Galaxy redshift surveys measure the expansion of the universe with a "standard ruler," a clustering feature at a known physical size of about 150 Mpc. To use it, they place optical fibers on thousands of galaxies at once. Two fibers cannot sit closer than a hardware limit (62 arcseconds in SDSS/BOSS), so in crowded groups some galaxies never get a redshift. For twenty years the standard fix has been to copy the missing galaxy's redshift from its nearest observed neighbor and double that neighbor's weight. This works for large-scale averages. It fails for the pairs that matter most, though: a chance alignment of a foreground and a background galaxy gets collapsed into one fake pair, and a real companion loses its random orbital velocity. We propose to treat the correction the way a large language model treats retrieval. A cheap internal guess resolves most collided galaxies. A spare fiber on a later pass, which is a scarce and expensive external lookup, goes only to galaxies whose guesses disagree. The correction sits in front of the survey's frozen analysis code, much as a denoiser sits in front of a frozen classifier.
The practical question is concrete. Surveys have a limited budget of repeat observations. Given that budget, which collided targets should be re-observed so that small-scale clustering and redshift-space distortions are fixed, while the standard-ruler distances stay provably untouched? Today the answer is "all of them if possible, otherwise none." That is the astronomical version of retrieving documents for every query regardless of need.
Core Claim
A small set of draws from a cheap imputation model can separate collided galaxies into two groups. The model is conditioned on the neighbor's redshift, the collided galaxy's photometric-redshift estimate, and color similarity within the group. When the draws agree, the galaxy is almost certainly a physical companion, and its line-of-sight position is set by the group's velocity spread. When the draws disagree, the galaxy is likely a chance projection or sits in an ambiguous environment. Our claim has two parts. First, spending re-observations on the disagreeing galaxies should remove most of the nearest-neighbor bias in the redshift-space power spectrum multipoles at a small fraction of the cost of re-observing everything. Second, because the corrector only reassigns redshifts along fixed angular positions and is averaged over draws, it should leave the standard-ruler scale unbiased to well below the survey's statistical error.
Each input direction plays a specific role. Fiber collision bias is the problem. The nearest-neighbor weight of Reid et al. is known to distort line-of-sight velocity dispersions and the elongated "Fingers of God" that virialized groups produce in redshift space. Standard-ruler cosmology is the constraint and the evaluation: any correction must not move the acoustic scale. Adaptive retrieval supplies the decision mechanism. TARG gates retrieval on the uncertainty of a short draft, using token margin or disagreement among a few sampled drafts. It reports cutting retrieval frequency by 70–90% while matching or beating always-retrieve on PopQA, TriviaQA and NQ-Open. Model-agnostic defense supplies the architecture and a warning. Denoised Smoothing and Diffusion Denoised Smoothing wrap a frozen model in an external correction layer and average over random draws. The second of these papers also found that heavy generative denoising hallucinates features, and so favored faithfulness over realism. We take that lesson directly: our imputer is deliberately low-capacity and local, not a generative model that invents group members.
Research Program
The first component is a conditional imputer. For each collided target $i$ it produces $K$ redshift draws $z_i^{(1)},\dots,z_i^{(K)}$. The draws mix two explanations. In the "companion" explanation, the redshift is the neighbor's redshift plus a velocity offset scaled to the group's multiplicity and richness. In the "projection" explanation, the redshift is sampled from the photometric-redshift posterior. The mixing probability comes from color and photometric-redshift agreement. The imputer is fit on galaxy pairs that both received fibers, which the survey observes directly. This is the analogue of a denoiser trained on clean data, and it needs no N-body simulation. The model's inputs are drawn only from what the imaging and the neighbor's spectrum provide. It never moves angular positions, and it never adds or removes objects.
The second component is the gate. Following TARG's small-sample-variance score, the uncertainty of target $i$ is
$$U_i = \frac{1}{K}\sum_{k=1}^{K} \mathbf{1}\left[\,|\chi(z_i^{(k)}) - \chi(\bar z_i)| > \Delta\,\right],$$
where $\chi$ is comoving distance, $\bar z_i$ is the median draw, and $\Delta$ is a few $h^{-1}$Mpc, roughly the scale of a Finger of God. $U_i$ is the fraction of draws that put the galaxy in a different structure from the consensus. Given a re-observation budget of $B$ fibers, the gate re-observes the $B$ collided targets with the highest $U_i$. Every other target keeps its $K$ draws.
