qanneal 0.6.1

Research-grade simulated quantum annealing toolkit

qanneal

Research-grade Ising/QUBO optimizer — version 0.6.1

Four annealing engines from classical SA to the closest available CPU simulation of quantum annealing, plus a theoretically-grounded optimal adaptive J⊥ schedule derived from the local adiabaticity condition.

Method	Short name	What it simulates
Simulated Annealing	sa	Classical thermal fluctuations
Simulated Quantum Annealing	sqa	Discrete-time path-integral (Trotterized QA)
SQA + Parallel Tempering	sqapt	QA with replica exchange on a (β, Γ) ladder
Continuous-Time PIMC	ctpimc	Continuous-time path-integral (worldline sampling)

All engines share a unified solve() Python API and a C++17 core with pybind11 bindings.

---

Install

# From repo root (recommended)
python -m pip install . --no-build-isolation

# Editable/development
python -m pip install -e . --no-build-isolation

# Convenience scripts
./setup.sh        # macOS / Linux
setup.bat         # Windows cmd
./setup.ps1       # Windows PowerShell

Requires: Python ≥ 3.11, numpy, C++17 compiler. Optional: OpenMP (parallel replicas), MPI (multi-node), matplotlib (plots).

---

Quickstart

Solve a QUBO in one call

import numpy as np
from qanneal import solve

# Encode a 3-variable QUBO: minimise x₀ + x₂ − 2x₀x₁ − 2x₁x₂
Q = np.array([
    [ 1.0, -2.0,  0.0],
    [-2.0,  0.0, -2.0],
    [ 0.0, -2.0,  1.0],
], dtype=float)

result = solve(Q, method="sqapt", reads=16, seed=0, return_bits=True)
print(result.best_sample)   # array of bits {0, 1}
print(result.best_energy)   # minimum QUBO energy found

Number partition (classic NP-hard)

import numpy as np
from qanneal import DenseIsing, solve

# Partition numbers [3, 5, 7, 11, 13] into two groups with equal sum
nums = np.array([3, 5, 7, 11, 13], dtype=float)
n = len(nums)

# Ising encoding: spin +1 → group A, spin -1 → group B
h = np.zeros(n)
J = np.zeros((n, n))
for i in range(n):
    for j in range(i + 1, n):
        J[i, j] = J[j, i] = 2.0 * nums[i] * nums[j]

ising = DenseIsing(h, J, c=float(np.dot(nums, nums)))
result = solve(ising, method="sqa", reads=20, seed=42)
spins = result.best_sample        # +1 or -1
diff  = abs(float(np.dot(nums, spins)))
print(f"Partition diff: {diff}")  # 0.0 = perfect split

---

The Four Methods — When to Use Which

sa — Simulated Annealing

Classical Metropolis algorithm with a temperature schedule.

• Use when: fast results, simple landscapes, baseline comparisons.
• Control: sweeps_per_beta (thermal equilibration per temperature step).

sqa — Simulated Quantum Annealing

Quantum Monte Carlo in the Suzuki–Trotter formulation. The spin system is replicated across trotter_slices imaginary-time slices; flips along the time dimension mimic quantum tunneling through energy barriers.

• Use when: landscapes have tall narrow barriers that classical SA misses.
• Key extra controls: trotter_slices, worldline_sweeps, cluster_sweeps.

sqapt — SQA + Parallel Tempering (recommended default)

Multiple SQA replicas run simultaneously at different (β, Γ) points on a ladder. Adjacent replicas periodically swap configurations, letting solutions discovered at high fluctuations (large Γ, low β) flow toward low-energy states at strong freezing.

• Use when: rugged/multi-modal landscapes, moderate to hard combinatorial problems.
• Key extra controls: replicas (ladder length), pt_steps, swap_interval.

ctpimc — Continuous-Time PIMC

Samples worldlines in continuous imaginary time using Swendsen–Wang cluster updates. No Trotter discretisation error; better mixing in the high-Γ regime.

• Use when: you want a closer analog to D-Wave sampling; density-matrix–level statistics.
• Key extra controls: ctpimc_qubits_per_update, ctpimc_qubits_per_chain.

---

Core Concepts

Energy conventions

QUBO (binary variables x ∈ {0, 1}):

E(x) = Σᵢ Σⱼ Qᵢⱼ xᵢ xⱼ

• Q[i,i]: linear (bias) term for variable i.
• Q[i,j] (i ≠ j): quadratic coupling. Include both Q[i,j] and Q[j,i] for symmetric problems.

