When a gravitational-wave template leaves out a piece of physics, louder signals make the parameter estimates worse, not better. The inference engine cannot fit the missing effect honestly, so it shifts the masses and spins it does have until the residual shrinks. The result is a tight posterior around a wrong answer. We propose to stop the template from being asked about the part of the signal it cannot model.
Concretely, we identify the small family of phase distortions that the omitted physics can produce. For truncated post-Newtonian terms or missing conservative self-force, this family is a low-dimensional span of known frequency power laws, or a learned span where those power laws break down. We then compare template and data only after projecting that family out. Parameters are then estimated from the information the model can actually account for. Their error keeps shrinking with signal-to-noise ratio (SNR) at a slower, honest rate, instead of hitting a floor of bias.
The same construction explains why the "butterfly" atom interferometer is blind to gravitational acceleration. That gives an exact, independent check that the method recovers a known design.
Core Claim
The post-Newtonian (PN) approximation expands the dynamics of a binary in powers of orbital velocity over light speed. Its errors are not arbitrary. A missing term at a given PN order changes the frequency-domain phase by a specific power of frequency, possibly times a logarithm. The derivatives of the phase with respect to the physical parameters (chirp mass, mass ratio, spins) are built from the same power-law basis.
This shared structure turns the vague question "can we be robust to unmodeled physics?" into an exact finite linear-algebra question. The question is whether the discrepancy span intersects the span of the parameter derivatives, and at what angle. If they are separated, projecting out the discrepancy span removes the bias to first order and costs a computable amount of information. If a parameter's derivative lies inside the discrepancy span, that parameter is provably unrecoverable without the missing physics, and the method says so instead of reporting a confident wrong number.
Write $\langle a, b\rangle$ for the usual noise-weighted inner product used in matched filtering. Write $D = \mathrm{span}\{d_1,\dots,d_r\}$ for the discrepancy subspace of phase or waveform perturbations, and $P_D^\perp$ for the projector onto its orthogonal complement under that inner product. The projected log-likelihood for data $s$ and template $h(\theta)$ is
$$\ln L_\perp(\theta) = -\tfrac12 \langle P_D^\perp (s - h(\theta)),\, P_D^\perp (s - h(\theta))\rangle .$$
In the linear-signal approximation of Cutler and Vallisneri, the bias $\Delta\theta$ that results from a true discrepancy $\delta$ is
$$\Delta\theta = \Gamma_\perp^{-1}\, \langle \partial_\theta h,\, P_D^\perp \delta\rangle ,$$
where $\Gamma_\perp$ is the Fisher matrix computed with projected derivatives. The bias vanishes when $\delta \in D$. It is bounded by the size of the part of $\delta$ that falls outside $D$, divided by the smallest principal angle between $D$ and the parameter tangent space. Because the bias is independent of SNR while statistical error falls as 1/SNR, the prediction is sharp. The projected estimator should keep roughly $\mathrm{SNR}^{-1}$ scaling with an inflated prefactor, up to the SNR at which the out-of-span residual of $\delta$ becomes visible.
Where the pieces come from
The problem is documented most clearly for extreme and intermediate mass-ratio inspirals, where a small body spirals into a supermassive black hole over hundreds of thousands of cycles.
Influence of conservative corrections on parameter estimation for extreme-mass-ratio inspirals showed that omitting 2PN conservative corrections in a numerical kludge model costs roughly one cycle of phase every 60 days. It also found the ratio of model error to statistical error below 1 at SNR 30. Importance of including small body spin effects in the modelling of extreme and intermediate mass-ratio inspirals found that above mass ratio about $10^{-4}$, omitting small-body spin or conservative terms makes systematic errors several times the statistical ones. The central black hole's mass and spin absorb the missing phase. Importance of including small body spin effects in the modelling of intermediate mass-ratio inspirals. II removed the bias at SNR 1000 by adding the missing 2.5PN spin and conservative terms.
