Large-scale implementation of quantum subspace expansion with classical shadows

arXiv:2510.25640 · quant-ph · Submitted 2025-10-29 · 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: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "Large-scale implementation of quantum subspace expansion with classical shadows".

Mira: Here's the gist of the paper. Quantum subspace expansion (QSE) offers promising avenues to perform spectral calculations on quantum processors but comes with a large measurement overhead Here, we report the first large-scale implementation of QSE with IC measurements.

Kai: In particular, we probe the quantum phase transition of a spin model with three-body interactions, for which we observe accurate ground state energy recovery and mitigation of local order parameters across system sizes of up to 80 qubits.

Mira: We achieve this by reformulating QSE as a constrained optimization problem, obtaining rigorous statistical error estimates and avoiding numerical ill-conditioning.

Kai: The QSE Framework Quantum subspace expansion (QSE) is defined by the state representation ρSE(⃗c) = W†ρ0W/ Tr W†ρ0W, parametrized by a coefficient vector ⃗c = (c1,..., cL).

Paper summary: Mira: So, looking at the paper "Large-scale implementation of quantum subspace expansion with classical shadows," what's the big picture takeaway for us?

Kai: The main point is that they successfully implemented a large-scale version of QSE using classical shadows and showed it can recover ground state energies accurately up to eighty qubits.

Lev: The core idea they bring in is turning QSE into a constrained optimization problem, which gives them a way to manage the statistical error by setting an external limit on it, epsilon max.

Mira: It’s about using those classical shadow measurements—those single-qubit X, Y, and Z measurements—to build the necessary information before they start optimizing the state expansion coefficients.

Kai: The implication is that we can get noise-agnostic error mitigation in ground state preparation circuits for systems up to eighty qubits using this method.

Lev: They also showed that this approach allows them to iteratively optimize the chosen expansion operators, meaning they can reuse their measurement data to improve the estimate over time.

Mira: That’s important because it shows a practical way to handle the measurement bottleneck in these spectral calculations on quantum processors.

Kai: The paper notes that they found up to six singular values need to be discarded out of a twenty-two dimensional subspace just to respect the variational principle, which is a concrete detail about the complexity involved.

Lev: And they pointed out that for each system size N, you end up needing about one thousand twelve times the total number of single-qubit Pauli traces just to run all these estimates.

Mira: So it’s a large-scale implementation that works, but it comes with a lot of computational overhead on the classical side.

Kai: Exactly. It shows the feasibility of this framework, even if running the whole thing takes about one hour of wall clock time for those eighty qubit tests.

Conclusion: Kai: So, we've seen how they actually built this thing up to eighty qubits and got those ground state energies back, right?

Mira: Yeah, but I want to talk about the title itself—"Large-scale implementation of quantum subspace expansion with classical shadows." It sounds pretty technical.

Lev: It is technically dense. But what it means in simple terms is they took a complex way to look at quantum states and made it work on a real, noisy machine.

Kai: Right. Think about the authors who did this work, they're showing how you can tackle those measurement headaches by using these classical shadows we talked about before.

Mira: I think the big idea here is that you don't need perfect quantum measurements to get good results anymore, if you use smart classical tools to fill in the gaps.

Lev: For someone building real hardware, this means we can use existing measurement setups instead of needing entirely new, expensive equipment just for error mitigation.

Kai: It’s about proving that the method is solid enough to handle those bigger systems without just breaking down into noise.

Mira: And it shows that you can control your own error budget by setting a limit on how noisy you let the calculation get, which is a really practical thing for anyone trying to make sense of these quantum simulations.

Lev: It points toward making these kinds of spectral calculations less dependent on perfect hardware and more dependent on smart post-processing.

Kai: And that’s what we need to keep looking at—how this works when you push it beyond eighty qubits, where the measurement overhead gets really big.

