Efficient reanalysis of events from GWTC-3 with RIFT and asimov

In a coalescing quasicircular orbit, a compact binary can be fully defined by 15 parameters. Among these, eight are the intrinsic parameters - signified by $\lambda$ - which encompass any factors describing the physical properties of the individual objects within the system - the binary’s masses, spin magnitudes, spin angles, and azimuthl spin angles. Extrinsic parameters - signified by $\theta$ - consist of the other seven values necessary to describe its spacetime position and orientation relative to the detectors.

RIFT interprets gravitational wave observations $d$ by comparing them with predicted gravitational wave signals $h(\bm{\lambda},\bm{\theta})$ in a two-step iterative process. In the first stage, RIFT has numerous workers concurrently calculate a marginal likelihood

$${\cal L}{({\bm{\lambda}})}\equiv\int{\cal L}_{\rm full}(\bm{\lambda},\bm{\theta})p(\bm{\theta})d\bm{\theta}$$

(1)

for many different values of $\bm{\lambda}$, where ${\cal L}_{\rm full}(\bm{\lambda},\theta)$ is the likelihood of the gravitational wave signal in the multi-detector network; see [28, 24] for further details. In the second stage, RIFT will use accumulated marginal likelihood evaluations $(\bm{\lambda}_{\alpha},{\cal L}_{\alpha})$ to construct an approximation to ${\cal L}(\bm{\lambda})$ which is then used to infer the detector-frame posterior distribution

$$p_{\rm post}=\frac{{\cal L}(\bm{\lambda})p(\bm{\lambda})}{\int d\bm{\lambda}{\cal L}(\bm{\lambda})p(\bm{\lambda})}.$$

(2)

where the prior $p(\bm{\lambda})$ represents the prior distribution on intrinsic parameters.

RIFT inference is performed using the spin-weighted spherical harmonic waveforms $h_{lm}(t)$ or $h_{lm}(f)$ [28, 29, 24, 30], usually computed from binary parameters through the lalsimulation or gwsignal library. In this work, we will present RIFT inferences performed using several families of models. For the SEOB family of models, we employ SEOBNRv4PHM [10, 11] and SEOBNRv5PHM [22, 27]. Among surrogate models calibrated to numerical relativity, we employ the NRSur7dq4 model [15]. While we will not use RIFT to analyze these events with IMRPhenomXPHM [12], we will compare our results to analyses performed by others, including analyses using this waveform.

The asimov infrastructure provides a framework for managing suites of parameter inference calculations. This framework in particular provides not only a mechanism for storing event-specific and algorithm-specific settings, but also a mechanism to implement these settings, launching and overseeing production analyses with a variety of parameter inference tools. Starting with the GWTC-3 and GWTC-2.1 papers, the LVK began performing its analyses using asimov; in particular, the asimov configurations needed to perform these analyses are available [23]. In this work, we minimally adapt these settings with pipeline- and approximant-specific adjustments.

The default analysis settings define a specific frequency range for each event, usually covering from $20{\rm Hz}$ up to some maximum frequency $0.875f_{\rm PSD}$ where $f_{\rm PSD}$ is the largest frequency used in the original LVK PSD.

Other groups, however, have performed their own large-scale reanalyses using their own framework and metadata storage. To facilitate comparison with specific sets of previous reanalyses [12] and [14], we implemented and provide tools to convert from their metadata format to a standard asimov configuration file. Since not all of the metadata used for previous analyses is compatible with asimov, some differences are expected between the original analyses and our attempt at reproducing a comparable analysis. For our analyses, we use asimov to get the PSDs by running Bayeswave for all events. However, we also provide tools to retrieve PSDs from previous publications and provide a demonstration on how this method also be used.

The GWTC-3 paper, covering the second half of O3 (henceforth denoted as O3b), presented several analyses performed by the LVK, including many events analyzed with SEOBNRv4PHM via RIFT. The default design choices underlying these analyses prioritized matching the configuration of corresponding analyses with another code, notably including the effects of calibration marginalization. Thus, after first using RIFT to analyze events with SEOBNRv4PHM, these analyses were postprocessed with rejection sampling, to account for the impact of calibration marginalization. Previous studies have demonstrated that calibration marginalization has little impact on intrinsic or extrinsic parameters, modulo a modest effect in some cases on the source sky location. By contrast, the decimation performed by the O3b rejection sampling provides surprisingly few samples to downstream users. For this reason, the reanalyses presented below will omit calibration marginalization entirely.

