⚛️ quantum physics

The exact information cost of randomized measurements in quantum metrology

This paper demonstrates that while 3-designs preserve a constant fraction of quantum Fisher information, 2-designs like MUBs and SICs suffer from dimension-dependent information loss due to aligned blind spots, whereas independent random bases or Clifford measurements can achieve dimension-independent performance by scattering these blind spots.

Original author: Wanqing Li, Jiaxin Liu

Published 2026-11-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original author: Wanqing Li, Jiaxin Liu

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 author. For technical accuracy, refer to the original paper. Read full disclaimer

Quantum sensors are devices that use the strange rules of the subatomic world to measure the physical universe with extreme precision. They work by encoding unknown information, such as a magnetic field or a change in time, into a delicate quantum state. To read this information, scientists must perform a measurement. However, the act of measuring a quantum system is tricky; it disturbs the state, and the type of measurement chosen determines how much useful information can be extracted. The ultimate limit of this precision is set by a theoretical quantity called the quantum Fisher information, which represents the maximum amount of information any possible measurement could ever pull from a specific state.

When a sensor is trying to measure just one thing, scientists can usually design a perfect measurement to reach this limit. But real-world sensors often need to track several changing parameters at once, or the exact state of the system might be unknown. In these cases, the best strategy is often to use a "randomized measurement." Instead of picking one specific way to look at the system, the experimenter randomly chooses from a set of different measurement settings, records the outcome, and repeats this process many times. This approach is powerful because a single set of data can be used to learn about many different properties of the system, revealing hidden patterns and dynamics without needing to know the state in advance. The critical question for engineers and physicists has been: how much of that ultimate precision do we lose when we rely on randomness instead of a perfect, custom-made measurement?

A new study by Wanqing Li and Jiaxin Liu answers this question by revealing that the geometry of the measurement settings matters more than the symmetry of the set. For years, the scientific community relied on highly symmetric sets of measurements, known as mutually unbiased bases, which are mathematically elegant and widely used in quantum computing and cryptography. These sets are designed so that if you know the outcome in one setting, the results in all other settings are completely random. The prevailing assumption was that this high level of symmetry would guarantee a consistent, reliable performance for sensing, regardless of the size of the system. The researchers, however, discovered that this assumption is wrong. They found that these perfectly symmetric sets actually contain "blind spots" that line up with each other, causing a catastrophic loss of information for certain types of changes in the system.

The researchers showed that in these symmetric sets, the measurements are collectively blind to specific changes. Imagine a quantum state as a point on a sphere. Some measurements are good at detecting changes in the "height" of that point (population changes), while others are good at detecting changes in its "rotation" or phase. In a perfectly symmetric set, the measurements that are good at detecting height changes are exactly the ones that are blind to phase changes, and vice versa. When the system changes in a way that aligns with these blind spots, the measurement fails to register the shift. The study proves that for a system with a large number of levels, this alignment causes the sensor to retain only a tiny fraction of the available information—specifically, a fraction that shrinks as the system gets larger. For a system with a certain number of levels, the information retained can drop to as little as one divided by the number of levels plus one. This means that as the system grows, the sensor becomes increasingly inefficient, requiring exponentially more repetitions to achieve the same precision.

This loss is not a theoretical curiosity; it happens in real, simple systems, such as a single quantum bit encoded within a larger, multi-level atom. The researchers demonstrated that even in these small, manageable setups, the symmetric measurements fail to capture the full picture. The problem is that the "blind spots" of the different measurements are all pointing in the same direction, creating a large gap in the sensor's vision. The study explicitly rules out the idea that second-order symmetry alone is enough to guarantee good performance. While these symmetric sets work well for typical, random states, they fail miserably for the structured states that are most common in actual sensors, such as superpositions of just two energy levels.

The solution, the authors found, is not to seek more symmetry, but to break it. By scattering the blind spots so that they do not align, the sensor can recover its efficiency. The researchers showed that using a small number of completely random, unstructured measurement settings is far superior to using a large, perfectly symmetric set. They calculated that just five random bases are enough to guarantee that the sensor retains a constant, dimension-independent fraction of the information, no matter how large the system becomes. In practical terms, a dozen random bases outperform the entire set of symmetric bases for systems with more than about eleven levels. This finding changes the design rules for quantum sensors. Instead of building complex, perfectly symmetric measurement arrays, engineers can achieve near-optimal precision with a much simpler, randomized approach that avoids the geometric traps of the symmetric sets.

The study also confirmed that this principle holds for the random circuits used in modern quantum computers. For systems built from standard quantum bits, the random measurements used in these processors retain exactly one-third of the maximum possible information, a constant value that does not degrade as more bits are added. This provides a precise error budget for experiments, allowing scientists to know exactly how many times they need to repeat a measurement to reach a desired level of accuracy. The research concludes that the geometry of the measurement settings, rather than their abstract symmetry, is the true controller of performance. By understanding and managing these blind spots, scientists can design sensors that are robust, efficient, and capable of reaching the fundamental limits of precision in the quantum world.

Technical Summary: The Exact Information Cost of Randomized Measurements in Quantum Metrology

Problem Statement
Quantum metrology relies on measuring unknown parameters encoded in quantum states. The precision of such measurements is bounded by the Quantum Fisher Information (QFI). While standard strategies involve choosing a measurement optimized for a known state, this is often impractical when the state depends on the very parameters being estimated or when multiple parameters conflict. Randomized measurements offer a state-independent alternative, where a random unitary rotates the state before a fixed projective measurement. However, it remains unclear whether specific classes of randomized measurements, particularly those forming "2-designs" (which reproduce Haar averages up to the second order), retain a constant fraction of the QFI across all states and directions. Previous work established that 3-designs retain at least a quarter of the QFI, but the performance of 2-designs—widely used in practice due to their experimental feasibility (e.g., Mutually Unbiased Bases (MUBs), Symmetric Informationally Complete (SIC) measurements, and random Clifford circuits)—has been an open question.

