WorksPhysics and Astronomy
Deglobalising the fit: how to search for physics beyond vacuum General Relativity in LISA catalogues
Abstract
The Laser Interferometer Space Antenna will observe up to overlapping signals, which have to be characterised simultaneously with a global fit. This is a challenge for model comparison since rerunning the global fit for every physical model of interest, including deviations from vacuum General Relativity, is computationally demanding. An alternative is a two-step post-processing strategy: a first global fit with vacuum models only, after which a selected group of target sources is reanalysed under a different model with all other sources subtracted. We show that in this strategy a beyond-vacuum effect in a target source can be partly absorbed by the overlapping nuisance sources in the first fit, subtracted with them, and never reach the reanalysis. In the linear-signal approximation, the loss is captured by a single number, the masking ratio . A deviation the global fit would detect at is reported by post-processing at , and bounds in null tests are also reported too tight by the same factor. We show how to compute from the posterior samples of the vacuum global fit, and give an algorithm that ranks the nuisance sources by how much of the effect each can absorb, so that only those most correlated with it need to be refitted with the target in a down-scaled reanalysis with a few sources. We validate our method with full Markov-chain Monte Carlo inference on a toy model of 21 overlapping chirping sinusoids, where one carries a small non-linear chirp that represents the missed physics, while the other 20 act as nuisance sources. This method can serve as a basis for tests of General Relativity in realistic LISA mock datasets and can be adapted to study the impact of systematics between different gravitational-wave sources.
Introduction
The current network of ground-based gravitational-wave (GW) detectors finds a new event every couple of days of operation [1, 2]. As instrument sensitivity improves, signals will begin to overlap and will have to be fitted jointly when sufficiently correlated. This is not a problem for the LIGO-Virgo-KAGRA network, but it is expected to happen for next-generation ground-based detectors, such as the Einstein Telescope and Cosmic Explorer [3, 4, 5, 6], where a typical binary neutron star signal will have tens of other signals overlapping it, and two events will merge within seconds of each other tens of thousands of times per year [7]. The challenge will be even greater for the space-based Laser Interferometer Space Antenna (LISA), scheduled to launch in 2035 [8]. LISA will probe the millihertz band, where many sources are long-lived: Galactic binaries, of which are individually resolvable while the rest contribute to an unresolved foreground; up to hundreds of massive BH binaries with total masses of – [9]; up to thousands of extreme-mass-ratio inspirals [10, 11], in which a stellar-mass compact object spirals into a massive BH while completing up to orbital cycles over months to years; a few stellar-mass BH binaries in their early inspiral, years before they enter the band where ground-based detectors may observe them [12]; and possibly stochastic GW backgrounds of astrophysical origin (e.g. unresolved extragalactic stellar-mass binaries) [13, 14, 15] or cosmological origin (e.g. phase transitions in the early Universe) [16]. Adding more complexity to the problem, LISA’s noise is not known a priori and is non-stationary, with gaps and glitches [17, 18, 19, 20], so it must also be inferred jointly with the signals.
This sheer number of overlapping signals requires a global fit to extract all sources from the data simultaneously [21, 22, 23, 24, 25]. Global fit algorithms are under active development using increasingly more realistic LISA mock datasets [26]. The current plan to tackle real data is for the first analysis to use waveform models assuming only two-body, vacuum General Relativity (GR) for all binary systems. Public data products will then be released that enable the scientific community to pursue science investigations that go beyond those initial catalogues [27], for example using more accurate vacuum waveform models, models with beyond-GR effects [28, 29, 30], or astrophysical environmental effects [31, 32]. How this will be achieved in practice is an open question. Rerunning a full global fit for every model the community might want to test is infeasible, both in complexity and computational cost. Instead, it is expected that the two-body, vacuum models will be accurate enough that any deviation from them is small, so that the public data products can be post-processed, reanalysing only a few target sources of interest [33]. To test GR, for example, the logical choice would be to pick sources with large signal-to-noise ratio (SNR), and/or with an anomalous accumulation of SNR [33, 34].
However, correlations between sources can mask physical effects already in the initial vacuum global fit and remain hidden in post-processing. Run under the vacuum hypothesis, that first fit can absorb any beyond-vacuum effect into its parameters. Part of the effect goes into the target source’s own “normal”, vacuum parameters, and a post-processing refit can (in principle) undo this. Another part is absorbed by overlapping nuisance sources, which are then subtracted from the data having already “eaten” part of the effect, which never reaches the post-processing step, and a reanalysis of the target source alone cannot recover it (note the same correlations matter when no deviation is present). Post-processing then quotes an uncertainty on the beyond-vacuum parameters that treats the nuisance sources as perfectly known, when in fact they were fitted and could themselves have absorbed a deviation had one been there. The upper bound it reports is therefore tighter than what it would be in the global fit. The propagation of biases between overlapping sources has been studied in [35], within the same linear-signal treatment we use here. That work studied how unfitted signals, inaccurate waveforms and imperfectly subtracted sources bias the recovered parameters, including a treatment of the unresolved foreground, and proposed iteratively correcting the peaks of a source-by-source analysis towards the global-fit answer. Here, we are instead concerned with how much of the evidence for a beyond-vacuum effect survives the two-step analysis, and which sources have to be refitted jointly for it to be recovered.
To achieve this, it would be useful to have a computationally cheap method to identify, after the first vacuum global fit, which sources are most able to mask a given unmodelled effect in another source, so that these can be analysed simultaneously in a ‘down-scaled’ reanalysis, a joint refit of the target and those sources only. This is what we construct here. As we will show, under the linear-signal approximation, the difference in the detection significance for a deviation from vacuum GR in the two strategies (global fit vs post-processing) can be captured by the masking ratio , which is a single number between 0 and 1. When a deviation is present, it is the factor multiplying the detection significance, so that a deviation the global fit would establish at is reported by post-processing at . When no effect is in the data, it is the factor by which the upper bound is overconfidently quoted. We also show a practical method to compute this ratio directly from the vacuum global-fit posterior samples, and an algorithm that, given a target source and a beyond-vacuum model, ranks the nuisance sources by how much of the effect each is able to absorb. Our treatment can also be applied to evaluate waveform systematics, tracking how they propagate to errors in the other overlapping sources and how they can lead to a false-positive deviation from GR. We validate our method with full Markov-chain Monte Carlo inference on a toy model of 21 chirping sinusoids.
The paper is organised as follows: in § 2, we summarise our results as a practical recipe. In § 3, we set up the notation and give a geometric and statistical picture of the problem. § 4 and § 5 derive the detection statistics of the global fit and of post-processing, including a variant that accounts for the uncertainty of the subtracted sources. In § 6, we compare them and define the masking and null ratios quantifying the loss in detection significance in the post-processing strategy, and in § 7 we show how to compute them from the vacuum global-fit posterior and how to select the sources to refit in order to minimise the loss in detection significance in post-processing. We validate the method with full Markov-chain Monte Carlo inference on a toy model in § 8, and discuss its applicability and limitations in § 9.
Recipe for tests of vacuum General Relativity in post-processing
For readers mainly interested in applying our method, we first summarise our results in the form of a recipe for post-processing a target source after a vacuum global fit. Each step refers to the section where it is derived, and every quantity it needs can be computed from the posterior samples of the vacuum global fit without rerunning it.
From the posterior samples of the vacuum global fit, compute the masking ratio and the null ratio (§ 6). The smaller of and bounds the loss for an effect of any size (Eq. (66)). For several beyond-vacuum parameters and no specific model, use instead the worst case over all directions, of Eq. (64), where is the smallest of the weights of Eq. (51).
If the number computed in Step 1 is close enough to 1, subtract all nuisance sources and refit the target source alone under the new model being studied. How close is close enough depends on how much detection significance one is willing to give up. For example, a value of 0.9 obtained in Step 1 means that a deviation the global fit would detect at is still detected at or more, and that bounds in null tests are at most 10% too tight.
If the number computed in Step 1 is below the determined treshold, rank the nuisance sources with the algorithm of § 7.3, and add them to a set until the same number, computed for the residual masking (Eqs. (82) and (84)) is close enough to one. If the effect is large, sweep the expansion point over the allowed range of the beyond-vacuum parameters and take the union of the selected sources, or re-evaluate the ranking at the value recovered by each refit until the selected set no longer changes.
Subtract only the sources outside , and refit the target sources jointly with the sources in .
The nuisance sources can be subtracted in two ways. Subtracting them at their best fit is cheaper and gives detection significances closer to those of the global fit, so it is the better choice for detecting a deviation. Averaging over the posterior of the first fit instead costs one refit per posterior draw and up to a factor in detection significance, but brings the upper limits of a null test closer to those of the global fit, so it is the better choice for setting bounds.
Setup and Notation
We take the observed GW data stream to be
where is zero-mean Gaussian noise with covariance and denotes an average over noise realisations. Here, and are the waveform models used in the analysis, so that the full template is
and we separate the modelled signals into two classes:
: the target source(s) of interest, possibly carrying a beyond-vacuum effect;
: nuisance (or confusion) sources, without a beyond-vacuum effect.
The parameter space is also divided into three blocks:
: the beyond-vacuum parameters of the source(s) of interest, with recovering vacuum;
: the vacuum parameters of the target source(s);
: the nuisance-source parameters.
