Where Model Error Hurts: Predicting and Removing High-SNR Parameter Bias from the Geometry of Template Residuals

When a physics template is slightly wrong, a cleaner measurement can make the answer worse. As the signal-to-noise ratio (SNR) grows, the statistical error bar shrinks while the template's error stays fixed. The fitting procedure then bends the parameters it controls to absorb the mismatch, and the posterior collapses tightly around the wrong value. Gravitational-wave astronomy shows this most clearly. On a loud GW150914-like injection, a standard effective-one-body waveform model returned a chirp-mass posterior that excluded the true value entirely (Incorporating waveform calibration error in gravitational-wave modeling and inference for SEOBNRv4). The usual fixes are to inflate the noise or to add nuisance parameters for the template error. Both treat the problem as one of error size. We argue it is mainly a problem of error direction.

The proposal is to compare two subspaces in the noise-weighted space of possible signals. The first is spanned by the template's residual errors, learned from pairs of approximate and high-fidelity waveforms. The second is spanned by the template's sensitivity to its physical parameters. The angles between these two subspaces should predict three things: the SNR at which bias overtakes noise, which parameters will be biased, and how much information any debiasing method must give up. The same geometry also yields a correction. Project out only the few residual directions that overlap the sensitivity subspace, choosing how many to remove as a function of SNR. We test whether this matches full nuisance-parameter marginalization at a fraction of the cost, and whether it beats both noise inflation and the existing diagonal calibration-error model.

Core Claim

Write the data as $d = h(\theta) + \delta + n$. Here $h(\theta)$ is the approximate template at parameters $\theta$, $\delta$ is the template's residual error, and $n$ is detector noise. All quantities are whitened by the noise spectrum. Let $J = \partial h/\partial \theta$ be the template's sensitivity matrix and $F = J^\top J$ the Fisher information. A first-order expansion of the log-likelihood around the true parameters gives the bias

$$\Delta\theta \approx F^{-1} J^\top \delta .$$

Only the component of $\delta$ lying in the span of $J$ shifts the parameters. Every other component, however large, only lowers the match between template and data. Both $\delta$ and $J$ grow linearly with signal amplitude, while the posterior width shrinks as $1/\text{SNR}$. The ratio of bias to width therefore grows linearly with SNR. The constant of proportionality is set by the principal angles between the residual subspace and the column space of $J$.

Our central empirical claim is that, for waveform-model error against numerical relativity, residuals are approximately low-rank in this noise-weighted space. A handful of their directions carries nearly all of the parameter-relevant overlap. This has two consequences. First, a projector $P_r$ that removes only the top $r$ overlapping residual directions,

$$\tilde d = (I - P_r)\, d ,$$

should remove most of the bias while discarding only the information carried by those $r$ directions. Second, the best $r$ should rise stepwise with SNR. At low SNR it is zero, because removing anything costs more than the bias it fixes. This is the crossover regime that Influence of conservative corrections on parameter estimation for extreme-mass-ratio inspirals identified for self-force corrections, now made quantitative and actionable.

The novelty lies in three pieces that existing work lacks. The first is the overlap geometry as a predictor. The second is the learned, correlated residual subspace in place of independent per-frequency errors. The third is the SNR-dependent choice of how much to remove.

Relation to Prior Work

Three papers supply the method. Choosing observation operators to mitigate model error in Bayesian inverse problems proves that an observation operator whose null space contains the model error makes a misspecified model perform as well as the true one. It also proves that inflating the noise widens every parameter's posterior without fixing a structured shift. However, it requires the analyst to know the form of the error analytically, it is limited to linear operators, and it leaves the information trade-off unquantified. We learn the null space from data, measure the trade-off through principal angles, and apply the operator locally around nonlinear waveform models.

Incorporating waveform calibration error in gravitational-wave modeling and inference for SEOBNRv4 is the closest existing method. It fits Gaussian-process models of amplitude and phase error at 10 frequency nodes, trained on 141 SXS simulations, and marginalizes over 20 nuisance parameters. It treats the errors as uncorrelated across nodes. We keep its training data and test setup but replace independent per-node errors with a correlated low-rank residual basis. We also ask a question it does not: which error directions actually matter for which parameters.

The Limitations of Model-Based Experimental Design and Parameter Estimation in Sloppy Systems shows that tightening parameter constraints under model error worsens prediction, but it offers diagnosis without a correction. We reuse its systems-biology setting to test whether the overlap geometry explains its failure cases.

Several papers motivate the work by documenting the bias at high SNR. Importance of including small body spin effects in the modelling of intermediate mass-ratio inspirals. II shows primary mass and spin compensating for an omitted companion spin. Measuring dark matter spikes around primordial black holes with Einstein Telescope and Cosmic Explorer shows vacuum templates losing up to 80% of SNR against signals altered by surrounding dark matter. Rapid gravitational wave parameter estimation with a single spin shows systematic shifts in a fast approximate waveform comparable to truncation error.

One paper constrains the design. Anomaly preserving contrastive neural embeddings for end-to-end model-independent searches at the LHC finds that representations tuned to suppress known variation discard unknown signal. A projector learned from one family of residuals could therefore erase physical effects outside that family. We test this directly rather than assume it away.