These papers motivate the proposal: bias is real and grows with SNR. They also supply the testbed: kludge models with and without the missing terms, giving a truth/template pair with known discrepancy. Their fix, however, requires already knowing the missing terms, which is exactly what is unavailable for generic eccentric, precessing orbits.
The mechanism comes from Choosing observation operators to mitigate model error in Bayesian inverse problems. That paper shows that if the model discrepancy lies in the null space of the linear map from state to data, it becomes invisible to the posterior. It bounds the resulting posterior error, and it demonstrates the idea on an advection-diffusion-reaction equation. Its stated open problem is the trade-off in how large that null space should be: too large and it swallows the signal. That paper never faced a problem where the discrepancy and the parameters share a structured basis. Gravitational waveforms have that structure, which is what makes the trade-off computable here.
The Limitations of Model-Based Experimental Design and Parameter Estimation in Sloppy Systems acts as a constraint on the design. It showed that designing experiments to pin down poorly constrained parameters can expose model discrepancy and destroy predictive power. Our method moves in the opposite direction: it deliberately discards the directions where the model is wrong.
The algebraic-topology direction enters as a methodological lens, not as a mechanism, and we want to be explicit about that. The homotopy classification results in Homotopy classification of 4-manifolds with finite abelian 2-generator fundamental groups and Homotopy classification of 4–manifolds whosefundamental group is dihedral have the same logical shape as our question. A reduced invariant is complete exactly when a specific obstruction (torsion in a derived module) vanishes. Whether it vanishes is settled by exact computation, as in the SageMath check over all groups up to order 16, rather than by numerical approximation. Equivariant K-homology for hyperbolic reflection groups makes the same point: exactness of reduction can be read from the integer structure of boundary matrices.
We borrow this discipline, not the topology. In the weak-field regime, PN phase coefficients are rational functions of the mass ratio multiplying fixed rational powers of frequency. So "does parameter $\theta_i$ survive projection?" has an exact rank answer over the rationals, which we compute symbolically before any noisy numerics. We make no claim that torsion or unimodularity appears in waveform inference.
The generalized clock direction supplies an independent correctness test. Universality-of-clock-rates test using atom interferometry with $T^{3}$ scaling designs a four-pulse "butterfly" sequence (π/2–π–π–π/2) whose closed path cancels the dependence on gravitational acceleration and on initial position and velocity, while keeping a proper-time signal that grows as the cube of the interrogation time. This is exactly a measurement whose null space contains the nuisance terms. Generalized gravity-gradient mitigation scheme closes the path against gravity gradients using computed momentum kicks. It reports that errors above 10% in the assumed local gradient spoil the mitigation.
Neither paper formalizes its cancellation as null-space design. Doing so gives us a system where the right answer is known analytically.
Research Program
We will build the projected likelihood in three stages of increasing difficulty, each tied to one regime of the PN approximation.
In the weak-field stage, the discrepancy subspace is analytic. It is spanned by the frequency powers and logarithms of the PN orders at and beyond the template's truncation. We compute, exactly and symbolically, which physical parameter derivatives lie in or near that span. This produces an identifiability table: which parameters remain recoverable when a given set of PN orders is treated as untrustworthy, and at what information cost. We expect, for instance, that treating the 2PN order as untrustworthy will sharply degrade spin–spin information, because that information lives at the same order. The table makes such losses explicit before any inference is run.
Near the innermost stable circular orbit, PN series stop converging, so the analytic basis no longer spans the discrepancy. In this strong-field stage we learn the subspace instead. We take differences between kludge waveforms with and without conservative self-force and small-body spin terms over a grid of source parameters. We take their dominant singular vectors, and choose the rank $r$ by minimizing the bias bound plus the variance cost from the projected Fisher matrix. The number of dimensions needed, plotted against the final orbital velocity, is itself a new measurement: it quantifies from the inference side how fast PN structure dissolves as the orbit tightens.
In the transfer stage we ask whether a subspace learned from one kind of omitted physics also neutralizes a different one it was not built from. Examples include truncated PN order versus missing conservative self-force, or circular-orbit discrepancies applied to mildly eccentric truth. This is where the method is most likely to fail, and we will report it as such.
