Machine learning the effects of many quantum measurements

arXiv:2509.08890 · quant-ph, cond-mat.stat-mech · Submitted 2025-09-10 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Quantum Radio. Generated commentary on the latest quantum physics and condensed matter papers.

Kai: We're looking at the paper: "Machine learning the effects of many quantum measurements".

Mira: Here's the gist of the paper. Measurements are essential for the processing and protection of information in quantum computers Here we demonstrate a new approach for detecting such measurement-induced entanglement Our results additionally demonstrate a transition in the ability of a classical agent to accurately model the experimental data, which is closely related to a measurement-induced phase transition This work can act as a basis for future experiments on quantum error correction and more general problems in quantum control.

Kai: Detecting Measurement-Induced Entanglement The authors create short-range entangled states of one- and two-dimensional arrays of qubits in a superconducting quantum processor They aim to characterize the long-range entanglement induced between distant pairs of qubits when they measure all of the others.

Mira: They use unsupervised training of neural networks on observations to create computational models for post-measurement states and, by correlating these models with experimental data, they reveal measurement-induced entanglement.

Kai: The strategy involves using a computational model that takes observed measurement outcomes as input and generates predictions for the corresponding post-measurement states.

Title and authors: Mira: Let's look at the title again "Machine learning the effects of many quantum measurements." It tells you right away that machine learning is the tool they are using to study these complex measurement effects.

Kai: Right and when we look at the authors, we see a mix of hardware people like Kai and theorists like Mira who bring different perspectives to this.

Lev: I'm curious about how many physical qubits they were actually working with when they ran these experiments. That fidelity is always a huge factor for me.

Kai: They used one-dimensional and two-dimensional arrays of superconducting qubits, which sounds like something you’d build on a cryostat.

Mira: And the paper suggests that this approach might be useful in broader areas, not just this specific setup but also in quantum control and error correction research.

The paper's summary: Kai: So the summary says they create short-range entangled states of these arrays and then try to characterize long-range entanglement between distant qubits when they measure everything else.

Mira: They use unsupervised training of neural networks on observations—the measurement outcomes—to build computational models for what those post-measurement states should look like.

Lev: So the model takes the measurements as input and predicts how the state changes after that measurement, right? That’s a generative model in action.

Kai: Right and they correlate these predictions with their actual experimental data to reveal this measurement induced entanglement. It's a bit like using the AI to figure out what happened behind the scenes.

Mira: The key finding seems to be that they observe this measurement induced entanglement by tuning things, like varying a measurement basis in the two-dimensional array, and seeing when their neural network starts struggling to learn accurately.

The paper's improvements: Lev: They mention a few ways this is better than what we’ve seen before. One is that they can observe these phase transitions without needing postselection of the states.

Kai: That’s a big deal because postselection usually means you have to discard most of your data, but they show MIE surviving even when the angle theta is quite large, up to theta over pi equal to zero point five.

Mira: And they also found that these lower bounds on entanglement negativity derived from learning from data are comparable to those we get when we already know the true quantum state.

Lev: That means this machine learning approach gives us an objective characterization of the effects of measurements by giving us a bound on mixed state entanglement measures like negativity.

Conclusion: Kai: So to wrap up the paper, they show that using generative neural networks can detect measurement induced entanglement in these short-range arrays.

Mira: The main implication is that this method connects our ability to learn how to model a quantum system with our ability to observe the effects of measurements in a more general way.

Lev: For me the real value is scalability because they suggest that the number of repeats needed to generate a particular state scales exponentially small with L squared, which implies this could be very useful without having full postselection.

Kai: So we can use AI to find entanglement that might otherwise be hidden by the measurement process itself.

Mira: It opens up possibilities for studying measurement induced collective phenomena in other systems too, like ultracold atomic and molecular systems.

Lev: And immediately for us it points toward quantum control, especially quantum error correction, because understanding these measurements is fundamental to correcting them.