Ising (spins s ∈ {−1, +1}):

E(s) = Σᵢ hᵢ sᵢ  +  Σᵢ<ⱼ Jᵢⱼ sᵢ sⱼ  +  c

• h: local magnetic fields.
• J: pairwise coupling matrix.
• c: constant energy offset (irrelevant for optimization, useful for energy tracking).

Conversion: x = (s + 1) / 2 — QUBO uses bits, Ising uses spins. solve(..., return_bits=True) converts the result for you.

Hamiltonian models

Class	Memory	Best for
DenseIsing(h, J, c)	O(n²)	n ≤ ~3 000, fully connected
SparseIsing(h, edges, n, c)	O(n + |E|)	sparse graphs, n up to 100 000+
QUBO(Q)	O(n²)	binary variables; call .to_ising()

---

Optimal Adaptive Schedule (calibrated since 0.5.0; SQA two-phase since 0.6.1)

The optimal schedule paces the transverse-field ramp using the bond susceptibility χ_B = Var(B), a proxy for how close the Suzuki–Trotter path integral is to its quantum critical point. It implements the local adiabaticity condition (Roland & Cerf, 2002) for the path integral: move J⊥ slowly where the system is critical, quickly where it is not.

χ_B is large near the quantum transition and collapses toward zero in the ordered phase (it does not diverge). The schedule therefore dwells in the low-J⊥ critical region and crosses the ordered region quickly — a fast-slow-fast trajectory.

How it works (budget-calibrated, the default)

Pass optimal_eps_tilde=0.0 (the default) to request per-instance budget calibration. A short pilot pre-pass measures χ_B(J) at a grid of J⊥ points, then the run follows the inverse-CDF of ∫ χ_B(J)^α dJ, which guarantees — by construction, independent of the χ_B shape — that the schedule (a) consumes exactly optimal_num_steps, (b) reaches j_perp_end, and (c) dwells where χ_B is large. This replaces the old fixed-ε̃ update, which could "freeze" (barely move J⊥) when χ_B was large.

• α (optimal_alpha) — universality-class exponent, default 15/14 (1-D quantum Ising).
• j_perp_end — auto-set to max(j_perp_start, 5·j_rms) from the coupling scale.
• All schedule parameters adapt to your problem's coupling scale; nothing is hard-coded.

SQA runs a two-phase protocol (optimal_beta_ramp_fraction, default 0.3): phase 1 ramps β from beta_ramp_start up to a cold β = max(4/j_rms, 1) at fixed J⊥ (thermal anneal), then phase 2 fixes β and runs the calibrated adaptive J⊥ ramp. Without phase 1, single- temperature SQA has no thermal annealing and underperforms.

When does it help? The advantage appears for genuinely hard, frustrated Ising problems whose coupling scale puts them near a quantum transition (j_rms ≳ j_perp_start ≈ 2.9). Use scripts/sanity_schedule.py to check a new instance is in the favorable regime first.

Quickstart

from qanneal import solve

result = solve(
    problem,
    method="sqa",             # or "sqapt"
    schedule_type="optimal",
    reads=15,
    replicas=4,
    optimal_eps_tilde=0.0,    # 0.0 = per-instance budget calibration (recommended)
    optimal_num_steps=150,
    # optional overrides — all auto-computed from the problem scale if omitted:
    # optimal_j_perp_end=25.0,          # default max(j_perp_start, 5*j_rms)
    # optimal_beta_ramp_fraction=0.3,   # SQA two-phase split
    # optimal_alpha=15/14,
    # optimal_debug_csv="schedule.csv", # per-step J_perp / chi_B diagnostics
)
print(result.best_energy, result.best_sample)

SQAPT + optimal schedule

result = solve(
    problem,
    method="sqapt",
    schedule_type="optimal",
    replicas=8,
    reads=15,
    trotter_slices=32,
    worldline_sweeps=3,
    cluster_sweeps=0,         # >0 adds Swendsen–Wang cluster updates (SQAPT-SW)
    swap_interval=1,
    optimal_eps_tilde=0.0,    # calibrated
    optimal_num_steps=150,
)

J⊥ schedule helpers

from qanneal import j_perp_from_beta_gamma, optimal_j_perp_params, j_rms_from_problem

jp = j_perp_from_beta_gamma(beta=3.0, gamma=0.5, trotter_slices=32)  # 0.5*ln(1/tanh(βγ/M))
beta, jp_start, jp_end = optimal_j_perp_params(problem, trotter_slices=32)  # β=max(4/j_rms,1)
jr = j_rms_from_problem(problem)   # RMS of the present couplings (ndarray/DenseIsing/BQM/graph)

