Identity Basins in High-Dimensional Systems

• Identity basins are high-dimensional regions where points converge consistently toward an attractor, ensuring structural identity across various technical domains.
• They are rigorously defined using frameworks like morphological graph theory and holomorphic dynamics, often employing flooding chains and FSM-based mapping.
• Empirical studies demonstrate that identity basins yield robust performance in image segmentation, generative control, and stability analysis, measured via metrics like ArcFace similarity.

An identity basin is a region of a high-dimensional space—such as the latent space of a generative model or the phase space of a dynamical system—whose points all exhibit consistent convergence or structural identity with respect to an attractor, concept, or morphological minimum. The term encompasses a range of technical domains, from graph-based morphological segmentation, complex holomorphic dynamics, and high-dimensional dynamical systems, to text-to-image diffusion models where semantic or structural cues define coherent identity clusters in the output distribution.

1. Formal Definitions and Theoretical Foundations

Morphological Graph Theory

In mathematical morphology, particularly as developed in watershed segmentation, identity basins are rigorously defined through equivalent formulations on edge-weighted and node-weighted graphs. Let $G=(N,E)$ be an undirected graph, with node set $N$ and edge set $E$. Assign a function $e:E\to\mathbb{T}$ for edge weights or $n:N\to\mathbb{T}$ for node weights, where $\mathbb{T}$ is a totally ordered set (“altitude lattice”):

• Edge-weighted model: A catchment basin (identity basin) for a regional minimum $M$ is

$$CB_e(M) = \{x \in N \mid \text{there exists a flooding chain from } x \text{ into } M\}$$

• Node-weighted model: The corresponding formulation is

$$CB_n(M) = \{x \in N \mid \text{there exists a flooding path from } x \text{ into } M\}$$

A constructive equivalence is established: every edge-weighted graph admits a transformation to a node-weighted graph and vice versa, such that the sets of minima and basins (as defined above) are preserved exactly. The notion of "identity basin" thus becomes representation-invariant at the level of the flooding graph (Meyer, 2013).

Holomorphic Dynamics

For germs of holomorphic maps $F:(\mathbb{C}^2,0)\to(\mathbb{C}^2,0)$ tangent to the identity (i.e., $F(0,0)=0$ and $dF(0,0)=\mathrm{Id}$), identity basins are open sets in a neighborhood of $0$ whose orbits under iteration converge to the origin along a particular characteristic direction $[v]$. Classification of directions (non-degenerate, Fuchsian degenerate, irregular degenerate) and the existence of associated basins depend on resonance conditions and residue indices (Vivas, 2011).

Neural Generative Models

In the domain of text-to-image diffusion models, an identity basin is an empirically defined region in the latent or conditioning space. All prompt encodings within the basin yield outputs exhibiting a common visual identity, even without explicit use of the target’s name or image. For example, the basin associated with a memorized celebrity in Stable Diffusion 1.5 consists of all prompt combinations that cause deterministic convergence to images resembling that individual, as quantified by high ArcFace or CLIP similarity (Fraser, 20 Feb 2026).

2. Equivalence and Transformation of Identity Basins

Graphical Watersheds

Meyer’s equivalence theorem establishes that the process of segmenting an image or network via watersheds on edge- or node-weighted graphs yields exactly the same identity basins once the system is recast as a flooding graph:

• Node and edge weights are linked by $e(i,j) = \max(n(i),n(j))$
• Minima and catchment basins coincide for both representations
• Flooding chains and flooding paths map bijectively

The transformation utilizes local operators: the "lowest-adjacent" node operator $\varepsilon_{ne}(i)=\min_{(i,j)} e(i,j)$ and its edge counterpart, yielding a perfectly coupled node-edge weight system suitable for direct morphological analysis (Meyer, 2013).

Dynamical Systems and Basins of Attraction

The concept of an identity basin extends to basins of attraction in nonlinear dynamics. Here, the basin encompasses all initial states whose time evolution converges to a specific attractor—fixed point, limit cycle, or chaotic set. Recent algorithmic advances deploy finite state machine (FSM) frameworks to partition high-dimensional state spaces, automatically identifying and labeling these basins without prior knowledge of attractor structure (Datseris et al., 2021).

Identity Basins in Generative Latent Spaces

In diffusion models, prompt-space navigation reveals that overlapping subregions associated with morphological or phonological cues define attractor regions. These identity basins exhibit sharp, often non-interpolative boundaries: morphologically related descriptors or specific prompt-level sound patterns deterministically yield consistent, recognizable visual output clusters (Fraser, 20 Feb 2026).

