Algoritmo di ottimizzazione approssimata quantistica

import matplotlib.pyplot as plt
import rustworkx as rx
from rustworkx.visualization import mpl_draw as draw_graph
import numpy as np
from scipy.optimize import minimize
from collections import defaultdict


from qiskit.quantum_info import SparsePauliOp
from qiskit.circuit.library import QAOAAnsatz
from qiskit.transpiler.preset_passmanagers import generate_preset_pass_manager

from qiskit_ibm_runtime import QiskitRuntimeService
from qiskit_ibm_runtime import Session, EstimatorV2 as Estimator
from qiskit_ibm_runtime import SamplerV2 as Sampler

from qiskit_addon_opt_mapper.applications import Maxcut
from qiskit_addon_opt_mapper.converters import OptimizationProblemToQubo
from qiskit_addon_opt_mapper.translators import to_ising

n_small = 5

graph = rx.PyGraph()
graph.add_nodes_from(np.arange(0, n_small, 1))
edge_list = [
    (0, 1, 1.0),
    (0, 2, 1.0),
    (0, 4, 1.0),
    (1, 2, 1.0),
    (2, 3, 1.0),
    (3, 4, 1.0),
]
graph.add_edges_from(edge_list)
draw_graph(graph, node_size=600, with_labels=True)

maxcut = Maxcut(graph)
maxcut_problem = maxcut.to_optimization_problem()
print(maxcut_problem.prettyprint())

qubo = OptimizationProblemToQubo().convert(maxcut_problem)
cost_hamiltonian, offset = to_ising(qubo)
print("Cost Function Hamiltonian:", cost_hamiltonian)
print("Offset:", offset)

circuit = QAOAAnsatz(cost_operator=cost_hamiltonian, reps=2)
circuit.measure_all()

circuit.draw("mpl")

circuit.parameters

service = QiskitRuntimeService()
backend = service.least_busy(
    operational=True, simulator=False, min_num_qubits=127
)
print(backend)

# Create pass manager for transpilation. Level 3 is the most aggressive
# preset: slower to transpile, but produces shorter circuits that are
# more robust to hardware noise.
pm = generate_preset_pass_manager(optimization_level=3, backend=backend)

candidate_circuit = pm.run(circuit)
candidate_circuit.draw("mpl", fold=False, idle_wires=False)

# QAOA doesn't prescribe principled default angles — any bounded choice
# works as a warm start for problems this small. beta and gamma are
# periodic (beta in [0, pi] and gamma in [0, 2*pi] modulo the underlying
# Pauli-rotation periods), and pi/2 and pi are just midpoints of those
# ranges. For harder problems you would typically warm start from known
# good angles or transfer parameters from smaller instances.
initial_gamma = np.pi
initial_beta = np.pi / 2
init_params = [initial_beta, initial_beta, initial_gamma, initial_gamma]

def cost_func_estimator(params, ansatz, hamiltonian, estimator):
    # transform the observable defined on virtual qubits to
    # an observable defined on all physical qubits
    isa_hamiltonian = hamiltonian.apply_layout(ansatz.layout)

    pub = (ansatz, isa_hamiltonian, params)
    job = estimator.run([pub])

    results = job.result()[0]
    cost = results.data.evs

    objective_func_vals.append(cost)

    return cost

objective_func_vals = []  # Global variable
with Session(backend=backend) as session:
    # If using qiskit-ibm-runtime<0.24.0, change `mode=` to `session=`
    estimator = Estimator(mode=session)
    estimator.options.default_shots = 1000

    # Set simple error suppression/mitigation options
    estimator.options.dynamical_decoupling.enable = True
    estimator.options.dynamical_decoupling.sequence_type = "XY4"
    estimator.options.twirling.enable_gates = True
    estimator.options.twirling.num_randomizations = "auto"
    estimator.options.environment.job_tags = ["TUT_QAOA"]

    result = minimize(
        cost_func_estimator,
        init_params,
        args=(candidate_circuit, cost_hamiltonian, estimator),
        method="COBYLA",
        tol=1e-2,
    )
    print(result)

plt.figure(figsize=(12, 6))
plt.plot(objective_func_vals)
plt.xlabel("Iteration")
plt.ylabel("Cost")
plt.show()

optimized_circuit = candidate_circuit.assign_parameters(result.x)
optimized_circuit.draw("mpl", fold=False, idle_wires=False)

# If using qiskit-ibm-runtime<0.24.0, change `mode=` to `backend=`
sampler = Sampler(mode=backend)
sampler.options.default_shots = 10000