Direct C++ / low-level API

from qanneal import SQAAnnealer, SQASchedule

dummy = SQASchedule.from_vectors([1.0], [1.0])  # only used for construction
ann = SQAAnnealer(ising, dummy, trotter_slices=32, replicas=4)
result = ann.run_optimal(
    beta=1.0,                # phase-2 (cold) inverse temperature
    j_perp_start=0.1, j_perp_end=25.0,
    eps_tilde=0.0,           # <=0 -> budget calibration
    alpha=15/14, num_steps=150, sweeps_per_step=20,
    worldline_sweeps=3, cluster_sweeps=0,
    calib_probes=12, calib_sweeps=10,   # pilot pre-pass grid
    beta_ramp_fraction=0.3,             # two-phase thermal ramp
)
# result.j_perp_trace, result.chi_B_trace — per-step schedule diagnostics
# result.calibrated_eps_tilde, result.final_j_perp — calibration read-outs
# result.best_state, result.best_energy

See docs/optimal_schedule.md for the full derivation, calibration math, and benchmark methodology, and benchmarks/schedule/ for the reproducible study.

---

Schedule Design

Every annealer needs a schedule — a sequence of (β, Γ) or just β values.

Problem-adaptive schedules (recommended)

from qanneal import auto_schedule_sa_tuned, auto_schedule_sqa_tuned, auto_ladder_sqa_tuned

# For SA
schedule = auto_schedule_sa_tuned(ising, mode="balanced")

# For SQA / CT-PIMC
schedule = auto_schedule_sqa_tuned(ising, mode="balanced")

# For SQAPT — returns a (β, Γ) ladder with `replicas` points
ladder = auto_ladder_sqa_tuned(ising, replicas=8, mode="balanced")

How tuning works: The helper probes ~256 random single-spin flips, computes the 75th-percentile |ΔE|, and derives β_start/β_end from physical acceptance-target heuristics. This avoids manual tuning across problem sizes and coupling scales.

mode	Steps	Exploration	Use when
"fast"	30–40	High	Prototyping, large sweeps
"balanced"	60–80	Medium	Default production
"accurate"	110–130	Low	Best solution quality

Fixed schedules (legacy)

from qanneal import AnnealSchedule, SQASchedule
import numpy as np

# SA: linear β ramp
schedule = AnnealSchedule.linear(beta_start=0.1, beta_end=5.0, steps=60)

# SQA: paired (β, Γ) — geometric decay for Γ
schedule = SQASchedule.from_vectors(
    betas=np.linspace(0.1, 5.0, 60).tolist(),
    gammas=np.geomspace(5.0, 0.01, 60).tolist(),
)

---

Parameter Reference

Common to all methods

Parameter	Default	Meaning
reads	1	Independent runs; best over all reads is returned
sweeps_per_beta	20	Metropolis sweeps per temperature step
schedule	auto	Schedule object; auto-selected per method if None
seed	None	RNG seed for reproducibility
progress	True	Show tqdm progress bar
backend	"cpu"	Compute backend
return_bits	False	Return {0,1} bits instead of {−1,+1} spins

SQA and SQAPT (method="sqa" or "sqapt")

Parameter	Default	Meaning
trotter_slices	32	Number of imaginary-time slices. More slices → less Trotter error, more memory/time
replicas	1	Parallel SQA chains (SQA) or PT ladder rungs (SQAPT)
worldline_sweeps	5	Flips one spin through all time-slices at once (improves mixing)
cluster_sweeps	0	Swendsen–Wang cluster updates along imaginary time
continuous_time_slices	0	If > 0, approximates continuous-time limit within SQA

SQAPT only (method="sqapt")

Parameter	Default	Meaning
pt_steps	50	Number of local-update + swap epochs
swap_interval	1	Attempt replica swap every N steps
pt_betas	None	Explicit β ladder (overrides schedule)
pt_gammas	None	Explicit Γ ladder (must pair with pt_betas)

CT-PIMC only (method="ctpimc")

Parameter	Default	Meaning
ctpimc_qubits_per_update	1	Spins updated per cluster proposal
ctpimc_qubits_per_chain	1	Chain length for multi-qubit proposals

---

Accessing Results

result = solve(problem, method="sqapt", reads=32, ...)