3. Methodologies for Basin Identification and Manipulation

Morphological Image Segmentation

• Flooding-graph-based watersheds: Partition images or networks using flooding graphs, transforming data as needed between node- and edge-weighted representations without loss of generality. Morphological operators prune ambiguous overlaps or refine granularity.
• Algorithmic implications: Any watershed, minimum-spanning-forest, or distance-based segmentation algorithm can operate interchangeably, as the identity basin structures are invariant post-flooding-graph conversion (Meyer, 2013).

Holomorphic Map Classification

• Blowing up at the singularity $(0,0)$ in $\mathbb{C}^2$ facilitates analysis of tangency and resonance along characteristic directions. Existence of identity basins relies on the residue index $\zeta$ associated with the direction, leading to classification as Fuchsian, irregular, or apparent, each with explicit basin existence results (Vivas, 2011).

FSM-based Automated Basin Mapping

• Partition state space into a finite grid; simulate system trajectories, updating grid-cell labels based on recurrence, proximity, and convergence.
• Attractor identification proceeds via Poincaré recurrence properties, while basin labels propagate efficiently via shortcut rules as the grid fills.
• This approach enables systematic computation of basin entropy, tipping probabilities, and stability fractions in arbitrarily high-dimensional systems (Datseris et al., 2021).

Prompt Engineering and Latent Space Navigation

• Morphological and phonological decomposition allows for identification and targeting of identity basins in diffusion models. Descriptor-only prompts or phonestheme-rich nonsense words can reliably induce convergence to visually coherent, unique clusters.
• Targeted self-distillation and LoRA training on synthetic outputs can sharpen and steer basin boundaries or even shape the identity’s "inverse" region via push–pull conditioning (Fraser, 20 Feb 2026).

4. Empirical Results and Quantitative Properties

Significant empirical findings across domains include:

• Catchment basin invariance: Watershed segmentation yields identical basins for node- and edge-weighted models after transformation; practical segmentation is identical regardless of the initial representation (Meyer, 2013).
• Explicit convergence in neural models: Morphological descriptor prompts without the target name or photo can yield 70% intra-basin hit-rate after several rounds of LoRA self-distillation. The process was quantitatively evaluated using ArcFace cosine similarity and CLIP-based clustering; convergence is non-linear and often exhibits phase transitions at critical LoRA weights. CFG-sweeps show identity basin stability across guidance strengths (Fraser, 20 Feb 2026).
• Phonestheme-induced basins: Prompt-level sound patterns (e.g., "snudgeoid," "crashax") yield perfect visual identity clustering (Purity@1 = 1.0, zero external contamination), while control prompts do not, confirming emergent identity basins from sublexical structure (Fraser, 20 Feb 2026).
• FSM efficiency in dynamical mapping: FSM algorithms outperform naive converge-to-attractor baselining by 2–10× even in simple cases, allowing efficient mapping of highly interlaced or fractal basins, and direct computation of basin-relevant statistical quantities (Datseris et al., 2021).

5. Structural Properties and Phase Behavior

Identity basins exhibit several noteworthy structural and dynamical properties:

• Phase transitions: LoRA intensity sweeps reveal discrete jumps into or out of identity basins rather than smooth interpolations, indicating the presence of sharp, high-dimensional decision boundaries. This phenomenon recurs across both morphological and phonological navigation regimes (Fraser, 20 Feb 2026).
• Inverse and coherence drag: Attempted maximal away-conditioning (“push–pull” prompts) produces either severe structural breakdown (in base models) or "uncanny valley" outputs (after LoRA training), indicating that the attractor's basin exerts a broad local influence—not just within, but also in the immediate complement (Fraser, 20 Feb 2026).
• CFG-invariance: In text-to-image diffusion models, identity basin stability is invariant to classifier-free guidance magnitude; identity consistency persists across a range of CFG values (Fraser, 20 Feb 2026).

6. Applications and Future Directions

• Morphological segmentation: Identity basins enable design and analysis of segmentation pipelines independent of pixel/region representation, and provide a unified basis for algorithmic comparison and improvement (Meyer, 2013).
• Complex-dynamical classification: Understanding the structure and invariants of identity basins along characteristic directions informs the global phase-space topology of holomorphic dynamical systems and their bifurcations (Vivas, 2011).
• Generative model probing and control: Explicit construction and navigation of identity basins in diffusion models opens pathways for watermark-free identity targeting, synthetic data augmentation, prompt safety assessment, and the systematic creation of novel visual concepts grounded in sub-lexical or morphological structure. Extensions may include full latent space topology mapping and classification of basin spines, boundaries, and "apparent regime" anomalies (Fraser, 20 Feb 2026).
• Dynamical system diagnostics: FSM and basin-labeling frameworks enable rigorous study of multistability, tipping points, and final-state sensitivity in both theoretical and applied high-dimensional systems (Datseris et al., 2021).

As demonstrated across distinct technical domains, identity basins stand as a unifying concept for invariant regions of structural attraction, applicable from topological graph theory to neural generative modeling and beyond.