# Set simple error suppression/mitigation options
sampler.options.dynamical_decoupling.enable = True
sampler.options.dynamical_decoupling.sequence_type = "XY4"
sampler.options.twirling.enable_gates = True
sampler.options.twirling.num_randomizations = "auto"

sampler.options.environment.job_tags = ["TUT_QAOA"]

pub = (optimized_circuit,)
job = sampler.run([pub], shots=int(1e4))
counts_int = job.result()[0].data.meas.get_int_counts()
counts_bin = job.result()[0].data.meas.get_counts()
shots = sum(counts_int.values())
final_distribution_int = {key: val / shots for key, val in counts_int.items()}
final_distribution_bin = {key: val / shots for key, val in counts_bin.items()}
print(final_distribution_int)

# auxiliary functions to sample most likely bitstring
def to_bitstring(integer, num_bits):
    result = np.binary_repr(integer, width=num_bits)
    return [int(digit) for digit in result]


keys = list(final_distribution_int.keys())
values = list(final_distribution_int.values())
most_likely = keys[np.argmax(np.abs(values))]
most_likely_bitstring = to_bitstring(most_likely, len(graph))
most_likely_bitstring.reverse()

print("Result bitstring:", most_likely_bitstring)

partition = maxcut.interpret(np.array(most_likely_bitstring))
print("Partition:", partition)

plt.rcParams.update({"font.size": 10})
final_bits = final_distribution_bin
values = np.abs(list(final_bits.values()))
top_4_values = sorted(values, reverse=True)[:4]
positions = []
for value in top_4_values:
    positions.append(np.where(values == value)[0])
fig = plt.figure(figsize=(11, 6))
ax = fig.add_subplot(1, 1, 1)
plt.xticks(rotation=45)
plt.title("Result Distribution")
plt.xlabel("Bitstrings (reversed)")
plt.ylabel("Probability")
ax.bar(list(final_bits.keys()), list(final_bits.values()), color="tab:grey")
for p in positions:
    ax.get_children()[int(p[0])].set_color("tab:purple")
plt.show()

# auxiliary function to plot graphs
def plot_result(G, x):
    colors = ["tab:grey" if i == 0 else "tab:purple" for i in x]
    pos, _default_axes = rx.spring_layout(G), plt.axes(frameon=True)
    rx.visualization.mpl_draw(
        G, node_color=colors, node_size=100, alpha=0.8, pos=pos
    )


plot_result(graph, most_likely_bitstring)

cut_value = maxcut_problem.objective.evaluate(np.array(most_likely_bitstring))
print("The value of the cut is:", cut_value)
print("Cut value implied by the optimized cost:", -(result.fun + offset))

# Classical baseline: enumerate all 2**n_small bitstrings and take the best cut.
def brute_force_max_cut(problem) -> tuple[int, list[int]]:
    n = problem.get_num_vars()
    best_cut = -1
    best_x: list[int] = []
    for i in range(2**n):
        x = np.array([(i >> k) & 1 for k in range(n)])
        cut = problem.objective.evaluate(x)
        if cut > best_cut:
            best_cut = int(cut)
            best_x = x.tolist()
    return best_cut, best_x


classical_best, classical_x = brute_force_max_cut(maxcut_problem)
print(f"Classical optimum (brute force): {classical_best}")
print(f"QAOA cut value:                  {cut_value}")

# Precomputed parity lookup table: _PARITY[b] = +1 if popcount(b) is even, else -1.
# We use this to vectorize expectation-value evaluation across all Pauli terms.
_PARITY = np.array(
    [-1 if bin(i).count("1") % 2 else 1 for i in range(256)],
    dtype=np.complex128,
)


def evaluate_sparse_pauli(state: int, observable: SparsePauliOp) -> complex:
    """Expectation value of a SparsePauliOp on a single computational-basis state.

    For a Z-only observable (which QAOA cost Hamiltonians are, after the
    QUBO-to-Hamiltonian mapping), the eigenvalue of each Pauli term on a
    computational-basis state is simply (-1)**popcount(z_mask AND state),
    i.e., the parity of the bitwise-AND of the term's Z-support and the
    measured bitstring.

    This routine packs the Z-support of every Pauli term into bytes, ANDs
    them against the measured state in a single vectorized op, and looks up
    the parity in _PARITY. For a 100-qubit / ~hundreds-of-terms Hamiltonian
    over 10_000 samples, this is dramatically faster than calling
    SparsePauliOp.expectation_value per sample.
    """
    packed_uint8 = np.packbits(observable.paulis.z, axis=1, bitorder="little")
    state_bytes = np.frombuffer(
        state.to_bytes(packed_uint8.shape[1], "little"), dtype=np.uint8
    )
    reduced = np.bitwise_xor.reduce(packed_uint8 & state_bytes, axis=1)
    return np.sum(observable.coeffs * _PARITY[reduced])


def best_solution(samples, hamiltonian):
    """Return the sampled bitstring (as int) with the lowest Hamiltonian cost."""
    min_cost = float("inf")
    min_sol = None
    for bit_str in samples.keys():
        candidate_sol = int(bit_str)
        fval = evaluate_sparse_pauli(candidate_sol, hamiltonian).real
        if fval <= min_cost:
            min_cost = fval
            min_sol = candidate_sol
    return min_sol


