New Algorithm Solves Quantum States Without Costly Penalties
Chemistry

New Algorithm Solves Quantum States Without Costly Penalties

By reparameterizing neural wavefunctions, scientists have bypassed long-standing computational bottlenecks in quantum chemistry calculations.

By Bilal Abbasi
Published:
Email this Article
New Algorithm Solves Quantum States Without Costly Penalties
Photo by Brecht Corbeel on Unsplash

Researchers have successfully adapted a low-rank approximation framework, known in machine learning as LoRA, to solve the electronic Schrödinger equation for multiple quantum excited states of atoms without requiring complex penalty weights or costly matrix inversions. The method computes accurate energy levels for the first-row atoms lithium through neon, matching the precision of existing state-of-the-art computational techniques while drastically cutting down computational overhead. Full details of the new approach are published on the arXiv preprint server.

Overcoming the Computational Bottlenecks in Quantum Chemistry

Computing the electronic structure of molecules involves solving the Schrödinger equation, a task that scales exponentially with the number of electrons due to the curse of dimensionality. While neural-network variational Monte Carlo methods have achieved remarkable precision for the ground states of small molecules using advanced architectures like FermiNet and Psiformer, extending this success to excited states has remained an uphill battle. Until now, scientists relied on two main paradigms, each carrying severe limitations.

Penalty-based methods add artificial overlap penalties to encourage orthogonality between different energy states, but these require tuning delicate penalty weights that depend on unknown energy gaps. Alternatively, natural excited states variational Monte Carlo transforms the problem into a grand extended system, successfully avoiding penalties but introducing a massive computational wall time. Evaluating local energy in that framework demands a quadratic number of Laplacian calculations per optimization step, scaling poorly as the number of target states grows.

Figure 3: Carbon runs of Section 4.4, with Eref,k=0E_{\mathrm{ref},k}=0 or with the shift Eref,k=−λk+5​σkE_{\mathrm{ref},k}=-\lambda_{k}+5\sigma_{k}, for training seed 11 (a, c) and seed 22 (b, d). (a, b) Batch mean of the clipped local energy of the five networks of the D1{}^{1}\!D term; the thick line is the network furthest above the others with Eref,k=0E_{\mathrm{ref},k}=0. (c, d) Normalized overlap from the unclipped batch mean, averaged over the 1010 pairs of these networks. Every curve is a mean over a sliding window of 100100 steps, taken before the absolute value in (c, d).
Carbon runs of Section 4.4, with Eref,k=0E_{\mathrm{ref},k}=0 or with the shift Eref,k=−λk+5​σkE_{\mathrm{ref},k}=-\lambda_{k}+5\sigma_{k}, for training seed 11 (a, c) and seed 22 (b, d). (a, b) Batch mean of the clipped local energy of the five networks of the D1{}^{1}\!D term; the thick line is the network furthest above the others with Eref,k=0E_{\mathrm{ref},k}=0. (c, d) Normalized overlap from the unclipped batch mean, averaged over the 1010 pairs of these networks. Every curve is a mean over a sliding window of 100100 steps, taken before the absolute value in (c, d). Credit: Minchan Jeong et al. / arXiv

How Nested Low-Rank Approximation Changes the Equation

To resolve this trade-off, a team of researchers led by Minchan Jeong, J. Jon Ryu, Se-Young Yun, and Gregory W. Wornell turned to a mathematical principle called low-rank approximation. Originally developed to decompose operators in low-dimensional systems, the low-rank approximation objective was previously incompatible with variational Monte Carlo because it relied on intractable high-dimensional integral norms.

The study authors overcame this hurdle through a straightforward reparameterization. By representing each wavefunction as a normalized neural network multiplied by a trainable scale parameter, they made the objective fully compatible with the probability density sampling used in Monte Carlo algorithms. Furthermore, they incorporated a sequential nesting technique that recovers ordered energy states one by one without requiring post-training matrix diagonalizations.