The third component is propagation into the frozen pipeline. The survey's estimators, density-field reconstruction (Padmanabhan et al.) and BAO/RSD fitting code are run unmodified on each of the $K$ imputed catalogs. The results are averaged, and the spread across draws is added to the covariance. This mirrors the Monte Carlo averaging of randomized smoothing. It costs $K$ runs of the clustering code, which is small compared with a mock suite.
The fourth component is a faithfulness audit, which answers the hallucination tension that Carlini et al. identified. The corrector touches only a few percent of objects, and only their radial coordinate. Any change to pair counts at the ruler scale is therefore at most proportional to the collided fraction times the error in the cross-correlation between collided galaxies and the rest. We check this empirically rather than asserting it, by running the corrector on mocks with the acoustic feature removed and confirming that it creates no feature.
Baselines
The naive reference is doing nothing: collided targets are dropped, and angular pair counts are cut below the collision scale. Following Zehavi et al., this blinds the analysis below roughly 1 $h^{-1}$Mpc. It is run on the same mocks and data, and it defines the floor.
The strongest existing method, and the one the proposal must beat, is nearest-neighbor completeness weighting ($w_{cp}$) as implemented in the BOSS DR12 catalogues (Reid et al.). It is used in the DR12 anisotropic analysis of Gil-Marín et al.. That analysis reports, for CMASS at $z=0.57$, $f\sigma_8 = 0.444 \pm 0.038$ and $D_A/r_s = 9.42 \pm 0.15$. Anderson et al. report $D_V/r_s = 13.67 \pm 0.22$. We found no result on Lacuna for the residual bias that $w_{cp}$ induces in $f\sigma_8$ or in the power spectrum quadrupole, so part of the work is measuring it in mocks where true redshifts are known. $w_{cp}$ is the strongest competitor because it is free, transparent and already validated at the ruler scale. We run it with the same collision emulator, the same mocks, and the same frozen estimator and fitter. Beating it means two things at zero extra fibers. The first is a lower absolute bias in the quadrupole on scales $k \gtrsim 0.1\,h\,\mathrm{Mpc}^{-1}$. The second is a bias in $f\sigma_8$ below 0.2 times the CMASS statistical error, roughly 0.008. That target is ours, derived from the quoted error. The acoustic-scale shift must also be no worse than under $w_{cp}$.
The mock-calibrated route of Gil-Marín et al. and the DESI DR2 validation applies the collision algorithm to simulated catalogs and corrects through transfer functions. We run it with an equal number of AbacusSummit mocks. We compare biases when the mock's assumed galaxy-occupation model is deliberately mismatched to the "data" mock. This is where a data-driven imputer should be more robust.
The remaining baselines test the fiber budget and isolate each ingredient. The "always re-observe" oracle assigns the true mock redshift to every collided target. It is the analogue of retrieving documents for every query and sets the ceiling. Random allocation spends the same $B$ fibers uniformly and isolates the value of the gate. Densest-group allocation spends them on the richest groups and is a sensible heuristic an astronomer would try. Neighbor-redshift-plus-velocity-kick imputation uses no gate and no photometric information, and isolates the projection component. The imputer-without-gate ablation uses $B=0$ with $K$ draws, and isolates imputation from re-observation. Finally, a flexible generative completion model in the style of Diffusion-Occ is allowed to place collided galaxies freely. It tests whether the faithfulness restriction matters. All methods get the same training pairs and the same tuning budget, and are selected on held-out mocks.
Why not just…?
Why not just keep $w_{cp}$ and fit small-scale nuisance parameters? For BAO distances this may genuinely be enough, and our faithfulness experiment would show that. Anderson et al. judged residual collision effects negligible at 1.7% precision. But Reid et al. document that $w_{cp}$ distorts line-of-sight dispersions. Nuisance parameters absorb that distortion at the cost of growth-rate precision, and they hide it rather than removing it. The comparison of $f\sigma_8$ bias and error bars at fixed scale cuts decides the question. If $w_{cp}$ plus nuisance terms matches our corrector at $B=0$, we will say so. In that case the contribution narrows to the fiber-allocation result.
