Dual Implication Point (DIP) in CDCL SAT Solvers

• Dual Implication Point (DIP) is defined as a minimal 2-vertex separator where neither individual node can block all paths from the decision node to the conflict node.
• DIP identification introduces extension variables that enable stronger clause learning and reduce proof sizes, thereby improving solver efficiency on complex CNF instances.
• Efficient O(|V|+|E|) algorithms detect DIP pairs, offering practical benefits in propagation and significantly decreasing median run times on combinatorially hard problems.

A Dual Implication Point (DIP) is a structural feature in the implication graph of a conflict-driven clause learning (CDCL) SAT solver, serving as a generalization of the classic Unique Implication Point (UIP). DIPs are formally defined as minimal 2-vertex separators—pairs of nodes that together, but not individually, must be traversed in every path from the current decision node to the conflict node. The identification and exploitation of DIPs enable the introduction of extension variables and the derivation of strictly stronger learned clauses via a restricted form of extended resolution. This approach provides substantial benefits in propagation, proof size, and solver efficiency, particularly on classes of combinatorially hard CNF instances (Buss et al., 2024).

1. Implication Graphs and Conflict Analysis

In a CDCL SAT solver, the implication graph $G = (V, E)$ at any point reflects the current partial assignment and propagation history. Each node $v \in V$ corresponds to a literal, identified either as a decision literal ($\dec(l)$), a propagated literal ($\prop(l)$), or the special conflict node ($\bot$). Edges model the logical implications induced by unit propagation: if a clause $C = l \lor l_1 \lor \cdots \lor l_k$ becomes unit under the assignment, edges are added from $\prop(l_i)$ (for all currently false $l_i$) to $\prop(l)$. The conflict node is introduced when a clause is falsified, and is pointed to by its last two falsified literals.

A decision level label is assigned to each node, distinguishing literals assigned at distinct decision steps versus those implied later. The structure of $G$ encodes both the local propagation dynamics and the global dependency structure necessary for effective backtracking and clause learning.

2. Classical Unique Implication Points (UIP) and Dominators

UIP-based clause learning is fundamental in modern SAT solvers. A node $v \in V$0 is said to dominate another node $v \in V$1 with respect to a designated start node $v \in V$2 if every path from $v \in V$3 to $v \in V$4 passes through $v \in V$5. Formally, $v \in V$6 holds if $v \in V$7.

In the context of conflict graphs, a UIP is a single literal (other than the most recent decision) that dominates $v \in V$8. The first UIP—closest to the conflict—is preferred, as it enables the extraction of backjumping clauses with low Literal Block Distance (LBD), enhancing propagation and pruning.

3. Formal Definition of Dual Implication Points

A DIP is defined as follows:

Let $v \in V$9 be an implication graph, $\dec(l)$0 the current (highest) decision node, and $\dec(l)$1 the conflict node. A pair $\dec(l)$2 is a Dual Implication Point if:

1. Separation: Every directed path $\dec(l)$3 from $\dec(l)$4 to $\dec(l)$5 contains at least one of $\dec(l)$6 or $\dec(l)$7 ($\dec(l)$8).
2. Minimality: Neither $\dec(l)$9 nor $\prop(l)$0 individually dominates $\prop(l)$1 (i.e., no singleton in the pair is a separator).

This property identifies $\prop(l)$2 as a minimal 2-vertex separator: any path is blocked unless it traverses at least one of the pair, but no member is individually sufficient. The DIP generalizes the UIP, and by Menger's theorem, in the absence of a UIP (a 1-vertex cut), a DIP (a 2-vertex cut) is the next minimal critical structure (Buss et al., 2024).

4. DIP Identification Algorithm

All minimal DIPs in a DAG can be computed in time $\prop(l)$3. The algorithm proceeds as follows:

1. Identify two vertex-disjoint paths $\prop(l)$4 and $\prop(l)$5 from $\prop(l)$6 to $\prop(l)$7.
2. Mark nodes on each path as bypassed if they can be reached via alternative paths not traversing themselves.
3. Mark forbidden pairs via crossing separators, i.e., the existence of paths that cross from one path to the other mid-way, invalidating the minimality of some pairs.
4. The remaining non-bypassed nodes on $\prop(l)$8 and $\prop(l)$9 yield the valid DIP pairs as minimal 2-vertex separators.

If the implication graph has a UIP, no DIP is present. If not, the presence of a DIP can be viewed as a fine-grained signal of combinatorial bottlenecks in the instance's logical structure.

5. Clause Learning and Extension Variables from DIPs

After identifying a DIP $\bot$0, a fresh extension variable $\bot$1 is introduced with the definition $\bot$2, encoded via three CNF clauses:

• $\bot$3
• $\bot$4
• $\bot$5

Conflict analysis then yields two essential learned clauses:

• Post-DIP clause: $\bot$6, where $\bot$7 collects lower-level literals implicating the conflict once $\bot$8 hold.
• Pre-DIP clause: $\bot$9, where $C = l \lor l_1 \lor \cdots \lor l_k$0 is the first-UIP and $C = l \lor l_1 \lor \cdots \lor l_k$1 comprises earlier-level literals leading up to $C = l \lor l_1 \lor \cdots \lor l_k$2.

The ability to abstract complicated conjunctions into a single variable not only reduces LBD and clause size but also frequently yields binary clauses, significantly improving propagation rates and proof conciseness. These properties are unattainable with ordinary 1-UIP learning (Buss et al., 2024).

6. Concrete Example of DIP Usage

For illustrative purposes, consider an implication graph excerpt where $C = l \lor l_1 \lor \cdots \lor l_k$3 is the decision node, $C = l \lor l_1 \lor \cdots \lor l_k$4, and two vertex-disjoint paths are identified:

• $C = l \lor l_1 \lor \cdots \lor l_k$5
• $C = l \lor l_1 \lor \cdots \lor l_k$6

By applying the DIP algorithm, nodes $C = l \lor l_1 \lor \cdots \lor l_k$7 (bypassed) and some pairs with crossing separators are excluded, and valid DIPs such as $C = l \lor l_1 \lor \cdots \lor l_k$8, $C = l \lor l_1 \lor \cdots \lor l_k$9, etc., are identified. Extension variables for these pairs are introduced in the clause database, and the resulting learned clauses substantially enhance solver performance.

7. Experimental Results and Practical Impact

Implementation of DIP-driven extended resolution clause learning in the MapleLCM solver, producing xMapleLCM, demonstrates:

• Significant increase in solved instances on hard classes such as random-regular and grid-based Tseitin formulas, and XORified formulas compared to standard MapleLCM and leading CDCL solvers (Kissat, CryptoMiniSat, SBVA+CaDiCaL).
• Order-of-magnitude reductions in median run time and proof size; for instance, median solving time halved on 4-regular Tseitin instances.
• Overhead of DIP detection and clause construction is consistently below 5% of total runtime for the majority of benchmarks.
• DIP learning is self-limiting on “easy” formulas, activating only when the solver detects that extended variables are beneficial in the branching heuristic.

This suggests that DIP-based learning can automate the benefits of extended resolution in practical solvers without substantial overhead, particularly for structurally complex problems (Buss et al., 2024).