⚛️ quantum physics

Adaptive detection of Rabi signals under composite hypotheses

This paper demonstrates that a myopic Bayesian adaptive policy for Rabi sensing, which dynamically selects measurement axes to maximize information gain, outperforms fixed non-adaptive strategies in detecting weak coherent drives with unknown amplitude and phase by achieving superior $n^{-1/2}$ sensitivity and Type-II error exponents under a fixed shot budget and calibrated false positive control.

Original authors: So Chigusa

Published 2026-09-16

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: So Chigusa

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Quantum sensing is a field dedicated to using the delicate properties of atoms and subatomic particles to measure the world with extreme precision. Imagine a tiny, controllable system, like a single atom, that acts as a sensor. When this system is exposed to a faint force or a weak signal, its behavior changes in a way that reveals the presence of that signal. Scientists use these sensors to hunt for everything from invisible dark matter to subtle shifts in magnetic fields. However, a major challenge arises when the signal is so weak that it is hard to distinguish from random background noise, and when the researchers do not know exactly what the signal looks like, such as its strength or its direction. In these situations, the goal shifts from simply measuring a known value to making a confident decision: is the signal there, or is it just noise?

A researcher at the Massachusetts Institute of Technology tackled this problem by designing a smarter way to listen for these faint whispers. They focused on a specific type of signal known as a Rabi drive, which is a rhythmic push that can be applied to a two-level quantum system, similar to how a gentle, rhythmic push can make a swing go higher. The difficulty lies in the fact that the researcher did not know the rhythm's strength or its timing relative to their sensor. To solve this, they treated the search as a series of repeated experiments with a fixed number of attempts, or "shots." In each shot, they prepared a fresh quantum sensor, let it interact with the environment, and then measured it. The crucial innovation was how they chose what to measure. Instead of sticking to a single, pre-planned measurement direction for every shot, they developed an adaptive strategy. After each measurement, they used the result to update their understanding of the signal's likely properties and then chose the most informative direction for the very next measurement.

The researcher compared this adaptive approach against simpler, non-adaptive methods where the measurement direction is fixed in advance. They found that the fixed methods had significant weaknesses. One common method, which measures the population of energy states, was robust but responded very weakly to the signal, requiring many more attempts to detect it. Another method, measuring a transverse direction, was more sensitive but would fail completely if the signal happened to be oriented in a direction the researcher did not guess. The adaptive strategy, however, learned the signal's orientation as it went. By using the information gathered from previous shots to guide the next one, the system effectively "learned" where to look. In computer simulations involving thousands of shots, this adaptive policy successfully detected the signal with a sensitivity that improved much faster than the fixed methods as the number of shots increased.

The study demonstrated that this learning process allowed the adaptive system to outperform even the best fixed strategies that were designed to measure in two different directions. While the fixed methods struggled to maintain sensitivity when the signal's direction was unknown, the adaptive policy used the accumulated history of outcomes to narrow down the possibilities and focus its efforts. The results showed that by the time the researcher reached the largest simulated number of shots, the adaptive method was significantly more powerful at detecting the signal while keeping the rate of false alarms strictly controlled. The findings suggest that in the real world, where signals are often weak and their properties are unknown, the ability to learn and adapt in real time offers a distinct advantage over static measurement plans. This work provides a concrete blueprint for how future quantum sensors can be programmed to be more efficient and reliable when hunting for the faintest traces of new physics.

Technical Summary: Adaptive Detection of Rabi Signals under Composite Hypotheses

Problem Formulation
This work addresses the detection of weak coherent drives acting on a two-level quantum sensor (qubit) when the signal parameters—specifically amplitude, detuning, and phase—are unknown. The problem is framed as an asymmetric composite hypothesis test with a fixed shot budget $n$ .

Null Hypothesis ( $H_0$ ): No signal is present ( $\Phi_R = 0$ ).
Alternative Hypothesis ( $H_1$ ): A signal exists with parameters $\theta = (\Phi_R, \Delta, \tilde{\phi})$ drawn from a prior distribution over a search region $\Theta$ .
The goal is to maximize detection power (minimize Type-II error) while strictly controlling the Type-I error (false positive rate) at a fixed level (e.g., $\alpha = 0.05$ ). The challenge lies in the "nuisance parameters" (phase and detuning) which are present only under $H_1$ and determine the most informative measurement basis.

