The Euler Characteristic Transform (ECT) describes a shape by sweeping a flat plane across it from many directions. At each step it records one integer that counts pieces minus holes plus enclosed voids. A theorem says this description is injective: two different shapes can never produce the same set of curves. That guarantee is why the ECT is used for landmark-free comparison of teeth, cells and CAD models. The theorem, however, assumes infinitely many directions, exact integer counting and noise-free shapes. Real pipelines use about 2 to 100 directions, smooth the counts with sigmoids so gradients flow, and run on jittery scans. Nobody has checked how much of the guarantee survives these choices.
We propose to measure this directly. We will search for collisions: pairs of shapes that are clearly different yet have nearly identical transforms under the settings practitioners actually use. The search runs by gradient descent through the differentiable transform. Any collision it finds is a concrete certificate that injectivity has failed at that setting. The largest geometric gap the search can open is an empirical "collision radius". This one number tells a practitioner how coarse a shape difference their chosen number of directions, smoothing and noise level can no longer see. The goal is a lower bound against which theoretical sufficiency results can be compared, in place of a theorem nobody has tested.
Core Claim
For a transform $T$ computed with $n$ directions, $m$ height thresholds, sigmoid sharpness $s$ and input noise level $\sigma$, define the collision radius of a shape $A$ as
$$r_\epsilon(A) = \sup_B \; d_H(A,B) \quad \text{subject to} \quad \|T(A) - T(B)\| \le \epsilon,$$
where $B$ ranges over shapes with the same mesh connectivity as $A$, $d_H$ is the Hausdorff distance between them, and $\epsilon$ is set to the transform distance produced by noise of level $\sigma$. Exact injectivity means $r_0 = 0$ in the limit of infinitely many directions.
We claim three things, and each can be tested. First, at the direction budgets in current use, $r_\epsilon$ is far from zero on standard benchmarks. Second, $r_\epsilon$ drops sharply at a data-dependent number of directions that is many orders of magnitude below the worst-case bound. Third, $r_\epsilon$ predicts downstream failures better than direction count alone. Those failures are confused class pairs in classification and topological artifacts in reconstruction.
Prior Work and Its Role Here
The theory we test comes from How Many Directions Determine a Shape. That paper proves the ECT and its persistence-based cousin are injective on constructible sets. It also gives a finite bound on the directions that suffice for piecewise-linear complexes, in terms of dimension, feature separation and simplex count. The paper contains no experiments, and the bound has never been evaluated on a real mesh.
We take our method from the differentiable ECT of Differentiable Euler Characteristic Transforms for Shape Classification. That work replaces the step functions with sigmoids so that gradients reach directions and vertex coordinates. It shows that 2 learned directions beat 16 fixed ones on MNIST point clouds. We reuse its layer, but to attack the representation rather than to classify. The same paper does not ask whether the sigmoid relaxation merges distinct shapes, and that is our second axis.
Two papers motivate the inversion side of the gap. Generative Topology for Shape Synthesis concedes that no explicit inversion algorithm exists. It trains a neural decoder from its injective Inner Product Transform back to 2048-point clouds. Randomness of Shapes … Smooth Euler Characteristic Transform states that inverting its smoothed transform remains open. A learned decoder's reconstruction error mixes decoder capacity with information lost in the transform. Collision search isolates the transform itself.
Other work sets constraints on the design. Topology Optimization in Medical Image Segmentation with Fast Euler Characteristic reports that Euler counts are sensitive to single stray voxels. Euler Characteristic Transform Based Topological Loss fixes 100 random directions and thresholds probabilities into binary volumes. Both warn that the noise and discretization axes must be varied separately. Frame theory adds an expectation: Invertibility and robustness of phaseless reconstruction shows that an injective measurement system needs redundancy beyond the minimum before reconstruction becomes stable. We therefore expect the noisy collision radius to fall later than the clean one.
Research Program
The core object is a collision-search procedure. Starting from a dataset shape $A$, we initialize $B$ as a perturbed copy. We then optimize B's vertex coordinates to minimize $\|T(A)-T(B)\|^2 - \lambda\, d(A,B)$. The distance $d$ is a smooth surrogate for Hausdorff distance, and $\lambda$ is swept. We use several random restarts. We also run a second family of starts from other shapes in the same dataset, so that collisions between different class members are found. Each candidate collision is then re-scored with the exact, unsmoothed ECT computed by integer counting. This separates two kinds of failure. Some collisions are genuine coarse-sampling failures of the true transform. Others exist only because of sigmoid smoothing. That separation is what turns the smoothing question (Gap 3 in the file) into a measured quantity.
We sweep four knobs over a grid. Directions run from 2 to 1024, both uniform on the sphere and learned as in DECT. Thresholds run from 16 to 256. Sigmoid sharpness runs from soft to effectively hard. Gaussian jitter $\sigma$ is set as a fraction of shape diameter. For each cell of the grid and each dataset we report the distribution of $r_\epsilon$ over shapes. We also report the elbow, meaning the direction count at which the median radius falls below a chosen fraction of shape diameter.
We will numerically evaluate the Curry–Mukherjee–Turner bound on the same meshes. This is a reference point, not a competitor. If the bound exceeds the empirical elbow by many orders of magnitude, that is the same pattern as conservative-constant results elsewhere in the file. An example is the Nibble clustering guarantees of Spielman et al.. In that case the gap itself is a finding worth reporting.
The search yields only lower bounds on the true supremum. We address this in two ways. On small 2D complexes (MNIST-scale), we approximate an upper bound by dense enumeration of bounded vertex perturbations, which checks how tight the search is. On larger meshes we report the radius together with the number of restarts at which it saturates.
