⚛️ quantum physics

Shallow randomized measurement in noisy quantum devices

This paper establishes a theoretical framework demonstrating that shallow-depth randomized measurements offer practical advantages over single-qubit measurements for learning quantum state properties on noisy devices, a finding validated through experiments on IBM quantum hardware with up to 40 qubits.

Original authors: Gyungmin Cho, Dohun Kim

Published 2026-07-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Gyungmin Cho, Dohun Kim

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

Imagine you have a very complex, delicate machine (a quantum computer) that is currently a bit "noisy" and prone to making mistakes. You want to understand what's happening inside it, but you can't look at every single gear and wire at once because the machine is too big and too fragile.

This paper is about a smarter, more efficient way to peek inside this noisy machine without breaking it or waiting forever.

The Problem: The "Blindfolded" Guess

Traditionally, to understand a quantum state, scientists used a method called Randomized Measurements. Think of this like trying to guess the contents of a sealed, opaque box by shaking it and listening to the sound.

The Old Way (Single-Qubit): You shake the box gently, one tiny part at a time. It's safe and easy, but you might miss how the parts interact with each other. It takes a lot of shaking (measurements) to get a clear picture.
The "Perfect" Way (Global Entanglement): You could shake the whole box violently at once to hear how everything rattles together. This gives a perfect picture, but on current noisy machines, shaking the whole thing at once is too much; the machine breaks (errors overwhelm the signal) before you get an answer.

The Solution: The "Shallow" Peek

The authors propose a middle ground called Shallow Randomized Measurements.

The Analogy: Instead of shaking the whole box or just one tiny corner, you shake the box in small, connected clusters (like shaking two or three adjacent compartments at once).
Why it works: This is "shallow" because it doesn't require deep, complex circuits that the noisy machine can't handle. But it's "entangling" because it captures how small groups of parts interact, which the single-corner shaking misses.

The New Toolkit: "Block Shadows"

The paper introduces a specific mathematical framework called Block Shadows. Imagine you are trying to reconstruct a puzzle, but the pieces are noisy.

The "Block" Strategy: Instead of looking at one puzzle piece at a time, you look at small 2x2 or 3x3 blocks of pieces together.
The Benefit: Because you are looking at these small blocks, you can mathematically "undo" the noise much more easily and accurately than if you tried to look at the whole puzzle or just single pieces. It's like having a special filter that cleans up the static in a small group of radio channels, making the music clear.

What They Discovered (The "Magic" Tricks)

The authors didn't just invent a new way to shake the box; they showed how to use this method to do several specific things better:

The "Deterministic" Shortcut:
Usually, you have to try random shakes over and over to get a good average. The authors showed a way to pick the best specific shakes in advance (like a chef choosing the perfect spices beforehand) so you need fewer attempts to get the same result. They tested this on a real IBM quantum computer and proved it works even with the machine's noise.
The "Multi-Shot" Trick:
Imagine you have a camera. Instead of taking one photo, changing the angle, and taking another, you take 10 photos from the same angle, then move. The paper shows that for these "shallow" measurements, taking multiple photos from the same angle (reusing the measurement) actually helps reduce errors and saves time, which wasn't always true for the old methods.
The "Bias" Helper:
Sometimes you have a rough guess of what the machine is doing (a "bias state"). The paper shows how to use this rough guess to correct your measurements. It's like having a blurry map and a GPS signal; you can combine them to get a clearer location than using the GPS alone. They showed this works even when the map is a bit wrong, as long as you correct for the machine's noise.
Teaching Computers to Learn:
They used these "shallow" measurements to feed data into a classical machine learning algorithm (a computer program trying to learn patterns).
- The Result: When the computer learned from data where small groups of qubits were measured together (the "block" method), it was much better at spotting a specific change in the quantum system (a "phase transition") than when it learned from single-qubit data. It's like the computer could "see" the pattern in the crowd better when it looked at small groups of people rather than individuals.

The Real-World Test

The authors didn't just do math on paper. They ran these experiments on a real quantum computer (IBM's ibm_marrakesh) with up to 40 qubits.

They found that even with the machine's natural errors, their "shallow" method was more efficient and accurate than the old "single-qubit" method.
They successfully measured properties of complex quantum states that would have been too noisy or too slow to measure using previous techniques.

Summary

In short, this paper says: "Don't try to shake the whole fragile machine at once, and don't just shake one tiny piece. Shake small, connected groups of pieces. It's faster, it handles the machine's noise better, and it gives you a clearer picture of what's actually happening inside."

They proved this works on real hardware, offering a practical way to get more useful data from the noisy quantum computers we have today.

