Hamiltonian-reconstruction distance as a success metric for the Variational Quantum Eigensolver

This article proposes the Hamiltonian reconstruction distance as a practical success metric for the Variational Quantum Eigensolver (VQE) and validates it by demonstrating through simulations and cloud-based trapped-ion experiments that it effectively assesses solution quality and prevents erroneous premature termination without requiring prior knowledge of the true ground state.

The Gist

The Big Problem: Guessing the Bottom

Imagine you are trying to find the deepest point of a dark, foggy valley. This valley represents a complex quantum system, and the deepest point represents the "ground state" (the most stable state with the lowest energy). You have a robot (the Variational Quantum Eigensolver or VQE) that can walk through the valley and tell you its height at any location.

The robot's goal is to find the absolute lowest point. It does this by taking steps, checking the height, and adjusting its path to go deeper.

The Catch: You have no map and do not know where the true lowest point actually is. You only know the robot's current height.

Normally, the robot stops when it feels it cannot go any lower. It says, "Okay, I am stuck; I must be at the bottom." But here lies the danger: The robot could get stuck on a small, flat patch of grass (a local minimum) that looks like the bottom but is not. If you stop too early, you think you have found the solution, but you are actually still stuck on a hill.

The New Tool: The "Shapeshifter" Test

The authors of this paper propose a new method to check whether the robot has truly found the real bottom without needing a map. They call this the Hamiltonian-Reconstruction (HR) Distance.

Here is the analogy:
Imagine the valley has a very specific, unique shape defined by a set of rules (the Hamiltonian). The robot tries to mimic this shape.

1. The Old Way: You only look at the robot's height (energy). When the height stops decreasing, you assume it is finished.
2. The New Way (HR Distance): You ask the robot: "Based on where you are standing right now, what do you think are the rules of this valley?"
   • The robot analyzes its surroundings and tries to reconstruct the rules that created the valley.
   • Then you compare the rules the robot guessed with the actual rules of the valley.
   • The Metric: If the robot is standing at the true bottom, its estimate of the rules will be perfect. The "distance" between its estimate and the truth will be zero.
   • If the robot gets stuck on a fake flat patch, its estimate of the rules will be wrong. The "distance" will be large, even if the robot thinks it is finished because the height is not changing.

What They Did

The researchers tested this idea on two specific types of quantum puzzles (called Spin Models), using both a real quantum computer (a cloud-based trapped-ion machine from IonQ) and computer simulations.

• The Test: They let the robot (VQE) run to find the bottom of the valley.
• The Result: In several cases, the robot's height (energy) stopped changing, making it look as if it were finished. However, the HR Distance was still high. This told the researchers: "Hey, the robot thinks it is finished, but it is actually still stuck on a fake patch. Keep going!"
• The Correlation: They found that the HR Distance decreased as the robot got closer to the true bottom. It acted like a reliable "progress indicator" that did not lie.

Important Limitations (The Fine Print)

The paper emphasizes very carefully that this tool is not magic. It works best under certain conditions:

• The Gap is Crucial: The valley must have a clear "drop" between the bottom and the next higher step. If the bottom is too flat or too close to the next step, the test gets confused.
• Noise is Crucial: Real quantum computers are "noisy" (like a radio with static). If the noise is too loud, the robot's estimate of the rules becomes blurred, and the HR Distance metric loses its accuracy.
• It Needs Practice: The robot must approach the end before this test is useful. If you check it at the very beginning of the run, the test might give a false sense of security.

The Conclusion

The paper claims that the Hamiltonian-Reconstruction Distance is a useful new "check engine light" for quantum computers.

Instead of just asking: "Are we deep enough?" (which can be misleading), it asks: "Do we understand the shape of the problem we are solving?" If the answer is "No," the computer knows it must keep searching, even if the energy values look as if it is finished. This helps prevent the algorithm from stopping too early and providing a wrong answer.

Technical Summary

Technical Summary: Hamiltonian-Reconstruction Distance as a Success Metric for the Variational Quantum Eigensolver

Problem Statement
The Variational Quantum Eigensolver (VQE) is a leading hybrid quantum-classical algorithm for simulating quantum systems on current noisy intermediate-scale quantum (NISQ) hardware. Its primary goal is to heuristically find the ground state and ground state energy of a target Hamiltonian. A critical challenge in VQE, shared by other heuristic ground-state search algorithms, is determining the quality of the output solution when the true ground state and its energy are unknown.

Standard stopping criteria often rely on the energy stabilizing (i.e., the energy no longer changes rapidly between iterations). However, the authors note that this can lead to erroneous premature termination, where the algorithm halts at a suboptimal local minimum or a "barren plateau" (a barren plateau region), and falsely interprets energy stagnation as convergence to the true ground state. Existing metrics, such as the Hamiltonian variance, often require case-by-case interpretation or prior knowledge of the ground state to be effective.

Methodology: Hamiltonian-Reconstruction (HR) Distance
To address the lack of a reliable, ground-state-agnostic metric, the authors propose and investigate the Hamiltonian-Reconstruction (HR) distance, denoted as $\Delta(\bar{\theta})$. This metric is derived from the theoretical framework of Hamiltonian reconstruction, which infers a parent Hamiltonian from an eigenstate.