A subscript “0” denotes the true value of a parameter and “bf” a best-fit value under a given hypothesis. The other term in Eq. (1), , is the model error, representing all signal components of the data that the template does not include when evaluated at the true parameters. These errors could come from unmodelled (quiet) sources, waveform systematics, or instrumental artefacts the noise model does not describe, like glitches and gaps.
We assume a Gaussian likelihood so that , where is the noise-weighted inner product
We caution the reader that the next subsections are particularly heavy on notation, but introducing it makes the presentation of our results later on much more tractable.
Geometric intuition
Before delving into any statistical analysis, we can use the inner product (3) to build a geometric picture of the problem. Varying the template parameters traces out a manifold in data space. The vacuum (sub)manifold is defined by , so it contains everything a vacuum global fit can represent. Near a reference point in parameter space, the model is characterised by its tangent directions , so its tangent space is
where is spanned by the vacuum directions of the target source and by the nuisance-source directions. We can project any vector onto this space with
where is the sub-block of
A beyond-vacuum effect displaces the model in data space by , which we assume is small enough that the curvature of the vacuum submanifold is negligible. The part of this displacement along can be reproduced by adjusting the vacuum and nuisance-source parameters, and is therefore indistinguishable from vacuum. Only the orthogonal component is in principle observable, which has length
with
The bracket defines the conditional matrix , whose statistical meaning is given in Sec. 3.2. We make this more precise below, but intuitively, to claim a deviation from vacuum GR we need this length to be large, so
where the number on the right-hand side is set by the confidence level at which the deviation is claimed (, , etc). In practice, the criterion is applied to the recovered best fit, , which noise and model error displace from the true value. This displacement depends on which parameters are fitted jointly with , so it differs between the global fit and post-processing, and we will compute it for each strategy in Secs. 4 and 5, respectively.
The same construction applies to any subspace. In the simplest post-processing strategy, the nuisance sources are subtracted from the data as if they were perfectly known and only the target source is reanalysed. Therefore, only is then available to absorb the effect in the reanalysis of the residual, and projecting out alone gives a second conditional matrix
Since , intuitively projecting out the larger subspace can only shorten the vector. Their difference is what nuisance sources are able to absorb. This has two consequences for post-processing (which we make precise in § 5). First, because the vector is longer, the posterior on is narrower than in the global fit, so when no effect is present post-processing is overconfident in the upper limit it sets. Second, when an effect is present, the part of it lying along is absorbed by the nuisance sources in the first fit and subtracted with them, so it never reaches the reanalysis, and the recovered value is biased towards vacuum. The extra length that makes post-processing look more sensitive is therefore precisely the part of the effect that the nuisance sources have already absorbed.
Statistical interpretation
We now give a statistical meaning to each object in the geometric picture above. The matrix (6) is the Fisher information matrix [36, 37]. It is the curvature of the (Gaussian) log-likelihood around the truth in the linear-signal approximation, in which the template is expanded to first order in the parameter displacement so that the log-likelihood becomes quadratic. In the same approximation, the posterior is Gaussian with covariance , so that correlations between parameters and the measurement uncertainty in the quantity are
Noise and model error do not affect the covariance at leading order, but bias the recovered parameters by the Cutler–Vallisneri bias [36]
which is the projection of the residual (model error and noise) onto the tangent vector of the corresponding parameter direction. The noise contribution averages to zero over many realisations, but has non-zero variance. The noise-induced fluctuations are statistically consistent with the width of the posterior distributions.
Projecting out a block of directions , as done above in the geometric picture, is what in statistics is called marginalisation. It corresponds to integrating out while retaining the uncertainty it induces on the remaining blocks. It gives the conditional Fisher matrix
with shorthand for the case. Marginalising over two blocks at once factorises,
so that marginalising over is the same as marginalising over and then (or vice versa). Marginalisation can only lead to a loss of information. Note also that since is positive-definite, for every , which we write .
The inverse of a two-block Fisher matrix is (in the same conditional notation)
Here, and can each stand for a group of parameters, so the same rule inverts a matrix with more than two blocks once its rows and columns are split into the respective subsets. Taking and shows . For a single beyond-vacuum GR parameter with its fully marginalised uncertainty, and is simply the distance from vacuum in units of that uncertainty. The geometric criterion (9) is then the statement that a confident deviation from GR requires a shift (much) larger than the statistical uncertainty in the measurement of the beyond-vacuum GR parameter.
Global Fit strategy detection statistic
The detectability criterion of Eq. (9) is what one obtains in the full global fit, in which the beyond-vacuum GR parameters are included in the model from the start. We rederive it here with the statistical machinery of Sec. 3.2, which also gives us the bias and the effect of systematic errors in the model.
In the global fit strategy, the model used to analyse the data is
The Fisher matrix splits into three blocks
The off-diagonal terms, and , quantify how much the vacuum parameters and the nuisance sources can absorb the beyond-vacuum effect. Marginalising over both, using Eq. (14), we have
This joint Fisher matrix over a target source and its overlapping neighbours, and the conditioning of its target block, were used in Ref. [35] to quantify how fitting the neighbours jointly degrades the precision on the target’s parameters.
The best-fit value is at a distance from vacuum given by
To compute , we apply Eq. (12) with residual , inverting Eq. (17) using Eq. (15) with the blocks grouped as and , so that the beyond-vacuum GR block of the inverse is . We find
with
and . Geometrically, is the overlap of the residual with . The operator projects into the space not spanned by changes to the vacuum and nuisance source parameters, while spans the space accessible by changing the beyond-vacuum parameters. This implies that the shift in the beyond-vacuum parameters is driven by the part of the residual that can be absorbed by changes to the beyond-vacuum parameters and not by changes to the vacuum and nuisance source parameters. The noise piece has zero mean and covariance by Eq. (15). Thus, averaging over noise, the term linear in vanishes and the quadratic one gives , with , so that
The first term is the signal contribution in the fully marginalised metric, the cross term an interference between model error and signal, and the last one is present even in vacuum (), where it can lead to a false-positive deviation from vacuum GR. Note that for no model errors, , the detection statistic becomes
while with model errors, the three terms of Eq. (22) form a perfect square by defining the effective beyond-vacuum parameter (which is the noise average of the recovered best fit value)
in terms of which noise-averaged detection statistic becomes
Maximum likelihood detection significance
is minus the argument of the Gaussian likelihood evaluated for vacuum-GR and, as such, can be directly interpreted as a significance of a departure from vacuum GR, in terms of (the sum over beyond-vacuum-GR parameters of the squares of) the number of standard deviations the best-fit parameters are away from zero. However, it can also be interpreted as a Bayes Factor or as the standard test statistic for identifying a deviation from vacuum-GR. If we maximise the (Gaussian) likelihood over all parameters vs with held at 0 (the null hypothesis), the two best-fit residuals differ by
so is the log-likelihood ratio between the beyond-vacuum and vacuum hypotheses. It follows a non-central distribution
with degrees of freedom and mean . A detection is then claimed at significance , which for a single beyond-vacuum parameter is the recovered value in units of its uncertainty, with . For a strong beyond-vacuum GR effect, it averages to . Conversely, in vacuum the distribution is central, so a false-alarm probability fixes the threshold , or equivalently sigmas. For one beyond-vacuum parameter the two coincide, , so three sigma is and five sigma . Note that the threshold at fixed grows with .
Also, since is built from instead of , the threshold is only meaningful if the systematics are controlled well enough such that . This is a requirement not on the overall accuracy of the waveform model but on its projection onto the beyond-vacuum direction, and it can be used in future studies on the impact of waveform systematics in beyond-vacuum GR effects.
Relationship to the Savage-Dickey ratio
We have formulated the detection statistic in terms of a comparison of maximum likelihoods. However, in gravitational-wave astronomy it is more usual to carry out model comparison using Bayesian evidence ratios. Within the linear signal approximation we employed, these are actually equivalent. For the comparison of nested hypotheses of the form
under the assumption that the priors on the common model parameters, , are consistent between the two hypotheses, the log evidence ratio (log Bayes factor) for over reduces to the Savage-Dickey ratio [38, 39]
which is simply the difference in the log-prior and log-posterior probability of the null hypothesis under the alternative hypothesis. If we assume that we place a flat prior on the beyond-vacuum parameters and that the allowed prior volume is large enough to contain the bulk of the posterior distribution, so that the integral may be approximated as an integral to infinity, we obtain
in which is the volume of the prior space for the beyond-vacuum parameters, so that the argument of the logarithm is the ratio of the prior volume to that of the posterior, . The prior volume must be bigger than a few times so that it completely contains the posterior. Making the volume even bigger reduces the Bayes factor. We conclude that the biggest log Bayes factor will be approximately . So, a threshold on computed from a distribution is equivalent to setting a threshold on the log-evidence to be greater than a few. The fact that the choice of impacts the Bayes factor simply reflects the usual prior dependence of Bayesian model comparison. In the following we will not worry about this further, but will base our analysis on and related quantities, as this is the key quantity in both frequentist and Bayesian analyses.
Post-processing strategy detection statistic
We now want to obtain a similar detection statistic in the post-processing strategy, where the data is first fitted with vacuum templates
and the target source of interest is then reanalysed on the residual left after subtracting the nuisance sources. Here, we assume that the adopted post-processing strategy involves subtracting the best-fit nuisance sources and using the single residual obtained in this way for the subsequent reanalysis. Later we will consider an alternative strategy in which we also marginalise over the uncertainty in the initial parameter estimation.
