Inspiral-Merger-Ringdown Consistency Test

• The IMR consistency test is a gravitational-wave analysis method that compares remnant black hole properties derived independently from the inspiral and merger-ringdown phases.
• It utilizes frequency-domain segmentation and Bayesian or Fisher-matrix techniques to map binary parameters to final mass and spin using numerical-relativity fits.
• Its sensitivity to waveform systematics and detector noise makes the test a critical tool for probing deviations from general relativity.

The inspiral-merger-ringdown (IMR) consistency test is a theory-agnostic gravitational-wave test of general relativity (GR) that asks whether the remnant black hole’s final mass and spin, inferred independently from the inspiral and from the merger-ringdown portions of a compact-binary coalescence signal, are mutually consistent. In its standard binary-black-hole form, the signal is split into a low-frequency inspiral part and a high-frequency post-inspiral part, each segment is analyzed under GR, and the component masses and spins inferred from each segment are mapped to remnant properties using numerical-relativity fitting formulas. If GR is correct and the waveform model is adequate, the two remnant estimates should agree within uncertainty; statistically significant disagreement is interpreted as either a deviation from GR or an unmodeled systematic (Ghosh et al., 2016, Ghosh et al., 2017, Carson et al., 2019).

1. Definition and physical basis

The IMR consistency test is built on the GR statement that a binary black-hole coalescence produces a single remnant black hole characterized, in the standard formulation, by a final mass $M_f$ and a dimensionless final spin $\chi_f$ or $a_f$. The early inspiral and the late merger-ringdown probe different dynamical regimes, but they describe the same source. The test therefore compares two logically distinct inferences of the same remnant: one obtained from the inspiral by inferring the binary parameters and mapping them to $(M_f,\chi_f)$, and one obtained from the merger-ringdown by inferring the remnant from the late-time waveform (Carson et al., 2019, Carson et al., 2019).

A standard implementation splits the data in the frequency domain. For GW150914-like analyses, one formulation adopts a transition frequency

$f_{\rm trans}=132~{\rm Hz},$

while another uses the Kerr ISCO frequency $f_{\rm ISCO}$ of the remnant estimated from the full IMR analysis as the cutoff between inspiral and post-inspiral (Carson et al., 2019, Ghosh et al., 2017). The frequency-domain split is used because it changes only the integration limits of the likelihood and avoids time-domain windowing artifacts (Ghosh et al., 2017).

The remnant is inferred independently from the two bands. In one common notation,

$$M_f^{\rm I,MR}=M_f^{\rm I,MR}(m_1,m_2,\chi_1,\chi_2),\qquad \chi_f^{\rm I,MR}=\chi_f^{\rm I,MR}(m_1,m_2,\chi_1,\chi_2),$$

where the superscripts denote inspiral and merger-ringdown analyses (Carson et al., 2019). In another notation,

$$M_f = M_f(m_1,m_2,a_1,a_2), \qquad a_f = a_f(m_1,m_2,a_1,a_2),$$

with the remnant mapping applied separately to the inspiral-only and post-inspiral posteriors (Ghosh et al., 2017).

The null hypothesis is that the inspiral-derived and post-inspiral-derived remnant properties are the same. This is expressed through difference variables such as

$$\Delta M_f \equiv M_f^{\rm I}-M_f^{\rm MR},\qquad \Delta \chi_f \equiv \chi_f^{\rm I}-\chi_f^{\rm MR},$$

or, equivalently,

$$\Delta M_f := M_f^i - M_f^{\rm mr}, \qquad \Delta a_f := a_f^i - a_f^{\rm mr}.$$

GR predicts consistency with the origin in the deviation plane (Carson et al., 2019, Ghosh et al., 2017).

2. Statistical formulation and remnant-variable posteriors

The test is implemented either with full Bayesian parameter estimation or with Fisher-matrix approximations for sufficiently loud events. In the Bayesian formulation, the signal is analyzed separately in the inspiral and merger-ringdown frequency ranges, producing posteriors over the intrinsic binary parameters and, after remnant fitting, over $\chi_f$0 (Ghosh et al., 2017, Carson et al., 2019). In the Fisher formulation, the posterior is approximated as Gaussian around the maximum-likelihood point, with inner product

$\chi_f$1

signal-to-noise ratio

$\chi_f$2

and Fisher matrix

$\chi_f$3

With Gaussian priors,

$\chi_f$4

or equivalently

$\chi_f$5

in the forecasting literature (Carson et al., 2019, Carson et al., 2019).

