Skip to content
Quantum.world

The Signal

Physical Review A publishes quantum Boltzmann energy method

A paper in Physical Review A presents a quantum Boltzmann-machine algorithm for estimating Hamiltonian ground-state energies. The authors prove convergence to an approximate stationary point using a sample count polynomial in the target precision, parameter count and Hamiltonian norm.

Physical Review A publishes quantum Boltzmann energy method

A published convergence result

A quantum Boltzmann-machine method for estimating Hamiltonian ground-state energies has been published in Physical Review A, with a proof that its training algorithm reaches an approximate stationary point using a polynomial number of thermal-state samples. The authors say that sample requirement grows polynomially with the inverse target precision, the number of model parameters and the norm of the Hamiltonian being optimized.

The paper, by Dhrumil Patel, Daniel Koch, Saahil Patel and Mark M. Wilde, appeared on November 17, 2026. Its earlier version was posted as an arXiv preprint on October 16, 2024; the journal publication is the latest development in the work.

Thermal states instead of circuit ansatzes

The proposed approach uses a hybrid quantum-classical optimization loop, but represents candidate states as parameterized thermal states rather than states generated by parameterized quantum circuits. The objective is the expected energy of the target Hamiltonian in each trial state, and the classical optimizer updates the model parameters using estimates of the objective’s gradient.

The authors frame the method as a variational approach: minimizing energy within the chosen family of trial states provides an upper bound on the ground-state energy. Their analysis assumes that samples of the parameterized thermal state can be prepared and that the target Hamiltonian can be efficiently measured on a quantum computer.

How the gradient is estimated

The central technical step is a quantum Boltzmann gradient estimator. The circuit construction combines classical random sampling with Hamiltonian simulation and the Hadamard test to estimate the gradient; the paper also describes estimating one part of the gradient from measurements on two copies of a thermal state.

The authors calculate the objective’s gradient and Hessian, then use bounds on the Hessian’s matrix elements to establish a Lipschitz bound for the gradient. That smoothness result is used in the convergence analysis for stochastic gradient descent.

A stationary point, not a global guarantee

The guarantee is for an approximate stationary point of the energy function over the selected parameter space, not a proof that the algorithm finds the global minimum or the exact ground state. The paper notes that the objective is generally non-convex, making a stationary point the target of its convergence result.

The authors describe quantum Boltzmann machines as an alternative to variational quantum eigensolvers, whose trial states are built from parameterized circuits. They also say the machine-learning approach is not known to suffer from the barren-plateau problem in this setting; that is a claim about the method, not a reported hardware demonstration.

Sources

  1. Physical Review A
  2. arXiv

Related stories

Explore

More articles