FoL++: Extensions of First-Order Logic

• FoL++ is a family of augmented formalisms that extend first-order logic through intensional semantics, recursive fixpoints, and algorithmic graph modifications.
• It incorporates two-step interpretation over PRP domains, least fixpoint reasoning for heap structures, and fixed-parameter meta-theorems for planarity constraints.
• The topic also includes a visual place recognition method that uses reliability-aware matching and adaptive candidate re-ranking to enhance performance.

FoL++ designates several technically distinct constructions that extend, reinterpret, or operationalize first-order logic. In one usage, it is a minimal intensional first-order logic that separates meaning from reference by mapping formulae to concepts in a PRP domain and then extensionalizing those concepts world by world (Majkic, 2011). In another, it denotes first-order logic with least fixpoints over heap structures, written FO+lfp, with recursive predicates, Park-style induction, and model-guided synthesis of inductive lemmas (Murali et al., 2020). A further algorithmic usage combines an FOL sentence $\phi$ with the non-FOL constraint of planarity after graph modification, yielding a fixed-parameter meta-theorem for four graph edit operations (Fomin et al., 2021). Outside logic proper, “FoL++” also names a visual place recognition method that combines reliability-aware region modeling with adaptive re-ranking (Chen et al., 24 Apr 2026). The term therefore denotes a family of augmentations rather than a single standardized formalism.

1. Distinct technical usages

The following usages are explicitly associated with FoL++ in the supplied literature.

Usage	Domain	Defining formulation
Minimal intensional FOL	FOL semantics	two-step semantics with $\mu$ and $e_w$, PRP ontology, recovery of Tarskian truth
FO+lfp in the same semantic territory as FoL++	Recursive heap reasoning	least fixpoint predicates, fixpoint abstraction, PFP obligations, FOSSIL
FoL++ viewpoint for FOL plus planarity	Parameterized graph algorithms	Graph $\boxtimes$-Modification to Planarity and $\phi$ in $f(k,|\phi|)\cdot n^2$ time
FoL++ as a VPR method	Visual place recognition	Reliability Estimation Branch, SAL/SCE, pseudo-correspondence, ACS

This multiplicity matters because the shared label does not imply a shared semantics, proof theory, or application domain. In the logical papers, FoL++ is associated with semantic enrichment of first-order logic or with algorithmic frameworks around it. In the VPR paper, by contrast, FoL++ is a model name whose components are region reliability, local matching, and candidate scheduling rather than logical operators or proof systems (Chen et al., 24 Apr 2026).

2. FoL++ as minimal intensional first-order logic

In "First-order Logic: Modality and Intensionality" (Majkic, 2011), FoL++ is a minimal intensional enrichment of FOL that keeps the FOL syntax intact while introducing a two-step interpretation. Its universe of discourse is a PRP ontology,

$$D = D_{-1} + D_I,\qquad D_I = D_0 + D_1 + D_2 + \dots,$$

where $D_{-1}$ are particulars, $D_0$ are propositions, and $D_n$ for $\mu$0 are $\mu$1-ary concepts. Functions are treated as relations via their graphs, and a constant-domain assumption is adopted.

The central semantic machinery is a single meaning relation implemented as a meaning map

$\mu$2

with $\mu$3, together with an extensionalization map

$\mu$4

for each world $\mu$5. The Tarskian interpretation at a world is recovered compositionally:

$\mu$6

For a world $\mu$7 corresponding to a Tarski model $\mu$8, the paper states that

$\mu$9

This yields a commutative homomorphic chain from the free syntax algebra of FOL, through an intensional algebra of concepts, to an extensional relational algebra.