The standard dimensionless discrepancy variables are fractional differences between inspiral and post-inspiral remnant estimates. One widely used convention is

$\chi_f$6

with

$\chi_f$7

GR predicts

$\chi_f$8

Another notation uses

$\chi_f$9

or the null variables $a_f$0 directly, with the same GR prediction at the origin (Carson et al., 2019, 2207.13761, Ghosh et al., 2017).

The joint posterior in the deviation plane is obtained by transforming the inspiral and merger-ringdown remnant posteriors and marginalizing over the averages. A standard expression is

$a_f$1

An equivalent formulation appears in the original Bayesian construction using $a_f$2 and the Jacobian $a_f$3 (Carson et al., 2019, Ghosh et al., 2017).

A practical significance metric is whether the credible region in the $a_f$4 plane contains $a_f$5. In Fisher forecasts, the area of the 90% credible ellipse is often taken as the operational measure of test power (Carson et al., 2019). A plausible implication is that the IMR test is best regarded as a remnant-consistency test in a reduced two-parameter space, although later work shows that this reduction can discard informative correlations (Zhong et al., 2024).

3. Development from “golden binaries” to catalog analyses

The modern IMR consistency test was introduced in the context of “golden” stellar-mass black-hole binaries, for which ground-based detectors can observe substantial inspiral, merger, and ringdown content (Ghosh et al., 2016). That early Bayesian implementation used stochastic sampling, defined the inspiral and merger-ringdown bands with

$a_f$6

for inspiral and

$a_f$7

for merger-ringdown, and emphasized that the posterior on deviation parameters can be combined across multiple observations (Ghosh et al., 2016).

A subsequent full formulation showed how the test can be applied to many moderate-SNR events rather than only rare loud signals (Ghosh et al., 2017). If the same deviation parameters apply to each event, the per-event posteriors can be combined hierarchically as

$a_f$8

That paper stressed that uncertainties shrink roughly as $a_f$9 for similar events, while also noting that this assumption can fail if the deviation depends strongly on masses or spins (Ghosh et al., 2017).

By the time of the 2019 review literature, the LIGO-Virgo Collaboration had performed IMR consistency tests on the observed binary-black-hole catalog and found that all events were statistically consistent with GR (Carson et al., 2019). For GW150914-like events, Fisher and Bayesian 90% credible-region areas were reported to agree at the $(M_f,\chi_f)$0 level, with area $(M_f,\chi_f)$1 in a LIGO O1 Fisher estimate and $(M_f,\chi_f)$2 in the Bayesian result (Carson et al., 2019). A parallel forecasting study gave the same qualitative conclusion for single-band and multiband observations, again emphasizing agreement between Fisher and Bayesian contours at the $(M_f,\chi_f)$3 level (Carson et al., 2019).

Later catalog-level work generalized the combination step. A multidimensional hierarchical analysis modeled the population of deviation parameters as a multivariate Gaussian,

$(M_f,\chi_f)$4

with hyperpriors on $(M_f,\chi_f)$5, standard deviations $(M_f,\chi_f)$6, and a correlation matrix $(M_f,\chi_f)$7 using an LKJ prior (Zhong et al., 2024). Applied to the classic 2D IMR test, that framework found consistency with GR at the $(M_f,\chi_f)$8 credible level without GW190814 and $(M_f,\chi_f)$9 with GW190814 included; in a 4D formulation, the corresponding values were $f_{\rm trans}=132~{\rm Hz},$0 and $f_{\rm trans}=132~{\rm Hz},$1 (Zhong et al., 2024). The same study argued that the classic 2D reduction is “under-dimensionalized” because the underlying variables are really

$f_{\rm trans}=132~{\rm Hz},$2

and introduced additional average coordinates

$f_{\rm trans}=132~{\rm Hz},$3

This suggests that correlations between deviation variables and average remnant properties can bias catalog-level inference if they are marginalized away implicitly (Zhong et al., 2024).

