Quantum Boltzmann Machine (QBM)

Overview

A Boltzmann machine is a neural network that learns a probability distribution by modeling it as the thermal distribution of a system of interacting binary units. A quantum Boltzmann machine, or QBM, replaces the classical energy function with a quantum Hamiltonian that includes terms that flip units, usually a transverse field. The model's distribution is then the measurement statistics of a quantum thermal state, which can include quantum correlations that a classical model cannot represent.

The idea connects machine learning to the quantum annealers sold by D-Wave, which naturally sample from approximately thermal states of Ising-type quantum Hamiltonians. The questions are whether such models learn richer distributions than classical ones and whether they can be trained efficiently. Neither has been settled.

The problem it solves

Classical Boltzmann machines require expensive Gibbs sampling for training. Quantum Boltzmann machines may sample from quantum Gibbs distributions more efficiently.

History and background

Classical Boltzmann machines were introduced by Hinton and Sejnowski in 1985, and restricted Boltzmann machines, which limit connections for efficient training, were popularized by Smolensky and Hinton in the 1980s and 2000s as building blocks of deep belief networks. Quantum versions were proposed by Wiebe, Kapoor, and Svore, who gave a quantum deep learning algorithm with a quadratic speedup in training in 2016, and by Amin and coauthors, whose Quantum Boltzmann Machine paper appeared on arXiv in 2016 and in Physical Review X in 2018.

Kieferová and Wiebe showed in 2017 how to train QBMs to learn quantum data and to perform tomography. Experimental work used D-Wave annealers as samplers: Adachi and Henderson at Google trained deep belief networks with quantum annealing in 2015, and Benedetti, Realpe-Gómez, Biswas, and Perdomo-Ortiz studied effective temperatures and hardware-embedded models in 2016 and 2017. The conclusions were modest: annealers sample from distributions with an effective temperature that must be estimated, and demonstrated gains over classical training are small.

How it works

Extends classical Boltzmann machines to quantum Hamiltonians. The quantum fluctuations (transverse field terms) help the network escape local minima and explore a richer energy landscape during training.

In a classical Boltzmann machine, each unit is a spin that is either up or down, and the energy of a configuration depends on biases for single spins and weights for pairs. The probability of a configuration is proportional to e to the minus energy. Some units are visible, matching the data, and some are hidden, giving the model flexibility. Training adjusts biases and weights so that the distribution over the visible units matches the data.

A quantum Boltzmann machine adds a transverse field that rotates each spin about a perpendicular axis. The Hamiltonian is then a sum of terms that do not commute with each other, so the thermal state is not diagonal in the basis of configurations. Measuring the visible units gives samples that include effects of quantum fluctuations, which might allow the model to represent distributions that classical Boltzmann machines need many more hidden units to match.

Training faces a technical obstacle. For a classical machine, the gradient of the log-likelihood is the difference between averages in the data-clamped and free-running ensembles, which are estimated by sampling. For a quantum machine the non-commuting terms make that formula fail for the transverse-field parameters. Amin and coauthors defined a trainable bound on the log-likelihood using the Golden-Thompson inequality, which allows sample-based gradients for the other parameters.

Step by step

1. Choose the architecture: visible and hidden units, with a quantum Hamiltonian containing biases, couplings, and a transverse field.
2. Prepare the thermal state of the Hamiltonian by quantum annealing, quantum Gibbs sampling, or classical simulation for small systems.
3. For each training example, prepare the thermal state of the Hamiltonian with the visible units clamped to the data, and record the relevant statistics.
4. Prepare the unclamped thermal state and record the same statistics.
5. Compute the gradient of the bound on the log-likelihood from the difference between the two sets of statistics, and update the biases and weights.
6. Repeat until the model's visible statistics match the data, then use the trained model to generate samples or to classify.

A worked example

Consider learning a distribution over four binary variables that has strong correlations, such as parity constraints. A classical Boltzmann machine with a few hidden units can fit it. A quantum version with a transverse field might fit it with fewer hidden units because the quantum terms allow tunneling between configurations, which reshapes the distribution.

In practice, the evidence for such advantages is limited to small simulations. On annealer hardware, the effective temperature drifts with the problem, so researchers estimate it from the samples and rescale the Hamiltonian to compensate, which is itself a source of error.

Complexity and resources

What hardware it needs

• For near-term experiments, a quantum annealer or an analog simulator that approximates thermal states.
• For gate-based versions, fault-tolerant machines with Gibbs-state preparation routines.
• A classical computer for parameter updates and temperature estimation.

Applications

• Generative modeling and sampling from complex distributions.
• Learning and reproducing quantum states and thermal ensembles.
• Pre-training of deep networks, in exploratory studies using annealers.
• Studying how quantum fluctuations affect learning.

Limitations and caveats

• Training is difficult because of the non-commuting terms and the need for Gibbs samples.
• Annealer hardware samples at an uncontrolled effective temperature.
• No clear advantage over modern classical generative models has been shown.
• Classical deep generative models such as diffusion models have largely replaced Boltzmann machines in practice.

Common misconceptions

• A quantum Boltzmann machine is not guaranteed to represent richer distributions than a classical one in a practical sense.
• Using an annealer as a sampler does not make the sampling exact or fast for all models.
• Boltzmann machines are not the dominant approach in modern machine learning.

How it compares

• Classical restricted Boltzmann machines train efficiently with contrastive divergence and are well understood.
• Quantum GANs and quantum circuit Born machines train circuits directly without needing thermal states.
• Quantum Gibbs sampling provides the subroutine that makes gate-based QBMs possible.

Real-world impact

Primarily demonstrated on D-Wave quantum annealers for generative modeling tasks. Ongoing research into whether quantum effects provide genuine advantage over classical Boltzmann machines.

Why it matters

QBMs were among the first proposals for quantum generative models and a natural use of annealing hardware in machine learning. They highlight the central difficulty of quantum learning, which is preparing and sampling from the right quantum states.

The field's experience, especially with effective temperature estimation, is a useful guide to the real-world problems that arise when theory meets noisy analog devices.

Frequently asked questions

Keep learning

• Related algorithm: Quantum Gibbs Sampling
• Related algorithm: Quantum Annealing
• Related algorithm: Quantum GAN (QGAN)
• Related algorithm: Quantum Metropolis Algorithm

This page is written for educational purposes. Complexity statements describe the best known results at the time of review (September 30, 2026), and research in this area moves quickly, so check recent papers before relying on a specific bound.

← Back to Algorithms Database