The possible bias in comes in two steps. First, the vacuum fit biases the recovered nuisance-source parameters, because those can absorb the unmodelled beyond-vacuum effect. Second, those biased sources are subtracted from the data, carrying the bias to the residual used for the beyond-vacuum inference in post-processing analyses. We compute these two contributions below to obtain and the corresponding detection statistic .
First step: the vacuum global fit
Let denote the vacuum parameters that best reproduce the true signal of the target source of interest, that is, the minimiser over of with
Minimising then gives , and hence
We now expand the vacuum global fit around and write for the displacement each block has from that point at the maximum likelihood solution. The residual is
The model error enters the analysis in the same way as , since both are pieces of data the vacuum templates cannot represent. Minimising gives
which is similar to the bias formula (12) applied to the first vacuum fit, with the difference that the displacement is measured from rather than from the true parameters. Substituting back into Eq. (34), we find that the best-fit residual is
so the fit removes whatever lies in the vacuum tangent space, in agreement with the geometric intuition of § 3.1. Now, inverting Eq. (35) blockwise using Eq. (15) (with , )
Isolating the term corresponding to the unmodelled beyond-vacuum effect
This is where post-processing loses relative to the global fit. The part of the effect absorbed by the target source’s own vacuum parameters, , can be recovered in the second fit of the post-processing stage, which frees again. But whatever correlates with the nuisance sources () is absorbed into their parameters in the first fit, and is subtracted from the data with them, never reaching the second step of the reanalysis. Note that Eq. (38) is just the multi-source (Cutler–Vallisneri) bias of Ref. [35], with their generic waveform error applied to the parametrised effect .
Second step: reanalysis of the residual
Subtracting the fitted, possibly biased nuisance sources from the data leaves the residual
We now fit to . This time, we expand around the true parameters instead of , so that what we obtain is the bias on the beyond-vacuum parameters, which is directly comparable with the of § 4.
Applying the bias formula (12) with the two-block rule (Eq. (15) with , ), in analogy with Eq. (20),
Note that now appears instead of , because this step marginalises over but has already been fixed at its biased value and subtracted from the data. Also, Eq. (40) is not the sum of independent contributions because is itself built from , and through Eq. (37). Expanding it and separating into these 3 terms
where are the global-fit overlaps of Eq. (21) and the last equality follows from Eq. (18). The last term, , is precisely the part of the effect absorbed by the nuisance sources in the first step. It is set by the difference between the two conditional Fisher matrices, and vanishes when the nuisance sources do not correlate with the beyond-vacuum direction of the target, so that and .
The distance from vacuum, the analogue of Eq. (19), is
By Eq. (41) the beyond-vacuum term can be absorbed into the prefactor, , so that
Averaging over noise realisations, the term linear in vanishes as before, and the quadratic one gives
The last inequality follows from and that Fisher matrices are positive semi-definite by construction (we show this in Appendix § 11). The expectation of the post-processing statistic in the null case, when no extra effect is present in the data, is therefore smaller than that of the global fit. This implies that the recovered parameters fluctuate less across noise realisations than the quoted uncertainty in post-processing. In this strategy, the second step assumes that only the target source is present in the data and therefore all noise degrees of freedom can impact parameter estimation. In reality, the noise degrees of freedom that also correlate with the nuisance sources have been removed in the first stage of the fit, and therefore the residual uncertainty is less than the model assumes. Therefore, the posterior uncertainty overestimates the true fluctuations, unless , i.e. unless the nuisance sources do not correlate with the beyond-vacuum direction of the target source.
Expressing Eq. (43) as done for the global fit strategy (22),
where the different terms have a similar interpretation and can also be combined into a perfect square at the same effective parameter of Eq. (24).
Comparing Eq. (43) with Eq. (20), we see that
Post-processing does not recover but
So, the best fit estimate in post-processing is biased by the fraction of the effect the nuisance sources absorbed. As we show below, this bias is always towards vacuum.
Detection significance
The best-fit residuals obtained by fitting and to the residual of Eq. (39) differ by
so is the right statistic to use for claiming a deviation from GR in the post-processing strategy. It also gives the Bayes factor for the presence of a beyond-vacuum GR effect, as discussed in § 4.2. The distribution of is not a , because the recovered value is not being compared against its covariance. Applying to Eq. (43)
For a Gaussian distribution , the quadratic form is a only when . In general, it follows a weighted sum of independent non-central , with weights given by the eigenvalues of . Here, and is Eq. (50), so that and
with and being, respectively, the eigenvalues and orthonormal eigenvectors of the symmetric matrix . We show that are real, positive and no greater than unity in Appendix § 11. The non-centrality parameters are given by
In the global fit strategy, and Eq. (51) becomes the of Eq. (27). Also, note that for a single beyond-vacuum parameter there is only one weight, , and the two statistics are then related by a constant, , for every noise realisation.
Accounting for the uncertainty of the subtracted nuisance sources
The two-step strategy above assumes that a single residual will be used in post-processing by subtracting the nuisance sources using the point estimate of the first fit. But the first fit also provides the uncertainty of that estimate, and so an alternative would be to propagate this uncertainty into the second stage. Instead of subtracting the best fit, one can draw samples from the posterior of the vacuum global fit, subtract each, repeat the second step of the post-processing analysis, and average the resulting posteriors. The additional cost compared to the simple post-processing strategy is then one low-dimensional refit per draw.
At linear-signal order, the effect of averaging can be inferred following Sec. 5.2. The nuisance parameters of a draw are , where is the fluctuation of the subtracted configuration around the first-step best fit. Its covariance is the block of the inverse of the first-step Fisher matrix in Eq. (35), which by Eq. (15) is . Subtracting instead of adds to the residual of Eq. (39), which in turn adds to in Eq. (40), and displaces the recovered value of Eq. (43) by
Since is drawn from a Gaussian, follows a zero-mean Gaussian with covariance
where the last equality follows from Eq. (18).
Each per-draw posterior is the Gaussian of Sec. 5.2 displaced by , which is itself a zero-mean Gaussian. Averaging them then leads to a Gaussian posterior with the same mean as simple post-processing, Eq. (43), and with covariance the sum of that of simple post-processing and that of the shifts
where the overbar denotes the average over the posterior draws.
The detection statistic to consider is now the distance from vacuum of the averaged posterior, measured with this new covariance
which is the analogue of Eq. (42) for simple post-processing. Substituting Eq. (43) and taking the noise average as for Eq. (46)
In the next section we compare the detection statistics under these different strategies.
The Masking Ratio
We start by quantifying what simple post-processing loses with respect to the full global fit. The difference between their detection statistics is
and both terms are non-negative, as we show in Appendix § 11. So, on average and in the linear-signal approximation, post-processing can only mask a beyond-vacuum GR effect. When no such effect is present, the bound it sets on a deviation from GR is spuriously tight. We consider both cases below, and then combine them into a single measure in Eq. (66).
It may seem counter-intuitive that the post-processing strategy is both too conservative, potentially missing deviations that are present, and too optimistic, setting bounds that are too tight. This is because it gets both the mean and the width of the posterior wrong. The width is determined by the number of degrees of freedom being fitted and this is fewer in the post-processing strategy as the nuisance sources are not included. The mean, however, is shifted by an amount that is determined from the correlations between the nuisance and target sources. The post-processing strategy assumes all degrees of freedom in the target source are available in the fit, while some of these are removed because they correlate with the nuisance sources. Thus, the mean is shifted closer to zero than the post-processing model assumes is possible. Overall, post-processing produces posteriors that are too narrow, but are also closer to zero and fluctuating less than the post-processing model expects. The result is that we are both over-confident in upper-limits on deviations from vacuum-GR, and less likely to detect a deviation that is present.
For a sufficiently strong beyond-vacuum effect, the significance of a global-fit detection approaches , the square root of the non-centrality of its distribution (§ 4.1). We then define the masking ratio
A deviation that the global fit would have detected at is then detected by post-processing at .
When no effect is present, can still be computed, since it is independent of the amplitude of the effect. However, its interpretation changes, as it is now related to the two possible failure modes of a null test. The first is that we think a deviation is present when it is not. Fully quantifying this requires a detection algorithm, but we assume for simplicity that it is based on . A threshold is placed on , and a value above it is deemed a detection. Under the post-processing model assumptions, follows a distribution, which would be used to set the detection threshold. In reality, it follows the weighted sum of distributions of Eq. (51), and we define the null ratio as the ratio of its actual mean to the assumed one
so the factor by which the null expectation of is below the on which the detection statistic would be based.
Since , post-processing is on average less likely than the global fit to claim a deviation that is not there, and, as we now show, this holds at every significance level. In terms of the eigenvectors of Eq. (51), the two statistics are and , with the same in both, because by Eq. (43) the two best fits are related by . Since , in every noise realisation (for they are simply proportional, as we show in Eq. (68) below). In the null case, the are independent standard Gaussians, so follows exactly the from which we set the detection threshold in post-processing. Hence for any threshold, so post-processing is a conservative strategy in the sense that never rejects vacuum GR more often than the nominal false-alarm probability.
