Truncation Closed Image: Theory & Applications

• Truncation closed image is defined as the preservation of elements under truncation operations, a central concept in Hahn fields and valued field embeddings.
• A tower of complements and structured truncation operators facilitate explicit embeddings into generalized power series, ensuring algebraic and analytic consistency.
• In applied imaging and deep learning, managing truncation artifacts with reconstruction priors and decoupled representations enhances image robustness and completeness.

A truncation closed image is a central concept in valuation theory, functional analysis, and applied imaging domains, capturing the idea that certain substructures (fields, subspaces, or solution sets) are preserved under "truncation" operations corresponding to initial segment extraction or restriction to an accessible field of view. The notion is formally developed in model theory, particularly in the study of Hahn fields, but is also a key structural challenge in computational imaging where finite sampling or limited detector size induce image truncation and the associated "closed image" problems.

1. Truncation Closed Subfields in Hahn Fields and Valued Fields

The classical mathematical framework defining a truncation closed image arises in the context of embeddings of valued fields into Hahn fields. For a valued field $(K, v)$ with residue field $k$ and value group $\Gamma$, a Hahn field $k((t^\Gamma))$ consists of formal series $f = \sum_{\gamma\in\Gamma} c_\gamma t^\gamma$ with well-ordered support $\{\gamma : c_\gamma \neq 0\}$ (Dries, 27 Dec 2025, Fornasiero et al., 2013). The $\alpha$-truncation operator $\tau_\alpha$ acts by discarding all terms of exponent $\gamma \geq \alpha$: $$\tau_\alpha\left(\sum_\gamma c_\gamma t^\gamma\right) = \sum_{\gamma < \alpha} c_\gamma t^\gamma.$$ A subfield $F \subseteq k((t^\Gamma))$ is truncation closed if it is preserved under every truncation: for any $f \in F$ and any $\alpha \in \Gamma$, $\tau_\alpha(f) \in F$.

An embedding $\varphi: K \hookrightarrow k((t^\Gamma))$ is truncation closed if $\varphi(K)$ is a truncation closed subfield. This property is intimately connected to the existence of an internal "truncation structure" on $K$, consisting of truncation maps and a compatible cross-section of $\Gamma$ in $K^\times$, satisfying a precise set of axioms (T1)-(T8) (Dries, 27 Dec 2025). This correspondence plays a crucial role in the model theory and classification of valued fields, providing both internal and external characterizations of embeddability into generalized power series fields.

2. Towers of Complements and Truncation Closed Embeddings

Fornasiero, Kuhlmann, and Kuhlmann (Fornasiero et al., 2013) generalize the theory to fields of generalized power series $k((G,f))$ twisted by a factor set $f: G \times G \to k^\times$. A subfield $F \subseteq k((G,f))$ is truncation closed if any finite initial segment of the support of a series in $F$ also lies in $F$. The key structural device is the notion of a tower of complements, a family $\mathcal{A} = \{A[A]\}_{A\in\breve G}$ of $k$-subspaces indexed by Dedekind cuts $A$ of the value group $G$, satisfying direct sum, multiplicativity, and monotonicity properties relative to valuation ideals. Existence of a $t$-compatible tower of complements is necessary and sufficient for the existence of a truncation closed embedding $K \hookrightarrow k((G,f))$.

The method is constructive for Henselian or algebraically maximal Kaplansky fields, where a Zorn's Lemma argument and combinatorial splitting show that truncation closed embeddings exist and extend to maximal immediate extensions. The tower-of-complements construction realizes an explicit correspondence between algebraic data in $K$ and the truncation behavior in its analytic representation.

3. Truncation Closed Image in Applied Imaging: Closed-Image Problems

In computed tomography (CT) and cone-beam CT (CBCT), the term "closed-image" is used to describe the set of points in an image that can be reliably reconstructed given a truncated field of view (FOV). When only a portion of the object is within the detector's reach, the measured sinogram is truncated, making certain regions (outside the direct FOV) unobservable—a paradigm known as the "closed-image" region (Park et al., 11 Aug 2025, Liman et al., 2024).

In CBCT, the forward model under FOV truncation masks the true projection data: $$p_{\text{trunc}}(\theta, s) = W_T(\theta, s) \cdot A[u](\theta, s),$$ where $W_T$ is the truncation mask and $A$ is the cone-beam transform (Park et al., 11 Aug 2025). The inverse problem is ill-posed on the closed-image region, leading to artifacts unless specialized priors or data completion methods are employed.

