Fractal Whitney Embedding Prevalence

• The paper's main contribution is demonstrating that generic smooth maps embed fractal sets into Euclidean space when the target dimension exceeds twice the box-counting dimension.
• It employs transversality conditions to show that failure maps form a shy set, ensuring that almost every map is both injective and immersive on fractal structures.
• The theorem underpins practical techniques in machine learning, enabling reliable latent embeddings for attractor reconstruction and enhancing anomaly detection methods.

The Fractal Whitney Embedding Prevalence Theorem is a generalization of the classical embedding theorems of Whitney and Takens for dynamical systems theory. It guarantees that generic smooth maps can embed compact sets of arbitrary (non-integer, "fractal") box-counting dimension into finite-dimensional Euclidean space, provided the target dimension exceeds twice the box-counting dimension. This theorem formalizes conditions under which almost every sufficiently smooth map is simultaneously injective on the set and immersive on each of its smooth pieces, thus ensuring a topological and differential embedding. Its significance extends to the design of latent embeddings in machine learning architectures for time-series and dynamical system data, where it underpins theoretical guarantees for both representation fidelity and anomaly detection robustness.

1. Formal Statement and Terminology

Let $A\subset\mathbb{R}^k$ be a nonempty compact set with box-counting dimension $d<\infty$, and let $C^r(\mathbb{R}^k,\mathbb{R}^n)$ denote the space of $r$-times continuously differentiable maps from $\mathbb{R}^k$ to $\mathbb{R}^n$, equipped with the $C^r$-topology. The theorem asserts that if $n>2d$, then the set

$$\mathcal{G} = \left\{ F\in C^r(\mathbb{R}^k,\mathbb{R}^n) : F|_A \text{ is injective on } A, \ F \text{ is an immersion on each smooth piece of } A \right\}$$

is prevalent: the complement (failure maps) is a shy set in $C^r$.

Essential definitions:

• Box-counting dimension: For a compact $A\subset\mathbb{R}^k$,

$$\operatorname{boxdim}(A) = k - \lim_{\varepsilon\to0} \frac{\log(\operatorname{vol}(A_\varepsilon))}{\log \varepsilon},\quad A_\varepsilon = \{x\in\mathbb{R}^k : \exists a\in A, \|x-a\|\le \varepsilon\}$$

• Prevalence: In an infinite-dimensional Banach space $E$, a subset $S\subset E$ is shy if for some compactly supported Borel measure $\mu$, $\mu(x+S)=0$ for all $x\in E$. A set is prevalent if its complement is shy. This serves as the infinite-dimensional analogue of "full measure."
• Immersion: A map $F:M\to N$ is immersive at $p\in M$ if its differential $dF_p:T_pM\to T_{F(p)}N$ is injective, thus embedding tangent directions at $p$ into the target.

Key hypotheses are compactness of $A$, finiteness of box-counting dimension $d$, target space dimension $n>2d$, and differentiability $r\geq1$.

2. Underlying Propositions and Proof Architecture

The proof constructs the set of "bad" maps—those failing injectivity or immersion—as a countable union of submanifolds of strictly positive codimension. By properties of prevalence (“shy” sets), this union is itself shy, so its complement is prevalent.

• Finite-Pair Transversality: For $x\neq y\in A$, define

$$E_{x,y}:C^r(\mathbb{R}^k,\mathbb{R}^n)\to\mathbb{R}^n,\quad E_{x,y}(F)=F(x)-F(y)$$

The set where $F(x)=F(y)$ is a submanifold of codimension $n$, and since $n>2d$, these "collision" sets are sufficiently high codimension to be shy in union.

• Jet-Transversality for Immersion: For each $p\in A$,

$$J^1_p: C^r(\mathbb{R}^k,\mathbb{R}^n) \to \mathrm{Hom}(T_p\mathbb{R}^k, \mathbb{R}^n), \quad J^1_p(F)=dF_p$$

The set of maps where $dF_p$ fails to have full rank corresponds to a submanifold of positive codimension.