A major claim of the paper is that modalization alone does not yield intensionality. In the modal reading of quantifiers, denoted $e_w$0, worlds are assignments $e_w$1 and, for each variable $e_w$2, accessibility is the S5 relation $e_w$3 defined by agreement on all variables other than $e_w$4. With rigid interpretations for non-logical symbols, the intension of each formula is constant across assignment-worlds. The paper therefore proves that $e_w$5 is modal but not intensional, and that it yields exactly the same truth conditions as Tarski’s FOL. Genuine intensionality arises only in $e_w$6, where worlds are explicit Tarskian interpretations $e_w$7 and predicate extensions may vary across worlds (Majkic, 2011).

The paper sharply distinguishes three notions of sameness. For concepts $e_w$8 of the same arity, extensional equality at a world is $e_w$9 iff $\boxtimes$0; intensional equality as necessary equivalence is $\boxtimes$1 iff $\boxtimes$2; intensional identity is ordinary equality $\boxtimes$3 inside the concept domain $\boxtimes$4. This allows distinct meanings to remain distinct even when they are necessarily equivalent. The canonical example is that “x has been bought” and “x has been sold” may be coextensional at every world without being mapped by $\boxtimes$5 to the same unary concept.

The algebraic side is equally central. The intensional algebra over concepts contains operations $\boxtimes$6, $\boxtimes$7, and $\boxtimes$8, which under extensionalization become natural join, complement, and projection:

$\boxtimes$9

This is explicitly contrasted with cylindric algebras: conjunction is rendered by natural join rather than intersection, and quantification by projection $\phi$0 indexed by free-variable position rather than by abstract cylindrifications.

Relative to Montague, the paper rejects the identification of intensions with world-to-extension functions, because that collapses necessarily equivalent but semantically distinct expressions. Relative to Bealer, it argues that intensional abstraction is not required: reification is already provided by $\phi$1 into the PRP domain. The resulting system is a conservative semantic enrichment: it preserves standard FOL truth and global consequence while adding a concept-level semantics that supports reasoning about meanings as distinct from extensions (Majkic, 2011).

3. FoL++ as FO+lfp for recursive heap reasoning

"Model-Guided Synthesis of Inductive Lemmas for FOL with Least Fixpoints" places FO+lfp in “the same semantic territory as what many authors call FoL++” (Murali et al., 2020). Here the object is a multi-sorted first-order logic over a signature $\phi$2 with a distinguished foreground sort $\phi$3 for heap locations and background sorts such as $\phi$4 and $\phi$5. Symbols involving the foreground sort are uninterpreted, while symbols restricted to a single background sort are constrained by a decidable background theory and combined by Nelson–Oppen.

The extension beyond pure FOL is a family $\phi$6 of recursively defined predicates, and in the tool also partial functions, each given by a definition $\phi$7 that uses only universal quantification over foreground variables and refers to recursive symbols either to strictly smaller layers or positively within the same layer. This stratification and positivity ensure monotonicity of the induced operator. For an induced monotone operator $\phi$8, the least fixpoint exists by Tarski’s theorem and is characterized as

$\phi$9

A model of FO+lfp interprets the background sorts by their theories, the foreground sort by an arbitrary finite or infinite domain $f(k,|\phi|)\cdot n^2$0, and evaluates recursive symbols to their least fixpoints.

The logic is used to encode pointer-based structures and mixed shape/data properties. The paper gives formulations for singly-linked lists, list segments, reachability, trees with heaplets, and binary search trees. For example,

$f(k,|\phi|)\cdot n^2$1

and

$f(k,|\phi|)\cdot n^2$2

The tree and BST encodings combine positive recursion with background set constraints such as disjointness of heaplets.

Inductive reasoning is organized through Park-style induction. For a lemma

$f(k,|\phi|)\cdot n^2$3

the paper generates a pre-fixpoint obligation

$f(k,|\phi|)\cdot n^2$4

and proves that if $f(k,|\phi|)\cdot n^2$5 then $f(k,|\phi|)\cdot n^2$6. The fixpoint abstraction $f(k,|\phi|)\cdot n^2$7 replaces least-fixpoint semantics by equational fixpoints,

$f(k,|\phi|)\cdot n^2$8

which is sound but not complete for FO+lfp. Pure FO obligations under this abstraction are discharged by SQI, a systematic quantifier instantiation procedure that is complete for foreground-only quantification at bounded term depth.

