Many regression targets in science and medicine are really counts of shapes: how many cells are in a micrograph, how many lesions are in a scan, how many voids are in a galaxy survey. Topological data analysis offers a principled way to count such structures. Persistent homology, a method that tracks how long each blob or loop survives as you sweep a threshold across the data, separates long-lived features (signal) from short-lived ones (noise). In practice, though, every topological counter hides choices: how much to smooth, which neighbourhood size to use, and where to cut the lifespan. Those choices decide whether a faint, short-lived feature is counted. Current methods either fix them by hand or bootstrap under one fixed choice. Neither tells you whether a borderline feature is doubtful because the measurements are noisy or because the method cannot know.
We propose to treat the whole family of reasonable filtering choices as an imprecise probability model. A credal set, meaning a set of candidate probability distributions rather than a single one, gives every topological feature a lower and an upper probability of being real. Turned into counts, this gives a lower and an upper count. The bet is that this interval keeps its coverage when the test data shift, which is exactly where learned regressors fail. On the Cells-Tails cell-counting benchmark, a Gaussian ensemble reaches only about 59% coverage against a 90% target (How Reliable is Your Regression Model's Uncertainty Under Real-World Distribution Shifts?). The reason is plausible: a model trained to map images to numbers cannot extrapolate to counts it never saw, whereas counting connected components does not care how many components there are.
Core Claim
A topological counter whose uncertainty is a credal envelope over filtering hyperparameters and resampling will do two things that neither precise topological filtering nor precise regression UQ does. First, it will hold near-nominal interval coverage under target-value shift without any calibration data from the shifted domain. Second, its two numbers will separate the two kinds of doubt. The width of a feature's existence interval (upper minus lower probability) will track epistemic ambiguity, meaning sparse support or sensitivity to the filtering choice. The location of the interval will track aleatoric measurement noise. Persistence diagrams on their own treat these two cases identically.
The second claim is the part with no precedent in the file. There are no papers combining persistent homology with credal sets. The closest topological work, SCHU (Finding cosmic voids and filament loops using topological data analysis), bootstraps confidence bands under a single, hand-tuned smoothing parameter $m_0$. The closest credal work, CreDRO (Learning Credal Ensembles via Distributionally Robust Optimization), builds box credal sets from ensembles trained with different robustness levels, but only for classification. We take CreDRO's idea that ensemble diversity should come from deliberately varied modelling assumptions rather than random seeds. We then apply it to the assumptions inside a filtration rather than inside a network.
Research Program
Precise construction. For an input image or point cloud, each configuration $\theta$ in a box $\Theta$ fixes three things: a smoothing bandwidth (or the distance-to-measure mass parameter $m_0$), the neighbourhood sizes $k, k'$ of the outlier-robust denoiser, and a lifespan cutoff. The denoiser is the Combined Noise Algorithm of Topological analysis of scalar fields with outliers, which uses distance-to-measure for misplaced points and a low-variance neighbour average for corrupted values. That paper supplies the method and the interleaving guarantee we rely on for each fixed $\theta$. Under each $\theta$, bootstrap or noise-perturbation resampling gives a distribution $P_\theta$ over persistence diagrams. The credal set is the convex hull of $\{P_\theta : \theta \in \Theta\}$.
Per-feature and count intervals. For the event $A_j$ that a matched feature $j$ survives the cutoff, we report
$$\underline{P}(A_j)=\min_{\theta\in\Theta}P_\theta(A_j),\qquad \overline{P}(A_j)=\max_{\theta\in\Theta}P_\theta(A_j).$$
For the count $N$, the interval runs from the smallest lower $\alpha/2$ quantile to the largest upper $1-\alpha/2$ quantile across $\theta$. The lower and upper envelopes are exact because $\Theta$ is a small grid, so we avoid the exponential-cost credal computations flagged in Integral Imprecise Probability Metrics.
Fitting the hyperparameter box. $\Theta$ is fit on in-distribution training images only. We choose the smallest box whose envelope covers 90% of training counts. Because the box governs image-level filtering rather than the count range, it should transfer to shifted count ranges. That transfer is the central empirical bet.
Three-way decision. The per-feature intervals give a keep / discard / flag rule. A feature is kept if $\underline{P}$ is high and discarded if $\overline{P}$ is low. Otherwise it is flagged as undetermined. This rule is our concrete answer to the "subtle feature versus noise" problem raised by Ayman Fakhouri et al.: short-lived features are not thrown away silently, but surfaced when the evidence cannot decide. It mirrors interval-width abstention in classification (Hüllermeier et al.).
Axioms. The count interval's width is non-negative and unchanged by shifting the output. It therefore avoids the negative-entropy and location-dependence failures that An Axiomatic Assessment of Entropy- and Variance-based Uncertainty Quantification in Regression proves for differential entropy. We will check its other axioms (zero width under full agreement, growth with spread) analytically for the envelope construction.
