Answers are grounded in this paper and cite their sources.
Singular fractional-order Lane–Emden equations, which govern the internal structure of self-gravitating polytropic gas spheres, are difficult to approximate near the origin owing to their intrinsic coefficient singularity. This work proposes a single-layer, mesh-free neural architecture in which the hidden layer is replaced by a functional expansion over shifted Fibonacci polynomials satisfying a two-term recurrence on the unit interval. The network output is a strictly linear combination of these basis functions, so the trial solution is polynomial in the trainable weights and admits fully analytical gradients; initial conditions are imposed exactly through the trial-solution construction, and training proceeds without labelled data by minimizing the governing-equation residual at a finite set of collocation points. Closed-form and recurrence expressions are derived for the conformable fractional derivatives of the basis, reducing their evaluation cost from quadratic to linear in the number of basis functions, while an Adam optimizer with gradient clipping is shown to overcome the ill-conditioning inherent to the non-orthogonal basis. A universal approximation theorem, an a priori error bound, a stability estimate, and a convergence guarantee for the training dynamics are established. On four singular linear and nonlinear conformable Lane–Emden problems of varying fractional order, the method attains a maximum absolute error of 1.4949e-4 against the known integer-order analytical solution and matches an existing Chebyshev-neural-network benchmark, using substantially fewer trainable parameters.
No takes yet. Share an insight, caveat, or question.
Hama et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: