The many-electron Schrodinger equation is the fundamental description of quantum chemistry. Solve it exactly and you know everything about a molecule: its geometry, its energy, its reactivity. But exact solutions are computationally intractable beyond a handful of electrons — the wavefunction lives in a space whose dimension grows exponentially with electron count. Every practical quantum chemistry method is an approximation.
Wu et al. (arXiv: 2603.23897) introduce a neural network architecture for this problem and prove it has spectral convergence — meaning the error decreases faster than any polynomial of the basis size. The architecture, SOG-TNN (sum-of-Gaussians tensor neural network), represents the wavefunction as a tensor product of Gaussian-based building blocks. Model reduction techniques compress the representation. The Slater determinant ansatz enforces antisymmetry (the requirement that the wavefunction changes sign when any two electrons are exchanged).
Spectral convergence means that doubling the basis size doesn't just halve the error — it crushes it exponentially (mixed algebraic-exponential decay). This is the gold standard for numerical methods. Polynomial convergence means more work buys proportionally more accuracy. Spectral convergence means more work buys disproportionately more accuracy, with each additional basis function doing more than the last.
The through-claim: the right representation makes the hard problem tractable. The many-electron Schrodinger equation hasn't gotten easier — the exponential scaling is fundamental. But the SOG-TNN architecture exploits the fact that physical wavefunctions are not arbitrary functions in the exponentially large space. They have structure — smoothness, antisymmetry, Gaussian-like decay — that can be captured by a compact tensor network. The curse of dimensionality is a property of the space, not the solution.
Wu, Zhou, Zheng, Xie & Xu, 2603.23897. Quantum chemistry / neural networks / Schrodinger equation / spectral convergence / tensor decomposition.