4. Sensitivity, detector dependence, and multiband forecasts

The principal limitation of present-day IMR consistency tests is detector noise (Carson et al., 2019). Current detectors were capable of performing the test, but with comparatively large credible regions (Carson et al., 2019). In forecast studies for GW150914-like systems, Cosmic Explorer (CE) improves the 90% area in the $f_{\rm trans}=132~{\rm Hz},$4 plane by about three orders of magnitude relative to LIGO O1, reaching

$f_{\rm trans}=132~{\rm Hz},$5

for CE (Carson et al., 2019). A multiband CE + LISA configuration reduces the area further to

$f_{\rm trans}=132~{\rm Hz},$6

with similar gains of about seven to ten reported for CE combined with TianQin, B-DECIGO, or DECIGO (Carson et al., 2019, Carson et al., 2019).

The multiband interpretation is physically straightforward. Space-based detectors constrain the low-frequency inspiral, while ground-based third-generation detectors constrain the merger-ringdown. In the benchmark GW150914-like forecasts, future single-band detections improve current tests by roughly three orders of magnitude, and combining a space-based inspiral measurement with CE yields an additional factor of $f_{\rm trans}=132~{\rm Hz},$7–$f_{\rm trans}=132~{\rm Hz},$8 improvement (Carson et al., 2019). The same study used a GW150914-like source with component masses $f_{\rm trans}=132~{\rm Hz},$9, spins $f_{\rm ISCO}$0, and an optimistic LISA observing baseline of four years before merger (Carson et al., 2019).

The improved sensitivity of third-generation and multiband measurements has a direct systematic consequence: effects that are negligible for current detectors can become limiting systematics for CE-class observatories. This point is explicit in eccentricity studies, which show that the eccentricity level at which the IMR test is biased decreases sharply with detector precision (2207.13761). A plausible implication is that the statistical success of future IMR tests depends increasingly on waveform completeness rather than only on SNR.

The test has also been used as a forecasting tool for specific beyond-GR frameworks. In Einstein-dilaton Gauss-Bonnet gravity, a Fisher-based IMR consistency analysis found that current ground-based detectors are not sensitive enough to probe the coupling below existing bounds, whereas CE and especially CE + LISA could improve constraints, with multiband observations potentially surpassing current limits by about an order of magnitude (Carson et al., 2020). For generic beyond-Kerr parameterizations, IMR consistency and direct parameterized tests were found to give very similar bounds, with future CE and LISA observations improving constraints by two to three orders of magnitude in the examples studied (Carson et al., 2020).

5. Waveform systematics and false violations

Because the IMR consistency test is a null test built from GR templates, it is sensitive not only to genuine beyond-GR effects but also to waveform-model mismatch. A recurring result in the literature is that inaccurate modeling can produce an apparent inconsistency that mimics a false violation of GR (2207.13761, Pompili et al., 14 Apr 2025).

Residual orbital eccentricity is a prominent example. A Fisher-matrix study using angle-averaged IMRPhenomD, with eccentricity included only as a leading-order correction to the inspiral phase through the 3PN eccentric phase term $f_{\rm ISCO}$1, showed that neglecting eccentricity systematically biases the inspiral-based estimate of $f_{\rm ISCO}$2, while leaving the merger-ringdown estimate essentially unchanged in the paper’s approximation (2207.13761). The systematic parameter bias was estimated with the Cutler-Vallisneri formalism,