Testing whether the credal set is right. We adapt the containment test of A calibration test for evaluating set-based epistemic uncertainty representations, whose authors list regression as open future work. The adapted test asks whether the empirical distribution of counts lies between the lower and upper count distribution functions.
Scope limit. The approach only applies to targets that are functions of topological summaries, such as counts, numbers of loops, or summed component areas. We do not claim a general credal regression method.
Baselines
Naive reference: precise topological counter. This is the topological counter with a single tuned $\theta$ and no interval, plus the same counter with a fixed $\pm$ interval sized on training data. It establishes how much of any gain comes from topological counting itself rather than from imprecision.
Strongest existing regression UQ. These are the regression methods evaluated by How Reliable is Your Regression Model's Uncertainty Under Real-World Distribution Shifts?. The Gaussian ensemble is the reported reference on Cells-Tails, at about 59% test coverage against a 90% target. On HistologyNucleiPixels the paper reports that most methods reach only 55%–70%. Split conformal regression is included in that study and is the method a reviewer will most expect. We will rerun these methods with the authors' splits and the same training images, compute budget, and hyperparameter search budget as our box fitting. To beat them, we need coverage at or near 90% on the shifted Cells-Tails test set, with mean interval width reported alongside, so that coverage is not bought with trivially wide intervals.
Strongest topological UQ: SCHU-style bootstrap. This baseline bootstraps diagrams under one tuned $\theta$, following Finding cosmic voids and filament loops using topological data analysis. It is the fair competitor for the disentanglement claim, because it is our method with the credal envelope removed.
Ablations. We separate the two sources of imprecision by building one interval from the hyperparameter envelope with no resampling, and another from resampling alone. We also run a conformalized topological counter, described below.
Why not just…?
Why not just conformalize a topological counter? This is the real threat. If a counter's residuals do not depend on the count range, then split conformal calibration on in-distribution data might already keep 90% coverage under target shift, and the credal machinery would add nothing to coverage. We run this exact baseline. If it matches our coverage at similar width, we concede the coverage claim. In that case the contribution narrows to instance-adaptive width and to the per-feature flagging, which conformal intervals of fixed width cannot give. We will test that remaining contribution by correlating interval width with per-image count error.
Why not just bootstrap the persistence diagram, as SCHU does? Bootstrap resampling measures sampling variability under one smoothing choice. It does not represent the hand-tuned $m_0$, which the file identifies as a known weakness. The synthetic experiment below varies sparsity and noise independently, so it shows directly whether bootstrap width can tell them apart. If bootstrap width tracks sparsity as well as credal width does, the imprecise layer is unnecessary for disentanglement.
Why not just train a better image regressor or a deep ensemble? The cited shift study found that no method reached 90% coverage on any real-world shifted dataset. CreDRO's analysis also suggests that seed-based ensembles capture optimization noise rather than shift. A learned direct-count regressor still has to extrapolate to unseen count values. Our comparison on Cells-Tails tests this directly.
Proposed Experiments
Experiment 1: controlled disentanglement check. This is the decisive test of the second claim. We generate point clouds with known loops and voids, in the style of the Voronoi-foam setup, where SCHU recovered 23 of 23 voids. We then independently vary local sampling density (epistemic) and coordinate or value noise (aleatoric). The prediction is that interval width rises with sparsity while staying roughly flat in noise, and that interval location moves with noise. Bootstrap-only width should not show this separation. If credal width responds to both factors equally, the disentanglement claim fails.
Experiment 2: coverage under target shift. On Cells-Tails, we report coverage and mean width at a 90% target against all baselines, with the ablations included. On HistologyNucleiPixels, where the target is a summed nuclei area, we apply the same construction to areas of superlevel components and treat this as a harder, secondary test.
Experiment 3: real-data plausibility check. We rerun SDSS void detection and ask how SCHU's 15 significant voids out of 311 candidates split under credal persistence into kept, discarded and flagged. We will inspect whether flagged voids sit in sparsely sampled survey regions.
References
- How Reliable is Your Regression Model's Uncertainty Under Real-World Distribution Shifts?
- Finding cosmic voids and filament loops using topological data analysis
- Learning Credal Ensembles via Distributionally Robust Optimization
- Topological Analysis of Scalar Fields with Outliers
- Integral Imprecise Probability Metrics
- ML Model Coverage Assessment by Topological Data Analysis Exploration
- Quantification of Credal Uncertainty in Machine Learning: A Critical Analysis and Empirical Comparison
- An Axiomatic Assessment of Entropy- and Variance-based Uncertainty Quantification in Regression
- A calibration test for evaluating set-based epistemic uncertainty representations