Two measurements from the same survey seem to disagree. Galaxies in the Sloan Digital Sky Survey point their long axes along the cosmic filaments that connect clusters, and this signal is detectable out to 30–40 Mpc. Yet the dark-matter halos around those same kinds of galaxies show no significant flattening along the galaxy's light. If gravity shapes structure in the same way across scales, both signals should come from one shared cause, namely how well a galaxy's shape tracks its host halo. Either the measurements are broken somewhere, or that shared cause fails at small scales because gas physics introduces a characteristic scale.
We propose to settle which explanation fits by encoding both measurement pipelines as one small Bayesian network. A Bayesian network is a probability model built from local conditional tables, one per variable. Because its probabilities are sums of products of table entries, a single backward pass returns the exact sensitivity of any conclusion to every table entry at once. We then ask a precise question: which single assumption, changed by the smallest amount, makes both published results consistent? The project also gives the first empirical test of sensitivity-screening bounds that have so far appeared only as theory.
Core Claim
In a discrete Bayesian network, the probability of evidence is a polynomial that is linear in each parameter separately. A differential approach to inference in Bayesian networks builds this polynomial as an arithmetic circuit and differentiates it in time linear in the circuit's size. That work also shows that the derivatives answer "how much must this parameter change to reach a target probability" without rerunning inference. Our proposal relies on this method.
The property matters because generic automatic differentiation only gives a local slope. Here, one forward and one backward pass fix each parameter's sensitivity curve over its whole range, since a single-parameter posterior is a ratio of two linear functions. That is exactly the "minimal change to reconcile" quantity we need.
Our claim is that a discretized model of the joint alignment-measurement process can rank the possible sources of the tension correctly. The candidate sources are filament position error, shape noise, residual point-spread-function shear, satellite contamination, and how tightly galaxies track their halos. "Correctly" means the ranking agrees with brute-force reruns of the real continuous pipelines on simulated catalogs with injected errors. If the ranking agrees, the same audit applied to real SDSS evidence tells us whether a shared, scale-invariant galaxy–halo coupling survives once errors are accounted for.
Prior Work and Its Role
Three papers motivate the project. Detecting galaxy–filament alignments in the Sloan Digital Sky Survey III measures alignment on 361,762 BOSS LOWZ galaxies. It finds alignment out to 30–40 Mpc using smooth curve fitting with no regression model assumed, after removing the influence of clusters in the redMaPPer catalogue. It does not trace how filament errors propagate into that result.
Ellipticity of dark matter haloes with galaxy-galaxy weak lensing uses a rotated-shear subtraction to cancel spurious shear from the telescope's blurring. It finds no overall halo-ellipticity detection: $f_h \approx 0.60 \pm 0.38$ for red galaxies and $-1.4 \pm 1.7$ for blue ones. It leaves the dominant uncertainty unidentified.
Cosmic web reconstruction through density ridges: catalogue supplies the filament maps. It also gives the 10–20 Mpc positional uncertainty that makes filament error a plausible culprit.
The closest existing work is Sensitivity Analysis of Simulation-Based Inference for Galaxy Clustering. It tests misspecification by training a separate neural posterior estimator on each forward-model variant and testing it on another. Its findings shape our design. The galaxy-occupation model mattered, and the choice of halo finder biased results, while the gravity solver barely mattered. We therefore model galaxy–halo connection and catalog-construction error explicitly and do not differentiate through N-body physics.
We differ from that paper in two ways. We study alignment rather than clustering. We also compute exact, all-parameter sensitivities on a compiled model instead of comparing a handful of retrained estimators.
Revealing the 3D Cosmic Web through Gravitationally Constrained Neural Fields differentiates through a lensing forward model to reconstruct density. It is a reconstruction method, not a sensitivity audit, and it does not address alignment. We do not compete with it.
The remaining work supplies methods or limits our design. A Differential Semantics for Jointree Algorithms shows that the outward messages of the standard Shenoy–Shafer jointree algorithm are these same derivatives. The paper also warns that the Hugin variant loses derivatives when a probability is zero, so we use Shenoy–Shafer. Making Sensitivity Analysis Computationally Efficient gives the coefficients of every sensitivity function from one inward and two outward passes.
Exploiting Evidence-dependent Sensitivity Bounds derives bounds that use the probability of the observed evidence. On an illustrative example these tighten a sensitivity bound from 2.67 to 0.95. The authors state that the work contains no empirical evaluation on datasets. That gap is the third direction of this proposal, and we treat it as a test target rather than as an established tool.
Research Program
We build one network shared by both measurements. Its root variables are the galaxy class (red or blue, bright or faint), the true distance to the nearest filament, cluster proximity, and satellite status. A latent variable for galaxy–halo misalignment angle, in binned form, feeds two branches. In the filament branch, the true distance and a filament-error variable produce the measured distance. The misalignment angle combined with shape noise produces the measured galaxy–filament alignment bin. In the lensing branch, halo shape, the misalignment angle, and a residual-blur shear variable produce the measured halo-ellipticity bin.
The scale-invariance hypothesis is the constraint that both branches read the same misalignment table. The scale-breaking alternative gives the lensing branch its own table, conditioned on galaxy class. The treewidth, a measure of how tangled the network is, stays below about eight. Exact compilation is therefore trivial, and the treewidth barrier of large cosmic-web graphs never arises, because we model the measurement process rather than the web itself.