Laurin E. Fischer, Daniel Bultrini, Ivano Tavernelli, Francesco Tacchino

IBM Quantum Research Europe Zurich · Ecole Polytechnique Férale de Lausanne · Physikalisch-Chemisches Institut Heidelberg University

quant-ph

Submitted: 2025-10-29

Updated: 2025-10-29

Comments: 10 pages, 5 figures, 2 tables

Journal ref: Physical Review Letters 137.16 (2026): 160602

DOI: 10.1103/2w21-qs38

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

Importance score: 82/100

The gist: Quantum subspace expansion (QSE) offers promising avenues to perform spectral calculations on quantum processors but comes with a large measurement overhead Here…

Key concepts

Quantum Subspace Expansion (QSE)
QSE is a method that represents quantum states using a set of basis vectors. By expanding the state space beyond what the hardware can directly simulate, it allows for accessing more complex quantum information. This expansion helps in finding properties of the system that are difficult to calculate directly on current quantum computers.
Classical Shadows (CS)
Classical shadows are a technique used to estimate all necessary matrix entries from a single set of measurements. Instead of needing many different types of measurements, CS allows researchers to derive information about all observables by using only one type of measurement data, significantly reducing the required experimental overhead.
Constrained Optimization
The authors framed QSE as an optimization problem where they seek the minimum expectation value. This framework is crucial because it allows them to obtain accurate statistical error estimates and control the trade-off between bias and variance when estimating quantum properties, leading to more reliable results.

Terminology

Summary

Quantum subspace expansion (QSE) offers promising avenues to perform spectral calculations on quantum processors but comes with a large measurement overhead Here, we report the first large-scale implementation of QSE with IC measurements. p. 0 In particular, we probe the quantum phase transition of a spin model with three-body interactions, for which we observe accurate ground state energy recovery and mitigation of local order parameters across system sizes of up to 80 qubits. p. 0 We achieve this by reformulating QSE as a constrained optimization problem, obtaining rigorous statistical error estimates and avoiding numerical ill-conditioning. p. 0

The QSE Framework

Quantum subspace expansion (QSE) is defined by the state representation ρSE(⃗c) = W†ρ0W/ Tr W†ρ0W, parametrized by a coefficient vector ⃗c = (c1,..., cL). This framework allows accessing additional states beyond the scope of the quantum processor and may also serve as noise-agnostic error mitigation. p. 0 The expectation value of an observable O over a state in the subspace becomes Tr[OρSE(⃗c)] = PL i,j=1 c∗i cjOij / PL i,j=1 c∗i cjSij. p. 0 For ground state search, the goal is to find the minimal expectation value min⃗c Tr[HρGSE(⃗c)] which corresponds to the smallest pseudoeigenvalue λ of the generalized eigenvalue problem H⃗c = λS⃗c. p. 0

Overcoming Measurement Bottlenecks

The main experimental cost of QSE stems from the number of measurements required to estimate all matrix entries Oij and Sij, which depends on the dimension L and the complexity of the expansion operators. To alleviate this, informationally complete (IC) measurements, such as classical shadows [7] (CS), can be used to estimate all underlying observables from one set of measurement samples. p. 0 The authors formulate subspace expansion as a constrained optimization problem to obtain accurate statistical uncertainties and provides an external tuning knob to control these errors, trading potential bias for variance. p. 1

Experimental Demonstration

The experimental demonstration involves recovering the ground state energy of a Hamiltonian with known approximate ground state preparation circuits. They consider a one-dimensional spin chain described by the Hamiltonian H = X sum i=1 N (−gzzZiZi+1 − gxXi + gzxzZiXi+1Zi+2). p. 2 They implement the MPS ansatz circuits with system sizes increasing from N = 16 with 108 entangling CZ gates to N = 80 with 540 CZ gates on the IBM Quantum superconducting qubit device ibm fez. p. 2 The measurement involves an IC POVM that consists of uniformly randomized single-qubit X, Y, and Z measurements (standard local classical shadows with canonical duals [28]). p. 2