def _plot_cdf(objective_values: dict, ax, color):
    x_vals = sorted(objective_values.keys(), reverse=True)
    y_vals = np.cumsum([objective_values[x] for x in x_vals])
    ax.plot(x_vals, y_vals, color=color)


def plot_cdf(dist, ax, title):
    _plot_cdf(dist, ax, "C1")
    ax.vlines(min(list(dist.keys())), 0, 1, "C1", linestyle="--")
    ax.set_title(title)
    ax.set_xlabel("Objective function value")
    ax.set_ylabel("Cumulative distribution function")
    ax.grid(alpha=0.3)


def samples_to_objective_values(samples, hamiltonian):
    """Convert the samples to values of the objective function."""
    objective_values = defaultdict(float)
    for bit_str, prob in samples.items():
        candidate_sol = int(bit_str)
        fval = evaluate_sparse_pauli(candidate_sol, hamiltonian).real
        objective_values[fval] += prob
    return objective_values

# Step 1: build the 100-node graph, cost Hamiltonian, and QAOA ansatz.
n_large = 100
graph_100 = rx.PyGraph()
graph_100.add_nodes_from(np.arange(0, n_large, 1))
elist = set()
for u, v in backend.coupling_map:
    # The coupling map lists both directions of each coupler, so sort
    # the endpoints to avoid adding the same edge twice.
    if u < n_large and v < n_large:
        elist.add((min(u, v), max(u, v), 1.0))
graph_100.add_edges_from(sorted(elist))

maxcut_100 = Maxcut(graph_100)
maxcut_problem_100 = maxcut_100.to_optimization_problem()
qubo_100 = OptimizationProblemToQubo().convert(maxcut_problem_100)
cost_hamiltonian_100, offset_100 = to_ising(qubo_100)

circuit_100 = QAOAAnsatz(cost_operator=cost_hamiltonian_100, reps=1)
circuit_100.measure_all()

# Step 2: transpile for hardware.
pm = generate_preset_pass_manager(optimization_level=3, backend=backend)
candidate_circuit_100 = pm.run(circuit_100)

# Step 3: run the QAOA optimization loop on the device, then sample the
# final distribution with the optimized parameters.
initial_gamma = np.pi
initial_beta = np.pi / 2
init_params = [initial_beta, initial_gamma]

objective_func_vals = []  # Global variable
with Session(backend=backend) as session:
    estimator = Estimator(mode=session)
    estimator.options.default_shots = 1000

    # Set simple error suppression/mitigation options
    estimator.options.dynamical_decoupling.enable = True
    estimator.options.dynamical_decoupling.sequence_type = "XY4"
    estimator.options.twirling.enable_gates = True
    estimator.options.twirling.num_randomizations = "auto"
    estimator.options.environment.job_tags = ["TUT_QAOA"]

    result = minimize(
        cost_func_estimator,
        init_params,
        args=(candidate_circuit_100, cost_hamiltonian_100, estimator),
        method="COBYLA",
    )
    print(result)

# Assign optimal parameters and sample the final distribution.
optimized_circuit_100 = candidate_circuit_100.assign_parameters(result.x)

sampler = Sampler(mode=backend)
sampler.options.default_shots = 10000

# Set simple error suppression/mitigation options
sampler.options.dynamical_decoupling.enable = True
sampler.options.dynamical_decoupling.sequence_type = "XY4"
sampler.options.twirling.enable_gates = True
sampler.options.twirling.num_randomizations = "auto"

# Add a unique tag to the job execution
sampler.options.environment.job_tags = ["TUT_QAOA"]

pub = (optimized_circuit_100,)
job = sampler.run([pub], shots=int(1e4))

counts_int = job.result()[0].data.meas.get_int_counts()
shots = sum(counts_int.values())
final_distribution_100_int = {
    key: val / shots for key, val in counts_int.items()
}

# Step 4: find the best-cost sample and evaluate its cut value.
best_sol_100 = best_solution(final_distribution_100_int, cost_hamiltonian_100)
best_sol_bitstring_100 = to_bitstring(int(best_sol_100), len(graph_100))
best_sol_bitstring_100.reverse()

print("Result bitstring:", best_sol_bitstring_100)

cut_value_100 = maxcut_problem_100.objective.evaluate(
    np.array(best_sol_bitstring_100)
)
print("The value of the cut is:", cut_value_100)

# Plot convergence
plt.figure(figsize=(12, 6))
plt.plot(objective_func_vals)
plt.xlabel("Iteration")
plt.ylabel("Cost")
plt.show()

# Visualize the cut
plot_result(graph_100, best_sol_bitstring_100)

# Plot cumulative distribution function
result_dist = samples_to_objective_values(
    final_distribution_100_int, cost_hamiltonian_100
)
fig, ax = plt.subplots(1, 1, figsize=(8, 6))
plot_cdf(result_dist, ax, backend.name)