Syntactic Cost Soundness Theorem

• The paper proves that a candidate abstract transformer is globally sound if and only if its total cost—combining syntactic and semantic measures—is zero.
• It introduces a unified, two-part cost function that integrates syntactic validation with quantitative semantic analysis using explicit aggregation of constraint violations.
• The theorem underpins automated synthesis in abstract interpretation, notably improving neural network verification by certifying soundness through optimization-driven techniques.

The syntactic cost soundness theorem establishes a precise correspondence between cost-based constraints on syntactic and semantic well-formedness and global semantic soundness for candidates synthesized in abstract interpretation frameworks. By reformulating soundness as a property certified by the vanishing of an explicitly constructed cost function, the theorem enables automated, optimization-driven synthesis of sound abstract transformers—most notably in domains such as neural network verification—using a unified cost-guided feedback mechanism (Gu et al., 17 Nov 2025).

1. Formal Framework

The synthesis setting consists of searching for sound abstract transformers expressed as programs over a domain-specific language (DSL):

• $H$: the set of all DSL programs (candidates).
• Validity predicate $\mathrm{Valid}(p)$: $p$ is syntactically and statically well-formed.
• $V \subseteq H$: set of all well-formed programs ($\{p \mid \mathrm{Valid}(p)\}$).
• Soundness predicate $\mathrm{Sound}(p)$: For all $z \in A$, $F(\gamma(z)) \subseteq \gamma(F^{\sharp}(z))$.
• $G \subseteq V$: set of valid, sound candidates ($\{p \in V \mid \mathrm{Sound}(p)\}$).
• $\gamma: A \rightarrow \mathcal{P}(C)$: concretization from abstract domain $A$ to concrete domain $C$.

To operationalize the search, the framework introduces a two-part cost function:

• Syntactic cost: $C_{\mathrm{syntactic}}(p) = 0$ if $\mathrm{Valid}(p)$, $\infty$ otherwise.
• Semantic unsoundness cost:

$$\Delta_S(p) = \bigoplus_{z \in A^*} \bigoplus_{x \in \gamma(z)} \bigoplus_{j=1}^k \varepsilon( f(x)_i, C_{ij} )$$

where $A^*$ is the set of violating abstract elements, $C_{ij}$ are constraints characterizing allowed outputs via Fourier–Motzkin elimination, and $\varepsilon(m,C)$ computes the violation magnitude relative to constraint $C$.

• Total cost: $C(p) := C_{\mathrm{syntactic}}(p) + \Delta_S(p)$.

Zero total cost is possible only if the candidate is both well-formed and semantically sound.

2. The Syntactic Cost Soundness Theorem

Theorem (Syntactic Cost Soundness):

Let $H$, $V$, $G$, $C_{\mathrm{syntactic}}$, $\Delta_S$, and $C$ be as above. For every $p \in H$,

$C(p) = 0 \quad \text{if and only if} \quad p \in G$

Specifically, $C(p) = 0$ ensures that $p$ is both syntactically valid and semantically sound.

3. Hypotheses and Structural Assumptions

Correctness of the theorem relies on the following conditions (Gu et al., 17 Nov 2025):

• A1: The validator precisely characterizes well-formedness: $C_{\mathrm{syntactic}}(p) = 0$ iff $\mathrm{Valid}(p)$.
• A2: The abstract domain $A$ allows finite constraint decomposition of transformer images using Fourier–Motzkin.
• A3: The unsoundness measure $\varepsilon(\cdot,\cdot)$ is nonnegative, zero exactly for satisfied constraints.
• A4: Aggregation operators ($\oplus$) are monotone, nonnegative, bounded, and only vanish on the empty set.
• A5: The sampling and weighting strategy for $\Delta_S$ preserves the property that $\Delta_S(p) = 0$ implies true global soundness, as confirmed by solver checks.

These guarantee that the cost function is tight: strictly zero exactly on the target set $G$.

4. Proof Outline

Consider $C(p) = C_{\mathrm{syntactic}}(p) + \Delta_S(p)$:

• If $C(p) = 0$, then both $C_{\mathrm{syntactic}}(p) = 0$ (hence $p \in V$ by A1) and $\Delta_S(p) = 0$ (no semantic violations, so $p$ is sound).
• Conversely, if $p \in G \subseteq V$, then $C_{\mathrm{syntactic}}(p) = 0$, and if $p$ is sound, every violation term is zero, so $\Delta_S(p) = 0$.
• The disjunctive ($+$) and aggregation structure ensures no cost "leakage": any departure from validity or soundness increases $C(p)$ strictly above zero.

No subtlety arises beyond this, as the cost function is constructed as an exact indicator for the semantically and syntactically correct region.

5. Consequences and Illustrative Example

A direct corollary ("Zero-Cost Soundness") states that achieving $C(p) = 0$ in the synthesis process certifies both validity and global soundness. This permits automated acceptance of candidate abstract transformers without further verification.

An explicit example occurs in neural network verification under the DeepPoly abstract domain: an unsound candidate $p_0$ for the $\mathit{HardSigmoid}$ operator satisfies $C_{\mathrm{syntactic}}(p_0) = 0$ but $\Delta_S(p_0) > 0$. Iterative refinements eventually yield $p_T$ with $\Delta_S(p_T) = 0$, $C_{\mathrm{syntactic}}(p_T) = 0$, thus $C(p_T)=0$ and consequently $p_T$ is a sound transformer (Gu et al., 17 Nov 2025).

6. Conceptual Significance in Cost-Driven Synthesis

The theorem transforms the global, universally quantified notion of semantic soundness—traditionally nonconstructive—into a checkable, optimization-compatible real-valued cost. Specifically:

• All candidates are filtered for syntactic validity ($C_{\mathrm{syntactic}}=0$).
• The framework incrementally minimizes $\Delta_S$, supported by sampling and solver confirmation.
• Once $C(p)$ reaches zero, global soundness and syntactic validity are certified by construction; no residual uncertainty remains.

This operationalization is critical for scaling automated synthesis in abstract interpretation, particularly for complex or nonlinear domains emergent in modern neural network analysis. The syntactic cost soundness theorem thus forms the logical foundation for cost-driven search-and-correction loops that guarantee correctness at termination (Gu et al., 17 Nov 2025).

7. Relationship to Broader Research and Methodological Context

The theorem represents a refinement of “syntactic soundness” arguments widely used in proof theory (e.g., cut-free completeness of tableaux (Bonichon et al., 2015)) and static complexity analysis of higher-order languages (Danner et al., 2012), but here integrated into a quantitative optimization framework for abstract interpretation. It contrasts with cost soundness theorems from typed resource logics and static analysis (Mannucci et al., 7 Dec 2025, Danner et al., 2012) by:

• Directly fusing syntactic validity and semantic soundness into a single measurable target.
• Applying quantitative aggregation to handle graded, real-valued or weighted violations.
• Leveraging the construction for LLM-guided or other automated transformer synthesis instead of solely for static verification or white-box analysis.

This theorem thereby generalizes and operationalizes syntactic soundness principles for the context of program synthesis with strict correctness guarantees.