$f_{\rm ISCO}$3

with amplitude corrections neglected, and the bias scales approximately as

$f_{\rm ISCO}$4

for eccentricity $f_{\rm ISCO}$5 defined at $f_{\rm ISCO}$6 (2207.13761). The study found that, for LIGO-band systems with total mass $f_{\rm ISCO}$7–$f_{\rm ISCO}$8, the bias becomes significant at $f_{\rm ISCO}$9, while for CE systems with total mass $$M_f^{\rm I,MR}=M_f^{\rm I,MR}(m_1,m_2,\chi_1,\chi_2),\qquad \chi_f^{\rm I,MR}=\chi_f^{\rm I,MR}(m_1,m_2,\chi_1,\chi_2),$$0–$$M_f^{\rm I,MR}=M_f^{\rm I,MR}(m_1,m_2,\chi_1,\chi_2),\qquad \chi_f^{\rm I,MR}=\chi_f^{\rm I,MR}(m_1,m_2,\chi_1,\chi_2),$$1 the corresponding threshold is $$M_f^{\rm I,MR}=M_f^{\rm I,MR}(m_1,m_2,\chi_1,\chi_2),\qquad \chi_f^{\rm I,MR}=\chi_f^{\rm I,MR}(m_1,m_2,\chi_1,\chi_2),$$2 (2207.13761). It also estimated eccentric corrections to the remnant fitting relations themselves and concluded that they are $$M_f^{\rm I,MR}=M_f^{\rm I,MR}(m_1,m_2,\chi_1,\chi_2),\qquad \chi_f^{\rm I,MR}=\chi_f^{\rm I,MR}(m_1,m_2,\chi_1,\chi_2),$$3 for $$M_f^{\rm I,MR}=M_f^{\rm I,MR}(m_1,m_2,\chi_1,\chi_2),\qquad \chi_f^{\rm I,MR}=\chi_f^{\rm I,MR}(m_1,m_2,\chi_1,\chi_2),$$4, so the dominant issue is the eccentric bias in inspiral parameter estimation rather than the remnant map (2207.13761).

A full injection campaign with eccentric compact-binary signals reached a related conclusion using Bayesian recovery (Shaikh et al., 2024). For a GW150914-like eccentric binary analyzed with quasicircular waveforms, the IMRCT is broken at $$M_f^{\rm I,MR}=M_f^{\rm I,MR}(m_1,m_2,\chi_1,\chi_2),\qquad \chi_f^{\rm I,MR}=\chi_f^{\rm I,MR}(m_1,m_2,\chi_1,\chi_2),$$5 confidence for

$$M_f^{\rm I,MR}=M_f^{\rm I,MR}(m_1,m_2,\chi_1,\chi_2),\qquad \chi_f^{\rm I,MR}=\chi_f^{\rm I,MR}(m_1,m_2,\chi_1,\chi_2),$$6

at an orbit-averaged reference frequency $$M_f^{\rm I,MR}=M_f^{\rm I,MR}(m_1,m_2,\chi_1,\chi_2),\qquad \chi_f^{\rm I,MR}=\chi_f^{\rm I,MR}(m_1,m_2,\chi_1,\chi_2),$$7, and the violation becomes $$M_f^{\rm I,MR}=M_f^{\rm I,MR}(m_1,m_2,\chi_1,\chi_2),\qquad \chi_f^{\rm I,MR}=\chi_f^{\rm I,MR}(m_1,m_2,\chi_1,\chi_2),$$8 for

$$M_f^{\rm I,MR}=M_f^{\rm I,MR}(m_1,m_2,\chi_1,\chi_2),\qquad \chi_f^{\rm I,MR}=\chi_f^{\rm I,MR}(m_1,m_2,\chi_1,\chi_2),$$9

at the same reference frequency (Shaikh et al., 2024). When eccentric waveforms are used, the IMRCT remains intact for all eccentricities considered (Shaikh et al., 2024).

Higher-order modes and spin precession are additional systematic channels. An O2 reanalysis using the NRSur7dq2 surrogate with higher modes up to $$M_f = M_f(m_1,m_2,a_1,a_2), \qquad a_f = a_f(m_1,m_2,a_1,a_2),$$0 found that all analyzed events remained consistent with GR, and that the posterior structures in the more interesting cases were compatible with noise fluctuations, parameter degeneracies, and prior effects rather than a GR violation (Breschi et al., 2019). By contrast, a 2025 parametrized waveform study showed explicitly that neglecting spin precession can lead to false detections of deviations from GR even at current detector sensitivity, while a precessing parametrized model recovers consistency with GR for a highly precessing numerical-relativity injection (Pompili et al., 14 Apr 2025). That work is not the classic IMR split test, but it is directly relevant because it demonstrates a specific route by which waveform incompleteness can generate apparent non-GR behavior in remnant-based null tests (Pompili et al., 14 Apr 2025).

The choice of analysis window can also matter. A time-domain study of phase-separated tests using gating and in-painting found that waveform systematics can overstate the confidence of related area-law tests, especially when simplistic ringdown models are used at early start times (Kastha et al., 2021). A plausible implication is that the same caution applies to any IMR-style test in which remnant inference depends sensitively on where the post-inspiral segment begins or on the structure of the late-time waveform model.

