Implication Graph Construction

• Implication Graph Construction is a method for systematically building directed hypergraphs that capture logical entailment relationships among propositional formulas.
• It involves extracting minimal hyperedges by pruning redundant and non-immediate implications, and constructing an adjacency matrix to compute information values.
• The approach underpins advanced clause learning in CDCL SAT solvers by leveraging Unique and Dual Implication Points to optimize conflict resolution.

Implication graph construction refers to the systematic process of building a directed structure that encodes logical entailment relationships among propositional formulas or, in the context of SAT solving, propositional literals. Such graphs serve as foundational objects in both quantitative frameworks for propositional information and in conflict-driven clause learning (CDCL) with advanced resolution strategies. The construction is central to quantifying informativeness in formal theories and to enabling efficient, generalized clause-learning techniques during propositional reasoning.

1. Formal Definition of an Implication Hypergraph

An implication hypergraph is defined as a directed hypergraph where each vertex denotes a propositional atom or formula, and each hyperedge represents a semantic implication between conjunctions of formulas. Formally, an implication hypergraph is a pair $H = (V, E)$, where:

• $V$ is a finite set of vertices, each $p_i \in V$ denoting a propositional entity.
• $E$ is a set of directed hyperedges; each $e = (T(e), H(e))$ with $T(e), H(e) \subseteq V$, encoding the logical relationship:

$$\bigwedge_{P \in T(e)} P \implies \bigwedge_{Q \in H(e)} Q$$

Construction is restricted to strict and minimal hypergraphs:

• Strictness prohibits mutual implications between any pair ($v \rightarrow w$ and $w \rightarrow v$).
• Minimality implies (i) absence of self-loops ($v \rightarrow v$), (ii) exclusion of implied edges if an indirect path exists, and (iii) no redundant tails—no $(T' \supset T) \rightarrow H$ if $(T \rightarrow H)$ exists.

2. Extraction of Minimal Hyperedges from a Propositional Theory

Given a base propositional theory $\Delta$, the minimal set of implication hyperedges is constructed by:

1. Parsing each implication of the form $$(A_1 \wedge \dots \wedge A_k) \implies (B_1 \wedge \dots \wedge B_m)$$ into a hyperedge $(T, H)$.
2. Pruning redundant structures via:
   • Elimination of self-loops.
   • Removal of non-immediate implications, such that for any $v \rightarrow u$ inferred via a path, direct $v \rightarrow u$ edges not representing immediate implications are excluded.
   • Omission of hyperedges with redundant tails by checking that no proper superset of $T$ suffices for the same implication. The resulting hypergraph encodes only the strongest, smallest-tail implications necessary.

Step	Description	Complexity
Parse base theory $\Delta$	Map each implication to a candidate hyperedge	$O(|\Delta|)$
Prune for minimality	Remove loops, indirect, and redundant edges	Up to $2^{|V|}$; exponential
Output minimal strict hypergraph	Set $E$ encodes direct, minimal implications

3. Construction of the Implication Adjacency Matrix and Information Equations

Once a minimal implication hypergraph $H = (V, E)$ is available, the $n \times n$ adjacency matrix $A$ is constructed, where $n = |V|$. For each $p_i$, let $E_i = \{ e \in E \mid p_i \in H(e) \}$. Entries are computed as:

$$A_{ij} = \sum_{e \in E_i} \frac{1}{|H(e)|} \mathbf{1}_{\{p_j \in T(e)\}}$$

The propositional information values $\mathcal{I}_{\nu, \epsilon}(p_i)$, with positive parameters $\nu, \epsilon$, are defined recursively:

• For leaf nodes $p_i$ (no outgoing hyperedges): $\mathcal{I}_{\nu, \epsilon}(p_i) = \nu$
• Otherwise:

$$\mathcal{I}_{\nu, \epsilon}(p_i) = \sum_{j=1}^n A_{ij} (\mathcal{I}_{\nu, \epsilon}(p_j) + \epsilon)$$

Collecting these into an $n$-vector $I$ and defining $\ell$ to indicate leaves, the solution in closed form is:

$$\mathcal{I}_{\nu, \epsilon} = (I - A)^{-1} (\epsilon A \mathbf{1} + \nu \ell)$$

Inversion of $(I - A)$ is feasible if $1 \not\in \mathrm{spec}(A)$, which holds if the hypergraph is acyclic (Dalal, 31 Jan 2025).

