Fix #120: [Rule] GraphPartitioning to MaxCut (#707)

* Add plan for #120: [Rule] GraphPartitioning to MaxCut

* test: add GraphPartitioning to MaxCut reduction tests

* feat: add GraphPartitioning to MaxCut reduction

* test: add GraphPartitioning to MaxCut fixtures

* refactor: harden GraphPartitioning to MaxCut reduction

* docs: add GraphPartitioning to MaxCut reduction

* chore: remove plan file after implementation

* Fix formatting and clippy issues after merge with main

- Fix mod.rs ordering for graphpartitioning_maxcut
- Fix rustfmt issues in test file and qbf.rs (from merge)
- Fix clippy needless_range_loop in closestvectorproblem_qubo.rs

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Xiwei Pan <xiwei.pan@connect.hkust-gz.edu.cn>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: 3 people

docs/paper/reductions.typ

@@ -4262,6 +4262,52 @@ Each reduction is presented as a *Rule* (with linked problem names and overhead
   _Solution extraction._ For VC solution $C$, return $S = V backslash C$, i.e.\ flip each variable: $s_v = 1 - c_v$.
 ]
 
+#let gp_mc = load-example("GraphPartitioning", "MaxCut")
+#let gp_mc_sol = gp_mc.solutions.at(0)
+#let gp_mc_source_edges = gp_mc.source.instance.graph.edges.map(e => (e.at(0), e.at(1)))
+#let gp_mc_target_edges = gp_mc.target.instance.graph.edges.map(e => (e.at(0), e.at(1)))
+#let gp_mc_weights = gp_mc.target.instance.edge_weights
+#let gp_mc_nv = gp_mc.source.instance.graph.num_vertices
+#let gp_mc_ne = gp_mc_source_edges.len()
+#let gp_mc_penalty = gp_mc_ne + 1
+#let gp_mc_side_a = range(gp_mc_nv).filter(i => gp_mc_sol.source_config.at(i) == 0)
+#let gp_mc_side_b = range(gp_mc_nv).filter(i => gp_mc_sol.source_config.at(i) == 1)
+#let gp_mc_weight_lo = gp_mc_target_edges.enumerate().filter(((i, e)) => gp_mc_weights.at(i) == gp_mc_penalty - 1).map(((i, e)) => e)
+#let gp_mc_weight_hi = gp_mc_target_edges.enumerate().filter(((i, e)) => gp_mc_weights.at(i) == gp_mc_penalty).map(((i, e)) => e)
+#let gp_mc_source_cross = gp_mc_source_edges.filter(e => gp_mc_sol.source_config.at(e.at(0)) != gp_mc_sol.source_config.at(e.at(1)))
+#let gp_mc_cut_lo = gp_mc_target_edges.enumerate().filter(((i, e)) =>
+  gp_mc_weights.at(i) == gp_mc_penalty - 1 and
+  gp_mc_sol.target_config.at(e.at(0)) != gp_mc_sol.target_config.at(e.at(1))
+).map(((i, e)) => e)
+#let gp_mc_cut_hi = gp_mc_target_edges.enumerate().filter(((i, e)) =>
+  gp_mc_weights.at(i) == gp_mc_penalty and
+  gp_mc_sol.target_config.at(e.at(0)) != gp_mc_sol.target_config.at(e.at(1))
+).map(((i, e)) => e)
+#let gp_mc_cut_value = gp_mc_target_edges.enumerate().filter(((i, e)) =>
+  gp_mc_sol.target_config.at(e.at(0)) != gp_mc_sol.target_config.at(e.at(1))
+).map(((i, e)) => gp_mc_weights.at(i)).sum(default: 0)
+#reduction-rule("GraphPartitioning", "MaxCut",
+  example: true,
+  example-caption: [6-vertex minimum bisection to weighted Max-Cut],
+  extra: [
+    Here $m = #gp_mc_ne$, so $P = m + 1 = #gp_mc_penalty$ \