Wanda Hou, Samuel J. Garratt, Norhan M. Eassa, Eliott Rosenberg, Pedram Roushan, Yi-Zhuang You, Ehud Altman

Department of Physics, University of California, San Diego · Department of Physics, University of California, Berkeley · Department of Physics Princeton University · Google Research · Materials Sciences Division, Lawrence Berkeley National Laboratory

quant-ph, cond-mat.stat-mech

Submitted: 2025-09-10

Updated: 2026-11-09

Comments: 8+4+8 pages

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 100/100

The gist: Measurements are essential for the processing and protection of information in quantum computers Here we demonstrate a new approach for detecting such measurement-induced entanglement Our results…

Key concepts

Measurement-Induced Entanglement (MIE)
This is a type of entanglement that appears between distant qubits only after a complete set of measurements is performed on other qubits. The study shows how this entanglement can be detected by comparing predictions from an AI model with actual experimental data, even without knowing the initial quantum state.
Generative Neural Networks (NNs)
These are neural networks trained without being told the correct answer beforehand. In this work, they are used to learn how to predict the quantum state of specific 'probe' qubits based only on the outcomes of measurements performed on all other qubits in an array.
Measurement-Induced Phase Transition
This is a critical change in behavior observed when tuning a system, such as changing measurement settings. The paper shows that this transition can be seen directly in experimental data, allowing researchers to observe entanglement effects without needing prior knowledge of the exact quantum state or postselection.

Terminology

Summary

Measurements are essential for the processing and protection of information in quantum computers Here we demonstrate a new approach for detecting such measurement-induced entanglement Our results additionally demonstrate a transition in the ability of a classical agent to accurately model the experimental data, which is closely related to a measurement-induced phase transition This work can act as a basis for future experiments on quantum error correction and more general problems in quantum control. p. 0

Detecting Measurement-Induced Entanglement

The authors create short-range entangled states of one- and two-dimensional arrays of qubits in a superconducting quantum processor They aim to characterize the long-range entanglement induced between distant pairs of qubits when they measure all of the others. They use unsupervised training of neural networks on observations to create computational models for post-measurement states and, by correlating these models with experimental data, they reveal measurement-induced entanglement. p. 0

Methodology for Model Construction

The strategy involves using a computational model that takes observed measurement outcomes as input and generates predictions for the corresponding post-measurement states. For the one-dimensional array, they use generative neural networks (NNs) trained without supervision to predict the quantum state of the probe qubits from the outcomes of measurements on all other qubits. p. 1 For two-dimensional arrays, they tune the system through a finite-size version of a measurement-induced phase transition by varying the measurement basis. p. 2

Observing Phase Transitions

In the one-dimensional array, cross-correlations between the predictions of the model and experimental data reveal MIE between the two ends of an array of 34 qubits. In the two-dimensional array, they demonstrate that a critical point is associated with a transition in the ability of the NN to learn, and thereby observe MIE. p. 2 They find that measurement-induced phase transitions can be observed without advance knowledge of the quantum state, and without postselection. p. 2

Bounding Entanglement Properties

Cross-correlations between the shadows and computational model allow them to bound the true measurement-averaged entanglement in the ensemble of physical states. They define bounds on mixed-state entanglement measures, such as negativity. p. 2 Specifically, they establish a lower bound on measurement-averaged negativity using cross-correlations. p. 5

Machine Learning Models Used

The authors focus on generative neural networks that feature an attention mechanism and are inspired by the BERT language model. They also construct models via variational optimization of tensor networks for one-dimensional qubit arrays. p. 4 The training involves minimizing a negative log-likelihood loss function, which is defined as the negative log-likelihood − log2⟨ψmρC mψm⟩ of observing (mA, mB) conditioned on the model predicting ρC m. p. 4

Experimental Verification and Limitations

The results show that lower bounds on negativity based on learning from data are comparable to those obtained given knowledge of the quantum state. In two dimensions, they find a peak in NQC m > 0 and hence Nm > 0 at intermediate θ, which is a key signature of a measurement-induced phase transition visible in experimental data alone. p. 6 They also show that the post-measurement states are sensitive to distant outcomes across the array. p. 9

