Any smooth wind field drawn on the globe must be zero somewhere. This is the hairy ball theorem: a continuous field of arrows on a sphere cannot point in a nonzero direction everywhere. More precisely, the Poincaré–Hopf theorem says the zeros, counted with a sign called their index (+1 for vortices and sources, −1 for saddles), must sum to 2. When we fit a model to scattered wind measurements, these zeros (calm points, cyclone centres, saddles between circulation cells) are among the most interpretable things the model produces. Nobody asks how much they depend on individual measurements. We propose a leave-one-out analysis whose target is the defects themselves, not the model's parameters or its pointwise predictions. For each observation we ask two things: if it were deleted, how far would each zero move, and would zeros be created or destroyed?
The opportunity is that this question has a clean and cheap answer in exactly the setting where it matters. Vector-valued Gaussian Processes on Riemannian Manifolds via Gauge Independent Projected Kernels builds Gaussian processes (GPs: probabilistic regressors defined by a covariance function) for tangent vector fields on the sphere that do not depend on an arbitrary choice of local axes. Deleting one observation from a GP is a closed-form rank-one update. Moving a zero is governed by the field's derivative at that zero, and that derivative is coordinate-free precisely because the field vanishes there. Put together, these give a defect-level influence score that does not change when the coordinate axes are rotated. The score also predicts its own failure: it diverges when a zero is about to split or annihilate. At that point we hand over to exact recomputation.
Core Claim
Let $\mu$ be the GP posterior mean field on the sphere and $x^\*$ one of its zeros. Let $J(x^\*)$ be the $2\times 2$ derivative of $\mu$ at $x^\*$, expressed in any tangent basis, and let $\Delta\mu_{(-i)}$ be the change in the mean field when observation $i$ is removed. The implicit function theorem gives the first-order displacement of the zero:
$$\delta x^\*_{(-i)} \approx -\,J(x^\*)^{-1}\,\Delta\mu_{(-i)}(x^\*), \qquad \operatorname{ind}(x^\*) = \operatorname{sign}\det J(x^\*).$$
At a zero, the connection terms that normally make derivatives of vector fields coordinate-dependent multiply $\mu(x^\*)=0$ and drop out. As a result, both the displacement and the index are gauge independent, meaning they do not change under a rotation of the local axes. The factor $1/|\det J|$ shows where influence concentrates. Zeros with a small $|\det J|$ are close to a bifurcation, such as a vortex–saddle pair about to merge. For those zeros, a single deletion can move the zero a long way or remove a pair entirely, which changes the defect count while leaving the index sum fixed at 2.
Our central empirical claim is that defect influence is a different quantity from standard leave-one-out diagnostics. Specifically, the observations that most change predictions, as measured by the usual GP leave-one-out residuals, are largely not the ones that move or annihilate defects. The latter should instead concentrate near low-$|\det J|$ zeros in sparsely observed regions.
Research Program
We would build three components on top of the projected-kernel GP of Vector-valued Gaussian Processes on Riemannian Manifolds via Gauge Independent Projected Kernels. That paper solves the problems of axis-choice dependence and seams at the 180° meridian, but it does not study how sensitive the fit is to individual data. The first component is a zero finder on the sphere. It searches for sign changes on an icosahedral mesh, refines candidates by Newton steps in the embedding space projected back onto the sphere, and certifies each zero's index with a winding-number check on a small loop around it. The second component computes the linearised displacement above for every (observation, zero) pair. After one Cholesky factorisation, the rank-one downdates cost $O(N)$ per observation.
The third component is a bifurcation guard. For any pair where the predicted displacement exceeds a fraction of the distance to the nearest other zero, or where $|\det J|$ falls below a threshold, we recompute the fit exactly with that observation removed and re-run zero finding and winding-number checks. This catches creation and annihilation events, which a first-order formula cannot represent by construction. The same limitation of first-order methods motivates Testing Most Influential Sets, which shows that first-order influence underestimates the effect of deleting groups of points.
We then extend the analysis from single observations to small sets and ask whether the largest defect displacement is excessive. Testing Most Influential Sets derives Fréchet and Gumbel limits for maximal influence, but only for ordinary least squares in flat space. We do not claim to extend that theory. Instead, we reuse its block-maxima testing procedure with a simulated null distribution. We draw synthetic datasets from the fitted GP at the same observation locations, compute the maximum defect displacement for each, and use the resulting distribution to calibrate a $p$-value for the real maximum. This makes the method a tool and not a theorem, and we say so.
The pieces we reuse come from specific papers. The projected kernel supplies gauge-free fields. The block-maxima test supplies the calibration logic. Addressing the Topological Defects of Disentanglement and Learning Stable Vector Fields on Lie Groups motivate the premise that defects are forced by topology and cannot be optimised away. The remaining question is only which data decide where the defects sit. Riemannian Continuous Normalizing Flows learns neural vector fields on the sphere that also must have stagnation points. We treat it as a secondary target, handled only by exact refitting, because it has no closed-form deletion update.
