Linear Guardedness in Logic, Databases & AI

• Linear guardedness is defined by constraining linear order predicates with base atoms to ensure decidable query answering in existential rule systems.
• It enables trade-offs between expressiveness and tractability in disjunctive existential rules, with precise complexity bounds like 2ExpTime and coNP data complexity.
• In neural representation, linear guardedness underpins methods (e.g., LEACE) that erase sensitive attributes while preserving overall utility against linear or log-linear probes.

Linear guardedness is a central concept in logic, knowledge representation, and AI systems, particularly in the study of constraint satisfaction, database theory, and neural representation bias mitigation. It refers to carefully syntactic or algorithmic constraints—often involving both linear and guarded fragments—that enable tractable inference or provably secure information erasure. The term arises in both symbolic (relational logic) and subsymbolic (machine learning) settings, each with a precise technical definition and pivotal role in their respective areas.

1. Linear Guardedness in Existential Rule-Based Systems

The foundational use of linear guardedness appears in the context of existential rules (also called tuple-generating dependencies, or TGDs) where additional semantic restrictions are imposed—such as the inclusion of a distinguished linear-order predicate. In these frameworks, a relational signature is extended with a special binary predicate (e.g., "$<$") that must be interpreted as a strict linear order. The presence of this order increases the expressive power of the system but threatens decidability of query answering unless rigorously controlled.

A rule is called "base-covered frontier-guarded" or "linear-guarded" if every occurrence of the linear order predicate in the rule body is covered by a base atom; that is, for any atom $x < y$ present, there must be some atom $R(\ldots, x, y, \ldots)$ from the base signature that includes both $x$ and $y$ as arguments. This ensures that the logical order is never used as a guard by itself. The main results, established by Amarilli et al., demonstrate that under this restriction—termed "base-coveredness" or "linear guardedness"—query answering remains decidable with 2ExpTime combined complexity and coNP data complexity. Any relaxation (e.g., allowing unguarded occurrences of $<$) immediately leads to undecidability, as such fragments can encode tiling or Post correspondence problems (Amarilli et al., 2016, Amarilli et al., 2022).

2. Linear and Guarded Fragments in Disjunctive Existential Rules

Beyond the classical TGD setup, finite axiomatizability of classes of relational structures by disjunctive existential rules (dexrs) is intimately linked to their linear and guarded fragments. A dexr is "linear" if its body contains at most one atom, and "guarded" if the body contains a single atom (the "guard") covering all universally quantified variables. The expressiveness and decidability of these logical fragments are characterized through model-theoretic properties: criticality, closure under repairable direct products, and variants of "diagrammatic compatibility" (linear or guarded).

It is proven that a class $\mathcal{C}$ is finitely axiomatizable by linear dexrs if and only if it is critical, closed under repairable direct products, and enjoys linear-diagrammatic compatibility, with analogous results for the guarded case. Notably, there exists a decision procedure (with elementary, triply exponential complexity) that can determine whether a finite set of guarded dexrs can be equivalently rewritten as a set of linear dexrs, thus establishing a powerful bridge between the two fragments (Calautti et al., 16 Aug 2025).

Rule Class	Body Atoms	Guarding Requirement	Decidability (QA)
Linear dexr	$\leq 1$	None (linear by default)	2ExpTime
Guarded dexr	1	Covers all vars	2ExpTime
Unrestricted	Arbitrary	None	Undecidable

3. Decidability, Base-Coveredness, and Complexity Results

The syntactic notion of "linear guardedness" is the chief enabler of tractability when incorporating strict orderings into expressive rule languages:

• Query answering with base-covered (linear-guarded) rules and queries admits a 2ExpTime-complete combined complexity bound, even in the presence of strict linear order predicates and their associated axioms (irreflexivity, transitivity, totality).
• Data complexity (with rules and queries fixed) drops to coNP-complete.
• The key proof techniques exploit the tree-like model property enforced by base-covering, allowing compilation of the linear order constraints into bounded-width fragments manageable by automata or tree-decomposition methods.
• Any relaxation permitting the distinguished order to serve as a guard or to occur unguarded immediately triggers undecidability, as shown via reductions from tiling problems (Amarilli et al., 2016, Amarilli et al., 2022).