Key Findings on Mitigation

The results for the error-mitigated energy densities of the MPS ground state circuits for g ∈ (−0.9, +0.9) are shown in the top row of Fig. p. 3 2. The Krylov subspace already significantly improves on the unmitigated energies, while the Krylov+ approach manages to accurately recover the true ground state expectation values while respecting the variational principle within statistical uncertainties. However, for complex observables like S ZY, statistical uncertainties in the “Krylov+” values dominate at larger scales. p. 3

Conclusion and Future Work

The work represents one of the most significant experimental realizations of classical shadows to date. The approach offers two additional advantages over traditional methods, including the ability to iteratively optimize the chosen expansion operators using reused measurements. p. 2 The method suggests that only sufficiently local observables can be estimated efficiently when single-qubit randomized measurements are employed. p. 4 This technique could be extended by incorporating symmetry verification and powers of the root state in the spirit of virtual distillation. p. 4 The total post-processing for the experimental demonstration took ≈ 20 h. p. 7

Appendix Details

The authors detail the Pauli operators selected to construct the “Krylov+" subspaces, noting that the total dimensions of these spaces are kept feasible for standard laptop processing. The overlap matrix S and solves the generalized eigenvalue problem by computing the lowest eigenvalue of S˜−1H with a regularized overlap matrix S˜. p. 8 p. 8 The problem is regularized by setting the diagonal entries in S−1 = diag(1/s1,. p. 8 p. 8.. 1/sL) to zero for the smallest singular values and computing S˜−1 = V T S − 1 U T. p. 8 p. 8 The authors find that up to six singular values need to be discarded (out of a 22-dimensional subspace) to respect the variational principle. p. 8 p. 8 The computational bottleneck of our method is the evaluation of the Pauli traces between subspace expansion operators and measurement dual operators. p. 7

Hardware and Optimization

The classical post-processing is run on a standard laptop with an Intel Core Ultra 9 185H CPU leveraging the Numba Python package. The central hyperparameter in our procedure is the allowed maximal statistical error ϵmax, which controls a bias-variance-tradeoff. p. 7 This allows us to explore only those regions of the subspace where the statistical error is well controlled. p. 1 The results for the energy-mitigated energy densities are shown in Fig. p. 3 4a and b. The allowed error is chosen as a fixed multiple of the unmitigated signal to account for the increase in noise with the system size. p. 7 The optimization converges with, on average, 1438 iterations. p. 7 The method does not provide tunable statistical error bars in contrast to the constrained optimization procedure. p. 8 The overlap matrix S and solves the generalized eigenvalue problem by computing the lowest eigenvalue of S˜−1H with a regularized overlap matrix S˜. The problem is regularized by setting the diagonal entries in S−1 = diag(1/s1,... 1/sL) to zero for the smallest singular values and computing S˜−1 = V T S − 1 U T. The authors find that up to six singular values need to be discarded (out of a 22-dimensional subspace) to respect the variational principle. The authors note that the total numbers of single-qubit Pauli traces needed are O(1012) for each system size N. p. 6 The total quantum circuit execution for all results shown in Fig. p. 6 2 took ≈ 1 h of wall clock time. 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. Constrained optimization for subspace energy calculation allows exploring only those regions of the subspace where statistical error is well controlled by solving arg min c in Hˆ (c) subject to ˆϵ(c) ≤ ϵmax. This provides a direct tuning knob to trade off statistical errors for potentially lower estimates of the energy.

  2. Informationally complete measurements, specifically classical shadows, are used to estimate all underlying observables from one set of measurement samples, which helps alleviate the main experimental cost of QSE by avoiding the need to measure each observable individually.

  3. The framework offers flexibility to estimate additional observables O for an obtained energy-optimized state without additional measurements by constructing an estimator Oˆ (copt) equivalently to Eq. (4).

Sources

Related papers