Quantum extreme learning skips the likelihood: on a transverse-field Ising chain, 15 training points bring maximum absolute error to 0.013
Synopsis
Quantum sensing usually infers an unknown parameter with Bayesian inference or maximum likelihood estimation, but both estimators require constructing a likelihood function, which needs full knowledge of the probe dynamics and is often impractical. A quantum extreme learning framework instead places the unknown parameter in the reservoir Hamiltonian, lets a fixed initial state encode it through parameter-dependent evolution, and trains a linear readout on local observable expectation values, avoiding explicit likelihood construction. Estimating the transverse field of a transverse-field Ising chain, 15 training points give a maximum absolute error of 0.013 over 200 test values, and combining measurements at two times into a larger feature vector improves accuracy further.
Interpretation
Estimating an unknown physical parameter no longer requires writing down a likelihood function first: the parameter sits in the Hamiltonian of the quantum reservoir itself, is carried by the evolution of a fixed initial state, and is recovered from local observables by a trained linear readout. Bayesian inference and maximum likelihood estimation are optimal estimators in the large-sample limit, but they depend on a likelihood function; constructing that function efficiently on a classical computer requires full knowledge of the probe and its evolution, while Hamiltonian parameters or decoherence effects are often not precisely known and classical simulation of large systems is limited by the curse of dimensionality. For transverse-field estimation on a transverse-field Ising chain at a fixed system size and evolution time, 200 test values across the interval were checked: with 5, 10 and 15 training points the maximum absolute error was 0.562, 0.335 and 0.013 respectively, and at 15 training points the estimates track the true values across the whole interval.
The approach puts the parameter only in the evolution and trains only a linear map at the readout. Conventional quantum extreme learning encodes input information in the initial state and uses the quantum reservoir to process externally supplied input; here the initial state is fixed and deliberately not an eigenstate of the Hamiltonian, so the parameter enters through nontrivial dynamics across its whole range. After evolving for a set time, expectation values of a set of local Pauli operators are measured and assembled into a feature vector, and ridge regression fits linear weights from features to parameter values; the weights are then fixed and the root-mean-square error is computed on test values not used in training.
The measurement time has to be optimized separately, because the best time for accuracy and the best time for sampling fluctuations do not coincide. The root-mean-square error falls with evolution time, reaches a minimum and then rises before saturating, while the single-shot statistical uncertainty varies in time but is nearly uniform in the parameter, so their minima fall at different times. The procedure first constrains the average single-shot measurement uncertainty below a prescribed threshold and then minimizes the root-mean-square error within that constraint; in the transverse-field Ising model the threshold is reachable in several time windows while keeping the number of measurement samples per qubit at the level typical of many quantum experiments.
Perspective
The demonstrated scope is transverse-field estimation on an exactly solvable transverse-field Ising chain, with measurements restricted to expectation values of local Pauli operators and conclusions drawn from numerical simulation rather than experiment; robustness to dephasing is also examined under a local dephasing model.
Local Pauli measurements and expectation-value readout access only part of the parameter information in the evolved state, so precision does not grow quadratically with evolution time; larger system size generally improves accuracy but not monotonically, and the advantage weakens and tends to saturate as the training set grows; dephasing is negligible on short time scales and only noticeably degrades accuracy when it becomes strong.