The second failure mode is that, having found no evidence for a deviation, we place an upper limit that is too high or too low. Again, exactly what this means depends on the scenario, since there can be several beyond-vacuum parameters, constrained individually or jointly, and the upper limit is really a statement about the full beyond-vacuum posterior. The simplest thing to do is to compare the widths of the global-fit and post-processing posteriors, and this again yields . Now, its numerator can be interpreted as the expected squared significance, measured with the global-fit posterior, of a point drawn at random from the post-processing posterior, if both were centred at the origin. A value then implies that the post-processing posterior is more compact than the global-fit one, and hence yields tighter upper limits. This simple comparison neglects that the mean is shifted as well, by an amount inconsistent with the posterior width, which also affects the upper limit. Inspired by the previous interpretation of , one way to compute a quantity that accounts for this is to consider the distance at which samples randomly drawn from the post-processing posterior lie from the centre of the global-fit posterior, in units of its width, now allowing for the fluctuations in the mean as well as the differences in the width. In the null case, the global-fit posterior is and the post-processing one is , so a sample of the latter is , with , and the relevant quantity is
For samples of the global-fit posterior itself, its expectation is . If it is smaller, the post-processing posterior is too concentrated and any upper limit is overly tight, and if it is larger than , upper limits are overly conservative. Taking the expectation over and over noise realisations, and noting that by Eq. (43), we find
where the first term comes from the shift of the mean, , and the second from the width, . The upper limits obtained from post-processing are therefore always overly tight. Additionally, since , we conclude that placing a criterion that the null ratio is sufficiently close to 1 ensure that neither of the two failure modes of non-detections occurs. We will henceforth use only to assess the potential of wrong inference through post-processing in the null case.
Note that the null ratio and the masking ratio are not independent, as they both involve averages of the weights
The weights depend only on the waveform models of the target and nuisance sources, through the Fisher matrices at the expansion point, and not on the noise realisation. Therefore, they, and with them, can be computed before anything is known about the effect. The masking ratio does not depend on the amplitude of the effect either, since it is invariant under . For , however, it does depend on the direction in the beyond-vacuum subspace along which the effect is excited. Even then, it is bounded by the smallest and largest weights
as we also show in Appendix § 11. Since is the mean of the weights, it lies in the same range, .
The masking and null ratios can be interpreted as the two limits of a single quantity, which we call the survival ratio, as it measures the fraction of the expected global-fit statistic that survives post-processing, with from Eq. (58). By Eq. (51),
This quantity tends to when no effect is present and to for a strong one, . Unlike the other two ratios, the survival ratio depends on the amplitude of the effect through , which is not known before the analysis. However, since it is a weighted average of and , it always lies between them, so that
for an effect of any size. We therefore use the smaller of and as a single, conservative measure of how much post-processing loses. When a specific beyond-vacuum model is tested, is evaluated along the direction of that it predicts. Otherwise, the worst case over all directions is , which bounds all three ratios from below, and is then the natural number with which to select the sources to refit.
For a single beyond-vacuum GR parameter there is only one weight, and all three ratios coincide
independently of the tested value (a consequence of the linear-signal approximation). A single number therefore describes both detections and null tests, and the detection significances of the two strategies are related in every noise realisation
In post-processing, the best-fit value of the beyond-vacuum GR parameter shrinks by , while the quoted uncertainty is tighter than the global fit’s by only , so the error bar on shrinks at half the rate at which the nuisance sources absorb the effect. Although this is the simplest scenario, the case with covers many physical setups of interest [11, 40, 41, 42], in which the extra effect enters the waveform at leading order through a single dominant term whose amplitude is the only new parameter. This is also the case for parametrised deviations from GR [43, 44], in which the deviation is expanded in a fixed basis (e.g. in weak-field post-Newtonian order) and its coefficients are fitted to the data. These are usually constrained one order at a time, so .
We can also compare the detection significance of the two post-processing strategies (simple vs averaged). They share the same point estimate for the beyond-vacuum parameters, , and hence its sampling covariance, Eq. (50). The only difference is the covariance against which it is compared in the detection statistic. In simple post-processing it is in Eq. (49), and in the average one it is in Eq. (56). The latter is a function of the same matrix whose eigenvalues define the weights in Eq. (51), so in that decomposition only the weights change
with the same non-centrality parameters . The masking and null ratios of the averaged strategy then follow from Eq. (63) with
with the same bracketing by the smallest and largest weights as in Eq. (64), and the survival ratio of Eq. (65) decreases likewise. Since , the significance of a detection decreases, but by at most a factor of and by much less when any . This happens because averaging broadens the posterior, while not shifting the peak, bringing the distribution statistically closer to zero. In simple post-processing, the shrinking of the recovered value by is partly compensated by the tighter uncertainty, as noted after Eq. (68), and averaging undoes part of this compensation by correcting the width but not the mean. This suggests that, for a detection, it is better to subtract only the best fit, which is also computationally cheaper. However, this might be less robust to mismodelling, and it would be the wrong strategy for setting an upper limit, as we now argue.
In the null case, the expectation of the statistic decreases likewise, to . This means we would be less likely to falsely identify a beyond-vacuum GR deviation when none is present. However, the posterior also widens, becoming more consistent with the global-fit one, since , where the middle term is the averaged covariance of Eq. (55) in units of the global-fit one. We can also evaluate , as defined in Eq. (62), for this case. The mean is unchanged, but the covariance of is now , Eq. (55), which modifies to
where only the second, width term has changed. By this metric, averaging recovers posteriors that are, on average, statistically consistent with those of the global fit, which suggests that averaging is the better post-processing strategy for setting upper limits when no deviation is detected. This is intuitively sensible, as when the correction is absent, the vacuum model of the first fit is correct, and its full posterior preserves the information about the correlations between the target and the nuisance sources, so we get close to the true upper limit in post-processing. Note that, unlike for simple post-processing, is then no guide to the upper limits since it moves away from one, while the upper limits become calibrated.
For a single beyond-vacuum parameter there is again only one weight, with , so all the ratios of the averaged strategy coincide as well, and
where the last two equalities hold in every noise realisation, as in Eq. (68). The recovered value is still , with a posterior wider than in simple post-processing by and narrower than the global-fit one by .
In summary, when a deviation is present, the averaging strategy gives results that are more conservative than those of simple post-processing, and further from the global fit, while when no deviation is present it gives upper limits that are statistically consistent with the global fit. Depending on the nature of the effect and its expected size, averaging can therefore be preferable for null tests, although neither strategy undoes the masking, which requires refitting the nuisance sources that absorb the effect, as we discuss next. Simple post-processing is also computationally cheaper, so it is better to start from it and add complexity as needed. For the remainder of the paper we will focus on the simple post-processing strategy, and add comments on the averaged strategy as needed.
Computing the ratio and selecting the sources to refit
To compute the masking and null ratios discussed in the last section, we need the two matrices and . The first only involves the target source of interest and can be computed and inverted directly, but the second requires marginalising over all sources in the data. For LISA, this means computing and inverting a Fisher matrix over resolvable sources, which is computationally infeasible, especially given the large dynamic range that is typical of GW Fisher matrices, which makes them ill-conditioned and unstable to invert numerically. As we now show, this direct computation is not needed as the total masking can be estimated from the posterior samples of the vacuum global fit.
Estimating the total masking from the vacuum global fit posterior
The vacuum global fit already explores the joint posterior, and at linear-signal order that posterior corresponds to the inverse Fisher matrix. Recall the chain rule for the conditional matrices in Eq. (18). As mentioned above, the first term can be computed directly. To obtain the second, we need to isolate the beyond-vacuum direction from the target’s own vacuum absorption
evaluated at and the best-fit vacuum parameters (we come back to this choice of evaluation point in § 7.3). Then, for each posterior sample of the vacuum global fit, we compute
Linearising around the posterior mean, , the vacuum contribution cancels because by construction, leaving
The covariance of marginalised over is , so again using
which is exactly the term subtracted in the conditional matrix (18). The fully marginalised matrix can then be obtained from
We stress that in this estimate no Fisher matrix over the source catalogue is ever computed or inverted. The cost is one waveform-derivative evaluation plus one inner product per posterior sample evaluating the template , which was already computed during the first inference (and can in principle be saved). These are also the samples from which the averaging strategy of Sec. 5.4 draws its nuisance configurations, so the two reuse the same product of the first fit.
Since Eq. (77) involves a sample variance, its Monte-Carlo error can also be estimated. For effective posterior samples (the chain length divided by twice its autocorrelation time), the relative error of the covariance is , which propagates to
for a single beyond-vacuum-GR parameter.
Attributing the masking to a subset of nuisance sources
If both and are very close to 1, no nuisance source can absorb the effect, and simple post-processing is a safe strategy, with essentially no loss of detection significance and bounds close to those of the global fit. How close to 1 is close enough depends, ultimately, on how much significance one is willing to give up in exchange for a cheaper refit. What if this is not the case? The masking ratio is a single number for the whole data set, so it quantifies how much detection significance post-processing loses, but not to which nuisance sources it is lost, which is what we need in order to choose the ones to refit jointly in a down-scaled reanalysis.
Writing for the parameters of a subset of nuisance sources, consider their directions after the target source’s vacuum directions are projected out,
The conditional matrix restricted to this subset is
from which we compute the subset masking ratio
for a single beyond-vacuum parameter, with the obvious generalisation from Eq. (59) for . It has the same interpretation as , as if the sources in were the only nuisance sources in the data, and reduces to itself when is the whole catalogue. We denote the masking of a single source by . Computing these ratios requires only the waveform derivatives of the target and of the sources in , so the matrix to be inverted has the dimension of and can, in principle, be directly computed.