Title: Shallow Randomized Measurement in Noisy Quantum Devices
Authors: Gyungmin Cho and Dohun Kim (Seoul National University)

Problem Statement

While quantum hardware is improving, near-term devices remain constrained by noise and limited circuit depth. Standard protocols for learning properties of quantum states, such as classical shadows, face a trade-off: Random Pauli Measurements (RPM) are experimentally simple but inefficient for low-rank observables, while Random Clifford Measurements (RCM) are efficient for low-rank observables but require deep circuits ( $O(n)$ ) that are infeasible on current noisy intermediate-scale quantum (NISQ) devices. Existing "shallow shadow" approaches, primarily based on brickwork circuits, offer a middle ground but often lack closed-form expressions for the shadow channel and its inverse, complicating post-processing and extensions to advanced protocols like derandomization or error mitigation. Furthermore, the systematic advantages of shallow measurements across diverse learning tasks (beyond specific applications) and their robustness to experimental noise require further theoretical and experimental validation.

Methodology

The authors develop a unified theoretical framework based on Clifford measurement ensembles, specifically focusing on block shadows. In this approach, the $n$ -qubit system is partitioned into blocks of size $k$ , and independent random Clifford unitaries from $Cl(k)$ are applied to each block. This structure allows for:

Explicit Analytical Formulas: Derivation of closed-form expressions for the shadow channel, its inverse, and the moments required for variance analysis, avoiding the numerical approximations often needed for brickwork circuits.
Integration of Advanced Protocols: The framework incorporates derandomization, multi-shot protocols (reusing measurement bases), common randomized measurements (CRM) with bias states, and error-mitigated estimators.
Noise Modeling: Theoretical analysis of noise models (layer-wise and post-unitary) assuming Pauli twirling converts errors into Pauli channels, allowing for the construction of error-mitigated estimators using the noisy shadow channel.
Experimental Validation: Implementation on IBM quantum hardware (up to 40 qubits, 46 layers of two-qubit gates) to test derandomized block shadows, common randomized measurements with error mitigation, and entanglement-enhanced machine learning.

Key Contributions and Results

1. Theoretical Framework and Derandomization

The authors combine derandomization with shallow measurements, obtaining deterministic measurement bases whose confidence bounds are no larger than the randomized average.
Result: Numerical and hardware experiments on the Cluster-Heisenberg model show that derandomized block shadows require fewer measurements than single-qubit derandomization to estimate energy uncertainty, even in the presence of hardware noise.

2. Sample Complexity and Typical-Case Scaling

The paper establishes that for block shadows with $k2^k = \Omega(n)$ , the shadow norm of observables with bounded Schatten 2-norm is constant in typical cases.
State-Dependent Observables: Unlike fixed-observable analyses, the framework handles correlated settings (e.g., state verification where the observable is the state itself). The authors show that in the typical-case regime, block shadows achieve the same scaling as global random Clifford measurements.
Purity Estimation: A multi-shot purity estimator is constructed. Theoretical bounds indicate that for typical states, the sample complexity scales similarly to RCM, with the $V_2/NS$ term suppressible by reusing measurement bases.

3. Common Randomized Measurements (CRM) and Error Mitigation

The framework extends CRM to shallow shadows, allowing the use of a bias state $\sigma$ (approximating $\rho$ ) to reduce variance.
Error Mitigation: The authors extend robust shadow techniques to block shadows. Experiments on 40 qubits demonstrate that error mitigation significantly improves the accuracy of CRM, even when the bias state is less accurate (lower bond dimension).

4. Entanglement-Enhanced Machine Learning

The study investigates whether classical machine learning (ML) models trained on data from entangling shallow measurements outperform those trained on single-qubit data.
Result: Using Kernel PCA on data from the Su–Schrieffer–Heeger (SSH) model, the authors show that two-qubit block measurements ( $k=2$ ) yield a sharper phase transition signature near the critical point ( $w=1$ ) compared to single-qubit measurements ( $k=1$ ). The derivative of the fitting function peaks at $w \approx 1.026$ , consistent with the theoretical transition point.

5. Post-Processing Efficiency

Due to the tensor-product structure of block shadows, the inverse shadow channel is exact and computable in closed form. This contrasts with brickwork shadows, which often require tensor-network approximations or numerical optimization for the inverse, offering a distinct advantage in post-processing efficiency.

Significance and Claims

The paper claims that shallow-depth measurements, specifically via the block-shadow framework, provide practical benefits on noisy quantum devices.

Robustness: The advantages of shallow measurements persist despite experimental noise, as validated by hardware experiments.
Versatility: The framework unifies various classical shadow variants (derandomized, multi-shot, CRM, error-mitigated) under a single theoretical umbrella based on Clifford ensembles.
Scalability: By using logarithmic-depth circuits ( $O(\log n)$ ), the approach bridges the gap between simple local measurements and global Clifford measurements, enabling the study of complex quantum states and physical quantities on current NISQ hardware.
ML Advantage: The results suggest that entangling measurements can improve the classical representation of quantum experimental data for machine learning tasks, potentially reducing the number of experiments required to learn ground state properties or detect phase transitions.

The authors maintain a modest stance, noting that their typical-case results do not imply worst-case guarantees (citing specific block-product counterexamples where scaling is unfavorable) and that the approach is most effective for systems where local correlations are significant or where bias states are available.