Learning mixed quantum states in large-scale experiments

This paper presents and experimentally validates an efficient protocol that utilizes classical shadows from local randomized measurements to learn the matrix-product operator representation of large-scale mixed quantum states, successfully demonstrated on a superconducting processor with up to 96 qubits.

The Gist

Imagine you have a giant, incredibly complex machine made of 96 tiny switches (qubits). This machine is supposed to perform a specific dance (a quantum state), but because the machine is noisy and imperfect, the dance gets a bit sloppy. Your goal is to figure out exactly what the machine is doing right now, despite the noise, so you can understand it and maybe even fix the mistakes.

The problem is that describing this machine perfectly is like trying to write down the position of every single atom in a galaxy. It's too much data for any computer to handle.

This paper presents a clever new way to solve this problem. Here is the breakdown using simple analogies:

1. The Problem: The "Blurry Photo"

Usually, to understand a quantum machine, scientists take thousands of photos from different angles (randomized measurements). This creates a massive dataset called "Classical Shadows."

• The Old Way: If you wanted to reconstruct the whole machine from these photos, you'd need to process a library's worth of data. It's slow and gets impossible as the machine gets bigger.
• The New Way: Instead of keeping the whole library of photos, the authors propose compressing that information into a single, smart "instruction manual."

2. The Solution: The "Lego Instruction Manual" (MPO)

The authors use a mathematical structure called a Matrix Product Operator (MPO).

• The Analogy: Imagine the quantum state isn't a giant, solid block, but a long chain of Lego bricks. Each brick (tensor) only needs to know how to connect to the bricks immediately next to it.
• Why it works: Even though the whole chain is 96 bricks long, you don't need to know the secret code for the whole chain at once. You just need to figure out how Brick 1 connects to Brick 2, then Brick 2 to Brick 3, and so on. This makes the data manageable, like a simple instruction manual rather than a chaotic pile of bricks.

3. The Process: The "Puzzle Solver"

How do they turn the blurry photos into this Lego manual? They use a step-by-step learning algorithm, similar to how a person solves a long puzzle.

• Step 1: The Guess. They start with a blank manual.
• Step 2: The Local Fix. They look at a small section of the machine (say, 5 bricks in the middle). They ask the "Classical Shadow" data: "If I change this specific brick, does the picture look more like the real machine?"
• Step 3: The Sweep. They move down the line, fixing one brick at a time, then going back and doing it again. This is like a "renormalization group" (a fancy term for a very efficient cleaning process).
• The Result: Eventually, the manual perfectly describes the noisy, real-world machine, not just the perfect theoretical one.

4. The Big Win: Seeing the Invisible

The authors tested this on a real superconducting quantum computer (IBM's Brisbane processor) with 96 qubits.

• Previous Records: Before this, scientists could only do this kind of detailed reconstruction for about 13 qubits. It was like trying to describe a whole forest by looking at a single tree.
• Current Achievement: They successfully mapped the entire "forest" of 96 qubits. They didn't just guess; they proved their map was accurate by checking if the "Lego manual" predicted the machine's behavior correctly.

5. The Superpower: Error Correction (The "Magic Filter")

Here is the most exciting part. Because they have this clean "Lego manual" of the noisy machine, they can use it to fix the errors.

• The Analogy: Imagine you have a blurry photo of a friend. Usually, you can't tell if they are smiling or frowning. But if you have a perfect 3D model of your friend (the MPO), you can use that model to "filter out" the blur and see the real smile underneath.
• The Result: They used their method to strip away the noise and recover the "pure" quantum state. They found that even though the machine was noisy, the underlying quantum state was actually very close to perfect (over 90% fidelity). This is a huge step toward making quantum computers useful for real-world problems.

Summary

Think of this paper as inventing a smart compression algorithm for quantum chaos.

1. Input: A messy, noisy quantum experiment with 96 qubits.
2. Process: A smart, step-by-step puzzle solver that builds a compact "instruction manual" (MPO) from the data.
3. Output: A clean, understandable description of the quantum state that allows scientists to see the signal through the noise and fix errors without needing to run the experiment again.

This moves us from "we can barely measure a few qubits" to "we can map and fix large-scale quantum systems," bringing us closer to practical quantum computers.

Technical Summary

Here is a detailed technical summary of the paper "Learning mixed quantum states in large-scale experiments."

1. Problem Statement

Probing the quantum state of large-scale systems ($N \gg 10$) is a fundamental bottleneck in quantum simulation and computation. Traditional quantum state tomography scales exponentially with system size, making it infeasible for large $N$. While Classical Shadows (randomized measurements) have emerged as a scalable tool to estimate specific properties (like entanglement or fidelities) with provable statistical guarantees, they are generally "state-agnostic." They do not provide a compact, global description of the state.

The specific challenge addressed in this work is: How can one efficiently reconstruct a global, compact representation (specifically a Matrix Product Operator, MPO) of a mixed quantum state generated in a noisy, large-scale experiment, using only classical shadow data, without requiring exponential measurement resources?

2. Methodology

The authors propose a protocol that combines Classical Shadows with Tensor Network optimization (specifically MPOs) to learn the state $\rho$.