The methodology proceeds as follows:

1. Operator Selection: A set of operators $\{\hat{H}_i\}$ is chosen such that their linear span contains the true target Hamiltonian $\hat{H}_{\text{True}}$. For the specific models investigated (1D Transverse-Field Ising Model and 2D $J_1-J_2$ TFIM), the authors select operators corresponding to the terms present in the true Hamiltonian (e.g., $\sum \sigma^x_i$ and $\sum \sigma^z_i \sigma^z_{i+1}$).
2. Measurement: During VQE optimization, the expectation values of these operators and their correlators are measured to construct a covariance matrix $Q(\bar{\theta})$.
3. Reconstruction: The eigenvector corresponding to the smallest eigenvalue of this covariance matrix provides the coefficients $\{\tilde{c}_i\}$ for a reconstructed Hamiltonian $\hat{H}_R = \sum \tilde{c}_i \hat{H}_i$. If the variational state is an exact eigenstate of $\hat{H}_{\text{True}}$, then $\hat{H}_R$ should equal $\hat{H}_{\text{True}}$.
4. Metric Calculation: The HR distance is defined as the $L_2$ distance between the coefficient vector of the reconstructed Hamiltonian and the known coefficient vector of the true Hamiltonian:
   $$\Delta(\bar{\theta}) = \| \vec{\tilde{c}}(\bar{\theta}) - \vec{c} \|_2$$
   Crucially, this calculation requires no knowledge of the true ground state energy or the true ground state wavefunction, but only the structure of the Hamiltonian itself.

Experimental and Simulation Setup
The authors evaluated the HR distance metric through:

• Cloud-based Hardware Experiments: Using the IonQ Harmony trapped-ion quantum computer.
  • System 1: 11-qubit 1D Transverse-Field Ising Model (1D-TFIM).
  • System 2: 6-qubit 2D $J_1-J_2$ Transverse-Field Ising Model ($3 \times 2$ lattice).
  • Ansatz: An Alternating Layered Ansatz (ALA) was used to minimize the number of CNOT gates and reduce errors in state preparation.
• Numerical Simulations: Comprehensive simulations were conducted to analyze the effects of depolarizing noise (1- and 2-qubit gate errors) and shot noise on the HR distance and its correlation with fidelity.

Key Results

1. Diagnosis of Premature Convergence: In both 1D and 2D models, the HR distance successfully identified cases where the VQE energy appeared to have converged (stabilized), yet the solution was suboptimal. In these scenarios, the HR distance remained significantly non-zero (e.g., fluctuating around 0.4 in the 11-qubit simulation) and correctly indicated that the state was not the ground state, while the energy metric signaled convergence.
2. Correlation with Fidelity and Energy:
   • The HR distance showed a strong negative correlation with the fidelity (state overlap) to the true ground state as the VQE approached convergence.
   • It showed a positive correlation with the energy difference between the VQE solution and the true ground state.
   • These correlations held under specific conditions: (1) the VQE optimization was near convergence, (2) the energy gap between the ground state and the first few excited states was sufficiently large, and (3) gate fidelities were sufficiently high.
3. Noise Sensitivity:
   • Gate Errors: As the probabilities for 1- and 2-qubit gate errors increased, the correlation between the HR distance and fidelity deteriorated. At high noise levels, the HR distance became less reliable as a sole indicator of fidelity.
   • Shot Noise: A large number of shots (approximately $10^4$) was required to achieve an adequate signal-to-noise ratio for the HR distance measurement.
   • Energy Gaps: When the energy gap between the ground state and the first excited state was small, the HR distance lost its specificity and also showed negative correlations with the fidelity to excited states, complicating the diagnosis.
4. Comparison with Variance: The authors compared the HR distance with the Hamiltonian variance. While the variance also decreases as the solution approaches the ground state, its range is Hamiltonian-dependent, making it difficult to establish universal stopping thresholds. In contrast, the HR distance is normalized between 0 and 1 and offers a more consistent metric across different Hamiltonians.

Significance and Claims
The work claims that the Hamiltonian-Reconstruction distance provides a practical, experimentally accessible metric for evaluating the success of VQE algorithms on NISQ hardware without requiring prior knowledge of the true ground state or ground state energy.

• Diagnostic Utility: It serves as a diagnostic tool to prevent erroneous premature stopping of VQE iterations, particularly in cases where energy plateaus are misleading.
• Efficiency: The metric is efficient to compute and requires only a polynomial number of measurements relative to the number of terms in the Hamiltonian.
• Future Application: The authors suggest that the HR distance could be integrated into multi-objective cost functions for VQE optimization. By simultaneously minimizing energy and HR distance, VQE could find states with higher fidelity to the ground state, even when the energy is identical, and could thus potentially mitigate issues such as barren plateaus.

The authors maintain a modest tone and acknowledge that the reliability of the metric depends on hardware fidelity and energy gaps, and that it should be regarded as a supplementary diagnostic tool rather than a perfect replacement for knowledge of the true ground state.