Hamiltonian Learning Algorithms

Short answer: Hamiltonian learning is the task of working out a quantum system's Hamiltonian, the rule governing its time evolution, from experimental measurements. The key question is how much total evolution time is needed to reach a given precision. Several recent theory papers show that, with the right kind of control, the time can scale as 1/epsilon (the Heisenberg limit), a quadratic improvement over the 1/epsilon-squared scaling set by the standard quantum limit. These are theoretical algorithm results, not experimental demonstrations.

Why learn a Hamiltonian

A 2025 arXiv paper (2502.20373) explains that characterizing a Hamiltonian helps derive accurate theoretical models, calibrate quantum devices and design error-mitigation algorithms. Hamiltonian learning reconstructs a system's Hamiltonian from measurements. The paper notes that by the central limit theorem, reaching a root-mean-square error epsilon requires total evolution time and measurements scaling as 1/epsilon squared, called the standard quantum limit. The uncertainty principle allows the optimal scaling in evolution time of 1/epsilon, called the Heisenberg limit.

Control makes the difference

A paper in Quantum (November 2024) studies learning the Hamiltonian of a many-body system and shows that the rate of learning depends on how much control the experimenter has. It considers three control models: time evolution augmented with instantaneous quantum operations, a Hamiltonian augmented with constant terms, and no control over the time evolution. With continuous quantum control the authors give an adaptive algorithm reaching the Heisenberg limit, which needs only product-state preparation, time evolution and product-basis measurement. Without control, they prove learning is standard quantum limited for large classes of many-body Hamiltonians, including any that thermalizes under the eigenstate thermalization hypothesis. They conclude this establishes a quadratic advantage in experimental runtime for learning with quantum control.

Hybrid systems

The 2025 arXiv paper focuses on spin-boson type Hamiltonians, which describe hybrid systems with different particle species. The authors prove their algorithm achieves Heisenberg-limited estimation of all coupling parameters up to error epsilon with total evolution time of order 1/epsilon, using only polylogarithmically many measurements in 1/epsilon, and say it is robust to small state preparation and measurement errors. They also give an alternative algorithm based on distributed quantum sensing that reduces the evolution time per measurement, and illustrate hybrid Hamiltonian learning and spectrum learning.

Without short-time control

A paper accepted in PRX Quantum notes that recent algorithms rely on arbitrarily short-time dynamics, which raises experimental difficulties because of finite control bandwidth and transient pulse errors. The authors show Heisenberg-limited learning can be achieved when every query lasts at least a minimum time T. The key step emulates the continuous control iterative algorithms need using only such lower-bounded evolution times, reducing the task to sparse pure-state tomography. For logarithmically sparse Hamiltonians they report the optimal 1/epsilon scaling for any constant minimum time. For polynomially sparse many-body systems they report a tradeoff in which the minimum time can be relaxed at a polynomial cost in total evolution time. They state that ultra-short pulses are not fundamentally necessary for optimal learning.

What this does not tell you

These are proofs about algorithms under stated assumptions. They do not report running the algorithms on hardware, and constants, noise and system size will matter in practice.

FAQ

What is Hamiltonian learning?

Reconstructing a quantum system's Hamiltonian from measurements.

What is the Heisenberg limit here?

Total evolution time scaling as 1/epsilon, against 1/epsilon squared for the standard quantum limit.

Does control matter?

Yes. In the 2024 Quantum paper, learning is Heisenberg limited with continuous control and standard quantum limited without it for many Hamiltonians.

Are very short pulses required?

One accepted PRX Quantum paper shows Heisenberg-limited learning with a minimum query time.

Has this been run on hardware?

The papers described here present theory and do not report that.