⚛️ quantum physics

Dynamical spectral functions from bitstring-sampled quantum subspaces: entanglement, not one-body magic, tracks the sampling cost

This paper demonstrates that bitstring-sampled quantum subspaces can accurately reconstruct dynamical spectral functions for both molecules and Hubbard chains, revealing that the sampling cost is fundamentally driven by entanglement rather than one-body magic or determinant support.

Original authors: Nicolás Bonilla Vargas

Published 2026-08-18

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Nicolás Bonilla Vargas

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

In the world of materials science, understanding how electrons move and interact is the key to unlocking new technologies, from better batteries to faster computers. For decades, scientists have relied on powerful classical supercomputers to simulate these tiny particles, but as the systems they study grow more complex, the calculations become impossibly difficult. The electrons do not move in isolation; they are deeply entangled, meaning the state of one instantly influences the others in a web of relationships that is hard to track. To solve this, researchers have turned to quantum computers, machines designed to mimic nature's own rules. The hope has been that these devices could find the lowest energy state of a molecule, a value that tells us how stable a material is. However, a surprising twist has emerged: for finding just that single energy value, clever new classical algorithms have caught up to, and in some cases surpassed, the current generation of quantum machines. The question now is not whether quantum computers can calculate a static number, but whether they can do something else that classical machines struggle to do: track how a material changes over time and responds to energy, revealing the dynamic behavior of its electrons.

A researcher has now taken a significant step in this direction by shifting the goalpost from static energy to dynamic motion. Instead of asking a quantum computer to find the resting energy of a molecule, they asked it to simulate the real-time evolution of electrons and then measure the resulting patterns. They developed a method that uses a quantum processor to generate a stream of simple data points, known as bitstrings, which represent snapshots of the system's configuration. These snapshots are collected from a very shallow, simple circuit that does not require the complex, error-prone operations that have plagued earlier quantum experiments. Once these snapshots are gathered, a classical computer takes over, assembling them into a detailed map of the material's spectral function. This map is essentially a fingerprint of the material, showing how it absorbs and emits energy at different frequencies and momenta. The researcher tested this approach on a chain of atoms and a suite of nineteen different molecules, and in every case, the reconstructed maps matched the exact theoretical results with high precision. They even ran the experiment on a real quantum processor, the IBM Heron, successfully reproducing the spectral function of a small chain of atoms, proving that the method works on actual hardware without needing the most advanced error-correction techniques.

The study also tackled a fundamental question that has confused the field: what exactly makes a problem hard for a quantum computer? For years, scientists assumed that a specific property of the electrons, often called "magic" or non-Gaussianity, was the primary driver of difficulty. This property was thought to measure how far a system was from behaving like simple, non-interacting particles. The researcher tested this idea rigorously across their suite of molecules and found that this assumption was incorrect. They discovered that this "magic" number, while useful for identifying how complex the electron relationships are, has no reliable connection to the actual cost of running the simulation. A system could have high "magic" but be easy to simulate, or low "magic" but be extremely difficult. Instead, the true cost is determined by a different factor: the entanglement structure of the system as seen through the specific lens of the quantum computer's measurement. They found that the number of distinct configurations the computer must sample to get an accurate answer is directly tied to the "bond dimension," a measure of how many connections the electrons have to each other in the specific frame of reference used by the machine. This finding clarifies that the difficulty lies not in a single, easy-to-measure number, but in the complex, high-dimensional web of connections that classical computers struggle to compress.

To ensure their method was robust, the researcher also explored the role of artificial intelligence in the process. They trained a machine learning model to try and predict the missing pieces of the data, hoping it could fill in the gaps left by the quantum computer's limited sampling. The results were sobering but informative: the learning model did not outperform a much simpler, classical statistical method. In fact, the classical method was just as good at completing the picture. This suggests that the quantum computer's role is not to be replaced by a smart algorithm, but to provide the raw, high-quality data that no classical method can generate on its own. The quantum machine acts as a unique sensor, capturing the raw material of the simulation, while the classical computer and simple statistical tools do the heavy lifting of assembling that material into a coherent story. The researcher also found that the method is surprisingly resilient to the noise and errors inherent in current quantum hardware. By using a self-consistent recovery process, they could filter out the corrupted data points and still arrive at an accurate result, even when the hardware was making mistakes. This noise resilience is a crucial feature, as it means the method can work on today's imperfect machines without waiting for the perfect, fault-tolerant computers of the future.