In parallel, and cheaply, we write the atom-interferometer phase as a linear functional of the perturbation coefficients: acceleration, initial position and velocity, gravity-gradient terms, and the proper-time coupling. We then run the same subspace-selection procedure with pulse timings and momentum kicks as design variables. This uses the perturbative phase expansions of Perturbative operator approach to high-precision light-pulse atom interferometry as the forward model.
Baselines
The do-nothing reference is standard matched filtering with the incomplete template, following the Cutler–Vallisneri bias analysis used by Importance of including small body spin effects in the modelling of extreme and intermediate mass-ratio inspirals. It is run on identical injections and noise realizations. It is the reference for how badly bias grows with SNR, and beating it means lower bias-to-error ratio and correct coverage at SNR at and above 1000.
The oracle is the full model that contains the missing terms, as in Importance of including small body spin effects in the modelling of intermediate mass-ratio inspirals. II. That paper reports systematic error pushed below statistical error at SNR 1000 for $5\times10^3$ and $10^4$ solar-mass bodies into a $10^6$ solar-mass black hole. We do not expect to beat the oracle. We aim to approach its coverage while paying a stated variance cost, without using its missing terms.
The strongest existing method for honest error bars is waveform-calibration marginalization, as in SEOBNRv4CE from Incorporating waveform calibration error in gravitational-wave modeling and inference for SEOBNRv4. That paper trains a Gaussian process on differences against 141 numerical-relativity simulations and marginalizes over 20 amplitude and phase parameters at 10 frequency nodes. It reports three numbers: chirp-mass bias reduced by a factor of 1.5 or more for spinning systems, posterior width scaling as $\mathrm{SNR}^{-0.75}$, and roughly a 2× slowdown. We run it two ways. On its own comparable-mass setting (SEOBNRv4 templates, NRHybSur3dq8 injections resembling GW150914), our projection uses a subspace learned from the same 141-simulation training residuals, with the same sampler, the same injections, and matched wall-clock budget. On the kludge testbed, we train an analogous Gaussian-process calibration on the same waveform differences we use to learn our subspace. We beat it if we achieve equal or better coverage with a posterior-width exponent closer to $-1$ and lower cost.
Noise inflation, adding a model-error covariance to the detector noise, is included because it is the default engineering response. The key ablation replaces our chosen subspace with three alternatives of equal rank: a random subspace, the naive "all orders above truncation" subspace, and a subspace chosen without the information-cost term. This isolates whether the gain comes from the choice of kernel, as claimed.
Why not just…?
Why not just compute the missing higher-order terms? When they exist, one should. The oracle baseline shows that including 2.5PN spin and conservative terms fixes the bias for circular, equatorial intermediate mass-ratio inspirals. But generic captures are eccentric and precessing, PN series do not converge near the innermost stable orbit, and first-order self-force loses validity as mass ratios approach comparability. The method targets the regime where the missing term is not available, and the oracle comparison shows how much we give up.
Why not just inflate the noise to cover model error? Choosing observation operators to mitigate model error in Bayesian inverse problems analyzes this strategy and finds it inadequate because model error is structured and correlated rather than independent noise. In our setting that means inflation widens every direction while leaving the bias direction partly intact. We include it as a baseline and expect it to either over-widen or under-cover. The experiment will show which.
Why not just add free PN deformation coefficients to the template and sample over them? This is the real threat. In the linear regime, marginalizing over unconstrained coefficients on a subspace with flat priors is mathematically equivalent to projecting that subspace out. We will confirm this numerically on the weak-field testbed. So the contribution cannot be the projection itself. It is the choice and certification of the subspace, which a free-coefficient approach leaves to hand-picking. That consists of the exact identifiability table, the rank chosen by the bias–variance bound, the learned strong-field basis, and the transfer tests. The ablation settles this directly. If sampling free coefficients on the naive subspace matches our selected subspace, then the selection adds nothing, and we will say so. Analytic projection also avoids sampling $r$ extra parameters, but we treat speed as secondary.