Methodology
The authors analyze the retained fraction of QFI, defined as $c = \inf_{\psi, u} I(u)/J(u)$ , where $I(u)$ is the Fisher information of a randomized measurement along direction $u$ and $J(u)$ is the QFI. The analysis decomposes state changes into "population" changes (modulus variations) and "phase" changes. A single measurement basis acts as a polarizer, responding to population changes but remaining "blind" to phase changes (changes orthogonal to the basis in the complex plane).

The study employs:

Information Budget Analysis: Deriving an exact identity for pure states showing that $I(u) + I(iu) = J(u)$ (for regular states), where $iu$ is the conjugate direction. This establishes that any loss in QFI retention is due to an imbalance between the information retained for a change and its conjugate.
Geometric Analysis of Blind Spots: Investigating how the "blind spots" (phase changes) of different measurement bases align or scatter. The authors focus on "structured" states, specifically qubits encoded in two levels of a higher-dimensional qudit, to identify worst-case scenarios.
Analytical Proofs and Numerical Simulations: Theoretical bounds are derived for MUBs, SICs, and Clifford measurements. Numerical simulations (including adversarial searches over states and directions) are used to evaluate the performance of random bases and verify theoretical predictions against maximum-likelihood estimation costs.

Key Contributions and Results

2-Designs Do Not Guarantee Constant QFI Retention:
The paper demonstrates that being a 2-design is insufficient to guarantee a dimension-independent fraction of the QFI. The loss arises from the alignment of blind spots in structured states.
- Complete MUBs: For a qubit encoded in two levels of a $d$ -dimensional system, complete MUBs retain exactly $1/(d+1)$ of the QFI in the worst case. This occurs because the "which-level" basis detects population changes, while all $d$ interference bases are blind to them (treating them as phase changes).
- Random Clifford Measurements (Qudits): Similarly, random Clifford measurements on qudits of odd prime dimension lose a factor of $d+1$ for population-type parameters.
- SIC Measurements: SICs retain exactly $(1 - 1/(\sqrt{d} + 1))/d$ , which is comparable to the MUB loss.
The Role of Blind-Spot Geometry:
The primary determinant of worst-case performance is the geometry of blind spots, not just the order of the design.
- Aligned Blind Spots: In complete MUBs and Clifford measurements on qudits, the blind spots of different bases align for specific structured states (e.g., superpositions of two levels), leading to a dimension-dependent loss.
- Scattered Blind Spots: Randomizing bases independently scatters these blind spots. The authors prove that five independent Haar-random bases are sufficient to retain a dimension-independent fraction of the QFI for every pure state and direction with high probability. Numerical results suggest that twelve random bases outperform complete MUBs for dimensions $d \approx 11$ .
Specific Bounds and Values:
- Qubits ( $d=2$ ): Random Clifford measurements (which form a 3-design) retain exactly $1/3$ of the QFI in every direction. This is an exact result for any number of qubits, sharpening previous 3-design bounds.
- Magic Gates: Adding a single "magic" gate (non-Clifford resource) to Clifford circuits restores a dimension-independent fraction (at least $1/15$ , numerically $\approx 0.33-0.44$ for $d \le 11$ ).
- Typical vs. Structured States: Averages over Haar-random states mask this loss, as typical states retain a fraction approaching $1/2$ even with MUBs. The loss concentrates on structured states (superpositions of two measurement vectors), which are common in sensors (e.g., clock transitions).
Connection to Shadow Estimation:
The paper establishes a quantitative correspondence between sensing and learning (shadow estimation). The "shadow norm" bound for learning worst-case observables translates to a sensing guarantee. For 2-designs, the sensing cost is bounded by the shadow norm factor $\kappa$ , confirming that 2-designs suffer a dimension-dependent overhead in the worst case.

Significance and Claims
The authors claim that the "blind-spot geometry" is the critical factor controlling worst-case metrological performance, superseding the simple classification of designs by their order (2-design vs. 3-design).

Design Rules: For qubit processors, random Clifford measurements are safe, retaining exactly $1/3$ of the QFI. For qudit sensors, complete MUBs and random Clifford measurements incur a cost proportional to the dimension ( $d+1$ ). To achieve dimension-independent performance on qudits, one must either use random bases (approx. 12) or introduce a single non-Clifford (magic) gate.
Benchmarking: The paper argues that benchmarks for randomized measurements must include structured states (like encoded qubits) rather than just random states, as the latter fail to reveal the dimension-dependent losses inherent in symmetric 2-designs.
Practicality: The results provide exact error budgets for quantum sensors and suggest that a small number of unstructured random bases (or a single magic gate) can achieve near-optimal precision across all parameters and dimensions, offering a practical alternative to the costly implementation of full 3-designs or Haar-random unitaries.

The paper concludes that while perfect symmetry (as in MUBs) is optimal for tomography, it is detrimental for worst-case metrology due to the alignment of blind spots. Scattering these blind spots through unstructured randomness or minimal non-Clifford resources is the key to robust, state-independent quantum sensing.

Technical Summary: The Exact Information Cost of Randomized Measurements in Quantum Metrology

More like this