The residual masking left once has been included in the post-processing fit is then
So the masking factorises into the part due to and the part that remains after is refitted, and since , both factors lie in .
The residual ratio is the quantity that a down-scaled fit actually gives. Suppose we subtract from the data only the sources outside , at the parameters found by the vacuum global fit, and then fit the target together with the sources in , now with free. This is the two-step analysis of § 5, except that the sources in are treated as the global fit would treat them and only the rest are frozen and subtracted. Every step of that derivation goes through with replaced by , and therefore with replaced by . The analogues of Eqs. (43) and (68) for a single beyond-vacuum parameter are
The recovered value is shrunk by and the quoted uncertainty is too tight by , exactly as in § 6 but with the residual ratio in place of the total one. The corresponding null statement is that the down-scaled analysis has null ratio
so a bound set in the absence of any effect is too tight by the same mechanism, and by the same factor when .
The two limits recover the strategies already studied. For we have and , and Eq. (83) is simple post-processing. Conversely, for equal to the whole catalogue, and , and it is the global fit.
Selecting the sources
We are then looking for the smallest set of sources worth refitting together with the target source of interest. Let denote the part of that the sources in cannot reproduce, so that . Starting from empty, at each step:
For each candidate not included in , project out from the directions already covered by , , and compute how much of it accounts for,
Add the candidate with the largest to and update by removing this contribution.
Evaluate and .
Stop when the single measure of Eq. (66), computed for the residual masking, with Eqs. (82) and (84), is close enough to 1. For it is simply .
This is a greedy algorithm since at each step it adds the source that recovers the most masking, without revisiting earlier choices. For , is an matrix, and the subset ratios generalise to the eigenvalues of , as in Eq. (51). When a specific beyond-vacuum model is tested, both are projected along the direction it predicts, which reduces the selection to the single-parameter case. For a generic reanalysis, we instead add at each step the candidate that most increases the smallest of these eigenvalues, and stop when its square root is close enough to one, so that the selected set is adequate for an effect in any direction. Once is selected, we subtract from the data only the sources outside it and refit the target source jointly with the sources in , now with free. Since is projected perpendicular to the sources already chosen, measures what candidate adds to the selected set. The projection can be updated sequentially as each source is added to , since only its own direction has to be projected out of each remaining candidate and of . Every step therefore involves only inverses of matrices the size of a single source’s parameter space not yet included in , as in Eq. (85), however large has become. The same selection can be used in the averaging strategy of § 5.4. The depend only on the target and nuisance directions, not on how the sources outside are subtracted, and for a single beyond-vacuum parameter is a monotonic function of , so the ranking is the same as in simple post-processing and only the stopping threshold changes.
In the first iteration of this algorithm, with empty, we obtain the individual ratios , with which we can rank the catalogue from the most to the least problematic source. Some extra care is needed when nuisance sources are strongly correlated among themselves. If two sources can reproduce the same part of , keeping one of them is enough. However, it can also happen that is only well reproduced by a combination of the two, and then each source is relatively harmless individually () even though the pair is not. This can only happen for sources that are correlated with each other. If their directions are orthogonal, , their contributions simply add, , whereas two nearly degenerate sources can together reproduce a direction that neither reproduces alone. To account for this, we can also include small groups of mutually correlated sources as single candidates in step 1, which can be identified from the correlations in the posterior of the first vacuum global fit. We will see an example of this in the next section where we apply our method to a simple toy model.
Finally, recall that all derivatives are evaluated at , since that is what the vacuum global fit provides. If the beyond-vacuum GR effect is very strong, it can enter the waveform non-linearly, and the ratios and the ranking of nuisance sources will then depend on the true value of . This is not in conflict with the linear-signal approximation, but means that a single tangent plane may not extrapolate well across the full range of allowed. In the toy model of § 8, for instance, the worst-masking sources change between and the value eventually recovered. The null ratio is not affected by this, since is the correct expansion point when no effect is present. For the masking ratio and the selection of , one can instead repeat the procedure above at a few values of , take for the union of the sources flagged anywhere in the sweep, and report the smallest obtained. Alternatively, the expansion point can be updated iteratively by evaluating the ranking at the current estimate of , refit with the selected sources, and repeat until the selected set no longer changes. We will see this in play in our toy model example.
A toy model
We now apply the procedure of the previous section to a toy model, and validate it against full Markov-chain Monte Carlo (MCMC) inference. The dataset is illustrated in Figure 1. It consists of a noiseless data stream of 21 linearly chirping sinusoids
Figure 1. The toy model dataset. Left: the (noiseless) time-domain data stream, with the target source, with SNR = 50, highlighted. Right: frequency tracks of the 21 chirping signals, coloured by SNR. The target is flanked by two pairs of nearly parallel tracks, group A below and group B above, while the remaining 16 sources cross its track during the observation. The dashed line is the target’s track when the beyond-vacuum effect of Eq. (87) is present with .
where and are the amplitude and initial phase of source , and and are constant for the 20 nuisance sources. The data is observed over , with all quantities in dimensionless units with , and with a flat noise power spectral density, . Source is the target source of interest, with SNR = 50, and its frequency evolution can carry a non-linear chirp term
in which case its phase is obtained by integrating Eq. (87). This plays the role of the beyond-vacuum GR effect in this toy model, with recovering vacuum. To keep the illustration simple we fix , so that the beyond-vacuum subspace has dimension one. We experimented with other choices and the qualitative picture below does not change. As discussed in § 6, this is similar in spirit to the parametrised tests of GR used in current GW analyses, where the order of the deviation is fixed and only its amplitude is fitted.
The remaining 20 sources form the nuisance population. Four of them lie close to the target and chirp almost in parallel with it. We divide them into two groups. Sources , which we label group A, at frequency offsets and with and ; and sources , group B, at and with and . The other 16 chirp much faster and cross the target’s track at some point during the observation. Their SNRs range from 9 to 38.
In all MCMC inference we sample the posterior with the Eryn sampler [45], using 128 walkers initialised in a Gaussian of width around the maximum-likelihood point of each analysis. We discard at least iterations as burn-in and check convergence with the integrated autocorrelation time, which for the slowest direction of the 84- and 85-dimensional fits is a few hundred stored steps, while the chains have stored steps. Here, is the marginal Fisher uncertainty of each parameter, obtained by inverting the full joint Fisher matrix of the whole population, which in this case can be computed directly given the simplicity of the source catalogue. The priors are uniform and span around the injection, picked wide enough that no posterior is truncated by them.
Vacuum Injection
We first consider the case where no deviation is present, so that . The question is then what upper bound each strategy gives, and whether the null ratio predicts the difference between them. Since no effect is in the data, is the correct expansion point and the caveats discussed at the end of § 7.3 do not apply. From the direct Fisher matrix, the total masking evaluates to , which for is also . The posterior estimator of Eq. (77), using effective samples of an 84-dimensional vacuum global fit of the same data is in good agreement, giving , where the uncertainty is the Monte-Carlo error of Eq. (78).
We then ran the two strategies. The global fit, with free together with all 20 nuisance sources for a total of 85 parameters, gives , to be compared with the Fisher prediction . Target-only post-processing, that is a vacuum global fit followed by a refit of the target alone with free, gives , matching . The corresponding posteriors are shown in the left panel of Figure 2, and the full 2-D marginals for the target source are given in Appendix § 12. Both are unbiased, as they must be when no extra effect is present in the data and there is no noise, but the ratio of the reported uncertainties is 0.74, in good agreement with the predicted .
Figure 2. Marginal MCMC posteriors of the beyond-vacuum parameter under the different strategies: simple post-processing, in which all nuisance sources are subtracted and only the target is refitted; the down-scaled fits in which group A, group B, or both are refitted jointly with the target source; and the full global fit. Left: vacuum dataset, . All posteriors are centred on the truth (vertical dashed line), and only their widths differ. Post-processing is the narrowest, so the upper bound it implies is tighter than the data support, and freeing each group widens the posterior by the of Table 1, until with both groups freed it is almost indistinguishable from the global fit. Right: with a beyond-vacuum effect present, . Post-processing is now shifted towards vacuum as well, by the factor of Eq. (47), so that it recovers a weaker effect at lower significance, in this case compatible with the null hypothesis. Freeing the groups moves the posterior back towards the truth, but note that here it is group B that recovers most of the effect, whereas in the vacuum case it was group A.
