Euler Characteristic Transform Overview

• ECT is a topological invariant that encodes multiscale geometric and topological features of shapes via Euler characteristics computed over half-space filtrations.
• It applies to varied data types like simplicial complexes, point clouds, and images, and supports powerful extensions such as weighted, lifted, and smooth transforms.
• With guarantees of injectivity, stability, and efficient computation, ECT is pivotal for shape analysis, statistical inference, and deep learning integration.

The Euler Characteristic Transform (ECT) is a topological data analysis (TDA) invariant that encodes the multiscale geometric and topological structure of a shape by reporting its Euler characteristic across a family of half-space filtrations. The ECT takes a shape such as a compact definable set or simplicial complex in $\mathbb{R}^n$, slices it in all directions on the unit sphere $S^{n-1}$, computes the induced sublevel sets at varying thresholds along each direction, and records the Euler characteristic of these sublevel sets as a function of direction and threshold. This process yields a multivariate, piecewise-constant integer-valued function,

$$\mathrm{ECT}(S)(v, t) = \chi\left(\{x \in S : \langle x, v \rangle \leq t\}\right), \qquad (v, t) \in S^{n-1} \times \mathbb{R},$$

which is provably injective: for broad classes of shapes, the ECT uniquely determines the original set up to definable homeomorphism (Ji, 2023, Ji et al., 2023, Curry et al., 2018). The transform admits extensions—including weighted, lifted, and smooth forms—and is accompanied by established stability results, explicit discretization strategies, and scalable algorithms for computation and inversion.

1. Formal Definition and Mathematical Properties

Let $S \subset \mathbb{R}^n$ be a compact definable set in an o-minimal expansion of the real field. For each direction $v \in S^{n-1}$ and threshold $t \in \mathbb{R}$, define the sublevel set

$$S^v_t = \{ x \in S : \langle x, v \rangle \leq t \}.$$

The Euler characteristic of $S^v_t$, denoted $\chi(S^v_t)$, is computed via definable cell decomposition and is always an integer. The Euler Characteristic Transform is then defined as

$\mathrm{ECT}(S)(v, t) = \chi\left(S^v_t\right).$

Alternatively, in the language of Euler calculus and constructible functions, for an indicator function $_{S}$,

$$\mathrm{ECT}(S)(v, t) = \int_{\mathbb{R}^n}~_{S}(x)~_{ \{x \cdot v \leq t\} }(x)~d\chi(x).$$

This function is right-continuous in $t$ for fixed $v$, with only finitely many jump discontinuities (Ji et al., 2023).

The ECT can be extended in several directions:

• Weighted ECT (WECT): Incorporates a weight function $\omega : S \to \mathbb{R}$, yielding

$$\mathrm{WECT}(S, \omega)(v, t) = \sum_{\sigma \subset S,\, H_v(\sigma) \leq t} (-1)^{\dim(\sigma)}\,\omega(\sigma).$$

• Quadric ECT (QECT): Replaces linear half-spaces with quadric hypersurfaces,

$$\mathrm{QECT}(S)(A, v, t) = \chi\left(\{ x \in S : x^T A x + v \cdot x \leq t \}\right),$$

with $A$ symmetric.

Smoothing the transform via integration produces the Smooth ECT (SECT), which lives in a Hilbert space and admits analytic stability arguments (Ji et al., 2023, Marsh et al., 2022).

2. Injectivity, Continuity, and Stability Theorems

The ECT is injective for compact definable (or constructible) sets:

• Injectivity (Schapira inversion): The map $S \mapsto \mathrm{ECT}(S)$ is injective for compact definable sets, i.e., $\mathrm{ECT}(S_1) = \mathrm{ECT}(S_2)$ forces $S_1 = S_2$ up to definable homeomorphism (Ji, 2023, Curry et al., 2018, Ji et al., 2023).
• Right-continuity: For fixed $v$, the function $t \mapsto \chi(S^v_t)$ is right-continuous with finitely many jumps (Theorems 3.1–3.2 in (Ji et al., 2023)).
• Deformation-retraction: For $S$ compact definable and $f:S\rightarrow\mathbb{R}$ continuous definable, the inclusion $S^f_t \subset S^f_{t+\delta}$ is a deformation retraction for small $\delta > 0$, yielding right-continuity of Betti numbers (Theorem 4.1 in (Ji et al., 2023)).
• Stability: For embedded simplicial complexes with bounded curvature, the $L^1$ distance between their ECTs is bounded above by the sum of vertex displacements, i.e.,

$$d_{\mathrm{ECT}}(\mathrm{ECT}(X), \mathrm{ECT}(Y)) \leq 2 C_K C_d \sum_{v \in V(K)} \|f(v) - g(v)\|_2,$$

where $C_K$ and $C_d$ are mesh and dimension-dependent constants (George et al., 24 Jun 2025, Marsh et al., 2023). This provides rigorous Lipschitz-type control over perturbations.

3. Computational Algorithms and Practical Discretization

Given $S$ or a simplicial complex $K \subset \mathbb{R}^n$:

• Discretization: Directions $v$ are sampled uniformly (e.g., via icosahedral meshes or spherical designs), and thresholds $t$ are discretized over the projection range of vertices (Munch, 2023, Rieck, 2024).
• Filtration and streaming update: For each direction, simplex values $p_v(\sigma)$ are computed, and the Euler characteristic is updated incrementally as each new filtration level is reached (Cisewski-Kehe et al., 5 Nov 2025, Jiang et al., 2020). This is $O(N \cdot T)$ for $N$ directions and $T$ simplices.
• Exact computation: The digital ECT algorithm avoids sampling bias by using spherical cell decomposition and symbolic integration to compute closed-form ECT values and shape distances. This is implemented in the Ectoplasm package (Kirveslahti et al., 2024).
• Vectorized GPU computation: PyTorch implementations using tensor primitives yield up to $180 \times$ speedup and scale to high-dimensional, high-complexity data (Cisewski-Kehe et al., 5 Nov 2025).