6. Variants, extensions, and related consistency frameworks

The standard IMR consistency test has generated a family of extensions that retain its central logic—compare independently inferred source summaries—but alter the segmentation, the remnant variables, or the objects being compared.

One direction is the “meta inspiral-merger-ringdown consistency test” (meta IMRCT), which replaces the inspiral-versus-post-inspiral split by a comparison between the outputs of any two GR tests $$M_f = M_f(m_1,m_2,a_1,a_2), \qquad a_f = a_f(m_1,m_2,a_1,a_2),$$1 and $$M_f = M_f(m_1,m_2,a_1,a_2), \qquad a_f = a_f(m_1,m_2,a_1,a_2),$$2 that can be mapped to remnant mass and spin (Madekar et al., 2024). The generalized deviation variables are

$$M_f = M_f(m_1,m_2,a_1,a_2), \qquad a_f = a_f(m_1,m_2,a_1,a_2),$$3

and consistency is summarized through a GR quantile

$$M_f = M_f(m_1,m_2,a_1,a_2), \qquad a_f = a_f(m_1,m_2,a_1,a_2),$$4

That framework found quasicircular GR signals consistent with GR, detected non-GR and eccentric signals, and in some cases produced stronger inconsistency than the individual tests being compared (Madekar et al., 2024).

A second direction is segment generalization. The “Multi-Segment Consistency Test” (MSCT) presents the classic IMRCT as the two-segment case of a broader time-domain construction in which different segments share common source parameters, particularly extrinsic parameters such as $$M_f = M_f(m_1,m_2,a_1,a_2), \qquad a_f = a_f(m_1,m_2,a_1,a_2),$$5, while allowing segment-specific intrinsic or derived parameters (Prasad, 6 Mar 2026). In that framework,

$$M_f = M_f(m_1,m_2,a_1,a_2), \qquad a_f = a_f(m_1,m_2,a_1,a_2),$$6

with joint likelihood

$$M_f = M_f(m_1,m_2,a_1,a_2), \qquad a_f = a_f(m_1,m_2,a_1,a_2),$$7

The paper argues that earlier two-band tests treated segments too independently for a single astrophysical source and that enforcing shared extrinsics yields stricter constraints while capturing covariances (Prasad, 6 Mar 2026).

A third direction is the extension of IMR-style logic beyond binary black holes. For binary neutron stars, a pre/post-merger consistency test compares the dominant post-merger frequency $$M_f = M_f(m_1,m_2,a_1,a_2), \qquad a_f = a_f(m_1,m_2,a_1,a_2),$$8 predicted from inspiral tidal information through an EOS-insensitive relation with the $$M_f = M_f(m_1,m_2,a_1,a_2), \qquad a_f = a_f(m_1,m_2,a_1,a_2),$$9 measured directly from the post-merger signal (Breschi et al., 2023). The authors emphasize that this is technically similar to IMR consistency tests but conceptually less informative, because it tests the breakdown of a quasi-universal relation rather than GR itself (Breschi et al., 2023).

Finally, several recent ringdown-only tests are explicitly described as IMR-consistency-like but not standard IMR tests. A greybody-factor-based post-merger model called GreyRing infers remnant mass and spin from the post-merger signal alone and compares them with standard black-hole spectroscopy or full-signal results, thereby creating a post-merger versus spectroscopy consistency check rather than an inspiral versus ringdown test (Rosato et al., 13 Apr 2026). Likewise, the “merger-ringdown consistency test” based on deep learning compares the remnant spin measured from ringdown with the spin inferred from ringdown-derived progenitor information, probing consistency between plunge-merger excitation and linear ringdown response (Bhagwat et al., 2021).

Across these variants, the standard IMR consistency test remains the reference construction: split the signal, infer the remnant twice, and ask whether the posterior supports the GR point $$\Delta M_f \equiv M_f^{\rm I}-M_f^{\rm MR},\qquad \Delta \chi_f \equiv \chi_f^{\rm I}-\chi_f^{\rm MR},$$0 in a remnant-deviation plane (Carson et al., 2019, Ghosh et al., 2017). The central lesson of the subsequent literature is not that this construction is obsolete, but that its interpretation is inseparable from waveform fidelity, parameter correlations, and the choice of how the signal is partitioned (2207.13761, Zhong et al., 2024).