Table 1 lists the events we have selected for reanalysis. In our demonstration, we chose to reanalyze the same set of O3b events recently reanalyzed with NRSur7dq4 [14]. The table also provides metadata needed to initiate a RIFT analysis: the specific event time and approximate mass range to investigate. By default, RIFT accesses this information from a single file, nominally produced by real search codes (a “coincidence” file commonly denoted coinc.xml), but alternatively generated from median search results. For complete reproducibility of pre-existing LVK analyses, we list meta-information enumerated by the event’s internal gracedb characteristic event identifier along with a secondary identifier identifying the characteristic event geocenter time and mass. These identifiers sometimes differ from the preferred released event, because of small but important differences between full inference and the event time or other parameters targeted by searches. For public reanalyses, we have also generated our own alternative characteristic information, disseminated along with our configuration files.

In this work, we will perform inference on either the publicly-released GWTC-3 data from GWOSC [31], or (to facilitate use of settings adopted in prior work) bit-equivalent internal data. Following the analyses used for LVK publications, we use GWOSC-released deglitched data when appropriate. Unless otherwise noted, we generate independent PSD estimates using Bayeswave [32, 33, 34]. However, while as a second alternative and for comparison, when alternatively gathered PSDs from existing analyses of the GWTC-3 catalog [5]; examples of this configuration are also disseminated in our configuraiton repository [25].

Following [14], Figures 1 and 2 illustrate a simple summary statistic – the Jensen-Shannon (JS) divergence – quantifying differences between marginal one-dimensional posterior distributions inferred for each event. The vertical dashed red line corresponds to a JS divergence of $0.02$, above which differences between one-dimensional marginal distributions are easily identified by eye. The top panels of Figures 1 and 2 show comparisons between inferences with three state-of-the-art waveform models compared to inferences derived with an older model (IMRPhenomPv2) known to omit important physics and higher-order modes. Except for a handful of notable exceptions (GW191109, GW200302, GW200129), most analyses with IMRPhenomPv2 are broadly consistent with analyses done with more sophisticated waveforms. We demonstrate how this counterintuitive result arises using concrete examples in the next section. By contrast, the bottom panels of Figures 1 and 2 show familiar differences when comparing results between different state-of-the-art waveform models. These comparisons suggest that almost every event has some parameter in which notable waveform differences occur. Likewise, these comparisons suggest that every pair of waveforms (including SEOBNRv4PHM and SEOBNRv5PHM) has at one or more event where conclusions derived using those two waveforms differ by more than our ad hoc JS threshold. These two bottom panels present qualitatively similar results as previous work [14] – see their Figures 3 and 4 – albeit now performed exclusively with RIFT and using a different set of waveform models for comparison (i.e., including SEOBNRv5PHM and IMRPhenomPv2, but omitting IMRPhenomXPHM).

Using the summary statistics provided above and in previous work [14], we have identified five events that merit further discussion: GW200129, GW191109, GW200216, and both events on the date GW200220. In an appendix and associated data release, we present posterior inferences for all events in our sample. In both the appendix and our own figures, for comparison and to normalize the scale of differences presented in this work, we also present the IMRPhenomXPHM results produced for GWTC-3 [5]. We also present NRSur7dq4 results provided by a previous study, for validation [14].

GW200129 has inspired considerable followup investigation, including claims of precession [35], eccentricity [36], and validation studies to assess stability of results under different data cleaning methods [37, 38]. As illustrated in Figure 3, in our benchmark analyses, using public cleaned frames and with a (common) independently generated PSD, we find our three waveform models draw dramatically different conclusions about this event, most notably the spin magnitudes of the two masses. For this event, even the two similar models SEOBNRv4PHM and SEOBNRv5PHM lead to notably different conclusions. Our analysis with NRSur7dq4 favors even more extreme mass ratios, primary spins, and $\chi_{p}$ than our two SEOBNR results. However, as discussed below (“comparison to previous results”), our analyses with NRSur7dq4 do not produce as extreme conclusions about the mass ratio and spin as found in another independent reanalysis [14], which employs slightly different settings (e.g., an independently-generated PSD).

