arXiv:2609.19239
Lazy training of quantum physics informed neural networks
Anderson Melchor Hernandez, Giacomo De Palma
Abstract
We study the gradient-flow training dynamics of quantum physics-informed neural networks (QPINNs) for the solution of second-order elliptic partial differential equations with Dirichlet boundary conditions. We consider parameterized quantum circuits as function approximators and analyze their overparameterized regime through the lens of the neural tangent kernel (NTK). Our contribution is a nonasymptotic lazy-training theory for QPINNs and their variational formulation: we prove that, for sufficiently large circuit width, the nonlinear gradient flow is quantitatively approximated by a linearized NTK model, with explicit bounds depending on the number of qubits, circuit depth, circuit light-cone geometry, and the dimension of the domain of the solution to the PDE.
AI Summary
The paper studies when gradient-flow training of quantum physics-informed neural networks (QPINNs) for second-order, variable-coefficient elliptic PDEs with Dirichlet boundary conditions is approximated by a frozen neural tangent kernel (NTK). It treats both pointwise PDE-residual losses and variational losses using a finite-dimensional test-function space. The circuit model has $m$ qubits, $L$ layers, independently initialized parameters, and a normalized sum of single-qubit observables as output. Under spatial-regularity assumptions and a quantitative lower bound on the expected training-kernel’s smallest eigenvalue, the main results give high-probability exponential decay of empirical loss, bounds on parameter displacement, and uniform-in-time nonlinear-versus-linearized prediction errors. The displayed error bounds contain width/depth factors $L^{12}m^2|\mathcal M|^5|\mathcal N|^2/[b_K(m)^2b(m)^5]$ for collocation and $L^8$ in place of $L^{12}$ for the variational formulation, alongside logarithmic and problem-dependent factors. Here $\mathcal M,\mathcal N$ describe circuit light cones. Sufficient kernel-concentration conditions scale as $mL^9|\mathcal M|^4|\mathcal N|^2/b(m)^4\to0$, improved to $L^5$ variationally. The analysis combines spatial-derivative estimates, parameter-shift identities, light-cone counting, McDiarmid concentration, and a bootstrap controlling kernel drift and parameter motion. Additional covariance and kernel-limit assumptions yield finite-dimensional Gaussian-process limits. The results concern continuous-time optimization of finite empirical losses, not approximation guarantees for the true PDE solution or quantum computational advantage. Function-space convergence remains open, as do quantitative sampling-versus-width guarantees, evolutionary PDEs, and extensions to mixed, time-dependent, or noisy boundary data.
Central claim may not hold: The variational lazy-training theorem claims a finite discrepancy bound uniformly over an unbounded linear test-function space, although its discrepancy scales linearly with the test function. The collocation discrepancy proof drops the factor b_K(m) multiplying the normalized NTK in the output dynamics, invalidating its derivation of the stated explicit normalization dependence. The variational concentration threshold has A_0 sqrt(vol(B)) where the displayed tail inequality requires A_0^2 vol(B), leaving its claimed high-probability guarantee unsupported for allowed coefficient and domain sizes.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.