+    Weight $#(gp_mc_penalty - 1)$ edges (original edges): {#gp_mc_weight_lo.map(e => $(v_#(e.at(0)), v_#(e.at(1)))$).join(", ")} \
+    Weight $#gp_mc_penalty$ edges (non-edges): {#gp_mc_weight_hi.map(e => $(v_#(e.at(0)), v_#(e.at(1)))$).join(", ")} \
+    Canonical witness $A = {#gp_mc_side_a.map(i => $v_#i$).join(", ")}$, $B = {#gp_mc_side_b.map(i => $v_#i$).join(", ")}$ cuts source edges {#gp_mc_source_cross.map(e => $(v_#(e.at(0)), v_#(e.at(1)))$).join(", ")} and attains weighted cut $#gp_mc_cut_lo.len() * #(gp_mc_penalty - 1) + #gp_mc_cut_hi.len() * #gp_mc_penalty = #gp_mc_cut_value$ #sym.checkmark
+  ],
+)[
+  @garey1976 Graph Partitioning minimizes cut edges subject to a perfect-balance constraint, while Max-Cut maximizes a weighted cut without any balance constraint. A standard folklore construction in combinatorial optimization removes that constraint by rewarding every cross-pair equally and then subtracting one unit on original edges. The resulting weighted complete graph forces every optimum to be balanced first, and among balanced cuts it exactly minimizes the original bisection width.
+][
+  _Construction._ Given a Graph Partitioning instance $G = (V, E)$ with $n = |V|$ and $m = |E|$, set $P = m + 1$. Build the complete graph $G' = (V, E')$ on the same vertex set, where $E'$ contains every unordered pair $\{u, v\}$ with $u != v$. Assign weight $w'_(u, v) = P - 1$ when $(u, v) in E$, and $w'_(u, v) = P$ otherwise. For any partition $(A, B)$ of $V$, the weighted cut in $G'$ is
+  $ "cut"_(G')(A, B) = P |A| |B| - "cut"_G(A, B). $
+
+  _Correctness._ ($arrow.r.double$) Let $(A, B)$ be a maximum cut of $G'$. If it were unbalanced, then $|A| |B|$ would be at least one smaller than for a balanced partition $(A', B')$. Hence
+  $ "cut"_(G')(A', B') - "cut"_(G')(A, B) >= P - ("cut"_G(A', B') - "cut"_G(A, B)) >= P - m > 0, $
+  because $0 <= "cut"_G(·, ·) <= m$ and $P = m + 1$. Therefore every maximum cut of $G'$ is balanced. Among balanced partitions, $P |A| |B| = P (n slash 2)^2$ is constant, so maximizing $"cut"_(G')(A, B)$ is equivalent to minimizing $"cut"_G(A, B)$. ($arrow.l.double$) Conversely, every minimum bisection of $G$ is balanced and therefore maximizes $P |A| |B| - "cut"_G(A, B)$ in $G'$.
+
+  _Solution extraction._ Read off the same partition vector on the original vertex set: the Max-Cut bit for vertex $v$ is already the Graph Partitioning bit for $v$.
+]
+
 #let mis_clique = load-example("MaximumIndependentSet", "MaximumClique")
 #let mis_clique_sol = mis_clique.solutions.at(0)
 #reduction-rule("MaximumIndependentSet", "MaximumClique",

src/models/formula/qbf.rs

@@ -153,7 +153,6 @@ impl QuantifiedBooleanFormulas {
             }
         }
     }
-
 }
 
 impl Problem for QuantifiedBooleanFormulas {

src/rules/closestvectorproblem_qubo.rs

@@ -163,8 +163,7 @@ impl ReduceTo<QUBO<f64>> for ClosestVectorProblem<i32> {
             matrix[u][u] =
                 gram[var_u][var_u] * weight_u * weight_u + 2.0 * weight_u * g_lo_minus_h[var_u];
 
