burn_quantum

Differentiable qubit and qudit simulation using Burn 0.22's native tensors, devices, modules, parameters, optimizers and weight records. Rust 1.95+.

QuantumBackend describes quantum execution; DenseSimulator implements it using supported Burn devices. The public API accepts native Tensor<D> and Device values. It does not require a separate tensor backend or a Burn fork.

Quick start

[dependencies]
burn_quantum = "0.3.0"
burn = { version = "~0.22.1", default-features = false, features = ["std", "flex"] }

use burn::tensor::Device;
use burn_quantum::{Circuit, DenseSimulator, QuantumBackend, Register, Unitary};

fn main() -> burn_quantum::Result<()> {
    let mut circuit = Circuit::new(Register::qubits(2)?, 0, 0);
    circuit.unitary(vec![0], Unitary::h())?;
    circuit.unitary(vec![0, 1], Unitary::cnot())?;
    let state = DenseSimulator::default()
        .execute(&circuit, None, None, 1, &Device::flex())?;
    state.validate(1e-5)?;
    // [0.5, 0.0, 0.0, 0.5] in |00>, |01>, |10>, |11> order.
    println!("{}", state.probabilities());
    Ok(())
}

Capabilities

Batched qubit, qudit and mixed-dimensional pure states and density matrices.
Validated local unitaries, differentiable rotations, observables, partial traces, Kraus channels and seeded terminal measurement.
Serializable circuits, native autodiff, qudit-aware parameter-shift references, and QuantumLayer integration with Burn modules and records.
Reusable encodings, circuit templates, probability losses and fidelity kernels.
Quantum natural SPSA with a regularized Fubini–Study metric and checkpoints.
Boolean CNF/DIMACS SAT search with reversible oracles and verified witnesses.
photonic adds dual-rail/vacuum encoding, loss, Fock optics, calibrated fusions, physical circuit compilation, finite fabric scheduling, Clifford/frame simulation, encoded resources, QEC memories, finite boundaries, logical/magic-state references and bounded event/process execution. A separate tensor_network module supplies mixed-qudit MPS and distributed entangled-state ownership, with explicit approximation and host-SVD limits. See the implemented models and remaining gates. These capabilities are included in version 0.3.0.

Default features are flex and autodiff. Optional wgpu, cuda and fusion features select accelerator capabilities; callers supply the execution device. F32/F64 availability depends on the device. GPU qualification uses F32.

Dense simulation uses exponential state storage: for batch B and Hilbert dimension N, pure states store 2BN real scalars and density matrices store 2BN². Allocation limits are preflight checks, not a total memory cap. SAT search may return Inconclusive when its budget is exhausted. QN-SPSA's dense host metric uses O(d²) storage and an O(d³) solve for d parameters. These implementations do not establish quantum advantage or efficient simulation of arbitrary circuits.

Package and repository

This package owns src/, integration tests/, simulator examples/, and the executable-only support/device.rs helper. From a repository checkout:

cargo run --locked -p burn_quantum --example bell
cargo test --locked -p burn_quantum
cargo check --locked -p burn_quantum --no-default-features
cargo run --locked -p burn_quantum --features wgpu --example wgpu_smoke

The sibling burn_quantum_example package contains single-file quantum applications, including VQE, QAOA, classifiers, a Born machine, hybrid regression, Grover SAT, QN-SPSA VQE, physical photonic optimization, finite-boundary QEC studies, magic-state distillation and distributed MPS workloads. It is run from a repository checkout:

cargo run --locked -p burn_quantum_example --bin qnspsa_vqe
cargo run --locked -p burn_quantum_example --bin fbqc_bell_swap

See the API documentation, example suite, guides and validation records, and changelog.

License

MIT OR Apache-2.0. Both license texts are included in this package.

Canonical algorithms and dynamic protocols

algorithms supplies reusable QFT/inverse QFT, phase estimation, Deutsch–Jozsa parity oracles, single-target Grover, generalized Toffoli, graph MaxCut QAOA, teleportation and coherent odd-repetition recovery. Builders retain the ordinary Circuit, symbolic bindings and operation labels. append_qft composes with an existing circuit; all wires follow the register's big-endian convention.