When seismologists locate an earthquake, they need a model of how fast waves travel through the rock between the source and each sensor. That model, the velocity model, is never known exactly. Yet most location pipelines fix one velocity model and report error bars that ignore the fact that it might be wrong. The resulting uncertainties are too small, often badly so for depth. Statisticians face the same structure whenever part of the information needed for an analysis is unobserved. Their standard remedy is multiple imputation: fill in the unknown part many times with plausible values, run the analysis on each completed version, and combine the results with Rubin's rules. Those rules add the ordinary within-analysis variance to the spread between analyses. This note proposes treating the velocity model, and the traces from dead stations, as missing data in exactly this sense.
Two things follow. First, it gives a cheap and principled way to get calibrated location uncertainty. It needs a few dozen inversions rather than the thousands of wave simulations a full Bayesian sampler requires. Second, the statistics of multiple imputation supply a quantity that geophysics lacks: the fraction of missing information. For every event and every coordinate, it says what share of the uncertainty comes from not knowing the velocity model and what share comes from missing stations. Operators can use that split directly to decide whether to fund a velocity survey or repair sensors.
Core Claim
Location uncertainty that pools over properly drawn velocity models and station-trace imputations will reach nominal coverage. Here "properly drawn" means sampled so that they reflect posterior uncertainty given calibration data, not just an arbitrary spread. Coverage means that 95% intervals contain the true hypocenter about 95% of the time. We expect fixed-model inversion to fall well short of nominal coverage. We also expect our method to approach the posterior spread of full Bayesian samplers at a small fraction of their cost.
The core quantity is Rubin's decomposition. For a location coordinate $\theta$ estimated from $M$ completed analyses, let $\bar U$ be the average within-analysis variance, for example from pick noise, and $B$ the between-analysis variance of the $M$ point estimates. The total variance and the fraction of missing information are
$$T = \bar U + \left(1 + \tfrac{1}{M}\right) B, \qquad \lambda = \frac{(1 + 1/M)\,B}{T}.$$
With nested imputation, where station traces are drawn inside each velocity draw, $B$ splits into a between-velocity part and a within-velocity, between-trace part. That split gives $\lambda_{\text{vel}}$ and $\lambda_{\text{sta}}$ separately.
One prediction makes the claim falsifiable. Locating earthquake epicenter without a seismic velocity model shows by dimensional analysis that $\log R = A\log t + B$, with epicenter information carried by the depth-independent slope. That work finds 97.2% of 1,209 ground-truth events within ±0.1° with no velocity model at all. If our decomposition is meaningful, $\lambda_{\text{vel}}$ should be near zero for horizontal coordinates and large for depth. A method that reported large horizontal $\lambda_{\text{vel}}$ would contradict Wei's result and signal improper draws.
How This Differs From Nearby Work
The obvious intersection is to impute missing traces with a generative model and then run waveform inversion. We deliberately do not center the proposal on that. Seismic Data Interpolation via Denoising Diffusion Implicit Models With Coherence-Corrected Resampling already fills missing traces with a diffusion model and reports SSIM 0.972 on consecutive gaps. A careful reviewer would also note that simply dropping missing receivers from the inversion's misfit is the correct likelihood when the physics is trusted. Imputed traces add prior information but no new data, so "impute then invert" is unlikely to beat masked inversion on accuracy. What SeisDDIMCR leaves open is uncertainty propagation: it uses sample variance only locally and never carries multiple realizations through a downstream inversion. We use its sampler as one ingredient, for dead-station traces, and make propagation and decomposition the contribution.
Three lines of work motivate the proposal. Methods for Large Scale Hydraulic Fracture Monitoring shows velocity errors map directly into mislocation, and leaves tolerating velocity uncertainty explicitly open. Real-time earthquake monitoring using a search engine method shows template-search reliability is bounded by the 3D model. First observations of core-transiting seismic phases on Mars shows how inferences about the Martian core are bounded by mantle-structure uncertainty. None of the three quantifies how much of the final uncertainty the velocity model contributes.