result.best_sample    # np.ndarray, shape (n,), dtype int — best spin/bit configuration
result.best_energy    # float — lowest energy found across all reads
result.samples        # list of n arrays — one per read
result.energies       # list of floats — one per read
result.trace          # list of floats — energy trace from the first read
result.var_order      # variable ordering (relevant for BQM/graph inputs)

# Convert spins to bits manually
bits = ((result.best_sample + 1) // 2).astype(int)

Observers (low-level diagnostics)

from qanneal import Annealer, MetricsObserver, AnnealSchedule
import numpy as np

ising = ...
schedule = AnnealSchedule.linear(0.1, 5.0, 60)
obs = MetricsObserver()
ann = Annealer(ising, schedule)
res = ann.run(sweeps_per_beta=40, observer=obs)

import matplotlib.pyplot as plt
plt.plot(obs.energy_trace)
plt.xlabel("Temperature step"); plt.ylabel("Energy"); plt.show()

SQA observers (SQAMetricsObserver, SQAStateTraceObserver) capture per-sweep replica/slice state snapshots — see docs/sqa_trace_parameters.md for the full field list.

---

Parallelism

OpenMP (automatic, within a run)

Enabled by default (QANNEAL_ENABLE_OPENMP=ON). Parallelizes:

• ReplicaAnnealer — across replicas
• SQAAnnealer — across replicas × slices (when no sweep-level observer is attached)
• SQAParallelTemperingAnnealer — across ladder replicas

Control at runtime:

export OMP_NUM_THREADS=8

Multi-process / multi-node

For large problems or many reads, use the HPC launcher:

# Single node, multiprocessing (no extra deps)
python examples/python/hpc_sqa_launcher.py \
  --n 5000 --method sqapt --reads 64 --workers 8

# Multi-node MPI (requires mpi4py)
mpirun -n 32 python examples/python/hpc_sqa_launcher.py \
  --n 50000 --method sqa --reads 8 --mpi

# SLURM job array (most portable, no mpi4py needed)
sbatch scripts/slurm/run_sqa_array.sh
python examples/python/merge_array_results.py results/run_*.json

C++ MPI (build-time)

cmake -S . -B build -DQANNEAL_ENABLE_MPI=ON
cmake --build build
mpirun -n 8 build/qanneal_mpi_example

---

Notebooks

Interactive Jupyter notebooks in notebooks/:

Notebook	What you'll learn
01_quickstart.ipynb	First problem, SA vs SQA comparison
02_sqa_physics.ipynb	Trotter slices, Γ schedules, worldline sweeps
03_sqapt_and_ctpimc.ipynb	Replica exchange, (β,Γ) ladders, CT-PIMC
04_large_problems.ipynb	SparseIsing, Chimera/MaxCut, HPC scaling

---

Examples

# Tuned SQAPT demo
python examples/python/tuned_sqapt_demo.py --mode balanced

# Number partition benchmark (SA / SQA / SQAPT / CT-PIMC vs D-Wave SDK)
python examples/python/number_partition_benchmark.py \
  --with-sqapt --with-ctpimc --schedule-mode balanced --jobs 8

# Full comparative benchmark vs D-Wave
python examples/python/dwave_comparative_benchmark.py --n 40 --reads 16

# HPC scaling benchmark
python examples/python/hpc_sqa_launcher.py --scaling --method sqapt --mode fast

# Interactive graph problem editor
python examples/python/graph_editor_gui.py

---

Documentation

File	Content
docs/api.md	Complete API reference (all classes, parameters, return types)
docs/optimal_schedule.md	Optimal adaptive J⊥ schedule — physics, algorithm, and parameter guide
docs/overview.md	Architecture, data flow, component map
docs/user_guide.md	Step-by-step usage guide with physical intuition
docs/ctpimc.md	CT-PIMC algorithm details
docs/sqa_trace_parameters.md	SQA observer field guide

---

Build Options

cmake -S . -B build \
  -DQANNEAL_ENABLE_OPENMP=ON   \  # parallel loops (default ON)
  -DQANNEAL_ENABLE_MPI=OFF     \  # distributed anneal (default OFF)
  -DQANNEAL_BUILD_TESTS=ON     \  # unit tests
  -DCMAKE_BUILD_TYPE=Release
cmake --build build -j
ctest --test-dir build

---

License

Apache-2.0. See LICENSE and NOTICE. The CT-PIMC engine uses D-Wave's localPIMC (Apache-2.0), credited in NOTICE.