Skip to main content

ruqu_algorithms/
qaoa.rs

1//! Quantum Approximate Optimization Algorithm (QAOA) for MaxCut
2//!
3//! QAOA is a hybrid classical-quantum algorithm for combinatorial optimization.
4//! This module implements the **MaxCut** variant: given an undirected weighted
5//! graph, find a partition of vertices into two sets that maximizes the total
6//! weight of edges crossing the partition.
7//!
8//! # Circuit structure
9//!
10//! A depth-p QAOA circuit has the form:
11//!
12//! ```text
13//! |+>^n --[C(gamma_1)][B(beta_1)]--...--[C(gamma_p)][B(beta_p)]-- measure
14//! ```
15//!
16//! where:
17//! - **Phase separator** C(gamma) = prod_{(i,j) in E} exp(-i * gamma * w_ij * Z_i Z_j)
18//!   is implemented with Rzz gates.
19//! - **Mixer** B(beta) = prod_i exp(-i * beta * X_i) is implemented with Rx gates.
20//!
21//! The 2p parameters (gamma_1..gamma_p, beta_1..beta_p) are optimized
22//! classically to maximize the expected cut value.
23
24use ruqu_core::circuit::QuantumCircuit;
25use ruqu_core::simulator::{SimConfig, Simulator};
26use ruqu_core::types::{PauliOp, PauliString};
27
28// ---------------------------------------------------------------------------
29// Graph representation
30// ---------------------------------------------------------------------------
31
32/// Simple undirected weighted graph for MaxCut problems.
33#[derive(Debug, Clone)]
34pub struct Graph {
35    /// Number of vertices (each mapped to one qubit).
36    pub num_nodes: u32,
37    /// Edges as `(node_i, node_j, weight)` triples. Both directions are
38    /// represented by a single entry (undirected).
39    pub edges: Vec<(u32, u32, f64)>,
40}
41
42impl Graph {
43    /// Create an empty graph with the given number of nodes.
44    pub fn new(num_nodes: u32) -> Self {
45        Self {
46            num_nodes,
47            edges: Vec::new(),
48        }
49    }
50
51    /// Add an undirected weighted edge between nodes `i` and `j`.
52    ///
53    /// # Panics
54    ///
55    /// Panics if `i` or `j` is out of range.
56    pub fn add_edge(&mut self, i: u32, j: u32, weight: f64) {
57        assert!(i < self.num_nodes, "node index {} out of range", i);
58        assert!(j < self.num_nodes, "node index {} out of range", j);
59        self.edges.push((i, j, weight));
60    }
61
62    /// Convenience constructor for an unweighted graph (all weights = 1.0).
63    pub fn unweighted(num_nodes: u32, edges: Vec<(u32, u32)>) -> Self {
64        let weighted: Vec<(u32, u32, f64)> = edges.into_iter().map(|(i, j)| (i, j, 1.0)).collect();
65        Self {
66            num_nodes,
67            edges: weighted,
68        }
69    }
70
71    /// Return the total number of edges.
72    pub fn num_edges(&self) -> usize {
73        self.edges.len()
74    }
75}
76
77// ---------------------------------------------------------------------------
78// Configuration and result types
79// ---------------------------------------------------------------------------
80
81/// Configuration for a QAOA MaxCut run.
82pub struct QaoaConfig {
83    /// The graph instance to solve MaxCut on.
84    pub graph: Graph,
85    /// QAOA depth (number of alternating phase-separation / mixing layers).
86    pub p: u32,
87    /// Maximum number of classical optimizer iterations.
88    pub max_iterations: u32,
89    /// Step size for gradient ascent.
90    pub learning_rate: f64,
91    /// Optional RNG seed for reproducible simulation.
92    pub seed: Option<u64>,
93}
94
95/// Result of a QAOA MaxCut run.
96pub struct QaoaResult {
97    /// Highest expected cut value found.
98    pub best_cut_value: f64,
99    /// Bitstring that achieves (or approximates) `best_cut_value`.
100    /// `best_bitstring[v]` is `true` when vertex `v` belongs to partition S1.
101    pub best_bitstring: Vec<bool>,
102    /// Optimized gamma parameters (phase-separation angles).
103    pub optimal_gammas: Vec<f64>,
104    /// Optimized beta parameters (mixer angles).
105    pub optimal_betas: Vec<f64>,
106    /// Expected cut value at each iteration.
107    pub energy_history: Vec<f64>,
108    /// Whether the optimizer converged.
109    pub converged: bool,
110}
111
112// ---------------------------------------------------------------------------
113// Circuit construction
114// ---------------------------------------------------------------------------
115
116/// Build a QAOA circuit for the MaxCut problem on `graph`.
117///
118/// The circuit starts with Hadamard on every qubit (equal superposition),
119/// then applies `p` alternating layers:
120///
121/// 1. **Phase separation**: `Rzz(2 * gamma * w)` on each edge `(i, j, w)`.
122/// 2. **Mixing**: `Rx(2 * beta)` on each qubit.