By contrast, for GW200216 (Figure 4), GW200220_061928 (Figure 5), and GW200220_124850 (Figure 6), our inferences using the two SEOBNR models are very consistent with each other, and differ from NRSur7dq4 primarily in the mass ratio and spin magnitude distributions. In all three cases, the NRSur7qd4 analysis favors a wider mass ratio distribution and slightly prefers larger spins, while our two SEOBNR analyses favor comparable mass ratios and lower spins.

For GW191109, our inferences once again depend more sentitively on the assumed waveform, as illustrated in Figures 7 and 8. As might be expected given the high mass of the event and the model’s simplified physics, interpretations drawn using IMRPhenomPv2 in particular produce strikingly different conclusions about the mass ratio, primary spin, and even $\chi_{p}$ than analyses performed with any other waveform. By contrast, though waveform systematics change the posterior distributions modestly, when , comparing analyses performed in this work with state-of-the-art waveforms (SEOBNRv4PHM, SEOBNRv5PHM, and NRSur7dq4), we find coarsely good overall agreement.

Despite the substantial and frequent differences seen between different calculations that both use state-of-the-art waveforms, analyses using IMRPhenomPv2 seem to be surprisingly comparable to the results derived using all other waveforms in this study; see the top panels of Figures 1 and 2. To investigate this numerical incongruity, in Figures 8 and 9 we illustrate the posteriors derived using all models, including IMRPhenomPv2. For many events and parameters, all one-dimensional posteriors are in good agreement, modulo truncation effects associated with the finite mass ratio range required for NRSur7dq4. However, more frequently than for other waveforms, the posteriors for IMRPhenomPv2 may differ substantially from the consensus conclusions derived with state-of-the-art waveforms.

Figures 8 and 9 show that for most of the events studied here, our estimates with NRSur7dq4, SEOBNRv4PHM, and SEOBNRv5PHM largely agree with results derived using previous analyses with NRSur7dq4 and IMRPhenomXPHM. This agreement is reassuring, given our analysis settings (e.g., the PSD) do not necessarily perfectly match the assumptions adopted in these previous works. However, for two events, these figures reveal striking differences between our analyses and these previous works.

For GW200129, the differences between previous studies and this work are frequent and significant. Interpreted using the assumptions of the previous study, both IMRPhenomXPHM and NRSurd7q4 would suggest a posterior extending to mass ratio down to and peaking near $0.4$, and primary spins that are preferentially very large. By contrast, our analyses generally prefer nearly comparable mass ratios and show no preference towards large primary spin; indeed, all our analyses point against primary spin. To validate the robustness of our conclusions, Figure 3 shows the underlying marginal likelhood data used to build our posterior distributions. In all cases, we don’t find points with high marginal likelihood at very asymmetric mass ratio $q<0.4$, nor (to a lesser extent) with near-extremal transverse spin.

For GW191109, the differences between previous studies and this work are less extreme but still noteworthy, with both external posterior results (IMRPhenomXPHM and NRSur7dq4) exhibiting unexpected structure relative to the consensus conclusions presented in this work. For example, the external NRSur7dq4 posterior favors a different total mass and strongly disfavors equal mass. The external IMRPhenomXPHM posterior favors a wide mass distribution and unexpected complexity in its conclusions about $\chi_{\rm eff}$ and mass ratio. By contrast, using the assumptions employed in this work, our different approximations produce qualitatively similar results, though differing in details (e.g., as noted for this event SEOBNRv4PHM and SEOBNRv5PHM ddraw slightly different conclusions).

For two more events (GW200220_124850 and GW200216) we see more modest differences between the results in this work and our reference analysis using NRSur7dq4. For other events (e.g., GW200224 and GW200220_061928), the external analysis performed with IMRPhenomXPHM seems to be an outlier relative to the consensus provided by other analyses, including the external NRSur7dq4 result. We point out these differences to help target future investigations such as head-to-head replication studies with different inference codes, which is outside the scope of this work.