The algorithmic pipeline is thus suitable for meshes, point clouds, graphs, and images, admitting incremental updates and vectorization for large datasets (Cisewski-Kehe et al., 2023).

4. Extensions: Weighted, Lifted and Local Transforms

• Weighted ECT (WECT): Encodes additional data via admissible weight functions. The weighted transform $\mathrm{WECT}(K, g)(v, r)$ aggregates the signed weights of all simplices in the subcomplex at $(v, r)$; injectivity is preserved (Jiang et al., 2020, Cisewski-Kehe et al., 2023).
• Lifted/Super-Lifted ECT (LECT/SELECT): Generalizes ECT to scalar fields $f : S \to \mathbb{R}$, producing

$$\mathrm{SELECT}(f)(v, h, t) = \chi\left(\{ x : x \cdot v \leq h,\; f(x) \geq t \}\right),$$

which is injective on large classes of piecewise-linear fields (Kirveslahti et al., 2021, George et al., 24 Jun 2025).

• Local ECT ($\ell$-ECT): Focuses on neighborhoods, yielding lossless local fingerprints for graphs and complexes. The $\ell$-ECT provides strong expressivity, surpassing message-passing GNN layers, with metrics that are rotation-invariant up to alignment by $SO(n)$ actions (Rohrscheidt et al., 2024).

5. Applications in Data Science and Machine Learning

• Shape analysis: ECT is widely used for shape comparison, classification, and morphometry in biology (e.g., bones, seeds), medical imaging (brain tumors), and molecular machine learning. Its injectivity guarantees lossless shape encoding (Turner et al., 2013, Cisewski-Kehe et al., 2023, Toscano-Duran et al., 4 Jul 2025, Marsh et al., 2022).
• Statistical inference: In exponential-family models, discrete ECT samples serve as sufficient statistics. ECT features integrate naturally into support-vector machines, random forests, kernel regression, and deep networks (Turner et al., 2013, Cisewski-Kehe et al., 2023).
• Deep learning integration: Differentiable ECT layers leverage smooth sigmoid approximations for end-to-end learning, retaining topological expressivity and delivering competitive or superior accuracy on point cloud and graph classification benchmarks (Roell et al., 2023, Rieck, 2024). Vectorized GPU computation and differentiable forms enable seamless deployment in modern pipelines.
• Temporal analysis: The DETECT framework averages smoothed ECTs over directions for rotationally invariant signatures, enabling quantification and classification of shape dynamics in time-series imaging (e.g., organoid growth regimes) (Marsh et al., 2022).

6. Theoretical and Empirical Guarantees: Sufficiency, Rotation-Invariance, and Open Problems

• Finite direction sufficiency: For piecewise-linear shapes with geometric bounds, only finitely many directions are necessary to recover the shape; explicit polynomial bounds in curvature and complexity are established (Curry et al., 2018, Kirveslahti et al., 2021).
• Rotation invariance: ECT-based metrics can be made invariant to $SO(n)$ alignment via spherical integration or explicit adaptive/grid search algorithms, supporting shape alignment and equivariance (Kirveslahti et al., 2024, Rohrscheidt et al., 2024).
• Stability under noise: Lipschitz-type stability holds under both geometric and curvature constraints, and consistent statistical estimators for the ECT can be constructed via Gaussian process smoothing on noisy samples (Marsh et al., 2023, George et al., 24 Jun 2025).
• Inversion and subshape selection: Theoretical inversion via Euler–Radon integral formulas is established, and digital ECT algorithms expose localized shape contributions, facilitating subshape selection and partial reconstructions (Kirveslahti et al., 2024).
• Open problems: Minimal direction counts for injectivity across classes, theoretical convergence bounds for finite samples, stability under extreme noise, and extensions to infinite-dimensional or non-definable settings remain active areas of investigation (Ji et al., 2023, Curry et al., 2018).

7. Summary Table: Principal ECT Variants and Properties

Transform	Domain	Definition	Injectivity
ECT	Compact definable set	$\mathrm{ECT}(S)(v, t) = \chi(S^v_t)$	Yes (Ji, 2023)
WECT	Weighted complex	$$\mathrm{WECT}(K, \omega)(v, t) = \sum_{\sigma \in K,\, H_v(\sigma) \leq t} (-1)^{\dim \sigma}\omega(\sigma)$$	Yes (Jiang et al., 2020)
QECT	Definable set, quadric slice	$$\mathrm{QECT}(S)(A, v, t) = \chi(\{x : x^T A x + v \cdot x \leq t\})$$	Yes, under bounds
SELECT (Lifted ECT)	Definable scalar field	$$\mathrm{SELECT}(f)(v, h, t) = \chi(\{ x : x \cdot v \leq h,\, f(x) \geq t \})$$	Yes (Kirveslahti et al., 2021)
$\ell$-ECT (local)	Graph/complex neighborhood	$\ell$-ECT$_k(x; X) := \mathrm{ECT}(N_k(x; X))$	Yes (Rohrscheidt et al., 2024)

The Euler Characteristic Transform and its extensions constitute a central, rigorously-characterized toolset for topological data analysis, enabling high-fidelity, multi-scale representation, alignment, and inference of shapes, fields, and structures in scientific and engineering contexts.