Because FO+lfp is highly expressive, complete procedures are unavailable. The paper states that over a discrete linear order with $f(k,|\phi|)\cdot n^2$9 and successor, one can lfp-define true addition and multiplication and encode reachability of Turing-powerful machines, so validity becomes undecidable and not even recursively enumerable. The practical response is FOSSIL, a framework that synthesizes inductive lemmas from three kinds of finite counterexample models: Type-I pseudomodels for failure to prove a goal under $$D = D_{-1} + D_I,\qquad D_I = D_0 + D_1 + D_2 + \dots,$$0, Type-II bounded lfp countermodels that truly refute a candidate lemma, and Type-III pseudomodels for failure to prove the corresponding PFP obligation.

The synthesis loop maintains Type-II and Type-III counterexamples per antecedent predicate, searches a grammar-bounded space of lemmas satisfying elimination and consistency constraints, attempts to prove $$D = D_{-1} + D_I,\qquad D_I = D_0 + D_1 + D_2 + \dots,$$1 by SQI, and either accepts the lemma or refines the search with new counterexamples. The paper’s exemplar lemma for list segments is the transitivity principle

$$D = D_{-1} + D_I,\qquad D_I = D_0 + D_1 + D_2 + \dots,$$2

This is presented as a typical short inductive invariant that FOSSIL can synthesize.

Empirically, the evaluation covers 50 real theorems and 673 synthetic theorems. On the 50-theorem suite, FOSSIL solved all within 5 minutes each, with average approximately 30 seconds and maximum 167 seconds, synthesizing up to 10 valid lemmas per theorem. On the 673 synthetic theorems, it proved all within 10 minutes and 628, approximately 93%, within 1 minute. Removing Type-II and Type-III counterexamples drastically degraded performance, producing 31/50 timeouts on the real suite and 672/673 timeouts on the synthetic suite (Murali et al., 2020).

4. FoL++ as an algorithmic framework for planarity plus FOL

"An Algorithmic Meta-Theorem for Graph Modification to Planarity and FOL" uses a FoL++ viewpoint in which first-order properties are combined with a non-FOL structural target handled algorithmically (Fomin et al., 2021). The setting fixes one of the four graph modification operations

$$D = D_{-1} + D_I,\qquad D_I = D_0 + D_1 + D_2 + \dots,$$3

namely vertex deletion, edge deletion, edge contraction, and edge addition, and studies the meta-problem Graph $$D = D_{-1} + D_I,\qquad D_I = D_0 + D_1 + D_2 + \dots,$$4-Modification to Planarity and $$D = D_{-1} + D_I,\qquad D_I = D_0 + D_1 + D_2 + \dots,$$5.

Formally, the input is an undirected finite simple graph $$D = D_{-1} + D_I,\qquad D_I = D_0 + D_1 + D_2 + \dots,$$6 and an integer $$D = D_{-1} + D_I,\qquad D_I = D_0 + D_1 + D_2 + \dots,$$7, and the question is whether there exists a set $$D = D_{-1} + D_I,\qquad D_I = D_0 + D_1 + D_2 + \dots,$$8 of exactly $$D = D_{-1} + D_I,\qquad D_I = D_0 + D_1 + D_2 + \dots,$$9 edits such that $D_{-1}$0 is planar and satisfies an FOL sentence $D_{-1}$1. The main theorem states that there exists a computable function $D_{-1}$2 such that, for every $D_{-1}$3 and every FOL sentence $D_{-1}$4, the problem is solvable in

$D_{-1}$5

time. This yields fixed-parameter tractability parameterized by $D_{-1}$6.

