Quantum Ml

Quantum Boltzmann Machines

A quantum Boltzmann machine is an energy-based generative model whose distribution comes from the thermal state of a quantum Hamiltonian.

From classical to quantum energy models

A classical Boltzmann machine defines a probability distribution over configurations from an energy function: low-energy configurations are more likely, weighted by exp(-E/T). A quantum Boltzmann machine (QBM) replaces the classical energy with a quantum Hamiltonian H, and the distribution comes from the thermal (Gibbs) state rho = exp(-H) / Z, where Z is the partition function. Crucially, H can include non-commuting terms that have no classical analogue.

What the quantum terms add

Kronos motion — three machines

The classical Ising energy uses only commuting Z-type interactions. A QBM Hamiltonian can add transverse field terms, such as X operators, that do not commute with the Z terms. These off-diagonal contributions let the model represent quantum correlations and distributions that a classical energy model cannot capture compactly, which is the source of any potential representational advantage.

The training obstacle

Sampling from a thermal state, quantum Gibbs sampling, is itself hard and is the central bottleneck.
The gradient of the log-likelihood involves expectation values under the thermal state that are difficult to estimate when terms do not commute.
Bound-based training uses a tractable surrogate loss to sidestep the intractable exact gradient for fully quantum Hamiltonians.

python

# QBM likelihood gradient, schematic structure
# grad_theta log p(v) = <H_theta'>_data-clamped - <H_theta'>_model
# Both terms need expectations under a thermal state exp(-H)/Z,
# which requires quantum Gibbs sampling -- the hard part.
for theta_i in parameters:
    g[i] = clamped_expectation(theta_i) - model_expectation(theta_i)

Preparing thermal states

Because the model needs Gibbs states, much QBM research focuses on preparing them efficiently: variational thermal-state preparation, imaginary-time evolution approximated on hardware, and quantum-assisted sampling. Each is costly and imperfect on NISQ devices, so demonstrations remain small. Thermal-state preparation is to QBMs what state loading is to other generative models: the step that decides feasibility.

Standing

QBMs are theoretically attractive because they extend a well-understood classical model into genuinely quantum territory and connect naturally to physics, where thermal states are the objects of interest. Their practical use is limited today by the cost of quantum Gibbs sampling. As with other quantum generative models, the strongest case is for quantum-native problems, modeling the thermal behavior of a quantum system, rather than for classical datasets.