The implications of this work extend beyond just a new algorithm; it redefines what we should expect from quantum computers in the near future. The ability to calculate the dynamic spectral function opens the door to studying phenomena that are currently out of reach, such as how electrons move in complex materials under the influence of light or magnetic fields. These are the kinds of measurements that experimental physicists make in the lab using tools like photoemission spectroscopy, and having a way to simulate them on a computer would revolutionize the design of new materials. The study demonstrates that the path forward is not to chase the ground-state energy, where classical methods are already dominant, but to focus on these dynamic, frequency-resolved observables where the quantum advantage is most defensible. By identifying the true resource that drives the cost—the basis-dependent entanglement rather than the orbital-invariant "magic"—the researcher has provided a clear map for where quantum computers will be most useful. They have shown that the real power of these machines lies in their ability to sample complex distributions that are hard for classical computers to generate, and that this power can be harnessed today to solve problems that matter, provided they ask the right questions and use the right tools to interpret the answers.

Technical Summary: Dynamical Spectral Functions from Bitstring-Sampled Quantum Subspaces

Problem Statement
Sample-based quantum diagonalization (SQD) and quantum-selected configuration interaction (QSCI) have emerged as leading near-term quantum algorithms for electronic structure, primarily targeting ground-state energies. However, the classical frontier for ground-state energy has closed; heuristic classical methods (e.g., matchgate simulators, autoregressive neural samplers, heat-bath configuration interaction) now match or exceed the performance of quantum-sampled subspaces for this specific observable. Consequently, the claim of a generic exponential quantum advantage in ground-state chemistry is increasingly untenable.

The paper identifies a critical gap: no published method reconstructs frequency-resolved single-particle spectral functions ( $A(\omega)$ ) or dynamical structure factors ( $S(q, \omega)$ ) from bitstring-sampled quantum subspaces, particularly in the charged ( $N \pm 1$ ) sectors required for Green's functions. Furthermore, there is a lack of clarity regarding which resource actually dictates the sampling cost. A natural hypothesis is that "one-body fermionic magic" (a measure of non-Gaussianity) drives the difficulty, but this remains unproven and potentially misleading.

Methodology
The author proposes a unified primitive that lifts the SQD/QSCI toolkit from static energies to continuous, frequency-resolved dynamics using only computational-basis measurements.

Sampling Primitive: The method utilizes a shallow, real-time evolution circuit without Hadamard tests, controlled unitaries, or overlap estimation.
- Seeds: For single-particle spectra, the system prepares response seeds $|\phi^\pm\rangle = \hat{c}^\dagger_p |0\rangle$ and $\hat{c}_p |0\rangle$ in the $(N \pm 1)$ sectors. For neutral dynamical structure factors, seeds are number-conserving density or spin operators $\hat{n}_q |0\rangle$ or $\hat{S}^z_q |0\rangle$ .
- Evolution: These seeds are time-evolved ( $e^{-i\hat{H}t_k}$ ) and measured in the computational basis to generate bitstrings.
- Subspace Construction: The union of sampled configurations across evolution times forms a determinant subspace $\mathcal{S}$ .
Classical Reconstruction: The quantum processor provides only bitstrings. The Hamiltonian is projected onto $\mathcal{S}$ and diagonalized classically. The retarded Green's function and spectral functions are assembled using the Lehmann representation:
$G_{pq}(\omega) = \sum_m \frac{\langle 0|\hat{c}_p|\tilde{m}^+\rangle \langle \tilde{m}^+|\hat{c}^\dagger_q|0\rangle}{\omega - (\tilde{E}^+_m - E_0) + i\eta} + \dots$
This yields $A(\omega)$ , momentum-resolved $A(k, \omega)$ , and dynamical structure factors $S(q, \omega)$ and $S^{zz}(q, \omega)$ .
Noise Mitigation: The method employs self-consistent configuration recovery (S-CoRe) to project noisy bitstrings back into the correct particle-number sector, improving subspace quality under realistic device noise.