Research Program

The first step builds the residual basis. For each SXS simulation, we compute the residual between the SEOBNRv4 waveform and the numerical-relativity waveform at matched physical parameters. Extrinsic time and phase are aligned so that only intrinsic modeling error remains. We whiten with the design-sensitivity noise spectrum. Principal components of these residuals are taken with the noise-weighted inner product, locally across bins in mass ratio and spin, with held-out simulations reserved for testing.

The second step measures the geometry. At each test point we compute $J$ for chirp mass, mass ratio and effective spin, then the principal angles between $J$ and the residual subspace. From these we form a predicted bias-to-width ratio and a predicted crossover SNR. The projector is the top-$r$ overlapping residual directions, with $r$ chosen by minimizing predicted mean-squared error (squared bias plus variance after projection). We also implement a soft variant that marginalizes over coefficients on the same basis with a Gaussian prior. This lets us separate the value of the basis from the value of hard removal.

The third step tests the idea outside gravitational waves. On the advection-diffusion-reaction benchmark from the observation-operator paper, the analytic null space is known, so the learned basis can be checked against it. On the EGFR and DNA-repair models from the sloppy-systems paper, we test whether the overlap between discrepancy and stiff parameter directions explains when tighter constraints hurt.

Baselines

The naive reference is standard inference with the uncorrected SEOBNRv4 template. This is the setting where the GW150914-like injection's chirp-mass posterior excluded the truth. Any method must restore coverage of the true value there.

The second baseline is noise inflation: adding a diagonal variance term sized to the typical residual norm. This is what most practitioners would try first. The observation-operator paper analyzes it formally, but no reported result was found on Lacuna for noise inflation on the SXS/NRHybSur3dq8 gravitational-wave test. We will run it with its variance tuned on the same training simulations.

The strongest existing method is SEOBNRv4CE, run as published: diagonal Gaussian-process calibration error and 20 nuisance parameters. Its reported results are a reduction in normalized chirp-mass bias by "a factor of 1.5 or more" in spinning systems, posterior width scaling as $\text{SNR}^{-0.75}$, and roughly 2x sampling slowdown (Incorporating waveform calibration error in gravitational-wave modeling and inference for SEOBNRv4). For fairness, all methods use the same 141 SXS training simulations, the same NRHybSur3dq8 injections, the same noise spectrum and sampler, and a matched wall-clock budget.

Beating SEOBNRv4CE means one of two outcomes. Either the projector achieves a lower normalized bias at equal posterior width, or it achieves equal bias and coverage with narrower posteriors, meaning less discarded information, at no more than its cost.

Two ablations isolate the key ingredients. The first is a full-covariance soft marginalization on our learned basis, which tests correlation without hard removal. The second is a projector built from random or top-variance residual directions rather than the overlap-ranked ones, which tests whether the geometric ranking matters.

Why not just…?

Why not just collect louder signals or more data? By the linear bias formula, bias-to-width grows with SNR, so more data makes the problem worse. The SEOBNRv4CE paper's high-SNR injection shows this directly. Our SNR sweep will reproduce the growth for the uncorrected baseline, and that reproduction is the motivating curve.

Why not just inflate the noise? Inflating the noise widens the posterior in every direction, including those carrying no bias. The observation-operator paper proves it leaves structured shifts uncorrected. Our comparison at equal coverage will measure how much posterior width it wastes. If it matches the projector on this test, the low-rank claim is wrong, and we will report that.

Why not just extend SEOBNRv4CE with a full covariance across frequency nodes? This is the real threat. Marginalizing over a correlated Gaussian error prior is a soft version of projection, and in the linear limit the two coincide. The soft-marginalization ablation settles it. If full-covariance marginalization matches hard projection, our contribution reduces to the learned basis plus the overlap diagnostic. The diagnostic would still predict crossover SNR and per-parameter bias, which no current method provides. We expect hard projection to win on cost, because $r$ is small, and on robustness, because it does not depend on tuning the prior variance. Neither expectation is certain.

Why not just use the more accurate surrogate as the template? NRHybSur3dq8 is used here as ground truth precisely because it is accurate. Accurate surrogates, however, cover limited regions of parameter space, and approximate waveforms are used wherever coverage, speed or new physics demands them. The question that matters is how to make inference trustworthy with the template one actually has.

Proposed Experiments

The first decisive check tests prediction. On held-out SXS configurations, we plot the predicted bias-to-width ratio from principal angles against bias measured from full injection-recovery runs, over SNR from about 10 to 200. A rank correlation near one, and crossover SNRs within a factor of two, support the geometric claim.

The second check tests correction. We compare normalized chirp-mass bias, true-value coverage measured by percentile-percentile plots across injections, posterior width scaling with SNR, and sampling cost against all baselines. The target is to clear SEOBNRv4CE's reported factor-1.5 bias reduction while keeping width scaling closer to $\text{SNR}^{-1}$ than its $\text{SNR}^{-0.75}$.

The third check tests failure under out-of-family error. Into signals that SEOBNRv4 never saw residuals for, we inject dephasing from dark-matter spikes, following Cole et al., and from companion-spin effects, following Huerta and Gair. We then measure whether the projector leaves such effects detectable or erases them. Erasure would confirm the risk raised by the LHC embedding work and would limit the method to modeling errors of known family.

The fourth check tests cross-domain transfer. On the PDE benchmark, the learned subspace should recover the analytic null space and reproduce the reported near-truth posterior mean. On the sloppy-systems models, the overlap angle should explain where optimal experimental design inflated model error.