A. The Learning Protocol

1. Input: The protocol takes as input two datasets of local randomized measurements (Classical Shadows) collected from the experimental state $\rho$.
   • Learning Set: Used to optimize the MPO parameters.
   • Testing Set: Used to benchmark the learned MPO $\sigma$ against the experimental state.
2. Ansatz: The target state is represented as a Matrix Product Operator (MPO) $\sigma$ with bond dimension $\chi'$. The MPO is defined by a set of $N$ local tensors $M^{(j)}$.
3. Optimization Strategy (Sequential Update):
   • The algorithm optimizes tensors sequentially, similar to the Density Matrix Renormalization Group (DMRG) algorithm.
   • Instead of maximizing the standard Max Fidelity (which is non-differentiable), the algorithm maximizes the Geometric-Mean (GM) Fidelity $F_{GM}(\rho, \sigma) = \frac{\text{tr}[\rho\sigma]}{\sqrt{\text{tr}[\rho^2]\text{tr}[\sigma^2]}}$.
   • Approximate Local Update: The core theoretical insight is that for states with finite correlation lengths (typical of noisy 1D circuits), the global optimization condition can be approximated by a local linear system. The optimal tensor $M^{(j)}$ approximately satisfies:
     $$\text{tr}_{I_j \setminus \{j\}}[\sigma_{I_j} \partial_{M^{(j)}} \sigma_{I_j}] = \text{tr}_{I_j \setminus \{j\}}[\rho_{I_j} \partial_{M^{(j)}} \sigma_{I_j}]$$
     where $I_j = [j-\ell, j+\ell]$ is a local window of size $2\ell+1$.
   • This equation is solved numerically as a linear system using the classical shadows from the learning set.
4. Efficiency: The number of measurements required to learn each tensor scales as $O(2^{2\ell}\chi' \epsilon^{-2})$, which is independent of the total system size $N$.

B. Scalable Benchmarking (AFC Fidelity)

To verify the learned state without re-processing the entire dataset or requiring exponential measurements:

• The authors utilize Approximate Factorization Conditions (AFC).
• They prove that for states with finite correlation lengths, the global fidelity and purity can be accurately estimated by factorizing them into products of local overlaps and purities over small regions of size $k$.
• The estimated fidelity is defined as:
  $$F^{(k)}_{AFC}(\rho, \sigma) = \frac{O^{(k)}_{AFC}(\rho, \sigma)}{\max(P^{(k)}_{2,AFC}(\rho), P^{(k)}_{2,AFC}(\sigma))}$$
• This allows for efficient benchmarking with a measurement budget scaling polynomially with $N$ (specifically $O(4^{2k}N^3)$), rather than exponentially.

C. Error Mitigation via QPCA

Once the MPO $\sigma$ is learned, the authors apply Quantum Principal Component Analysis (QPCA):

• They treat the MPO as a Hamiltonian $H = -\sigma$.
• Using DMRG, they find the dominant eigenvector $|\psi^\sigma_0\rangle$ (the principal component) of the mixed state.
• This effectively "purifies" the state, mitigating experimental noise by projecting onto the most probable pure state component.

3. Key Contributions

1. First Large-Scale MPO Tomography: The protocol successfully learns the MPO representation of entangled mixed states up to $N=96$ qubits, significantly surpassing previous limits (typically $N \le 13$ for randomized measurements and $N \le 20$ for MPS tomography).
2. Provable Efficiency: The authors provide rigorous proofs that the learning and benchmarking steps are efficient (polynomial in $N$) under the assumption of short-range correlations (finite Rényi-Markov length).
3. Hybrid Framework: They successfully bridge the gap between the statistical robustness of Classical Shadows and the expressive power of Tensor Networks.
4. Noise Characterization and Mitigation: The method not only reconstructs the state but also quantifies the effect of noise (via entropy analysis) and provides a pathway to error mitigation (via QPCA) without additional experimental runs.

4. Experimental Results

The protocol was demonstrated on the IBM Brisbane superconducting quantum processor (96 qubits) using a Kicked Ising Model circuit at depths $D=1$ and $D=2$.

• State Reconstruction:
  • For $N=96, D=1$, the learned MPO achieved a fidelity $F_{max} \approx 75\%$ with the experimental state.
  • For $N=96, D=2$, the fidelity was $\approx 50\%$.
  • Crucially, the learned MPO captured the mixed nature of the experimental state, whereas the fidelity with the ideal pure target state was much lower.
• Physical Properties:
  • The learned MPO accurately reproduced local observables (magnetization $\langle Z_j \rangle$) and two-body correlations $\langle Y_j Y_{j+d} \rangle$, matching the experimental data better than the ideal theoretical prediction (which ignores noise).
  • Entropy Analysis: The second Rényi entropy $S_2$ was measured, showing a weak volume law ($S_2 \propto \alpha N$ with $\alpha \approx 0.14$), indicating the presence of significant but not maximal noise.
• Error Mitigation:
  • Applying DMRG-QPCA to the learned MPO yielded a purified state $|\psi^\sigma_0\rangle$ with >90% fidelity to the ideal target state for $D=1$.
  • Local observables of the mitigated state showed dramatic improvement over the raw experimental data.

5. Significance

• Scalability: This work demonstrates that full quantum state tomography (in the form of a compressed MPO) is feasible for systems with nearly 100 qubits, a regime previously thought to be accessible only for specific properties or small systems.
• Data Efficiency: By leveraging the structure of noisy quantum states (short-range correlations), the method avoids the exponential scaling of traditional tomography.
• Practical Utility: The ability to convert experimental data into a classical MPO object enables:
  • Repeated estimation of various physical properties without re-running experiments.
  • Quantum Error Mitigation via principal component analysis.
  • Quantum Circuit Cutting: Saving intermediate states as MPOs to resume algorithms on different hardware.
  • Hybrid Quantum-Classical Workflows: Using a classical MPO to prepare states on a fault-tolerant quantum computer.

In summary, the paper presents a robust, scalable, and experimentally validated framework for learning and analyzing mixed quantum states in the noisy intermediate-scale quantum (NISQ) era, effectively turning "noisy" experimental data into a high-fidelity, error-mitigated classical description.