The proof hybridizes two classic techniques. The first is the irrelevant vertex method from Graph Minors, needed because planarity is not FO definable. The second is Gaifman locality, used to normalize $D_{-1}$7 into an equivalent Boolean combination of basic sentences

$D_{-1}$8

where each $D_{-1}$9 is $D_0$0-local. This reduces the logical side to scattered local witnesses whose neighborhoods can be tracked inside bounded-radius regions.

On the structural side, large treewidth implies the existence of large walls. The algorithm seeks a wall whose compass is planar, has bounded treewidth, and is disjoint from the neighborhood of previously selected edits. When many disjoint subwalls have identical $D_0$1-characteristics, a replacement argument identifies an irrelevant vertex in one compass, and deleting that vertex preserves YES/NO status. If the graph does not have sufficiently large treewidth, the recursion returns a bounded-width tree decomposition, after which Courcelle’s theorem is applied to the remaining MSOL-expressible property.

The technical handling of the four operations differs in detail. For $D_0$2, the algorithm may either find an irrelevant vertex or identify a forced vertex in the planarizer. For $D_0$3 and $D_0$4, the paper uses the proposition that if there is an edge-deletion or edge-contraction planarizer of size at most $D_0$5, then there is also a vertex-deletion planarizer of size at most $D_0$6, so a Planar Vertex Deletion initializer suffices as a front end. For $D_0$7, the graph must already be planar, since adding edges cannot reduce non-planarity, and planarity after the additions is encoded by an MSOL formula $D_0$8. The annulus-gluing lemma then states that, for an annulus-embedded separator, the whole graph is planar iff both the inner and outer components are planar, which allows local and external edits to be composed safely.

The parameter dependence is explicitly enormous. The paper notes non-elementary blow-up in the conversion of $D_0$9 to Gaifman form and introduces a repetition threshold

$D_n$0

which dominates $D_n$1. The significance of the meta-theorem is therefore not practical small-parameter efficiency but a uniform FPT recipe that integrates locality-based model checking with a genuinely non-FOL global constraint (Fomin et al., 2021).

5. FoL++ as a region-aware VPR architecture

In "Region Matters: Efficient and Reliable Region-Aware Visual Place Recognition," FoL++ is a two-stage VPR framework rather than a logical formalism (Chen et al., 24 Apr 2026). The problem is perceptual aliasing: visually similar but geographically distinct scenes can yield similar descriptors because of irrelevant regions, transient occluders, and repetitive structures. The paper identifies two specific culprits, irrelevant regions and rigid candidate scheduling, and addresses them through reliability-aware local matching and adaptive re-ranking.

The architecture uses a DINOv2 ViT backbone, either ViT-B or ViT-L. For an image $D_n$2, the backbone yields

$D_n$3

tokens $D_n$4. Global retrieval is built from Sinkhorn-based assignment to $D_n$5 learnable clusters:

$D_n$6

After adding a dustbin cluster and applying Sinkhorn, token-wise saliency is defined by

$D_n$7

This map is reshaped spatially and used, as in SALAD, to aggregate a global descriptor $D_n$8.

Local features are produced by a lightweight two-layer deconvolutional decoder with channels $D_n$9, kernel $\mu$00, and stride $\mu$01. The novel component is the Reliability Estimation Branch, which fuses backbone features at scales $\mu$02, $\mu$03, and $\mu$04 into an $\mu$05 representation and regresses a reliability map

$\mu$06

where each element is interpreted as the unconditional probability that the corresponding local region is reliably matchable. After upsampling, the reliability map is combined with the global saliency map by

$\mu$07

Training uses four losses. The Extraction–Aggregation Spatial Alignment Loss is a symmetric Kullback–Leibler divergence between normalized versions of $\mu$08 and $\mu$09,

$\mu$10

The Saliency Contrast Enhancement Loss defines salient and irrelevant descriptors from the mask $\mu$11 and encourages salient descriptors to be close across positive pairs and far from irrelevant background. A pseudo-correspondence strategy supplies dense local supervision without manual annotation: within cluster-consistent candidate pools, correspondences are accepted when