Why not just use Gaussian-process calibration as in SEOBNRv4CE? It needs a high-fidelity training catalog. For comparable masses that catalog exists, and with informative priors calibration may give tighter posteriors than a hard projection, which discards information in the kernel entirely. For extreme mass ratios no equivalent catalog exists for eccentric, precessing orbits. The head-to-head on the GW150914-like setting decides whether projection is competitive where calibration is strongest.
Proposed Experiments
The primary testbed follows the kludge setups of the Huerta–Gair papers: circular, equatorial inspirals of $10$ solar-mass, $5\times10^3$ and $10^4$ solar-mass bodies into a $10^6$ solar-mass black hole, with LISA-like noise. The truth waveform includes conservative self-force and 2.5PN small-body spin terms; the template omits them. We sweep SNR from 30 to 3000.
At each SNR we measure four quantities: the ratio of bias to statistical error, the frequency with which 90% credible intervals contain the truth, the fitted posterior-width exponent against SNR, and wall-clock cost. Published anchors are the model-to-statistical error ratio below 1 at SNR 30 Influence of conservative corrections on parameter estimation for extreme-mass-ratio inspirals, and systematic errors several times statistical above mass ratio $10^{-4}$ without the missing terms Importance of including small body spin effects in the modelling of extreme and intermediate mass-ratio inspirals. No reported coverage fractions for these kludge setups were found on Lacuna, so we compute all baselines ourselves.
The first decisive check is the SNR-scaling claim. At SNR 1000 the projected estimator must keep coverage near nominal, with a width exponent near $-1$ rather than flattening. The measured bias must also lie under the predicted bound. A bias above the bound falsifies the linear analysis.
The second is the comparable-mass comparison against SEOBNRv4CE's reported $\mathrm{SNR}^{-0.75}$ scaling and bias reduction of 1.5× or more, on the same kind of injections.
The third is the strong-field dimension curve. If the learned subspace rank does not grow as the inspiral ends closer to the innermost stable orbit, our picture of PN breakdown is wrong.
The fourth is a transfer failure test: injecting mildly eccentric truth. Here we expect partial failure and will report the residual bias honestly.
The fifth is the atom-interferometer check. Given only the linear phase model and pulse timings as design variables, the subspace procedure must recover a sequence that nulls acceleration and initial-kinematics dependence while keeping the $T^3$ proper-time term. That is, it must rediscover the butterfly geometry of Universality-of-clock-rates test using atom interferometry with $T^{3}$ scaling. We then ask the open question of whether timings alone can also null the leading gravity-gradient term without destroying the signal, or whether momentum kicks as in Generalized gravity-gradient mitigation scheme are unavoidable. If the procedure fails to recover the butterfly, the method is wrong before it ever touches a waveform.
References
- Influence of conservative corrections on parameter estimation for extreme-mass-ratio inspirals
- Importance of including small body spin effects in the modelling of extreme and intermediate mass-ratio inspirals
- Importance of including small body spin effects in the modelling of intermediate mass-ratio inspirals. II. Accurate parameter extraction of strong sources using higher-order spin effects
- Choosing Observation Operators to Mitigate Model Error in Bayesian Inverse Problems
- The Limitations of Model-Based Experimental Design and Parameter Estimation in Sloppy Systems
- Homotopy classification of 4-manifolds with finite abelian 2-generator fundamental groups
- Homotopy classification of 4–manifolds whosefundamental group is dihedral
- Equivariant K-homology for hyperbolic reflection groups
- Universality-of-clock-rates test using atom interferometry with <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>T</mml:mi><mml:mn>3</mml:mn></mml:msup></mml:math> scaling
- Generalized gravity-gradient mitigation scheme
- Perturbative operator approach to high-precision light-pulse atom interferometry
- Incorporating waveform calibration error in gravitational-wave modeling and inference for SEOBNRv4