123///
124/// `gammas` and `betas` must each have length `p`.
125pub fn build_qaoa_circuit(graph: &Graph, gammas: &[f64], betas: &[f64]) -> QuantumCircuit {
126    assert_eq!(gammas.len(), betas.len(), "gammas and betas must have equal length");
127    let n = graph.num_nodes;
128    let p = gammas.len();
129    let mut circuit = QuantumCircuit::new(n);
130
131    // Initial equal superposition
132    for q in 0..n {
133        circuit.h(q);
134    }
135
136    // QAOA layers
137    for layer in 0..p {
138        // Phase separator: Rzz for each edge
139        for &(i, j, w) in &graph.edges {
140            circuit.rzz(i, j, 2.0 * gammas[layer] * w);
141        }
142        // Mixer: Rx on each qubit
143        for q in 0..n {
144            circuit.rx(q, 2.0 * betas[layer]);
145        }
146    }
147
148    circuit
149}
150
151// ---------------------------------------------------------------------------
152// Cost evaluation
153// ---------------------------------------------------------------------------
154
155/// Compute the classical MaxCut value for a given bitstring.
156///
157/// An edge (i, j, w) contributes `w` to the cut if `bitstring[i] != bitstring[j]`.
158pub fn cut_value(graph: &Graph, bitstring: &[bool]) -> f64 {
159    graph
160        .edges
161        .iter()
162        .filter(|(i, j, _)| bitstring[*i as usize] != bitstring[*j as usize])
163        .map(|(_, _, w)| w)
164        .sum()
165}
166
167/// Evaluate the expected MaxCut cost from a QAOA state.
168///
169/// For each edge (i, j) with weight w:
170/// ```text
171/// C_{ij} = w * 0.5 * (1 - <Z_i Z_j>)
172/// ```
173///
174/// The total expected cost is the sum over all edges.
175pub fn evaluate_qaoa_cost(
176    graph: &Graph,
177    gammas: &[f64],
178    betas: &[f64],
179    seed: Option<u64>,
180) -> ruqu_core::error::Result<f64> {
181    let circuit = build_qaoa_circuit(graph, gammas, betas);
182    let sim_config = SimConfig {
183        seed,
184        noise: None,
185        shots: None,
186    };
187    let result = Simulator::run_with_config(&circuit, &sim_config)?;
188
189    let mut cost = 0.0;
190    for &(i, j, w) in &graph.edges {
191        let zz = result.state.expectation_value(&PauliString {
192            ops: vec![(i, PauliOp::Z), (j, PauliOp::Z)],
193        });
194        cost += w * 0.5 * (1.0 - zz);
195    }
196    Ok(cost)
197}
198
199// ---------------------------------------------------------------------------
200// QAOA optimizer
201// ---------------------------------------------------------------------------
202
203/// Run QAOA optimization for MaxCut using gradient ascent with the
204/// parameter-shift rule.
205///
206/// The optimizer maximizes the expected cut value by adjusting gamma and beta
207/// parameters. Convergence is declared when the absolute change in cost
208/// between successive iterations drops below 1e-6.
209///
210/// # Errors
211///
212/// Returns a [`ruqu_core::error::QuantumError`] on simulator failures.
213pub fn run_qaoa(config: &QaoaConfig) -> ruqu_core::error::Result<QaoaResult> {
214    let p = config.p as usize;
215
216    // Initialize parameters at reasonable starting values.
217    let mut gammas = vec![0.5_f64; p];
218    let mut betas = vec![0.5_f64; p];
219    let mut energy_history: Vec<f64> = Vec::with_capacity(config.max_iterations as usize);
220    let mut best_cost = f64::NEG_INFINITY;
221    let mut best_bitstring = vec![false; config.graph.num_nodes as usize];
222    let mut converged = false;
223
224    for iter in 0..config.max_iterations {
225        // ------------------------------------------------------------------
226        // Evaluate current expected cost
227        // ------------------------------------------------------------------
228        let cost = evaluate_qaoa_cost(&config.graph, &gammas, &betas, config.seed)?;
229        energy_history.push(cost);
230
231        // ------------------------------------------------------------------
232        // Track best solution: sample the most probable bitstring
233        // ------------------------------------------------------------------
234        if cost > best_cost {
235            best_cost = cost;
236            let circuit = build_qaoa_circuit(&config.graph, &gammas, &betas);
237            let sim_result = Simulator::run_with_config(
238                &circuit,
239                &SimConfig {
240                    seed: config.seed,
241                    noise: None,
242                    shots: None,
243                },
244            )?;
245            let probs = sim_result.state.probabilities();
246            let best_idx = probs
247                .iter()
248                .enumerate()
249                .max_by(|a, b| a.1.partial_cmp(b.1).unwrap_or(std::cmp::Ordering::Equal))
250                .map(|(i, _)| i)
251                .unwrap_or(0);