Why not just add a Gaussian velocity kick to the neighbor's redshift? This is the real threat to the imputer, and it is a named baseline. It fixes the Finger-of-God collapse for true companions but cannot identify chance projections. Our prediction is that the gap between the two appears specifically in the pair counts at line-of-sight separations above 20 $h^{-1}$Mpc and small transverse separation, where projections land. If the gap does not appear, projections are rare and the gate has little to find.
Why not just gate on photometric-redshift disagreement alone, without draws? This could perform nearly as well as $U_i$, because photometric information drives much of the disagreement. We include it as a gating baseline. If it matches, we adopt it. The contribution is then the budget-allocation study and the frozen-pipeline audit, not the particular score.
Why not just re-observe everything, or rely on simulations? Re-observing everything is the oracle ceiling. Its cost is exactly what we are trying to avoid, just as TARG shows that retrieving for every query wastes compute and can even hurt accuracy. Simulation-only correction inherits the galaxy-occupation assumptions baked into the mocks. The mismatched-mock experiment measures that sensitivity directly.
Proposed Experiments
The main testbed is AbacusSummit and MD-Patchy mocks. A 62-arcsecond collision emulator is applied, and true redshifts are retained for every collided target. For each method we measure four quantities: the power spectrum monopole and quadrupole, computed with the estimators used by Gil-Marín et al.; the projected correlation function below 2 $h^{-1}$Mpc; $f\sigma_8$; and the post-reconstruction distances $D_A/r_s$ and $Hr_s$. The central figure plots bias in the quadrupole and in $f\sigma_8$ against the re-observation budget $B$, expressed as a fraction of collided targets. It includes gated, random, densest-group and oracle curves, and is the direct counterpart of TARG's plot of accuracy against retrieval frequency. Our hypothesis is supported if the gated curve reaches most of the oracle's bias reduction at a budget of a quarter or less of collided targets. Random allocation needs a proportionally larger budget. It is refuted if gated and random curves coincide.
The faithfulness check runs every corrector on mocks built from a power spectrum with no acoustic feature. We require that the fitted acoustic amplitude stay consistent with zero. We also require that, on standard mocks, the shift in $D_A/r_s$ and $Hr_s$ relative to the truth-redshift catalog be below one tenth of the DR12 CMASS errors quoted above. The unconstrained generative baseline is expected to fail this check, and it is there to show whether the restriction is necessary.
A robustness check trains the imputer on pairs where both members received fibers, then tests it on mocks with shifted group richness and tracer type. This probes the domain-generalization failure that TARG reports for gating policies tuned on one distribution. The last step is a real-data check on BOSS DR12. Collided pairs that were resolved in regions of overlapping observations serve as held-out truth, and the full pipeline is rerun to report the change in $f\sigma_8$ relative to the published $w_{cp}$ result.
References
- SDSS-III Baryon Oscillation Spectroscopic Survey Data Release 12: galaxy target selection and large-scale structure catalogues
- TARG: Training-Free Adaptive Retrieval Gating for Efficient RAG
- Denoised Smoothing: A Provable Defense for Pretrained Classifiers
- (Certified!!) Adversarial Robustness for Free!
- Efficient reconstruction of linear baryon acoustic oscillations in galaxy surveys
- Galaxy Clustering in Early Sloan Digital Sky Survey Redshift Data
- The clustering of galaxies in the SDSS-III Baryon Oscillation Spectroscopic Survey: RSD measurement from the LOS-dependent power spectrum of DR12 BOSS galaxies
- The clustering of galaxies in the SDSS-III Baryon Oscillation Spectroscopic Survey: baryon acoustic oscillations in the Data Release 9 spectroscopic galaxy sample
- Validation of the DESI DR2 measurements of baryon acoustic oscillations from galaxies and quasars
- Diffusion-Occ: 3D Point Cloud Completion via Occupancy Diffusion