Other work supplies the machinery or sets constraints. Self-Flow-Matching assisted Full Waveform Inversion inverts for velocity by learning an online transport field from wavefield misfit. It reports relative L2 error 0.0130 versus 0.0213 for standard inversion on Marmousi. We reuse it as the engine that turns resampled calibration data into velocity draws, but it produces a single model and no uncertainty. Signal-based Bayesian Seismic Monitoring patches a 1D Earth model with Gaussian-process travel-time residuals and reduces location error fourfold. We reuse its residual processes as a velocity-imputation model for real regional data, where full waveform inversion is impractical. Its handling of dead stations by marginalization is the principled competitor for the station part. Microseismic event detection in large heterogeneous velocity models using Bayesian multimodal nested sampling constrains the design. It needed 4,000 wave simulations to train emulators, which is the cost we aim to avoid, and it provides the reference posterior we compare against.
On the statistical side, HyperImpute: Generalized Iterative Imputation with Automatic Model Selection is the cautionary case. It outputs point imputations that understate variance. Multiple imputation: review of theory, implementation and software supplies the pooling rules and the requirement that imputations be proper, meaning they include parameter uncertainty and do not merely add noise.
Research Program
We will build a two-level imputation loop around a standard location solver. The outer level draws $M_v$ velocity models. On synthetic media, each draw comes from running self-flow-matching inversion on a Bayesian bootstrap resample of calibration shots, the known sources used to constrain velocity, with an independent random initialization. The bootstrap is the classical device for making imputations approximately proper: it injects the uncertainty in the velocity model's own estimate. On real regional data, velocity draws are sampled from Gaussian-process travel-time residual fields on a 1D reference model, fit to ground-truth events held out from evaluation.
The inner level draws $M_s$ completions of the dead-station traces with the diffusion trace sampler, conditioned on surviving traces. Travel times for the imputed traces are then re-predicted under the current velocity draw, so an imputed arrival stays kinematically consistent with that model. This addresses the non-physical-arrival weakness of purely image-domain imputation. Each completed dataset is located with an ordinary grid-search or linearized solver that returns a point estimate and a within-analysis covariance. Results are pooled with nested Rubin's rules.
The analysis side has three parts. We will check properness with simulation-based calibration of the velocity draws. We will study how $M$ trades against interval stability; Rubin's $1/M$ inflation term predicts diminishing returns beyond a few tens of draws. We will also study whether $\lambda$ can be inverted to rank candidate interventions, such as adding a station at a given site versus shooting additional calibration sources.
Baselines
The naive reference point is fixed-model location. We use the single best self-flow-matching velocity model on synthetics and the uncorrected 1D model on real data, with dead stations deleted and error bars from pick noise alone. On synthetics every method shares the same shots, events, noise level, missing-station masks, and forward solver. Beating this reference means reaching near-nominal coverage where it undercovers, while keeping median location error no worse.
The strongest competitor for calibrated uncertainty is full Bayesian joint inference over velocity and source. We run it as multimodal nested sampling with emulators, following Das et al., on the same synthetic volumes. It gets a matched wall-clock budget, and also a generous budget so it can serve as the reference posterior. The file reports a posterior odds ratio of about 4168 for event detection but no reported result found on Lacuna for coverage or interval width of nested-sampling location. We will therefore measure it ourselves. Beating it means matching its coverage within a few points using at least ten times fewer forward simulations.
The strongest horizontal-accuracy reference on real data is Wei's velocity-free search, at 97.2% within ±0.1° on the IASPEI ground-truth list. We run it on the identical event subset. We do not expect to beat it on epicenter point accuracy. Our claim is to match it horizontally while adding calibrated depth intervals, which it cannot produce.
For the station component, the competitor is likelihood masking or marginalization, which drops dead stations as SIGVISA-style systems do. It is run through the same pooling so only the station treatment differs.
The key ablation isolates properness. It replaces bootstrap-conditioned velocity draws with an ensemble of inversions that differ only in random initialization, with identical $M$ and compute. If that ensemble calibrates as well, the statistical machinery adds nothing beyond ensembling.
