninic and Alpine chain. We stacked residuals for all analysed events to identify regions of stable negative or positive travel-time residuals. It is important to understand that after stacking of the demeaned travel-time residuals, the resulting residuals are relative to an unknown one-dimensional earth model defined by all events used for stacking and not to the standard earth model used to calculate travel-time differences in the first place e.g.. Hence, negative or positive residuals indicate higher or lower velocities, respectively, relative to this average model and not relative to a standard earth model.

As the azimuthal distribution of the events in our database is strongly uneven (Sect. ), it is important to balance out the influence of events from different directions when stacking. Otherwise, the influence of back-azimuths with high event density (e.g. NE in Fig. ) on travel-time residuals would dominate over data from poorly covered directions. Hence, we create 30∘ wide back-azimuth bins and, for each station, form a weighted average of all travel-time residuals associated with events in this bin with weights given by the inverse uncertainties of the residuals. The full stack over all events (Fig. ) is finally obtained by averaging over the individual 30∘ back-azimuth stacks. The value of 30∘ was chosen as a good compromise between angular resolution and smooth event distribution. The distribution of available measurements for different back-azimuth bin sizes can be found in the supplementary material (Fig. ).

We refrained from binning according to epicentral distances because an examination of residuals of different individual events (Fig. ) showed that the travel-time residual patterns vary much more strongly with back-azimuth than with epicentral distance. This may be explained by the fact that the incidence angle at the surface of mantle phases between 35 and 135∘ distance differs by a maximum of only ∼13∘ in a 1D earth model (AK135).

For an interpretation of mantle features in the residual pattern, we chose to correct the stacked residual patterns for influences of the strongly heterogeneous Alpine crust. We assembled a crustal model from different studies in the greater Alpine region, which we will show in more detail in the upcoming travel-time tomography. To create the model, we start with the generic crustal background model for Europe EuCrust-07 , which was compiled for the correction of crustal influences on seismic studies that analyse deeper structures such as a teleseismic tomography. The layer model contains information on sediment thickness and upper and lower crustal average velocity and thickness (and thus Moho depth) discretised on a 15′×15′ regular grid. It was created from various seismic reflection, refraction and receiver function studies. For the Alpine region, we improve information on the Moho depth using a more recent study of . Lastly, for the western and central Alpine region, we replace this model with the more detailed, fully 3D regional tomographic model of hereafter referred to as D09a. We use the information on the model resolution supplied to us by D09a to assess which model to favour at a certain point in space. To only account for crustal influences in our dataset, we remove velocity perturbations associated with the uppermost mantle below the suggested Moho proxy by D09a of 7.25 kms-1. The resulting travel-time differences (calculated by assuming a planar, vertically incident wave front passing through the crust) between our crustal 3-D model and the crustal minimum 1D model of D09a can be found in Fig.  in the Appendix. Crustal contributions to travel-time residuals related to the Ivrea body or sedimentary basins (e.g. Po plain) are of the order of a second and comparable to the corrections derived by  and need to be removed for an interpretation of mantle anomalies.

The most striking features of the stacked travel-time residuals after crustal correction are the negative residuals following the Alpine arc from 45∘ N, 7.5∘ E to 46∘ N, 14.5∘ E (Fig. ). For later reference and interpretation, we group these residuals into three major anomalies: a western negative anomaly (W) following the Alpine mountain chain to the east and bending south towards the Po plain, which can be clearly discerned from a large zone of positive residuals to the west; a central negative anomaly (C) attached to anomaly W in the south but following a very different strike; and an eastern anomaly (E) bending circularly to the south towards the Dinarides, whose separation from anomaly C an be recognised by virtue of the travel-time residuals observed on the very dense SWATH-D array (inset in Fig. ). In addition, we define anomaly A located at the western Italian coast and penetrating into central Italy with a strike of 130∘.

We already showed travel-time residuals for individual wave fields. To give a more stable impression of the azimuthal variation of the residuals, we stacked three neighbouring 30∘ averages subjected to crustal correction to cover the four major azimuthal sectors NE, SE, SW and NW (Fig. ). All of them exhibit the negative residuals following the Apenninic and the Alpine chain, the generally normal-to-negative residuals in the northern foreland, and the generally normal-to-positive residuals in southeastern France and the Pannonian basin. However, the location of residual anomalies varies depending on the major azimuthal direction.

This fact is easily demonstrated for the four anomalies defined in the previous section. For example, for waves incident from the northeast (Fig. a), anomaly A is shifted to the southwest and can even be tracked by the OBS stations off the Italian coast. For waves incident from the southwest (Fig. c), it is shifted to the northeast and into the Adriatic Sea where we lose its track due to missing seismological stations. For waves arriving from perpendicular directions (Fig. b and d), it remains mostly in place.

Shifting and change of appearance is also observed for the anomalies C and E located between 12 and 15∘ E. Both appear as strong, merging negative travel-time residuals for waves incident from NE (Fig. a) but seem to be weaker for waves arriving from other azimuths. For waves incident from SE (Fig. b) anomalies C and E appear to shift to the NW. The negative residuals of anomaly E almost completely disappear within the drawn outline. For illumination from the SW (Fig. c), anomaly C remains in place while, for illumination from the NW sector (Fig. d), anomaly C appears with a lower amplitude while anomaly E is shifted to the east.

The western Alpine anomaly (W) shows negative residuals for illumination from the SE that are shifted to the NW (Fig. b) but appears weak and partially positive for waves incident from the NW.

Owing to the high noise on the OBS records in the higher-frequency band, we assembled a low-frequency dataset with a maximum frequency of 0.1 Hz. As for the 0.5 Hz dataset, we determined absolute travel times and travel-time residuals using the same procedures as for the high-frequency data (including azimuthal binning, crustal corrections, etc.). We find that the obtained maps of travel-time residuals differ systematically between the considered frequencies (Fig. )

To illustrate the differences between both frequency bands, we directly show travel-time residuals determined for the 0.1 Hz dataset (Fig. a), and we form differences of the travel-time residuals; i.e. we subtract the values for the low-frequency dataset from those of the high-frequency dataset (Fig. b).

There are no striking differences between residuals of db0.5 (Fig. ) and db0.1 (Fig. a), which in general show very similar patterns of negative and positive residuals. However, the negative residuals of the 0.1 Hz dataset seem smaller (up to 0.3 s) and less well-defined around the central (C) and eastern (E) anomalies. In contrast to that, there is a marked increase in amplitude of the negative residuals (up to 0.5 s) in the Po plain. In the remaining regions, deviations are rather small with a weak tendency to negative values, especially in the Alpine foreland.

There is a massive increase in the total number of picks for the ocean bottom seismometers reflecting the increase in onsets and onset quality described in Sect. . Also the variation of residuals with back-azimuth is lower for the OBS, because there are more picks available, even from the poorly covered back-azimuths. The number of OBS picks for all events increases from 421 to 1113 (+164 %), compared to a 25 % increase only for all stations. This might have a notable effect on resolving structures below the Ligurian Sea when performing a teleseismic tomography. In the case of the stacked travel-time residuals, no strong effect is visible because of the applied azimuthal binning.