-            for v in (u + 1)..total_bits {
-                let (var_v, weight_v) = bit_terms[v];
+            for (v, &(var_v, weight_v)) in bit_terms.iter().enumerate().skip(u + 1) {
                 matrix[u][v] = 2.0 * gram[var_u][var_v] * weight_u * weight_v;
             }
         }

src/rules/graphpartitioning_maxcut.rs

@@ -0,0 +1,108 @@
+//! Reduction from GraphPartitioning to MaxCut on a weighted complete graph.
+
+use crate::models::graph::{GraphPartitioning, MaxCut};
+use crate::reduction;
+use crate::rules::traits::{ReduceTo, ReductionResult};
+use crate::topology::{Graph, SimpleGraph};
+
+/// Result of reducing GraphPartitioning to MaxCut.
+#[derive(Debug, Clone)]
+pub struct ReductionGPToMaxCut {
+    target: MaxCut<SimpleGraph, i32>,
+}
+
+#[cfg(any(test, feature = "example-db"))]
+const ISSUE_EXAMPLE_WITNESS: [usize; 6] = [0, 0, 0, 1, 1, 1];
+
+impl ReductionResult for ReductionGPToMaxCut {
+    type Source = GraphPartitioning<SimpleGraph>;
+    type Target = MaxCut<SimpleGraph, i32>;
+
+    fn target_problem(&self) -> &Self::Target {
+        &self.target
+    }
+
+    fn extract_solution(&self, target_solution: &[usize]) -> Vec<usize> {
+        target_solution.to_vec()
+    }
+}
+
+#[cfg(any(test, feature = "example-db"))]
+fn issue_example() -> GraphPartitioning<SimpleGraph> {
+    GraphPartitioning::new(SimpleGraph::new(
+        6,
+        vec![
+            (0, 1),
+            (0, 2),
+            (1, 2),
+            (1, 3),
+            (2, 3),
+            (2, 4),
+            (3, 4),
+            (3, 5),
+            (4, 5),
+        ],
+    ))
+}
+
+fn complete_graph_edges_and_weights(graph: &SimpleGraph) -> (Vec<(usize, usize)>, Vec<i32>) {
+    let num_vertices = graph.num_vertices();
+    let p = penalty_weight(graph.num_edges());
+    let mut edges = Vec::new();
+    let mut weights = Vec::new();
+
+    for u in 0..num_vertices {
+        for v in (u + 1)..num_vertices {
+            edges.push((u, v));
+            weights.push(if graph.has_edge(u, v) { p - 1 } else { p });
+        }
+    }
+
+    (edges, weights)
+}
+
+fn penalty_weight(num_edges: usize) -> i32 {
+    i32::try_from(num_edges)
+        .ok()
+        .and_then(|num_edges| num_edges.checked_add(1))
+        .expect("GraphPartitioning -> MaxCut penalty exceeds i32 range")
+}
+
+#[reduction(
+    overhead = {
+        num_vertices = "num_vertices",
+        num_edges = "num_vertices * (num_vertices - 1) / 2",
+    }
+)]
+impl ReduceTo<MaxCut<SimpleGraph, i32>> for GraphPartitioning<SimpleGraph> {
+    type Result = ReductionGPToMaxCut;
+
+    fn reduce_to(&self) -> Self::Result {
+        let (edges, weights) = complete_graph_edges_and_weights(self.graph());
+        let target = MaxCut::new(SimpleGraph::new(self.num_vertices(), edges), weights);
+
+        ReductionGPToMaxCut { target }
+    }
+}
+
+#[cfg(feature = "example-db")]
+pub(crate) fn canonical_rule_example_specs() -> Vec<crate::example_db::specs::RuleExampleSpec> {
+    use crate::export::SolutionPair;
+
+    vec![crate::example_db::specs::RuleExampleSpec {
+        id: "graphpartitioning_to_maxcut",
+        build: || {
+            crate::example_db::specs::rule_example_with_witness::<_, MaxCut<SimpleGraph, i32>>(
+                issue_example(),
+                SolutionPair {
+                    source_config: ISSUE_EXAMPLE_WITNESS.to_vec(),
+                    target_config: ISSUE_EXAMPLE_WITNESS.to_vec(),
+                },
+            )
+        },
+    }]
+}
+
+#[cfg(test)]
+#[path = "../unit_tests/rules/graphpartitioning_maxcut.rs"]
+mod tests;