Baselines
The naive reference point is random-pair distortion. We sample many pairs of real dataset shapes and plot transform distance against Hausdorff distance. We then read off the largest geometric gap among pairs whose transforms fall within $\epsilon$. This uses the same transforms, grid and compute accounting as the search. Collision search beats it only if it finds radii clearly larger than random pairs reveal at the same settings, say at least twice as large at the operating points used in published pipelines. If it does not, adversarial search adds nothing.
The strongest existing competitor for "how much shape information survives" is learned inversion with the IP-Encoder of Generative Topology for Shape Synthesis. That paper reports 0.001-second inference and ranks second in Earth Mover's Distance on ShapeNet categories. The numeric EMD values are not in the file, so we have no reported number on Lacuna for its per-category reconstruction error. We retrain it on the identical ShapeNet splits at each direction count, with a matched GPU-hour budget, which the paper reports as under an hour on an RTX 4070. We report its Chamfer distance and EMD next to our radius.
Our measure beats this baseline in a specific sense. It must rank direction settings by downstream failure more accurately than decoder error does. We test this with a rank correlation against the classification and reconstruction failures described below. Decoder error confounds capacity with information, so we expect it to rank settings worse.
For direction-count claims the published anchor is DECT's finding that 2 learned directions outperform 16 fixed ones on MNIST point clouds. We test whether learned directions also shrink the collision radius, or only improve accuracy on the classes being trained.
The key ablation removes the separation term ($\lambda=0$). This reduces the method to plain gradient inversion of the transform. It shows whether deliberately pushing shapes apart is what exposes collisions. A second ablation drops the exact re-scoring step, which shows how many apparent collisions are artifacts of smoothing.
Why not just…?
Why not just compute the theoretical bound and stop? The bound from How Many Directions Determine a Shape is a worst-case figure over a whole class of complexes. It says nothing about how often collisions occur, or how large they are, at budgets below it. Every published pipeline sits below it. We will compute it anyway, and the comparison with the empirical elbow is one of the results.
Why not just sweep direction count and measure reconstruction error of a learned decoder? This is the cheapest serious alternative, and we run it as a baseline. Its weakness is interpretive. A decoder can fail because its architecture is too small, and it can succeed by memorizing the training distribution even when the transform has lost the information. A fixed 2048-point decoder like the IP-Encoder cannot separate these cases. If decoder error and collision radius turn out to rank settings identically across all datasets, we will say so, and the cheaper method wins.
Why not just sample random shape pairs? Collisions are worst-case objects and are likely rare among natural pairs. Random sampling therefore estimates typical distortion rather than failure of injectivity. This alternative is a real threat on homogeneous datasets such as RBC cells, where natural variation may already cover the dangerous directions. The comparison of radii at matched compute described in the Baselines section settles it.
Why not just use persistent homology, which has stability theorems? Stability theorems bound how much the summary changes when the shape moves. They do not bound how far apart two shapes can be when their summaries agree, which is the question asked here. Persistent homology is also reported to be 72.7 times slower than fast Euler counting in Topology Optimization in Medical Image Segmentation with Fast Euler Characteristic, which makes a dense sweep impractical. We include the persistence transform only on the small 2D sets.
Proposed Experiments
The first experiment uses MNIST point clouds and the MPEG-7 silhouettes. Here we calibrate the search against near-exhaustive enumeration and measure how tight the lower bound is.
The second experiment runs the full sweep on ModelNet40 and on the ShapeNet airplane, chair and car splits used by the IP-VAE. Its decisive check is predictive. For each direction setting, we test whether pairs of classes whose shapes collide at a small radius are the pairs DECT confuses in classification. A significant correlation validates the radius as a practical diagnostic. No correlation would mean that collisions exist but do not matter for these tasks, and we would report that negative result.
The third experiment uses the RBC and Nuclei voxel volumes. There the ECT loss reduced volume error by 11.2% and 19.6% over the SHAPR baseline, per Euler Characteristic Transform Based Topological Loss. We ask whether 100 directions lie past the elbow. Concretely, we retrain with the direction count set at, below and above the measured elbow. If gains vanish below the elbow and saturate above it, the radius is useful for choosing budgets.
The fourth experiment adds jitter matched to ScanNet-like conditions. Here Deep Point Cloud Reconstruction reports a Chamfer distance of 1.91 for its denoiser, against 2.66 for the best baseline. We measure how much the elbow shifts under noise, as a direct estimate of the extra redundancy that frame theory predicts.
As a qualitative check, we apply the search to the 59 primate molars. We ask whether collisions at the settings used by SINATRA can move the paraconid cusp, which is the feature SINATRA identified. If they can, its landmark-free conclusions hold only above some direction budget, and we can say which.
References
- How Many Directions Determine a Shape and other Sufficiency Results for Two Topological Transforms
- Differentiable Euler Characteristic Transforms for Shape Classification
- Generative Topology for Shape Synthesis
- Randomness of Shapes and Statistical Inference on Shapes via the Smooth Euler Characteristic Transform
- Topology Optimization in Medical Image Segmentation with Fast Euler Characteristic
- Euler Characteristic Transform Based Topological Loss for Reconstructing 3D Images from Single 2D Slices
- Invertibility and robustness of phaseless reconstruction
- A Local Clustering Algorithm for Massive Graphs and Its Application to Nearly Linear Time Graph Partitioning
- Deep Point Cloud Reconstruction
- A statistical pipeline for identifying physical features that differentiate classes of 3D shapes