Let the measurements choose: adapting quantum entropy estimates

The same number of measurements can be spent in two very different ways: try many measurement settings once, or revisit a few settings repeatedly. Which choice best reveals how entangled part of a quantum system is? The answer depends on the state, which is awkward when the state is precisely what an experimenter is trying to characterize.

The second-order Rényi entropy is a measure of the purity of a subsystem. Purity is one for a pure state and smaller for a more mixed state; the entropy is its negative base-two logarithm. For many-body experiments, this quantity can reveal correlations without reconstructing the entire quantum state. But each measurement consumes a limited experimental shot, so allocating shots matters.

Mu Niu, Yi Hong Teoh, Roger G. Melko and Rajibul Islam build on a single-copy randomized-measurement protocol. The experimenter applies randomly chosen local rotations, measures qubits in a standard basis and repeats the process. Different rotations expose different views of the subsystem. Pairwise comparisons of observed bit strings, weighted by how many bits differ, provide an estimate of purity. Using pairs of distinct shots avoids the bias that would arise from simply squaring observed frequencies.

The protocol has two adjustable counts: the number of random unitary settings, NU, and the number of measurements per setting, NM. More settings help capture how outcomes change with the measurement basis; more repetitions at a setting reduce shot noise. The paper’s illustrative extremes explain why neither allocation is universally best. A maximally mixed state has probabilities that are basis-independent, so repetitions can be especially useful. A product state can look very different in different bases, so exploring more settings can matter more.

Rather than choose NU/NM in advance, the proposed method starts with at least two settings and four shots per setting. Those minimums let it estimate both sources of uncertainty from the data. It repeatedly estimates how much the total error might fall by adding settings versus adding shots to existing settings, then takes the direction with the larger estimated benefit per measurement cost. A warm-up guard forces some early exploration: otherwise a few accidentally similar settings could persuade the algorithm to keep repeating them. The procedure stops when its measurement budget is used up.

The authors test this strategy in simulations, sampling synthetic outcomes from exactly computed probabilities. In 500 runs on the half of several eight-qubit states, a budget of 20,000 measurements led the allocation to shift as expected: the mode for a random mixed state was 20 settings with 1,000 shots each, while for a product state it was 2,000 settings with 10 shots each. GHZ and random pure states fell between those extremes. A separate comparison on six-qubit examples found that the adaptive estimate’s standard deviation was comparable to the best point in a fixed-parameter grid for product and GHZ states; for random pure and random mixed states it was within about 1.2 and 2.5 times the grid’s best, respectively.

The paper also studies how many total measurements are needed to reach a fixed purity error as system size grows. For the tested all-up, equal-mixture and maximally mixed states, the adaptive method had the slowest measured growth among the fixed-ratio alternatives examined. These are finite-size numerical scaling results, not proof that the same performance holds for arbitrary states or hardware.

There is also a statistical caveat: the standard estimator is exactly unbiased when NU and NM are fixed in advance. Because the adaptive choices depend on observed data, the authors do not claim exact unbiasedness of the final estimate, though they expect the effect to be negligible when allocations concentrate across runs. The logarithm used to convert purity into entropy is not exactly unbiased either. And while local randomized measurements avoid preparing multiple coherent copies, the work tests synthetic data rather than noise, drift or other constraints of a live device. Its contribution is a way to let early data guide the resource split—not a universal optimal setting or an experimental demonstration.