src/rules/mod.rs

@@ -11,6 +11,7 @@ mod closestvectorproblem_qubo;
 pub(crate) mod coloring_qubo;
 pub(crate) mod factoring_circuit;
 mod graph;
+pub(crate) mod graphpartitioning_maxcut;
 mod kcoloring_casts;
 mod knapsack_qubo;
 mod ksatisfiability_casts;
@@ -93,6 +94,7 @@ pub(crate) fn canonical_rule_example_specs() -> Vec<crate::example_db::specs::Ru
     specs.extend(closestvectorproblem_qubo::canonical_rule_example_specs());
     specs.extend(coloring_qubo::canonical_rule_example_specs());
     specs.extend(factoring_circuit::canonical_rule_example_specs());
+    specs.extend(graphpartitioning_maxcut::canonical_rule_example_specs());
     specs.extend(knapsack_qubo::canonical_rule_example_specs());
     specs.extend(ksatisfiability_qubo::canonical_rule_example_specs());
     specs.extend(ksatisfiability_subsetsum::canonical_rule_example_specs());

src/unit_tests/models/formula/qbf.rs

@@ -255,10 +255,7 @@ fn test_qbf_quantifier_clone() {
 #[test]
 fn test_qbf_empty_clause() {
     // An empty clause (disjunction of zero literals) is always false
-    let problem = QuantifiedBooleanFormulas::new(
-        1,
-        vec![Quantifier::Exists],
-        vec![CNFClause::new(vec![])],
-    );
+    let problem =
+        QuantifiedBooleanFormulas::new(1, vec![Quantifier::Exists], vec![CNFClause::new(vec![])]);
     assert!(!problem.is_true());
 }

src/unit_tests/rules/graphpartitioning_maxcut.rs

@@ -0,0 +1,65 @@
+use crate::models::graph::{GraphPartitioning, MaxCut};
+use crate::rules::test_helpers::assert_optimization_round_trip_from_optimization_target;
+use crate::rules::{ReduceTo, ReductionResult};
+use crate::topology::{Graph, SimpleGraph};
+
+fn issue_example() -> GraphPartitioning<SimpleGraph> {
+    super::issue_example()
+}
+
+#[test]
+fn test_graphpartitioning_to_maxcut_closed_loop() {
+    let source = issue_example();
+    let reduction = ReduceTo::<MaxCut<SimpleGraph, i32>>::reduce_to(&source);
+
+    assert_optimization_round_trip_from_optimization_target(
+        &source,
+        &reduction,
+        "GraphPartitioning->MaxCut closed loop",
+    );
+}
+
+#[test]
+fn test_graphpartitioning_to_maxcut_target_structure() {
+    let source = issue_example();
+    let reduction = ReduceTo::<MaxCut<SimpleGraph, i32>>::reduce_to(&source);
+    let target = reduction.target_problem();
+    let num_vertices = source.num_vertices();
+    let penalty = i32::try_from(source.num_edges()).unwrap() + 1;
+
+    assert_eq!(target.num_vertices(), num_vertices);
+    assert_eq!(target.num_edges(), num_vertices * (num_vertices - 1) / 2);
+
+    for u in 0..num_vertices {
+        for v in (u + 1)..num_vertices {
+            let expected_weight = if source.graph().has_edge(u, v) {
+                penalty - 1
+            } else {
+                penalty
+            };
+            assert_eq!(
+                target.edge_weight(u, v),
+                Some(&expected_weight),
+                "unexpected weight on edge ({u}, {v})"
+            );
+        }
+    }
+}
+
+#[test]
+fn test_graphpartitioning_to_maxcut_extract_solution_identity() {
+    let source = issue_example();
+    let reduction = ReduceTo::<MaxCut<SimpleGraph, i32>>::reduce_to(&source);
+    let target_solution = super::ISSUE_EXAMPLE_WITNESS.to_vec();
+
+    assert_eq!(
+        reduction.extract_solution(&target_solution),
+        target_solution
+    );
+}
+
+#[test]
+fn test_graphpartitioning_to_maxcut_penalty_overflow_panics() {
+    let result = std::panic::catch_unwind(|| super::penalty_weight(i32::MAX as usize));
+    assert!(result.is_err());
+}