$\mu$12

repeating until $\mu$13 pairs are obtained. The overall training objective is

$\mu$14

At inference time, FoL++ first retrieves the top-$\mu$15 candidates by global similarity. It then computes an adaptive candidate count from the mean of the top-$\mu$16 similarities with $\mu$17, a percentile statistic over $\mu$18, and the formula

$\mu$19

with $\mu$20. Local re-ranking uses only the top $\mu$21 of positions under $\mu$22, performs mutual nearest neighbor matching, and computes a reliability-weighted local score

$\mu$23

Global and local evidence are then fused by

$\mu$24

with $\mu$25 in the implementation.

The experimental section reports seven benchmarks: Pitts30k-test, Pitts250k, MSLS-val, Tokyo24/7, Nordland, Nordland⋆, and Eynsham. For the two-stage ViT-B model at $\mu$26, reported Recall@1/5/10 values include 94.7/97.4/98.1 on Pitts30k-test, 94.1/97.8/98.1 on MSLS-val, 96.1/98.6/99.2 on Nordland, 98.1/99.1/99.1 on Tokyo24/7, and 89.7/96.7/97.8 on Nordland⋆. The paper also reports that inference speed improves by about 40% over FoL, with latency dropping from $\mu$27 to $\mu$28 on A100 at $\mu$29, and memory footprint decreasing from $\mu$30 to $\mu$31 (Chen et al., 24 Apr 2026).

6. Comparative themes, misconceptions, and scope

Across these works, FoL++ consistently denotes an augmentation of a base formalism or pipeline rather than a wholesale replacement. In the intensional semantics paper, the base is ordinary FOL syntax, and the augmentation is a two-step intension/extension semantics over a PRP ontology that still recovers Tarski at each world (Majkic, 2011). In FO+lfp, the base is first-order reasoning over heaps, and the augmentation is least-fixpoint semantics together with an inductive proof discipline based on PFP obligations and counterexample-guided lemma synthesis (Murali et al., 2020). In the planarity meta-theorem, the base is an FOL sentence $\mu$32, and the augmentation is a graph-minor and locality framework that enforces a non-FOL structural target after editing (Fomin et al., 2021). In VPR, the base is a two-stage global-retrieval/local-verification system, and the augmentation is region reliability, adaptive candidate scheduling, and cross-stage fusion (Chen et al., 24 Apr 2026).

A recurrent misconception in the semantic literature is that modal predicate logic is automatically intensional. The minimal intensional FOL paper explicitly rejects this: when worlds are assignments and all symbols are rigid, the induced intensions are constant, so the logic is modal but not intensional (Majkic, 2011). A second misconception would be to treat FoL++ as a single standardized name across fields. The documented usages show otherwise. One family concerns semantics and proof systems for first-order logic; another concerns algorithmic meta-theorems around FOL; and one usage is a computer-vision architecture unrelated to logical intensionality or fixpoint semantics.

The limitations are equally domain-specific. The intensional formulation assumes constant domains and a PRP ontology, though varying domains can be simulated by an existence predicate (Majkic, 2011). FO+lfp remains undecidable, excludes same-layer negative recursion unless stratified, and only offers a semi-decision procedure relative to the grammar of candidate lemmas (Murali et al., 2020). The planarity meta-theorem has an $\mu$33 polynomial factor but an extremely large parameter dependence due to Gaifman translation, characteristic repetition thresholds, and Courcelle-style constants (Fomin et al., 2021). The VPR method remains challenged by severe illumination change and highly repetitive patterns, and it exposes the usual accuracy–latency trade-off in candidate count and mask tightness (Chen et al., 24 Apr 2026).

Taken together, these usages indicate that FoL++ functions as a label for principled enrichments: meaning over extension, fixpoints over first-order descriptions, algorithmic enforcement of non-FOL graph structure, or reliability-aware local evidence over global retrieval. This suggests a stable naming intuition even where the technical content is heterogeneous.