Recent solutions address the closed-image truncation artifact by constructing priors—effectively "completing" the unobserved data. For example, an implicit neural representation (INR) trained over an extended region provides a coarse solution whose forward projections fill in the missing rays; its discrepancy with the measured projections is then used to correct a standard iterative reconstruction over the visible ROI (Park et al., 11 Aug 2025). Diffusion-based outpainting models are similarly used for field-of-view truncation recovery in chest CT imaging (Liman et al., 2024).

4. Truncation Robustness in Learning and Estimation

Truncation closedness also arises in the robustness of learned representations to data truncation. For instance, in metric-scale 3D human pose estimation, truncation-robust heatmaps are constructed by decoupling the metric 3D heatmap volume from the input image's field of view. This enables the network to infer joint positions for body parts that are outside the image boundary, effectively "hallucinating" or reasoning globally even under strong truncation (Sárándi et al., 2020). The metric-based representation thus resists the information loss caused by image truncation, enhancing the effective "closed-image" region that is reconstructible via the model.

Similarly, in information retrieval, the "truncation problem" is formalized as the task of selecting a cutoff in a ranked list of results. Extreme Value Theory and the Generalized Pareto Distribution are used to estimate per-query calibrated relevance scores and define statistically robust truncation points, ensuring that relevant items are not omitted due to the uncalibrated tail behavior of similarity scores (Bahri et al., 2020).

5. Characterization Theorems and Structural Conditions

The fundamental characterization of truncation closed images is established via the equivalence between:

• (i) The existence of a valued field embedding $K \hookrightarrow k((t^\Gamma))$ with truncation closed image,
• (ii) The existence of a truncation structure (resp. tower of complements) on $K$ (Dries, 27 Dec 2025, Fornasiero et al., 2013).

The truncation structure consists of:

• Truncation operators $(f, \alpha) \mapsto f|_\alpha$,
• A group section $\gamma \mapsto \tau^\gamma$ of the value group in $K^\times$,

satisfying axioms (T1)-(T8) such as preservation of order, support well-ordering, and compatibility with both addition and multiplication. Towers of complements are equivalently characterized through direct sum decompositions relative to valuation ideals at all Dedekind cuts of the value group, with respect to a given section $t:G\to K^\times$. These structures are also shown to persist through extensions, ensuring the tractability of truncation closedness in maximal immediate extensions.

6. Practical Implications and Extensions

The concept of truncation closed image governs structural stability and reconstructibility under truncation operations across theoretical and applied domains. In model theory, it is essential for understanding the expressive power of valued fields and their expansions (including differential and o-minimal analogs) (Dries, 27 Dec 2025). In imaging, algorithms that implicitly rebuild missing measurements over the closed-image region enhance artifact suppression and robustness in truncated projection geometries (Park et al., 11 Aug 2025, Liman et al., 2024). In learning, decoupling architectural representations from data truncation mitigates catastrophic failure in partially observed domains (Sárándi et al., 2020).

Corollaries highlight that algebraically closed fields, or algebraically maximal Kaplansky fields of positive characteristic, always admit the appropriate truncation structures, guaranteeing truncation closed embeddings. The bulk of the first-order axiomatizability of these structures, modulo a global well-ordering condition, further strengthens their utility in logical and computational applications.

7. Summary Table: Truncation Closed Image Across Domains

Domain	Definition of Closed Image	Structural Criterion
Hahn fields/Valued fields	Image under Hahn embedding closed under truncation	Existence of a truncation structure/Tower of complements (Dries, 27 Dec 2025, Fornasiero et al., 2013)
Computed Tomography (imaging)	Reconstructible region given FOV truncation	Model/data extension with prior or outpainting (Park et al., 11 Aug 2025, Liman et al., 2024)
Deep learning for pose estimation	Heatmap region predicting parts beyond crop	Decoupled metric heatmap volume (Sárándi et al., 2020)
Information retrieval	Truncated result list	Calibrated cutoff via EVT/GPD (Bahri et al., 2020)

This framework demonstrates the unifying abstraction and technical depth of the truncation closed image concept, spanning model theory, algebra, reconstruction algorithms, learning, and statistical inference.