The resulting algorithm, named NestedLoRA-VMC, requires only a linear number of Laplacian evaluations per training step, matching the efficiency of penalty-based methods while remaining entirely penalty-free. The authors also introduced asymmetric clipping strategies to handle heavy-tailed statistical noise near nodal surfaces and deployed a per-network operator shift to stabilize training convergence.

Figure 4: Overlap ratio R=ψℓ/ψkR=\psi_{\ell}/\psi_{k} of networks kk and ℓ\ell of the Li ansatz, on the walkers of ψk\psi_{k}. Panels (a) and (b) are at training step 1,0001{,}000 and number the networks by ascending energy at that step; panels (c) and (d) are at step 100,000100{,}000 and number them as in Table 9. Dotted lines in (a) mark the clip widths of (b); error bars in (b) are standard errors over six batches; the axis of RR in (d) is logarithmic beyond ±0.01\pm 0.01. Panel (b) uses the restarted chains (Appendix F.2.2).
Overlap ratio R=ψℓ/ψkR=\psi_{\ell}/\psi_{k} of networks kk and ℓ\ell of the Li ansatz, on the walkers of ψk\psi_{k}. Panels (a) and (b) are at training step 1,0001{,}000 and number the networks by ascending energy at that step; panels (c) and (d) are at step 100,000100{,}000 and number them as in Table 9. Dotted lines in (a) mark the clip widths of (b); error bars in (b) are standard errors over six batches; the axis of RR in (d) is logarithmic beyond ±0.01\pm 0.01. Panel (b) uses the restarted chains (Appendix F.2.2). Credit: Minchan Jeong et al. / arXiv

Testing the Limits on First-Row Atoms

The research team evaluated their algorithm on the first-row atoms lithium, beryllium, boron, carbon, nitrogen, oxygen, fluorine, and neon, targeting the lowest ten eigenstates for each element. Using Psiformer neural network architectures, the per-atom mean absolute errors against experimental term centroids from the National Institute of Standards and Technology stayed below standard chemical accuracy thresholds.

On carbon atoms, the team demonstrated that their operator shift successfully controlled curvature imbalances, bringing degenerate energy levels into tight alignment. Despite these successes, the authors note certain limitations. The current validation is restricted to first-row atoms, and scaling the approach to larger molecules remains bottlenecked by the inherent computational cost of calculating second-order derivatives for multi-electron systems. Future work aims to incorporate advanced optimization techniques, such as Wasserstein gradient flows, to push quantum neural network calculations even further.

The study was posted to the arXiv preprint server on November 16, 2026.

Fact Checked

This article has been fact checked for accuracy, with information verified against reputable sources. Learn more about us and our editorial process.

Last reviewed on .

Article history

  • Latest version

Reference(s)

  1. Jeong, Minchan. “Variational Monte Carlo for Quantum Excited States via Nested Low-Rank Approximation.” arXiv.org <http://arxiv.org/abs/2611.21434v1>.

Cite this page:

Abbasi, Bilal. “New Algorithm Solves Quantum States Without Costly Penalties.” BioScience. BioScience ISSN 2521-5760, 17 November 2026. <https://www.bioscience.com.pk/en/subject/chemistry/variational-monte-carlo-for-quantum-excited-states-via-nested-low-rank-approximation>. Abbasi, B. (2026, November 17). “New Algorithm Solves Quantum States Without Costly Penalties.” BioScience. ISSN 2521-5760. Retrieved November 17, 2026 from https://www.bioscience.com.pk/en/subject/chemistry/variational-monte-carlo-for-quantum-excited-states-via-nested-low-rank-approximation Abbasi, Bilal. “New Algorithm Solves Quantum States Without Costly Penalties.” BioScience. ISSN 2521-5760. https://www.bioscience.com.pk/en/subject/chemistry/variational-monte-carlo-for-quantum-excited-states-via-nested-low-rank-approximation (accessed November 17, 2026).
End of the article