252            best_bitstring = (0..config.graph.num_nodes)
253                .map(|q| (best_idx >> q) & 1 == 1)
254                .collect();
255        }
256
257        // ------------------------------------------------------------------
258        // Convergence check
259        // ------------------------------------------------------------------
260        if iter > 0 {
261            let prev = energy_history[iter as usize - 1];
262            if (cost - prev).abs() < 1e-6 {
263                converged = true;
264                break;
265            }
266        }
267
268        // ------------------------------------------------------------------
269        // Gradient ascent via parameter-shift rule
270        // ------------------------------------------------------------------
271        let shift = std::f64::consts::FRAC_PI_2;
272
273        // Update gamma parameters
274        for i in 0..p {
275            let mut gp = gammas.clone();
276            gp[i] += shift;
277            let mut gm = gammas.clone();
278            gm[i] -= shift;
279            let cp = evaluate_qaoa_cost(&config.graph, &gp, &betas, config.seed)?;
280            let cm = evaluate_qaoa_cost(&config.graph, &gm, &betas, config.seed)?;
281            gammas[i] += config.learning_rate * (cp - cm) / 2.0;
282        }
283
284        // Update beta parameters
285        for i in 0..p {
286            let mut bp = betas.clone();
287            bp[i] += shift;
288            let mut bm = betas.clone();
289            bm[i] -= shift;
290            let cp = evaluate_qaoa_cost(&config.graph, &gammas, &bp, config.seed)?;
291            let cm = evaluate_qaoa_cost(&config.graph, &gammas, &bm, config.seed)?;
292            betas[i] += config.learning_rate * (cp - cm) / 2.0;
293        }
294    }
295
296    Ok(QaoaResult {
297        best_cut_value: best_cost,
298        best_bitstring,
299        optimal_gammas: gammas,
300        optimal_betas: betas,
301        energy_history,
302        converged,
303    })
304}
305
306// ---------------------------------------------------------------------------
307// Graph construction helpers
308// ---------------------------------------------------------------------------
309
310/// Create a triangle graph (3 nodes, 3 edges, all weight 1).
311///
312/// The optimal MaxCut is 2 (any partition has exactly one edge within a
313/// group and two edges crossing).
314pub fn triangle_graph() -> Graph {
315    Graph::unweighted(3, vec![(0, 1), (1, 2), (0, 2)])
316}
317
318/// Create a 4-node ring graph (cycle C4, all weight 1).
319///
320/// The optimal MaxCut is 4 (bipartition {0,2} vs {1,3} cuts all edges).
321pub fn ring4_graph() -> Graph {
322    Graph::unweighted(4, vec![(0, 1), (1, 2), (2, 3), (3, 0)])
323}
324
325// ---------------------------------------------------------------------------
326// Tests
327// ---------------------------------------------------------------------------
328
329#[cfg(test)]
330mod tests {
331    use super::*;
332
333    #[test]
334    fn test_graph_construction() {
335        let g = triangle_graph();
336        assert_eq!(g.num_nodes, 3);
337        assert_eq!(g.num_edges(), 3);
338    }
339
340    #[test]
341    fn test_graph_add_edge() {
342        let mut g = Graph::new(4);
343        g.add_edge(0, 1, 2.5);
344        g.add_edge(2, 3, 1.0);
345        assert_eq!(g.num_edges(), 2);
346    }
347
348    #[test]
349    #[should_panic(expected = "node index 5 out of range")]
350    fn test_graph_add_edge_out_of_range() {
351        let mut g = Graph::new(4);
352        g.add_edge(0, 5, 1.0);
353    }
354
355    #[test]
356    fn test_cut_value_triangle() {
357        let g = triangle_graph();
358        // Partition {0} vs {1,2}: edges (0,1) and (0,2) are cut, (1,2) is not.
359        assert_eq!(cut_value(&g, &[true, false, false]), 2.0);
360        // All same partition: no cut.
361        assert_eq!(cut_value(&g, &[false, false, false]), 0.0);
362    }
363
364    #[test]
365    fn test_cut_value_ring4() {
366        let g = ring4_graph();
367        // Optimal: alternate partitions {0,2} vs {1,3} -> cut all 4 edges.
368        assert_eq!(cut_value(&g, &[true, false, true, false]), 4.0);
369    }
370
371    #[test]
372    fn test_build_qaoa_circuit_gate_count() {
373        let g = triangle_graph();
374        let gammas = vec![0.5];
375        let betas = vec![0.3];
376        let circuit = build_qaoa_circuit(&g, &gammas, &betas);
377        assert_eq!(circuit.num_qubits(), 3);
378        // 3 H + 3 Rzz + 3 Rx = 9 gates
379        assert_eq!(circuit.gates().len(), 9);
380    }
381
382    #[test]
383    fn test_cut_value_weighted() {
384        let mut g = Graph::new(3);
385        g.add_edge(0, 1, 2.0);
386        g.add_edge(1, 2, 3.0);
387        // Partition {0,2} vs {1}: cuts both edges -> 2.0 + 3.0 = 5.0
388        assert_eq!(cut_value(&g, &[true, false, true]), 5.0);
389    }
390}