Why not just…?
Why not just run an ensemble of inversions from different starting models and report the spread? This is the most serious threat, because it costs the same as our method and looks similar. The difference is that initialization spread measures optimizer sensitivity, not uncertainty about the velocity model given the data. On benchmarks where inversion converges reliably, the spread can collapse toward zero; the constant-start results in the IG-FWI paper suggest modern inverters often converge from poor starts. The properness ablation settles this directly. If the plain ensemble reaches nominal coverage on Marmousi and Overthrust, we will report that and narrow the contribution to the decomposition $\lambda$.
Why not just do full Bayesian joint inversion? It is the gold standard, and we use it as the reference posterior. It needed thousands of wave simulations plus emulators that can fail in high-contrast media, as in Microseismic event detection in large heterogeneous velocity models using Bayesian multimodal nested sampling. It also has no built-in, interpretable split of uncertainty by source. The matched-budget comparison will show whether multiple imputation gives a comparable posterior spread far more cheaply.
Why not just mask dead stations in the misfit instead of imputing them? For point accuracy, masking is probably as good as imputation, and we concede this; it is why trace imputation is not the headline. Masking alone, however, cannot report how much of an event's uncertainty is caused by the missing stations. That quantity comes from the between-imputation variance. Our experiments compare masking against imputation on location error. We expect a tie there, and the imputation arm's value to show in $\lambda_{\text{sta}}$ predicting the variance reduction from actually restoring those stations.
Why not just use a velocity-free method? For epicenters, Wei's method works and our framework predicts near-zero horizontal velocity dependence, consistent with it. The same work shows it is blind to depth, and the direct network in Locating induced earthquakes with a network of seismic stations in Oklahoma via a deep learning method breaks when its fixed 30-station layout loses sensors. Its error also rises from 2.0 to 4.9 km under label noise. Depth and station dropout are exactly where our method is aimed.
Proposed Experiments
The synthetic study embeds sets of point sources in Marmousi, the SEG/EAGE Overthrust and Salt models, and the 2004 BP model. We simulate surface and borehole arrays, remove 10–40% of stations in random and block patterns, and include an MNAR pattern where the stations nearest large events are clipped. Calibration shots constrain velocity through bootstrapped self-flow-matching inversion. We measure 95% coverage and interval width separately for horizontal coordinates and depth, continuous ranked probability score, median location error, and number of forward solves. The same metrics are reported for fixed-model location, the initialization ensemble, masking, and nested sampling.
The real-data study uses the IASPEI ground-truth list, with Gaussian-process residual velocity draws fit on held-out ground-truth events. We report coverage of the reference hypocenters, including depth, alongside Wei's epicenter accuracy on the same events.
Three checks are decisive. First, fixed-model intervals must undercover depth while ours sit near 95%, and the initialization-only ensemble must fall measurably short of ours; otherwise properness does not matter. Second, horizontal $\lambda_{\text{vel}}$ must be small and depth $\lambda_{\text{vel}}$ large, in line with Wei's dimensional argument. Third, on synthetics we will actually perform both interventions, adding the flagged stations and adding calibration shots. The ranking implied by $\lambda_{\text{sta}}$ versus $\lambda_{\text{vel}}$ must predict which one shrinks the intervals more, for most events. If the third check fails, the decomposition is descriptive only, and we would say so.
References
- Locating earthquake epicenter without a seismic velocity model
- Seismic Data Interpolation via Denoising Diffusion Implicit Models With Coherence-Corrected Resampling
- Methods for large scale hydraulic fracture monitoring
- Real-time earthquake monitoring using a search engine method
- First observations of core-transiting seismic phases on Mars
- Self-Flow-Matching assisted Full Waveform Inversion
- Signal-based Bayesian Seismic Monitoring
- Microseismic event detection in large heterogeneous velocity models using Bayesian multimodal nested sampling
- HyperImpute: Generalized Iterative Imputation with Automatic Model Selection
- Multiple imputation: review of theory, implementation and software
- Locating induced earthquakes with a network of seismic stations in Oklahoma via a deep learning method