4. Application in Clause Learning: Implication Graphs in CDCL

In CDCL SAT solvers, the implication graph $G = (V, E)$ dynamically traces the flow of reasoned assignments under the current partial assignment ("trail"):

• Nodes represent assigned literals (at various decision levels), decision sources, and a special conflict node $\perp$.
• Propagations due to falsified clauses introduce edges from antecedent literals to propagated literals.
• Conflicts correspond with edges into the sink $\perp$.
• This digraph is acyclic by construction, encoding the sequence of logical inferences leading to each literal. Key concepts:
• Unique Implication Point (UIP): A node $u$ through which every path from the most recent decision literal $d$ to $\perp$ passes.
• Dual Implication Point (DIP): A minimal pair $\{x, y\}$ such that every path from $d$ to $\perp$ meets at least one of $x$ or $y$, but neither alone is sufficient (Buss et al., 2024).

5. Algorithms for Implication Graph Analysis and DIP Detection

Extraction of DIPs in runtime implication graphs can proceed by:

• Classic Dominator Tree: Computes dominator sets, then enumerates two-element separators on paths from $d$ to $\perp$. For each candidate pair $\{w_i, w_j\}$ from the dominator chain, check minimal-separator properties.
• Two-Vertex Bottleneck (TVB) Algorithm: Linear-time routine for directed acyclic graphs. It constructs two vertex-disjoint $d \to \perp$ paths, identifies nodes not bypassed by alternate paths, and excludes node pairs separated by "crossing separators." Valid pairs constitute DIPs. xMapleLCM implements this method, yielding practical scalability (Buss et al., 2024).

6. Introducing Extension Variables and Clause Learning

Upon identification of a DIP $\{a, b\}$:

• Introduce a new variable $z$ with the definition:

$z \leftrightarrow (a \wedge b)$

captured by the Tseitin clauses

$$\neg z \vee a,\quad \neg z \vee b,\quad z \vee \neg a \vee \neg b$$

• Learn the post-DIP clause $\neg z \vee \bigvee_{d \in D} \neg d$ and the pre-DIP clause $\neg f \vee \bigvee_{c \in C} \neg c \vee z$, where $f$ is the 1-UIP, and $C,D$ partition lower-level literals by connectivity relative to the DIP. This process realizes an extended-resolution proof step, potentially shortening subsequent derived clauses (Buss et al., 2024).

7. Computational Complexity, Limitations, and Refinements

Extracting all minimal hyperedges for implication hypergraph construction is, in the worst case, exponential in $|V|$ due to consideration of all possible antecedent subsets. Practical approaches utilize restriction to known atomic implications or SAT-based closure checks. For large $|E|$, constructing $A$ is $O(n \cdot |E|)$ and solving $(I - A)I = \epsilon A\mathbf{1} + \nu \ell$ is $O(n^3)$; for larger $n$, iterative methods such as Jacobi iteration are effective under $\rho(A) < 1$.

Limitations and possible refinements include:

• The uniform $1/|H(e)|$ credit assignment may be too coarse; it can be replaced with weighted hyperedges derived from corpus statistics.
• Parameter $\epsilon$ is a crude buffer for incompleteness; data-driven calibration strategies may be warranted.
• The framework does not quantify joint information of conjunctions.

Potential extensions include weighted hyperedge frameworks, nonconstant leaf-value distributions, and generalization from propositional to first-order formulae (Dalal, 31 Jan 2025).

---

For detailed frameworks, algorithms, and further worked examples on implication graph construction and its applications to both propositional information quantification and clause learning in CDCL SAT solvers, see "Formalising Propositional Information via Implication Hypergraphs" (Dalal, 31 Jan 2025) and "Extended Resolution Clause Learning via Dual Implication Points" (Buss et al., 2024).