Future Applications

The method highlights a deep connection between our ability to learn how to model a quantum system, and our ability to observe the effects of measurements. This approach allows for the study of measurement-induced collective phenomena in general settings, including ultracold atomic and molecular systems. p. 6 An immediate application is in quantum control, particularly quantum error correction. p. 6

Classification Approach

An alternative approach involves 'binning', or classifying, observed shadows according to a computational model for the system. This involves partitioning the space of valid K-qubit density matrices into disjoint classes Mi based on a model mapping sets of measurement outcomes m to estimates for post-measurement states ρC m. p. 5

Scalability and Conclusion

The results suggest that machine learning provides a scalable way to detect measurement-induced entanglement. The experimental results indicate that the number of repeats necessary to generate a particular ρm is exponentially small in L squared, suggesting that machine learning can detect entanglement without postselection. This indicates that the structure of the post-measurement state has a highly nonlocal dependence on outcomes across the array. p. 9

Detailed Experimental Setup

The experiments involve preparing short-range entangled cluster states of one- and two-dimensional arrays on superconducting quantum processors They perform all measurements simultaneously, including those used to prepare the states ρm on A and B. The random bases in which they measure the probe qubits are defined by the unitary operations VA and VB applied in step (ii). p. 9

One-Dimensional Array Details

In the one-dimensional array, measurements of Zj on qubits j = 2,..., L−1 create a state ρm of the probe qubits A and B. p. 1 The form of ρm depends on the set of measurement outcomes m observed on the L − 2 ‘preparation’ qubits and, in the absence of noise, ρm is one of four (mutually orthogonal) pure maximally entangled two-qubit states. p. 2

Two-Dimensional Array Details

In the two-dimensional array, they apply Hadamard gates followed by a sequence of two-qubit gates exp[i(π/4)P⟨j,k⟩Z jZ k] between all neighboring pairs of system qubits. They then apply single-qubit rotations to change the measurement bases. p. 3 The entanglement between the pair of spatially separated probe qubits is expected to onset non-analytically at the critical point in an infinite system. p. 2

Cross-Correlations for Bounds

The cross-correlations are used to bound the measurement-averaged quantities P mpmSm, P mpmIm, and P mpmNm. These bounds are constructed by choosing weights Wm in Eq. p. 2 (8) that are based on approximate computational models ρC m for the true (and inaccessible) post-measurement states ρm. p. 3

Conclusion on Detectability

If the mean I SC r > 0 and the mean is much larger than the standard error in the mean, their cross-correlations can show that P mpmIm > 0 with high probability. This demonstrates that post-measurement states are entangled. p. 0 The results show that MIE survives at larger values of θ, extending all the way up to θ/π = 0. p. 65, but this structure is invisible to the NN. p. 6

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Unsupervised learning of post-measurement states allows an interactive observer to detect measurement-induced entanglement without needing prior knowledge of how the system was prepared, as cross-correlations between the predictions of the model and the experimental data reveal MIE.

  2. The system can be improved by using attention-based neural networks inspired by BERT to generate computational models for post-measurement states, enabling efficient verification of properties through unsupervised training of neural networks on observations to create computational models for post-measurement states.

  3. A key application is the ability to observe measurement-induced phase transitions in experimental data alone, as the paper demonstrates that a key signature of a measurement-induced phase transition is visible in experimental data alone, and its observation requires neither postselection nor advance information on the structure of the unmeasured state.

  4. The AI system can be used to construct lower bounds on measurement-averaged entanglement negativity by correlating predictions from computational models with experimental shadows, which provides an objective characterization of the effects of measurements.

  5. For two-dimensional arrays, the system can distinguish between learnable and unlearnable states by observing a transition in the ability of a neural network to generalize from data as it relates to measurement-induced phase transitions.

Sources

Related papers