Key Contributions

Dynamical Spectral Reconstruction: The paper demonstrates the first reconstruction of frequency-resolved single-particle spectral functions and dynamical structure factors from bitstring-sampled subspaces. This extends the SQD/QSCI paradigm from discrete energy levels to continuous dynamical observables, validated against exact diagonalization on Hubbard chains and a suite of 19 molecules.
Decoupling of Magic and Cost: The author rigorously proves that one-body fermionic magic (quantified by the fermionic AntiFlatness $F_1 = 4 \text{tr}[\gamma(1-\gamma)]$ $F_{1} = 4 tr [γ (1 - γ)]$ ) is decoupled from the sampling cost.
- $F_1$ is an orbital-rotation invariant (Gaussian invariant), whereas the sampling cost (determinant support $|S|$ ) is basis-dependent.
- While $F_1$ faithfully diagnoses static, multireference correlation (e.g., rising during bond dissociation), it fails to predict the sampling cost. For instance, increasing on-site repulsion in Hubbard chains raises $F_1$ but lowers $|S|$ due to localization.
- The true cost driver is identified as basis-dependent entanglement, specifically the minimal bond dimension $\chi$ (tensor-network bond dimension), which empirically tracks $|S|$ with high correlation ( $\rho \approx 0.90$ ).
Resource Diagnostics: The work establishes that genuine quantum advantage, if it exists, resides in the high-rank, non-Gaussianity of connected higher-body cumulants in the fixed sampling basis, rather than in the orbital-invariant one-body magic.

Results

Validation: The method reproduces exact spectral functions for Hubbard chains and 19 molecules (verified against Full Configuration Interaction to $<10^{-5}$ Ha) with relative- $L_1$ errors below $10^{-3}$ .
Hardware Demonstration: The algorithm was executed on the IBM Heron processor ( $ibm\_fez$ $ibm_f ez$ for $L=6$ $L = 6$ Hubbard model; $ibm\_marrakesh$ $ibm_ma r r ak es h$ for stretched $N_2$ $N_{2}$ ).
- The $L=6$ Hubbard spectral function was reconstructed from 50,000 shots, matching the exact result within broadening resolution.
- For stretched $N_2$ , noise-assisted configuration recovery achieved a ground-state energy error of $0.59 \pm 0.14$ mHa (below chemical accuracy), outperforming noiseless simulations of the same shallow circuit.
Scaling and Comparison:
- The sampling cost (fraction of the sector needed) decreases exponentially with system size for fixed accuracy, though the absolute determinant support $|S|$ grows exponentially.
- The time-evolution selector outperforms deterministic classical Krylov and CIPSI selectors in strong-coupling regimes ( $U/t \ge 12$ ) once the subspace exceeds $\sim 120$ determinants, though it does not surpass heat-bath CI at accessible sizes.
- A controlled comparison between a Hubbard chain and a two-leg ladder (matched one-body magic) showed that the ladder requires a significantly larger sampling fraction due to higher bond dimension $\chi$ , confirming that cost tracks $\chi$ , not $F_1$ .
AI/ML Limits: A learned generative model (GFlowNet) used as a "tail-completer" to supplement device samples did not statistically outperform a cheap classical CIPSI baseline, suggesting that current generative models do not yet replace the need for quantum sampling in this context.

Significance and Claims

The paper reframes the quantum advantage debate in electronic structure:

Shift in Target: It argues that the defensible quantum advantage lies not in ground-state energies (where classical methods have caught up) but in dynamical observables (spectral functions, structure factors) which are central to materials discovery and correlated-electron physics but remain hard for classical tensor networks in dimensions $>1$ .
Clarification of Resources: It corrects the misconception that one-body magic certifies sampling hardness. Instead, it provides a precise resource statement: one-body magic is a diagnostic for correlation type (multireference character), while the cost is determined by basis-dependent entanglement ( $\chi$ ).
Practical Utility: The method offers a hardware-native approach (no ancillas, no complex gates) to compute observables that current energy-only hardware cannot provide. It serves as a "triage" tool: the efficiently measurable fermionic magic can flag strongly correlated targets where mean-field methods fail, but the decision to use a quantum processor must rely on estimating the basis-dependent cost (bond dimension), which one-body magic cannot predict.

The author concludes that while the method does not claim a "beyond-classical" advantage for the specific instances tested (as classical DMRG remains effective for 1D and narrow ladders), it successfully defines the frontier where quantum sampling becomes necessary: in the high-rank, non-Gaussian sector of the fixed sampling basis, beyond the reach of free-fermion estimation and low-bond-dimension tensor networks.

Technical Summary: Dynamical Spectral Functions from Bitstring-Sampled Quantum Subspaces

More like this