We learn the tables from simulated catalogs. Quijote halo catalogs with the 5- and 10-parameter halo-occupation models used in the SBI sensitivity study provide galaxies. Halo shapes come from the halo finder. Galaxy shapes are painted on with a controlled misalignment distribution. We apply the ridge-finding filament algorithm and the alignment estimator to these mocks exactly as the SDSS papers apply them to data. The continuous estimator outputs are then binned to fit the tables.
On real data, the evidence is the two published outcomes, broken down by galaxy class. The queries are the probability of each detection outcome and the posterior over misalignment. For every parameter we report three quantities: the exact derivative, the full single-parameter sensitivity curve, and the minimal change needed to make both measurements jointly likely. We also apply the evidence-dependent bounds to screen out insensitive parameters before computing exactly.
Baselines
The naive reference point is to do no sensitivity analysis: report the two published results and attribute the gap to statistical noise. Our audit must identify a cause that this reference cannot, and confirm it on mocks, to be worth anything.
The strongest competitor is the standard astronomy practice of injection–recovery. One perturbs a single systematic on the mocks, such as shifting filaments by a set error or adding blur shear. One then reruns the full continuous pipeline, filament finding plus local polynomial regression or rotated-shear lensing, and records how the detection statistic moves. This is the ground truth our surrogate must match. It is also the method of the closest prior work, Sensitivity Analysis of Simulation-Based Inference for Galaxy Clustering, adapted to alignment. That paper quotes no posterior-width figures in the summary, and no reported result was found on Lacuna for injection–recovery ranking of alignment systematics.
We give injection–recovery the same mocks and the same set of perturbed quantities, and run it at three step sizes. The surrogate beats it if two conditions hold. Its top-three ranking of systematics must match the pipeline's, with Kendall rank correlation of at least 0.8. Its compute must be at least an order of magnitude lower, counting the cost of learning the tables.
The bound baseline compares two kinds of bounds that ignore the evidence: the general ones, and the evidence-dependent ones from Exploiting Evidence-dependent Sensitivity Bounds. That paper's illustrative figure is 2.67 against 0.95. The evidence-dependent bounds win if they screen out a larger share of parameters at no loss, meaning zero parameters wrongly dismissed when checked against exact values.
The key ablation removes the shared misalignment table. It tests whether the scale-invariant coupling, rather than extra free parameters, is what lets the model explain both measurements. We compare the two variants by held-out log-likelihood on independent mock realizations, and by the size of the minimal reconciling change on real evidence.
Why not just…?
Why not just finite-difference the network or the pipeline? On a network this small, finite differences cost almost nothing. Speed is therefore not our argument. The argument is that one exact pass gives each parameter's entire curve and the minimal change to reach a target, which finite differences only approximate after a grid search. Finite differences are already included where they matter, as the injection–recovery baseline on the real pipeline. For the network itself, we will report agreement between exact and finite-difference derivatives as a sanity check, not as a contest.
Why not just write the model in a differentiable probabilistic programming language and backpropagate through a continuous likelihood? This is the most serious threat. If the continuous model's local gradients rank the systematics as well as our method, discretization buys nothing beyond the global sensitivity curves. We will build that continuous version with the same structure and compare its rankings to injection–recovery. If its rankings match equally well, we will say that the contribution reduces to the minimal-change analysis and the bound validation.
Why not just run variance-based global sensitivity (Monte Carlo over all parameters jointly)? It handles simultaneous shifts in many parameters. The hyperbolic sensitivity analysis does not, a limitation the bounds paper acknowledges. Two parameters from different tables enter the polynomial bilinearly, so we can evaluate pairwise shifts exactly. We will compare pairwise interaction rankings against Monte Carlo global indices computed on the same mocks. If strong higher-order interactions appear that pairwise analysis misses, we will report that as a limit of the method.
Proposed Experiments and Decisive Checks
We run three experiments. The first is a fidelity check on mocks. For each of five injected systematics at three strengths, does the network's minimal-change ranking match the pipeline's injection–recovery ranking? This gates everything else. If rank correlation falls below 0.6, binning has destroyed the information that matters, and we stop and report that negative result.
The second validates the bounds. Across all tables and all evidence settings, we measure the share of parameters screened out by each bound type and the miss rate against exact values. This gives the first dataset-level evaluation of evidence-dependent bounds.
The third applies the audit to real SDSS evidence. The decisive outcome comes from comparing the shared-misalignment network and the scale-breaking network. Suppose the shared network can reconcile both results with a small change, in the tables for filament error or blur shear, that lies inside the 10–20 Mpc uncertainty or within plausible blur residuals. Then the tension is a measurement artifact and scale-invariant coupling survives. Suppose instead it requires implausible changes and only the class-dependent lensing table fits. Then the data favor baryonic scale-breaking at halo scales, consistent with the red and blue sign difference reported for $f_h$. As a final check, we will test whether the audit predicts that bright galaxies within 10 Mpc show stronger alignment, a result reported for LOWZ that the network was not fit to.
References
- A Differential Approach to Inference in Bayesian Networks
- Detecting galaxy–filament alignments in the Sloan Digital Sky Survey III
- Ellipticity of dark matter haloes with galaxy-galaxy weak lensing
- Cosmic web reconstruction through density ridges: catalogue
- Sensitivity Analysis of Simulation-Based Inference for Galaxy Clustering
- Revealing the 3D Cosmic Web through Gravitationally Constrained Neural Fields
- A Differential Semantics for Jointree Algorithms
- Making Sensitivity Analysis Computationally Efficient
- Exploiting Evidence-dependent Sensitivity Bounds