⚛️ quantum physics

Quantum Folding in Chebyshev-Based Quantum Subspace Diagonalization

This paper proposes "quantum folding," a method that seeds Chebyshev-based Quantum Subspace Diagonalization with the product of subsystem ground states to overcome the exponential overlap decay of conventional product-state initialization, thereby enabling scalable, noise-resilient ground-state preparation with polynomial cost.

Original authors: Seung Park, Dongni Chen, Stefano Chesi, Kyunghyun Baek, Mahn-Soo Choi

Published 2026-10-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Seung Park, Dongni Chen, Stefano Chesi, Kyunghyun Baek, Mahn-Soo Choi

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 quest to understand the behavior of matter at its most fundamental level, physicists often turn to quantum mechanics, the set of rules that governs the subatomic world. While these rules are precise, applying them to large collections of particles—like the atoms in a solid material or the electrons in a molecule—is notoriously difficult. The complexity grows so rapidly with the number of particles that even the most powerful supercomputers struggle to keep up. To solve this, scientists are developing quantum computers, machines designed to mimic these quantum systems directly. However, these machines are still in their early stages, prone to errors and limited in size. A major challenge in using them is finding the "ground state," the lowest energy configuration of a system, which determines its stability and properties. Traditional methods for finding this state on a quantum computer often rely on starting with a simple guess. If that guess is too far from the true answer, the method fails, especially as the system gets larger.

A team of researchers has proposed a new strategy called "quantum folding" to overcome this hurdle. Instead of guessing the state of a large system from scratch, they suggest building the answer piece by piece. Imagine trying to understand a long chain of magnets. Rather than guessing the arrangement of every magnet at once, the researchers propose first finding the perfect arrangement for a short segment of the chain. They then take two of these perfect segments and join them together to form a longer chain. Because the connection between the two segments is small compared to the whole, the combined state is already very close to the true answer for the longer chain. This process can be repeated, doubling the length of the chain at each step, allowing the computer to solve for increasingly large systems without getting lost in the complexity.

The researchers tested this idea using a specific mathematical technique known as Chebyshev-based quantum subspace diagonalization. This method works by projecting the problem onto a smaller, manageable space and solving it with high precision. In their simulations, they compared their "folded" approach against the standard method, which uses a simple, rigid starting guess known as the Néel state. The results were striking. For systems governed by local interactions, where particles only affect their immediate neighbors, the folded starting point was vastly superior. In simulations of antiferromagnetic chains, the folded method reduced the error under noise by two to five orders of magnitude compared to the standard method. This means the folded approach could find the correct answer with far fewer computational steps and could tolerate much higher levels of experimental error.

The success of this method hinges on a concept called "overlap," which measures how close the starting guess is to the true ground state. For standard methods, this overlap shrinks exponentially as the system grows, making the problem impossible to solve for large sizes. The researchers proved that by using the ground states of smaller subsystems as the starting point, this overlap remains strong. For systems with an energy gap—a property that makes them stable against small disturbances—the overlap stays constant regardless of size. For systems without such a gap, the overlap still holds up remarkably well, decaying only very slowly. This allows the recursive "folding" process to work efficiently, turning an exponentially difficult problem into one that scales polynomially, meaning it grows at a manageable rate as the system size increases.

To make this practical, the team also addressed how to turn the mathematical output of the quantum computer back into a physical state that can be used for the next step. They showed that the final state can be prepared with a high probability of success, provided the computer is calibrated correctly and the data is processed to filter out noise. Their analysis revealed that the cost of preparing these states grows slowly with the size of the system, making the entire framework scalable. They calculated the exact efficiency of this "folding" for different types of magnetic chains, finding that for some, the efficiency settles at a constant value, while for others, it follows a predictable, slow decline.

The study concludes that this approach offers a viable path forward for solving large quantum problems on future fault-tolerant quantum computers. By avoiding the exponential decay of accuracy that plagues current methods, quantum folding allows researchers to tackle systems that were previously out of reach. The work provides a clear roadmap for how to initialize these machines, how to manage the inevitable noise in the data, and how to build up complex solutions from simpler, well-understood parts. It transforms the challenge of finding the ground state from a task that becomes impossible with size into one that remains solvable, opening the door to simulating complex materials and chemical reactions with unprecedented accuracy.

Technical Summary: Quantum Folding in Chebyshev-Based Quantum Subspace Diagonalization

1. Problem Statement

Quantum Subspace Diagonalization (QSD), also known as the quantum Krylov method, estimates the ground-state energy of a Hamiltonian $H$ by projecting it onto a low-dimensional subspace generated on a quantum computer and solving the resulting generalized eigenvalue problem classically. While variants using Chebyshev polynomials offer exact realization without Trotter errors, their theoretical guarantees—specifically convergence rates, noise floors, and the cost of state preparation—are critically dependent on the initial overlap $|\gamma_0|^2 = |\langle g | \psi_0 \rangle|^2$ between the reference state $|\psi_0\rangle$ and the true ground state $|g\rangle$ .

For any fixed product state (e.g., the Néel state for antiferromagnets), this overlap decays exponentially with system size for non-trivial Hamiltonians. This "orthogonality catastrophe" renders conventional QSD impractical for large systems, as the required measurement budget and circuit depth scale exponentially to resolve the ground-state signal above the noise floor. The paper identifies the choice of reference state as the primary bottleneck determining whether QSD scales.

2. Methodology: Quantum Folding

The authors propose Quantum Folding, a strategy to construct a reference state that avoids exponential overlap decay for local Hamiltonians.