We then ranked the nuisance sources with the algorithm of § 7.3, and the results are collected in Table 1. All 16 crossing sources are individually harmless, with regardless of their SNR, although collectively they still reach . The four nearly parallel sources are more problematic, with (group A) and (group B). Group B also illustrates the joint masking discussed in § 7.3. Its two sources are strongly correlated with each other, and although each is individually almost harmless (), together they absorb almost twice as much as the two separately. These subset ratios predict how much of the masking each down-scaled fit recovers. Freeing group A together with the target in the refit should raise the residual ratio from to , while freeing group B to 0.82, and freeing both to 0.96. We checked this by repeating the second step of post-processing in each case, shown also in the left panel of Figure 2 (and in Appendix § 12 for the marginal 2D posteriors). The resulting 95% bounds are with group A freed, with group B, and with both, against for target-only post-processing and for the global fit. The bounds therefore go towards the global-fit value in the order the ratios predict, with the posterior widths following the expected inflation (to within 5%), and the refit with both groups freed is almost indistinguishable from the global fit. We highlight that in this case it is group A that carries most of the masking compared to group B. As we show next, this ordering is inverted once a strong beyond-vacuum GR effect is actually present.
| SNR | |||||
| group A | 0.85 | ||||
| source 1 | 0.97 | 12 | 0.87 | ||
| source 2 | 0.94 | 25 | 0.98 | ||
| group B | 0.96 | ||||
| source 3 | 0.88 | 18 | 0.99 | ||
| source 4 | 0.85 | 9 | 0.99 | ||
| crossers (16 sources) | — | 2.4–9.9 | 9–38 | 0.97 | |
| groups A B | 0.81 | ||||
| all 20 sources (= ) | 0.78 |
Table 1. Ranking of the nuisance catalogue at the null point (). is the initial frequency offset from the target and the chirp-rate ratio, is the individual masking ratio of each source, and the collective ratio of the group it belongs to. The four nearly parallel sources dominate, with no correlation with SNR, while the 16 crossers are individually harmless () but reach together. The last row is the full catalogue, recovering the total .
Beyond-vacuum GR injection
We now inject , which corresponds to 2.8 radian of accumulated dephasing relative to vacuum over the full observation window. The global fit recovers , containing the truth and excluding vacuum at , while post-processing recovers . biased towards vacuum, excluding the true value at , and inconsistent with vacuum at only . This is the significance loss of Eq. (68), which predicts . Evaluated at , the masking ratio is , so post-processing should recover , compatible with the measured . Freeing the different subsets of nuisance sources in the second inference step brings the recovered value back towards the truth according to the residual ratio of Eq. (82). Table 2 compares the predictions with the MCMC results for the different strategies, and the right panel of Figure 2 shows the corresponding posteriors for . Every step agrees with the prediction to within a small fraction of the posterior width. Note that freeing only group B already almost recovers the global-fit result, while freeing group A alone changes little, the opposite of what the ranking of Table 1 suggested.
| subset | predicted | measured | |||
| post-processing () | 1.00 | 1.00 | 0.66 | 0.043 | |
| + group A | 0.85 | 0.94 | 0.70 | 0.049 | |
| + group B | 0.96 | 0.71 | 0.93 | 0.087 | |
| + groups A B | 0.81 | 0.69 | 0.96 | 0.091 | |
| global fit () | 0.78 | 0.66 | 1.00 | 0.100 |
Table 2. Down-scaling ranking for the toy model at . Each row corresponds to freeing the subset of Table 1 together with the target source in the second refit step of post-processing, from simple post-processing () to the full global fit. The subset ratio is quoted at both expansion points, taken from Table 1 and used here. The residual masking follows from Eq. (82) and the predicted from Eq. (83). The measured values are posterior means and standard deviations of the corresponding MCMC analyses. Groups A and B exchange roles between the two expansion points in terms of which is the stronger masker.
The bias in the post-processing strategies is not limited to . Figure 5 in Appendix § 13 shows the full 2D-posteriors of the target source for each strategy. In target-only post-processing, the target’s own chirp rate is displaced as well, by , a bias more than its own posterior width, along the – degeneracy visible in the corresponding panel. The initial frequency follows along the same degeneracy, by about two of its own widths, while the amplitude moves by less than one. This is expected, as the part of the effect absorbed by the nuisance sources is subtracted from the residual, and the second fit compensates for what is missing by moving whichever vacuum parameters of the target can mimic it. Freeing group B removes almost all of these displacements at once, whereas freeing group A leaves and essentially where simple post-processing left them.
Figure 5. Same as Figure 4, for the beyond-vacuum GR injection . The post-processing recovery is biased from the truth along the –– degeneracy, so the target’s vacuum parameters are biased together with . Freeing group B in the second step of post-processing (or both groups) recovers almost the global-fit posterior, while freeing group A alone leaves it essentially unchanged.
Figure 4. Vacuum injection: marginalised posterior of the target source’s parameters for the five analyses of § 8.1. Contours enclose 68% and 95% of the posterior probability, and the injected values are marked by the dashed lines.
The nuisance sources are affected in the same way, as shown in Figure 6 and Figure 7 of Appendix § 13. In target-only post-processing their parameters are never refitted. They stay frozen at the values returned by the vacuum global fit, which absorbed part of the beyond-vacuum GR effect and are biased by Eq. (38). In the first step, the chirp rates of the two sources of group B are displaced by and respectively, the parameters of group A by at most , and those of the 16 crossing sources by at most , with a mean of . The sources that do the masking are therefore the same ones whose parameters are most biased by it, and the ranking of § 7.3 is able to identify them in this toy model.
Figure 6. Same as Figure 5 for the nuisance sources in group A, plotted as displacements from the injected values. Once these sources are refitted alongside the target they are recovered on their injected values, to within , with uncertainties around 10–15% tighter than the global-fit ones because the remaining nuisance sources stay fixed in the refit.
Figure 7. Same as Figure 6 for the nuisance sources in group B. The chirp rate of source 4, the faintest of the four parallel sources in the catalogue, at SNR , has a skewed posterior, which does not appear in the down-scaled post-processing refits.
This example also illustrates the role of the expansion point. The ranking obtained at the null point in § 8.1 identified group A as the dominant masker ( vs ), whereas evaluated at the two groups exchange roles ( vs ), and the MCMC results agree with it. The same non-linearity is visible in the posteriors which are not completely Gaussian. Nonetheless, the linear-signal predictions of Table 2 recover the measured means to within a fraction of the posterior width, so the approximation is adequate for ranking and selecting sources even where it is no longer exact for the shape of the posterior.
Section 7.3 gives two ways of dealing with the unknown expansion point, and here both lead to the same selection. Sweeping the expansion point over the range of allowed by the vacuum-injection analysis flags both groups, and the union gives both at and at . Updating the expansion point iteratively instead also illustrates the dangerous scenario in which the null-point ranking suggests that few sources are needed. Taking as the threshold, at the null point refitting group A alone already gives , and the corresponding down-scaled fit recovers , still compatible with vacuum at . Re-evaluating the ratios at this value, however, group A alone leaves only , while adding group B gives . At the value recovered with both groups the selected set no longer changes (), so the iteration converges after two refits, to the same set as the sweep strategy. With 16 nuisance parameters refitted instead of 80, this down-scaled fit recovers , within of the full global fit.
Averaged post-processing
Finally, we validate the averaged post-processing strategy of § 5.4 on the same two datasets, for the target-only refit and for the down-scaled fit with groups A and B freed. In each case, we draw 128 nuisance configurations from the first-step vacuum posterior chains, subtract each, rerun the second inference step, and concatenate the resulting chains, for 512 per-draw analyses in total. The averaged posteriors are shown in Figure 3, and the ratios between their widths and the simple post-processing ones follow Eq. (72) to for both the vacuum and the beyond-vacuum injections. For the target-only refit, averaging widens the posterior from to 0.027 in the vacuum injection (predicted 0.026) and from 0.021 to 0.025 in the beyond-vacuum one (predicted 0.026), while with both groups freed the averaged width is within of the global-fit one in both cases. The two consequences anticipated in § 6 are visible directly. In the vacuum injection, the bound becomes better calibrated, and the more nuisance sources are freed the less there is left to correct. The width difference with respect to the global fit falls from 26% to 11% for the target-only refit and from 4% to 2% with groups A and B freed. In the beyond-vacuum injection the significance of the detection degrades, with the recovered deviation dropping from to , slightly above the predicted because the averaged mean sits above the best-fit-subtraction one, which we attribute to a small departure from Gaussianity of the first-step posterior.
Figure 3. The averaged post-processing strategy of § 5.4 in the toy model, for the target-only refit () and the down-scaled fit with groups A and B freed, against the global fit, with the same colour scheme as in Figure 2. Solid curves are the averaged posteriors, obtained by drawing 128 nuisance configurations per analysis from the first-step vacuum posterior, rerunning the second step for each draw, and concatenating the chains. Dashed curves are the corresponding simple (best-fit-subtraction) posteriors of Figure 2. Left: vacuum injection. Averaging widens each posterior towards the global-fit width by , Eq. (72) with , so the reported bound becomes better calibrated, and the more sources are freed the smaller the correction. Right: beyond-vacuum injection. The means are essentially unchanged, Eq. (55), so averaging does not undo the masking of Eq. (47). The recovered value stays at while the detection significance drops by , Eq. (72).
Discussion
We have presented a criterion to decide which overlapping signals must be reanalysed jointly when one wants to test a different model for some target source than the one used in an initial global fit so that any extra effect in the new model is not absorbed and masked by other nuisance signals. This criterion is obtained by computing the overlap between the directions that a group of nuisance sources spans in data space and the direction of an extra effect absent from the first model. The absorption is captured by two numbers. The masking ratio of § 6 is the factor multiplying the detection significance when an effect is present, so that a deviation the global fit would establish at is reported by post-processing at . The null ratio quantifies the miscalibration when no effect is present. It can be interpreted as the factor by which the posterior shrinks, so the bounds that post-processing reports are tighter than those from the full global fit by . For an effect of unknown size, the loss lies between the two, so the smaller of and serves as a single, conservative measure. For a single beyond-vacuum parameter, the two coincide. Both can be computed by reusing the posterior samples of the initial vacuum global fit. We also construct an algorithm that ranks the nuisance sources by how strongly they mask a specific beyond-vacuum effect in a given target source. The rationale behind this is to keep the smallest subset that reproduces the inference of the full global fit, and to refit only those sources jointly with the target in a down-scaled reanalysis.