4. Linear Guardedness in Neural Representation Erasure

Linear guardedness also describes a property of neural representations where certain "concepts" (e.g., gender, part-of-speech, or demographic attributes) are sought to be irrevocably erased so that no downstream linear classifier can recover them. In this setting, perfect linear guardedness (as realized in methods like LEACE) means finding an affine transformation $r(e) = Pe + b$ such that for any weight vector $w$, the "erased" features carry zero linear information about the target concept. Mathematically, this is equivalent to enforcing $x < y$0 for all concept labels $x < y$1. The LEACE construction gives a closed-form, least-squares-optimal projector $x < y$2, guaranteeing that no linear probe can recover the erased attribute while minimally distorting other information (Belrose et al., 2023).

5. Log-Linear Guardedness and Its Limitations

A refinement known as log-linear guardedness focuses on information leakage against the broader class of log-linear models (e.g., logistic regression, softmax classifiers). A guarding function $x < y$3 is said to $x < y$4-guard an attribute $x < y$5 if the maximal log-linear information $x < y$6. This formalizes adversarial resistance to recovery of $x < y$7 using log-linear models.

Key theoretical findings highlight a dichotomy:

• For binary downstream classifiers and discretized probabilities, log-linear guardedness is closed under composition: any sequence of guarded logistic layers preserves attribute erasure.
• For multiclass softmax heads, log-linear guardedness collapses: a $x < y$8-class softmax classifier can partition the representation space into arbitrarily fine Voronoi regions, fully reconstructing the original binary attribute in many cases. This exposes a critical limitation of linear (and log-linear) erasure methods for multiway classification tasks (Ravfogel et al., 2022).

6. Implications, Open Problems, and Interplay with Bias/Fairness

The distinction between linear guardedness and log-linear guardedness exposes the gap between intrinsic and extrinsic bias mitigation. Intrinsic bias pertains to information encoded in representation spaces, while extrinsic bias concerns the behavior of downstream models. Merely applying linear erasure may guarantee that linear or logistic probes reveal no concept information, but does not guarantee fairness or independence if downstream model expressiveness exceeds that of linear or binary log-linear functions—most notably in multiclass classification.

Future research directions include:

• Development of stronger erasure notions, possibly requiring non-linear, distribution-aware interventions that surpass the limitations of orthogonal or oblique projections.
• Imposing fairness constraints directly on downstream classifier families instead of relying solely on representation-level interventions.
• Characterizing the expressiveness of adversary model classes for which a guardedness property (as measured by $x < y$9-information) is preserved under arbitrary composition (Ravfogel et al., 2022).

7. Summary and Representative Examples

Linear guardedness enables robust, tractable reasoning over expressive logical systems and offers security guarantees in neural representation erasure pipelines—provided that the expressiveness of downstream models does not exceed the class of linear or (in the binary case) log-linear models. In database theory, it delineates the maximal natural fragments where strict orders can be introduced without forfeiting decidability. In modern machine learning, it places crucial boundaries on the effectiveness of bias mitigation by concept erasure.

Example of a base-covered linear-guarded rule in a temporal event domain (Amarilli et al., 2022):

$R(\ldots, x, y, \ldots)$0

Here, every use of $R(\ldots, x, y, \ldots)$1 is “covered” by a base atom, manifesting linear guardedness and guaranteeing decidability of queries involving such rules.

In neural representation learning, application of LEACE to LLMs reduces the accuracy of linear gender probes to chance, significantly diminishes true positive rate gaps, and avoids harming non-targeted downstream utility, e.g., in part-of-speech-based language modeling tasks (Belrose et al., 2023).

Linear guardedness thus encapsulates a cross-disciplinary principle linking logic, databases, and representation learning: only by carefully constraining the interaction between expressive features (such as orders or concepts) and the guarded context can both expressiveness and tractability, or erasure and utility, be simultaneously preserved.