Quantum EigenGame for excited state calculation

David A. Quiroga, Jason Han, Anastasios Kyrillidis

TL;DR

This paper adapts EigenGame—a game-theoretic framework originally designed for decentralized Principal Component Analysis (PCA)—to quantum computers for calculating molecular excited states. By replacing ad-hoc deflation penalties with an energy-aware, self-normalizing objective, the proposed QuantumGame finds excited states without hyperparameter tuning or matrix deflation. The authors also prove convergence and error accumulation rates under noisy, derivative-free optimization settings.

What Problem Does This Paper Address?

Calculating the ground and excited states of molecular Hamiltonians is a core challenge in quantum chemistry and materials science, governing everything from photochemical reactions to electronic conductivity. The standard quantum approach, the Variational Quantum Eigensolver (VQE), natively finds only the ground state (the lowest eigenvalue).

To calculate higher "excited" energy levels, existing variational methods generally rely on deflation:

  1. Explicit deflation, which modifies the Hamiltonian directly, requiring expensive decomposition into long sums of Pauli observables.
  2. Implicit deflation (such as Variational Quantum Deflation, or VQD), which adds an orthogonality penalty to the cost function. However, VQD introduces an arbitrary penalty weight ($\beta$) that requires delicate hyperparameter tuning; improper tuning can severely slow down convergence or yield invalid states.

Additionally, near-term quantum processors (NISQ hardware) only provide noisy, zeroth-order (function value) evaluations, making optimization fragile and error propagation between successive states difficult to quantify.

Methods

The authors build on EigenGame, where calculating $k$ eigenvectors is formulated as a competitive game among $k$ "players" (each seeking to capture variance while penalizing alignment with earlier players):

  • Parameterized Quantum Formulation (QuantumGame): Each player $r$ optimizes the parameters $\theta^{(r)}$ of a quantum ansatz $U(\theta^{(r)})$ to maximize: $$\langle \psi(\theta^{(r)}) | M | \psi(\theta^{(r)}) \rangle - \sum_{j < r} \frac{|\langle \psi(\theta^{(r)}) | M | \psi(\theta^{(j)}) \rangle|^2}{\langle \psi(\theta^{(j)}) | M | \psi(\theta^{(j)}) \rangle}$$ Because the penalty is automatically normalized by the parents' calculated energies, it is completely adaptive and eliminates the tunable penalty weight $\beta$ used in VQD.

  • Efficient Cross-Term Circuit: Evaluating the overlap term $\langle \psi(\theta^{(r)}) | M | \psi(\theta^{(j)}) \rangle$ typically requires a Swap Test (demanding $2q+1$ qubits and restricted unitaries). The authors propose an interference-based circuit using a single ancilla qubit ($q+1$ qubits total) and Hadamard/Phase gates to evaluate both real and imaginary parts using two Hamiltonian measurements.

  • Theoretical Guarantees: The authors analyze both a classical 0th-order finite-difference EigenGame and the parameterized quantum version. They prove:
  • Global convergence bounds using Riemannian descent on manifolds.
  • That parental angular errors $\epsilon$ propagate into child gradient errors bounded by $\mathcal{O}(\epsilon)$ for the 0th-order case and $\mathcal{O}(\epsilon \sqrt{\ell q})$ for the parameterized case (where $\ell$ is ansatz depth and $q$ is the number of qubits).

Datasets

  • Molecular Hamiltonian ($H_2$): A 2-qubit representation of the hydrogen molecule ($H_2$) generated via second quantization and a ParityMapper reduction.
  • Synthetic Covariance Matrices: Classical random matrices of sizes up to $p = 1024$ constructed via QR decomposition ($M = P^{-1} D P$), with non-degenerate eigenvalues sampled from a power-law distribution in $(0, 1)$ to benchmark scaling.

Key Insights and Ideas

  • Game Theory Replaces Deflation: Instead of altering the physical operator $M$ after each state is discovered, orthogonality is enforced dynamically through a Nash equilibrium.
  • Self-Scaling Penalty: Traditional VQD requires hand-tuning $\beta$. If $\beta$ is too small, the optimizer collapses into lower states; if too large, the landscape becomes jagged and slow to traverse. QuantumGame's denominator naturally scales the penalty by the parent's energy expectation, eliminating manual tuning overhead.
  • Hardware-Friendly Overlap Measurement: By preparing an equal superposition of two parameterized states with an ancilla, the off-diagonal expectation $\langle \psi_r | M | \psi_j \rangle$ is measured directly without needing double the register size.

Key Results

  • $H_2$ Molecule Benchmark: Under noiseless conditions and 10,000-shot statistical noise, QuantumGame successfully converged to the true ground and excited energy levels. Compared to VQD across multiple penalty values ($\beta \in [0.1, 5.0]$), QuantumGame achieved higher accuracy without hyperparameter searches.
  • Ablation on Dimension Scaling: Evaluating the 0th-order finite-difference EigenGame against exact analytical EigenGame for up to $k = 1024$ eigenvalues demonstrated nearly identical iteration-scaling trends, verifying that finite-difference noise does not hinder scaling.

Limitations and Future Work

  • Sequential Dependencies: Currently, players optimize in a directed acyclic hierarchy where child players wait for parents to converge. The authors note that exploring fully parallel implementations (e.g., EigenGameR across distributed quantum processors) is an open direction.
  • Scale of Quantum Experiments: Empirical quantum validation is currently demonstrated on a small 2-qubit system ($H_2$ with 3 layers). Testing on larger molecules with deeper ansatzes will be necessary to observe barren plateau effects and shot-noise degradation in higher dimensions.
  • Ansatz Expressibility Assumptions: Theoretical convergence guarantees rely on assuming a sufficiently expressive ansatz initialized within $\pi/4$ of the target eigenstate.

Summary updated

Abstract

Computing the excited states of a given Hamiltonian is computationally hard for large systems, but methods that do so using quantum computers scale tractably. This problem is equivalent to the PCA problem where we are interested in decomposing a matrix into a collection of principal components. Classically, PCA is a well-studied problem setting, for which both centralized and distributed approaches have been developed. On the distributed side, one recent approach is that of EigenGame, a game-theoretic approach to finding eigenvectors where each eigenvector reaches a Nash equilibrium either sequentially or in parallel. With this work, we extend the EigenGame algorithm for both a $0^\text{th}$-order approach and for quantum computers, and harness the framework that quantum computing provides in computing excited states. Results show that using the Quantum EigenGame allows us to converge to excited states of a given Hamiltonian without the need of a deflation step. We also develop theory on error accumulation for finite-differences and parameterized approaches.
BibTeX
@inproceedings{quiroga2025quantum,
  author = {David A. Quiroga and Jason Han and Anastasios Kyrillidis},
  title = {{Quantum EigenGame for excited state calculation}},
  booktitle = {CPAL (Proceedings Track)},
  year = {2025},
  url = {https://openreview.net/forum?id=XFjwzut6c5},
}

Loading related content...

Figure