The linear-signal predictions for the uncertainties, the ratios, the recovered values and the down-scaled bounds were all confirmed by full MCMC analyses on a toy model that included one target source with a possible deviation and 20 nuisance sources.
Our methods are directly applicable to LISA data-analysis pipelines. The LISA Science Ground Segment currently plans to release catalogues of sources characterised with a global fit using vacuum GR only. The method described here could be used to assess for which sources and which types of beyond-vacuum GR deviations robust constraints can be obtained by post-processing only the target source, or the target source together with a small number of nuisance sources. We only need the overlaps of Eq. (74) between the beyond-vacuum direction of the target and the templates at the posterior samples of the vacuum global fit. For a set of standard beyond-vacuum (e.g. scalar-tensor theories, environmental effects), these could be computed during the first inference and released alongside the catalogue, and for any other model they can be computed from the released posterior samples, so that any user can compute the masking of their own beyond-vacuum model without rerunning the global fit. It will be possible to test this using forthcoming mock datasets. The LISA Distributed Data Processing Centre (DDPC) has generated a Mojito dataset, which contains the major LISA source classes together with non-stationary noise. This dataset will be analysed with the global-fit pipelines currently under development by the DDPC and the NASA Ground Segment, and the resulting posteriors will be released. The Mojito dataset contains no beyond-vacuum GR physics, so an analysis of such effects falls precisely into our null case, and the quantities to compute are and the per-source rankings. Once the vacuum pipelines are robust and their results available, our estimator can be applied to assess whether post-processing is a viable strategy for the scientific interpretation of the public data products, for which sources a down-scaled reanalysis is required instead, and under what circumstances the only option is to redo the entire global fit. Through this experience, and the analysis of future common datasets, we will learn how to obtain robust constraints on beyond-vacuum GR effects in a computationally efficient way from the real LISA data.
Although our toy model did not include systematic errors, the formalism accounts for them. Waveform model error enters both strategies through the same effective parameter , displacing what both tests measure and driving the false-positive term of Eq. (22). The condition for a threshold to remain meaningful, , is then a requirement not on the overall accuracy of the waveform model but on its projection onto the beyond-vacuum direction. This is an important topic to explore further, but will be sensitive to the assumed waveform error model, and in particular its projection onto the beyond-vacuum-GR space, and so it must be done carefully.
One of the main limitations of our method is the use of the linear-signal approximation [37], which enters both in approximating the vacuum manifold by its tangent plane and in characterising the first fit by a point estimate at the maximum of the likelihood. Curvature, priors, and subtraction at a posterior point estimate rather than at a maximum all degrade the exact relation between the two strategies, and we already saw in § 8.2 that the ranking itself depends on the expansion point once the effect is large enough. Sweeping the expansion point, or updating it iteratively at the value of the effect recovered by each refit accounts for this, and in the toy model both recover the right sources even though the null-point ranking alone points to the wrong group. It is also important to note that the final reanalysis itself does not rely on the linear-signal approximation, which only enters the selection of the sources. If the approximation failed to identify an important nuisance source, this could be compensated by a more stringent threshold. In addition, estimating the masking ratio from the posterior samples themselves, as described in § 7.1, should make it robust to some deviations from the linear-signal approximation. This is one of the points that can be tested using the DDPC common datasets.
Another aspect to explore more carefully with the common datasets is when and how to average in post-processing, i.e. to average over samples from the first fit instead of subtracting its best fit. We discussed this strategy in § 5.4 and § 6, where we showed that it reduces the significance of a detection by at most a factor of (6.12), while bringing the upper limits of non-detections closer to those of the global fit (6.14). This comes at the price of increased computational cost (one refit per posterior draw). However, it may have the advantage of increased robustness against the modelling approximations we were just discussing, and of better-calibrated uncertainties on the vacuum parameters, whose means it leaves unchanged at linear order. Using the common dataset analyses, it will be possible to compare the simple and averaged strategies in a realistic setting, and hence identify the right approach to take with real data.
Our toy model included only 20 nuisance sources in total, whereas the LISA data will contain individually resolvable Galactic binaries. We expect that a small number of sources will still dominate the masking of any given target, in which case the greedy ranking of § 7.3 will identify them for a refit. It could also be that no source dominates, but the cumulative effect of many still leads to an appreciable masking, as the 16 crossing sources of our toy model illustrate in a mild form. This would not go unnoticed, since the total masking estimated with Eq. (77) includes all sources, and the greedy algorithm would still terminate, but the selected set could become large, making the down-scaled reanalysis expensive. This is another point that should be tested using the DDPC common datasets.
Finally, we also assumed stationary Gaussian noise with a known spectral density, whereas in practice the noise will have to be inferred jointly with the sources. Formally, the noise parameters then act as one more nuisance block, but we expect them to matter less for the masking, because they couple in a different way to the residual. A biased nuisance source is subtracted, so its imprint is a waveform-like piece of the residual that the second step projects onto the beyond-vacuum direction, linearly in . Instead, a biased noise estimate cannot be subtracted as it enters only through the weighting of the inner product, and at linear-signal order the expected Fisher cross-terms between waveform and noise-model parameters vanish at the true parameters, because they are proportional to the mean of the residual. What the noise sector does see is the power of the residual, including that of the imprint left by the biased subtraction, and this enters only at second order in the residual. It increases the inferred spectral density in the frequency band where the residual lives, which does not displace along a degeneracy but down-weights that band, lowering the significance of every source there, including that of the effect itself. If the inflation were frequency independent, all Fisher matrices would scale by the same factor and the masking ratio would be unchanged. For a residual localised in frequency, the ratios have to be recomputed with the inflated spectral density, although the fractional inflation is set by the residual’s signal-to-noise ratio per frequency bin, and is therefore small for effects that are not too strong. The population of unresolved sources below a certain threshold can also be included in the analysis in the same way, either as an effective bias or as extra confusion noise [35]. Again, this is one more aspect we can study in more detail in realistic mock datasets to ensure that by the time LISA flies, we have a robust inference strategy to do precision tests of General Relativity.
Acknowledgments
We thank the organisers and participants of the meetings on EMRI Search and Inference within the LISA Global Fit - Part I (Paris) & Part II (Singapore) for very fruitful discussions that helped construct and improve the ideas present in this manuscript. These meetings were supported by Université Paris Cité - National University of Singapore (Paris-NUS) research grant 2024-01-RESEARCH/UPCité-NUS (IdEx code ANR-18-IDEX-0001). We also thank Arnab Dhani for insightful comments that helped improved this manuscript. The work of F.D. is supported by the European Research Council (ERC) Advanced Grant De Tenebris (grant agreement No. 101198798), funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Council Executive Agency. Neither the European Union nor the granting authority can be held responsible for them. J. G. acknowledges support from the Simons Foundation International via Grant No. SFI-MPS- BH-00012593-06.
The masking inequality
We prove here that the two terms of Eq. (58) are separately non-negative, so that and , and that the weights of Eq. (51) are real and lie in .
Let us write and , both symmetric, and assume . We only need to use that a symmetric positive-definite matrix has a symmetric, invertible square root , and that for a positive semi-definite matrix , the matrix is also positive semi-definite, since . First, by Eq. (18), , so for any vector
because is itself a conditional Fisher matrix and hence positive definite. Then, consider the symmetric, positive-definite matrix
If with , then taking ,
by Eq. A.1 and . Moreover, is similar to and to , and similar matrices have the same eigenvalues, so the weights of Eq. (51) are real and between . The trace is also invariant for similar matrices
which is the second term of Eq. (58) and gives . Then
The matrix is symmetric with eigenvalues , so positive semi-definite, and so is the right-hand side of the equation above. Evaluated at , this is the first term of Eq. (58). Explicitly, writing and for the orthonormal eigenvectors of ,
so . Since is a weighted average of the , we also get , as stated in Eq. (64).
Vacuum injection: 2D posteriors
We show here the marginalised 2D-posterior over the full parameter space of the target source of interest for the vacuum injection, .
Beyond-vacuum GR injection: 2D posteriors
The same posterior as in Appendix Appendix § 12 for the beyond-vacuum GR injection, , and also for the groups A and B of the nuisance sources.
References
- [1]LIGO Scientific, Virgo, KAGRA collaboration, GWTC-5.0: An Introduction to Version 5.0 of the Gravitational-Wave Transient Catalog, 2605.27223.
- [2]LIGO Scientific, Virgo, KAGRA collaboration, GWTC-5.0: Observations from the Second Part of the Fourth LIGO-Virgo-KAGRA Observing Run and Updates to the Gravitational-Wave Transient Catalog, 2605.27225.
- [3]ET collaboration, The Science of the Einstein Telescope, JCAP 03 (2026) 081 [2503.12263].