Combined, the set of non-injective or non-immersive maps is shy; thus, most ("prevalent") $C^r$ maps are topological and differential embeddings for $A$.

3. Connection to Classical Embedding Theorems

Whitney’s classical theorem states that any smooth $m$-dimensional manifold can be embedded in $\mathbb{R}^{2m+1}$. Takens’ delay-coordinate theorem (1981) extends this to the reconstruction of attractors in dynamical systems, showing that a delay map embeds an attractor of box dimension $d$ into $\mathbb{R}^{2d+1}$ for generic observables.

The Fractal Whitney Embedding Prevalence Theorem relaxes the requirement of integer Hausdorff dimension, requiring only the existence of a finite box-counting dimension $d$. It guarantees the existence of a prevalent set of smooth maps embedding any compact $A$ with $\operatorname{boxdim}(A)=d$ into $\mathbb{R}^n$, $n>2d$, preserving both topological structure and local geometry on smooth pieces. Thus, it provides the foundational guarantee for data-driven embeddings of non-manifold, fractal, or highly non-uniformly distributed sets.

4. Illustrative Corollaries and Case Studies

Two primary corollaries highlight the theorem's practical implications:

• Takens–Sauer–Yorke–Casdagli Corollary: For a compact invariant set $A$ of a dynamical system $\varphi$ with box-counting dimension $d$, and for almost every observable $h$, the delay-coordinate map

$$\Phi_h(x) = (h(x), h(\varphi(x)), \ldots, h(\varphi^{n-1}(x)))$$

is injective on $A$ provided $n>2d$.

• Lorenz Attractor Application: Numerical studies estimate $$\operatorname{boxdim}(\mathrm{Lorenz}) \approx 2.06$$, so the theorem guarantees that for $n > 4.12$, any generic smooth embedding (e.g., $n=5$) will be a topological embedding.

These results enable dimensionality reduction and attractor reconstruction for sets with fractal geometry.

5. Role in Anomaly Detection and Representation Learning

The theorem establishes theoretical conditions for latent representation schemes in complex dynamical systems. Specifically, it justifies the following key methodology choices (Somma et al., 26 Feb 2025):

• Embedding dimension selection: Estimate the box-counting dimension $d$ of the dynamical attractor; select $n > 2d$ as the latent space dimension to ensure generic smooth encoders are embeddings.
• Encoder genericity: The requirement of prevalence means a trained neural encoder of sufficient smoothness and width can generically act as the embedding, obviating the need for hand-crafted features.
• State-derivative pairs and immersion: Ensuring that the encoder is an immersion on smooth parts of the attractor guarantees that local tangent structures (and thus dynamic causal relations) are preserved in the latent space.

In the anomaly detection framework, deviations from the nominal attractor change the effective box-counting dimension or cause violations of injectivity/immersion, which can be detected algorithmically via loss metrics such as the Temporal Differential Consistency (TDC) loss or Jacobian rank deficiency.

6. Theoretical and Practical Implications

The Fractal Whitney Embedding Prevalence Theorem extends the toolbox of data-driven dynamical systems analysis beyond manifold settings, accommodating highly non-uniform and possibly fractal systems. It provides a rigorous basis for:

• The construction of low-dimensional, faithful embeddings of complex dynamical phenomena.
• The use of generic neural network encoders for time-series anomaly detection.
• The expectation that changes in the system's underlying dynamics—such as the occurrence of anomalies—will manifest as detectable degeneracies (loss of injectivity or immersion) in the latent space.

The theorem’s prevalence formulation ensures that these properties hold robustly under generic conditions, not merely as pathologies or exceptions. This mechanistic guarantee is fundamental in the design and theoretical validation of machine learning models for monitoring, predicting, and diagnosing behaviors in industrial and cyber-physical systems (Somma et al., 26 Feb 2025).