Methodology
The author proposes a Bayesian adaptive sensing protocol where the measurement basis for each shot is selected based on the accumulated classical history of previous outcomes.

Model: The sensor undergoes Rabi dynamics under a weak transverse drive. The readout probabilities are modeled including relaxation ( $T_1, T_2$ ), finite contrast, and classical bit-flip errors.
Decision Statistic: The test statistic is the log Bayes factor, $\log B_{10,n}$ , which updates sequentially as data is collected. A threshold $c_\alpha$ is calibrated via Monte Carlo pseudo-experiments under $H_0$ to enforce the target Type-I error.
Measurement Policies:
- Non-adaptive Baselines: Fixed population readout (fixed- $z$ ), fixed transverse readout (fixed- $x$ ), and alternating transverse readout (fixed- $x/y$ ).
- Oracle Benchmark: A transverse readout aligned with the true signal phase (unattainable in practice but serves as a lower bound for adaptive performance).
- Adaptive Policy: A "myopic" Bayesian policy that selects the readout axis $n_k$ to maximize the expected one-step information gain (mutual information) regarding the hypothesis and signal parameters. This policy utilizes the posterior distribution of the unknown phase to guide subsequent measurements.

Key Contributions and Analytic Benchmarks
The paper establishes analytic benchmarks for weak signal sensitivity scaling:

Population Readout (Fixed- $z$ ): The response is phase-independent but quadratic in signal amplitude ( $\delta p \propto \Phi_R^2$ ). This results in a sensitivity scaling of $\Phi_{req} \propto n^{-1/4}$ .
Transverse Readout: The response is linear in amplitude ( $\delta p \propto \Phi_R$ ) but phase-dependent. With a known or learned phase, sensitivity scales as $\Phi_{req} \propto n^{-1/2}$ .
Crossing Point: Analytic derivations show that while fixed- $z$ is superior for very small $n$ (due to the lack of phase learning overhead), transverse strategies eventually outperform it as $n$ increases. The adaptive policy aims to achieve the $n^{-1/2}$ scaling without prior knowledge of the phase.

Numerical Results
Monte Carlo simulations were performed with a fixed shot budget and calibrated Type-I error ( $\alpha \approx 0.05$ ):

Posterior Learning: The adaptive policy successfully learns the unknown signal phase, causing the posterior distribution to contract around the true parameters over time.
Detection Power: At a representative shot count ( $n=1024$ ) and signal amplitude, the adaptive policy achieved a phase-averaged power of 0.79, significantly outperforming the fixed- $z$ baseline (0.44) and fixed transverse baselines (0.60–0.67).
Sensitivity Scaling: The adaptive policy exhibits a sensitivity scaling consistent with $n^{-1/2}$ for large $n$ . While it outperforms fixed transverse schedules (like fixed- $x/y$ ) at the largest simulated shot counts, it remains less sensitive than the transverse oracle, reflecting the finite budget cost of learning the nuisance direction.
Error Exponents: The phase-averaged Type-II error exponents for the adaptive policy exceed those of non-adaptive references, indicating faster suppression of false negatives as the shot count increases.

Significance and Claims
The paper claims to demonstrate the "finite budget value of exploiting nuisance parameter information under calibrated false positive control."

It shows that adaptive control, specifically using posterior information to guide readout axes, can overcome the phase-blindness of fixed transverse measurements and the quadratic response limitation of population measurements.
The results validate that a myopic information-gain strategy is sufficient to achieve near-optimal asymptotic sensitivity ( $n^{-1/2}$ ) in a composite hypothesis testing scenario, even with a fixed total number of shots, though it does not fully reach the performance of an oracle with known signal direction.
The work provides a concrete framework for detecting weak coherent drives (motivated by wave-like dark matter searches) where signal parameters are unknown, establishing that adaptive learning of the signal direction is a viable strategy to enhance detection power within a fixed experimental budget.

More like this