- [4]D. Reitze et al., Cosmic Explorer: The U.S. Contribution to Gravitational-Wave Astronomy beyond LIGO, Bull. Am. Astron. Soc. 51 (2019) 035 [1907.04833].arxiv.org/abs/1907.04833
- [5]F. Iacovelli, M. Mancarella, S. Foffa and M. Maggiore, Forecasting the Detection Capabilities of Third-generation Gravitational-wave Detectors Using GWFAST, Astrophys. J. 941 (2022) 208 [2207.02771].DOI
- [6]J. Janquart, T. Baka, A. Samajdar, T. Dietrich and C. Van Den Broeck, Analyses of overlapping gravitational wave signals using hierarchical subtraction and joint parameter estimation, Mon. Not. Roy. Astron. Soc. 523 (2023) 1699 [2211.01304].DOI
- [7]A. Samajdar, J. Janquart, C. Van Den Broeck and T. Dietrich, Biases in parameter estimation from overlapping gravitational-wave signals in the third-generation detector era, Phys. Rev. D 104 (2021) 044003 [2102.07544].DOI
- [8]LISA collaboration, LISA Definition Study Report, 2402.07571.
- [9]A. Klein et al., Science with the space-based interferometer eLISA: Supermassive black hole binaries, Phys. Rev. D 93 (2016) 024003 [1511.05581].DOI
- [10]S. Babak, J. Gair, A. Sesana, E. Barausse, C.F. Sopuerta, C.P.L. Berry et al., Science with the space-based interferometer LISA. V: Extreme mass-ratio inspirals, Phys. Rev. D 95 (2017) 103012 [1703.09722].DOI
- [11]L. Speri, F. Duque, S. Barsanti, A. Santini, S. Kejriwal, O. Burke et al., Quantifying the Scientific Potential of Intermediate and Extreme Mass Ratio Inspirals with the Laser Interferometer Space Antenna, 2603.17072.
- [12]A. Klein et al., The last three years: multiband gravitational-wave observations of stellar-mass binary black holes, 2204.03423.arxiv.org/abs/2204.03423
- [13]S. Babak, C. Caprini, D.G. Figueroa, N. Karnesis, P. Marcoccia, G. Nardini et al., Stochastic gravitational wave background from stellar origin binary black holes in LISA, JCAP 08 (2023) 034 [2304.06368].arxiv.org/abs/2304.06368
- [14]F. Pozzoli, R. Buscicchio, A. Klein, V. Korol, A. Sesana and F. Haardt, Cyclostationary signals in LISA: A practical application to Milky Way satellites, Phys. Rev. D 111 (2025) 063005 [2410.08274].DOI
- [15]F. Pozzoli, S. Babak, A. Sesana, M. Bonetti and N. Karnesis, Computation of stochastic background from extreme-mass-ratio inspiral populations for LISA, Phys. Rev. D 108 (2023) 103039 [2302.07043].arxiv.org/abs/2302.07043
- [16]LISA Cosmology Working Group collaboration, Cosmology with the Laser Interferometer Space Antenna, Living Rev. Rel. 26 (2023) 5 [2204.05434].
- [17]A. Boumerdassi, M.C. Edwards, A. Vajpeyi and O. Burke, First-time assessment of glitch-induced bias and uncertainty in inference of extreme mass ratio inspirals, Phys. Rev. D 113 (2026) 104044 [2512.16322].DOI
- [18]O. Burke, M. Muratore and G. Woan, Impact of missing data on the construction of LISA time-delay-interferometry Michelson variables, Phys. Rev. Applied 25 (2026) 034041 [2510.06406].arxiv.org/abs/2510.06406
- [19]M. Muratore, J. Gair, O. Hartwig, M.L. Katz and A. Toubiana, Pipeline for searching and fitting instrumental glitches in LISA data, Phys. Rev. D 112 (2025) 063041 [2505.19870].
- [20]O. Burke, S. Marsat, J.R. Gair and M.L. Katz, Addressing data gaps and assessing noise mismodelling in LISA, Phys. Rev. D 111 (2025) 124053 [2502.17426].arxiv.org/abs/2502.17426
- [21]M.L. Katz, N. Karnesis, N. Korsakova, J.R. Gair and N. Stergioulas, Efficient GPU-accelerated multisource global fit pipeline for LISA data analysis, Phys. Rev. D 111 (2025) 024060 [2405.04690].arxiv.org/abs/2405.04690
- [22]T.B. Littenberg and N.J. Cornish, Prototype global analysis of LISA data with multiple source types, Phys. Rev. D 107 (2023) 063004 [2301.03673].DOI
- [23]S. Deng, S. Babak, M. Le Jeune, S. Marsat, E. Plagnol and A. Sartirana, Modular global-fit pipeline for lisa data analysis, Phys. Rev. D 111 (2025) 103014.arxiv.org/abs/2501.10277
- [24]S.H. Strub, L. Ferraioli, C. Schmelzbach, S.C. Stähler and D. Giardini, Global analysis of LISA data with Galactic binaries and massive black hole binaries, Phys. Rev. D 110 (2024) 024005 [2403.15318].arxiv.org/abs/2403.15318
- [25]J. Alvey, U. Bhardwaj, S. Nissanke and C. Weniger, What to do when things get crowded? Scalable joint analysis of overlapping gravitational wave signals, 2308.06318.arxiv.org/abs/2308.06318
- [26]Mock LISA Data Challenge Task Force collaboration, The Mock LISA Data Challenges: From Challenge 3 to Challenge 4, Class. Quant. Grav. 27 (2010) 084009 [0912.0548].
- [27]European Space Agency, “LISA Science Implementation Requirements Document.” Internal document, private communication.
- [28]E. Barausse et al., Prospects for Fundamental Physics with LISA, Gen. Rel. Grav. 52 (2020) 81 [2001.09793].
- [29]LISA collaboration, New horizons for fundamental physics with LISA, Living Rev. Rel. 25 (2022) 4 [2205.01597].
- [30]LISA Consortium Waveform Working Group collaboration, Waveform modelling for the Laser Interferometer Space Antenna, Living Rev. Rel. 28 (2025) 9 [2311.01300].
- [31]N. Yunes, B. Kocsis, A. Loeb and Z. Haiman, Imprint of Accretion Disk-Induced Migration on Gravitational Waves from Extreme Mass Ratio Inspirals, Phys. Rev. Lett. 107 (2011) 171103 [1103.4609].DOI
- [32]E. Barausse, W. Cardoso and P. Pani, Can environmental effects spoil precision gravitational-wave astrophysics?, Phys. Rev. D 89 (2014) 104059 [1404.7149].DOI
- [33]S. Kejriwal, F. Duque, A.J.K. Chua and J. Gair, Bias-corrected importance sampling for inferring beyond-vacuum-GR effects in gravitational-wave sources, Phys. Rev. D 112 (2025) 024005 [2503.01120].
- [34]L. Copparoni, R.S. Chandramouli and E. Barausse, When Vacuum Breaks: A Self-Consistency Test for Astrophysical Environments in Extreme Mass Ratio Inspirals, Phys. Rev. Lett. 137 (2026) 021405 [2510.06948].DOI
- [35]A. Antonelli, O. Burke and J.R. Gair, Noisy neighbours: inference biases from overlapping gravitational-wave signals, Mon. Not. Roy. Astron. Soc. 507 (2021) 5069 [2104.01897].DOI
- [36]C. Cutler and M. Vallisneri, LISA detections of massive black hole inspirals: Parameter extraction errors due to inaccurate template waveforms, Phys. Rev. D 76 (2007) 104018 [0707.2982].DOI
- [37]M. Vallisneri, Use and abuse of the Fisher information matrix in the assessment of gravitational-wave parameter-estimation prospects, Phys. Rev. D 77 (2008) 042001 [gr-qc/0703086].DOI
- [38]J.M. Dickey, The weighted likelihood ratio, linear hypotheses on normal location parameters, The Annals of Mathematical Statistics 42 (1971) 204.DOI
- [39]N.J. Cornish and T.B. Littenberg, Tests of Bayesian Model Selection Techniques for Gravitational Wave Astronomy, Phys. Rev. D 76 (2007) 083006 [0704.1808].DOI
- [40]L. Speri, S. Barsanti, A. Maselli, T.P. Sotiriou, N. Warburton, M. van de Meent et al., Probing fundamental physics with extreme mass ratio inspirals: Full Bayesian inference for scalar charge, Phys. Rev. D 113 (2026) 023036 [2406.07607].DOI
- [41]G. Lara et al., Signatures from metastable oppositely-charged black hole binaries in scalar Gauss-Bonnet gravity, 2505.14785.
- [42]F. Fantoccoli, F. Duque and J. Gair, Gotta light? Illuminating AGN disks with LISA EMRIs, 2606.12531.
- [43]N. Yunes and F. Pretorius, Fundamental Theoretical Bias in Gravitational Wave Astrophysics and the Parameterized Post-Einsteinian Framework, Phys. Rev. D 80 (2009) 122003 [0909.3328].DOI
- [44]A.K. Mehta, A. Buonanno, R. Cotesta, A. Ghosh, N. Sennett and J. Steinhoff, Tests of general relativity with gravitational-wave observations using a flexible theory-independent method, Phys. Rev. D 107 (2023) 044020 [2203.13937].DOI
- [45]N. Karnesis, M.L. Katz, N. Korsakova, J.R. Gair and N. Stergioulas, Eryn: a multipurpose sampler for Bayesian inference, Mon. Not. Roy. Astron. Soc. 526 (2023) 4814 [2303.02164].DOI