The Folded State: For a system partitioned into subsystems $A$ and $B$ with a local Hamiltonian $H_\Lambda = H_A \otimes \mathbb{1}_B + \mathbb{1}_A \otimes H_B + H_\partial$ (where $H_\partial$ represents interface terms), the folded reference state is defined as $|\psi_0\rangle = |g_A\rangle \otimes |g_B\rangle$ , the product of the exact ground states of the isolated subsystems.
Physical Justification: This state is the exact ground state of the full Hamiltonian with interface terms removed. Consequently, its difference from the true ground state is confined to the interface. For a 1D chain cut into halves, the interface is a single bond. The overlap is therefore governed by the physics of this single bond rather than the bulk volume, remaining size-independent for gapped systems and decaying only as a slow power law for gapless systems.
Recursive Ladder: The method employs a recursive construction ( $L \to 2L \to 4L \dots$ ). The output of a QSD calculation at size $L$ is prepared as a physical state and used as the seed for the QSD calculation at size $2L$ .

3. Key Contributions and Theoretical Developments

A. Error Bounds for Chebyshev QSD
The authors transplant the effective-Hamiltonian analysis (previously developed for real-time evolution) to the Chebyshev-Krylov basis.

Linear Noise Scaling: They derive a non-asymptotic error bound for the ground-state energy that is linear in the measurement noise ( $\eta$ ), improving upon previous sublinear scalings.
Filter Norm Control: A critical component is the construction of a Chebyshev low-energy filter with a coefficient vector of $O(1)$ Euclidean norm, ensuring that noise and thresholding errors are not exponentially amplified.
Rescaling Dependence: The analysis reveals that the block-encoding rescaling of the Hamiltonian is a parameter of the guarantee. Placing the ground state at the edge of the Chebyshev interval yields a convergence exponent proportional to $\sqrt{\Delta}$ (gap), whereas interior placement (typical of fixed block-encoding normalizations) yields a linear dependence on $\Delta$ .

B. State Preparation and Post-Selection
The paper addresses the challenge of converting the classical Ritz vector (a coefficient vector in a non-orthogonal basis) back into a physical quantum state.

Implementation: The state is prepared via Linear Combination of Unitaries (LCU) or Quantum Singular Value Transformation (QSVT) using powers of the walk operator.
Thresholding as a Floor: The authors prove that the thresholding required to stabilize the energy eigenvalue problem also imposes a lower bound (floor) on the post-selection probability. Without thresholding, noise in the near-null subspace of the overlap matrix leads to spurious coefficient vectors with vanishingly small success probabilities.
Certified Filters: To ensure the ground-state generation probability is not suppressed by poor filter efficiency, the authors propose using a "certified filter" (a polynomial constructed classically from the Ritz values) or $\ell_1$ gauge fixing. This ensures the generation probability scales as $\Theta(|\gamma_0|^2)$ , making the recursive folding ladder scalable.

C. Scaling Analysis

Gapped Systems: For gapped chains, the fold fidelity $F(L) = |\langle g_{2L} | g_L \otimes g_L \rangle|^2$ converges to a non-zero constant $F_\infty$ . The total cost of the recursive ladder scales polynomially with system size $L$ .
Gapless Systems: For gapless spin-1/2 chains, the overlap decays as a slow power law $L^{-0.107}$ . The resulting cost is quasi-polynomial, effectively polynomial for realistic system sizes.

4. Results

The authors validate the method through exact numerical simulations of antiferromagnetic Heisenberg chains ( $S=1/2$ and $S=1$ ) and the AKLT chain.

Overlap Advantage: The folded reference state maintains an overlap $O(1)$ (or slow power-law decay), whereas the Néel state overlap decays exponentially. For $S=1/2$ chains up to 24 sites, the folded overlap remains $>0.5$ , while the Néel overlap is $\sim 10^{-3}$ .
Convergence Speed: In noiseless simulations, the folded state reduces the Krylov dimension $D$ required for a given accuracy by a factor of 4–5 compared to the Néel state.
Noise Robustness: Under measurement noise, the folded state achieves errors 2–5 orders of magnitude lower than the Néel state. The Néel state's error saturates at a high floor determined by its tiny overlap, while the folded state's error continues to decrease with increasing $D$ .
Fold Fidelity:
- Gapless Spin-1/2: $F(L) \approx 1.09 L^{-0.107}$ (Open BC).
- Gapped Spin-1 (Heisenberg): $F(L) \to F_\infty \approx 0.22$ .
- AKLT Chain: $F(L) \to 1/4$ exactly, with closed-form convergence derived from the matrix-product structure.

5. Significance and Claims

The paper claims to provide a scalable framework for quantum ground-state preparation that circumvents the exponentially small overlaps associated with conventional product-state initialization.

Scalability: By replacing the bulk-dependent overlap with an interface-dependent one, quantum folding transforms the scaling of QSD from exponential to polynomial in system size.
Universality: The approach is not limited to 1D chains or spins; it applies to any local Hamiltonian where the interface is small compared to the subsystems.
Practicality: The method relies on standard block-encoding and qubitization techniques but introduces specific protocols (recursive folding, certified filtering, and thresholding) to manage the trade-offs between convergence, noise, and state preparation costs.
Theoretical Insight: The work clarifies the role of block-encoding rescaling in Chebyshev methods, demonstrating that interior placement of the ground state incurs a significant penalty in resolving power and filter degree that cannot be recovered by post-processing.

The authors conclude that quantum folding, combined with the derived error bounds and state preparation techniques, offers a viable path toward scalable quantum simulation of correlated many-body systems on fault-tolerant quantum computers.