Baselines
The naive reference point is to ignore defects. We report standard GP leave-one-out predictive residuals and a Cook's-distance-style score for each observation, both computed in closed form from the same factorisation. We then take the zeros' movement as whatever those rankings imply. If this reference ranks observations by their effect on defects as well as our score does, then the defect-level target adds nothing. We measure this by Spearman correlation with exact displacement and by precision at detecting creation and annihilation events.
The strongest competitor is exact deletion with the same projected-kernel GP. For every observation, we refit with fixed hyperparameters and track zeros by the same finder. This is the ground truth, not something to beat on accuracy. The question is whether the linearised score plus the guard reproduces it at much lower cost.
No reported result exists on Lacuna for defect-level influence on any benchmark. For the wind benchmarks, no reported number was found on Lacuna for the GP itself either. The nearest fidelity bar comes from general influence estimation. Clustered Influence Functions reports Spearman ρ > 0.95 against exact influence while giving 15.3× to 32.9× speedups on CIFAR-10 and ImageNet. We adopt the same bar: Spearman ρ ≥ 0.95 against exact displacement, with recall of at least 0.9 on events that change the defect count, at a wall-clock cost at least an order of magnitude below exact deletion on the same hardware.
A second strong competitor is a flattened latitude–longitude GP with Euclidean leave-one-out analysis, the practice most users would default to. It is run on the same observations with an equal budget for hyperparameter tuning.
An ablation removes the bifurcation guard to isolate how much of the event recall comes from exact fallback rather than from linearisation. A second ablation replaces $J$ with the plain Jacobian in latitude–longitude components. This isolates the contribution of gauge independence.
Why not just…?
Why not just refit exactly for every deletion and skip the linearisation? For moderate $N$ this works, and it is our ground truth. The contribution is not mainly speed. It is the choice of target (defects rather than parameters) and the $1/|\det J|$ account of where instability lives. That account tells the analyst in advance which zeros are fragile, which exact refitting does not. Exact deletion becomes costly when zero finding over a fine mesh is repeated $N$ times on satellite-scale data. It becomes very costly for subset searches, where the number of candidate subsets grows combinatorially. Our cost comparison will show where the crossover lies. If exact deletion is fast enough at realistic Aeolus sizes, we will report the method as a diagnostic lens and not as an acceleration.
Why not just flatten to latitude–longitude and run standard influence analysis? The source paper already shows that flattened GPs produce spurious uncertainty jumps at the 180° meridian and distortion at the poles. Riemannian Continuous Normalizing Flows reports a stereographic baseline degrading to NLL 63.60, against −1.78 for the native model on a sphere task. More specifically to influence, the east and north components of a derivative change under a rotation of the local frame near the poles. We will test this directly by applying random rotations of the tangent frame and reporting the variance they induce in each score. Our score should show zero variance up to numerical precision; the flattened score should not.
Why not just use posterior samples to get uncertainty on defect locations? This is a real alternative, and it answers a different question. Posterior spread describes what the unobserved field might look like given all the data. Leave-one-out analysis asks whether a conclusion rests on one particular measurement, for example a single dropsonde or a single faulty satellite retrieval. The two can disagree. A zero can be tightly determined in the posterior because one high-precision observation sits on it, which is exactly the case where deleting that observation matters most. We will report the correlation between posterior location spread and maximal leave-one-out displacement. If the correlation exceeds 0.9 across benchmarks, the leave-one-out view is largely redundant and we will say so.
Proposed Experiments
We start with synthetic fields on the sphere with known zeros, including a tunable saddle–vortex pair whose separation controls $|\det J|$. The fields are sampled at random or clustered locations with noise. This tests whether linearised displacement matches exact displacement away from bifurcations and whether the guard catches annihilation events as the pair's separation shrinks.
The pendulum-with-friction benchmark on the cylinder, from Vector-valued Gaussian Processes on Riemannian Manifolds via Gauge Independent Projected Kernels, provides known ground-truth equilibria: a stable bottom point of index +1 and an upright saddle of index −1. We test whether deleting trajectory samples near the upright position causes the learned dynamics to grow spurious equilibrium pairs. This is the physically meaningful failure.
The main application uses ERA5 reanalysis subsampled at Aeolus-like observation tracks. Because the dense ERA5 field is available, we can check whether a defect identified as fragile by our score also disagrees with the true field. As a secondary, exact-refit-only study, we fit the earthquake density flow of Riemannian Continuous Normalizing Flows and ask whether its stagnation points are driven by a handful of isolated events.
The decisive checks are four:
- Spearman correlation of at least 0.95 with exact displacement, and recall of at least 0.9 on events that change the defect count, at a large cost reduction.
- Zero change under frame rotation, compared with a clear change for the flattened baseline.
- Low overlap between the top-ranked observations by defect influence and by residual influence, which is the core claim. Overlap near one would falsify it.
- A calibrated block-maxima test that returns uniform $p$-values on GP-simulated data and flags deliberately planted corrupted observations near fragile zeros.