We study the problem of approximating the principal square root A1/2 of a nonsingular matrix A ∈ ℂn×n with no eigenvalues on ℝ≤0. Classical iterative methods-Newton, Denman-Beavers, Halley, productform, Padé, and the Schur method of Björck-Hammarling-each require O(n3) work per step and may converge slowly near the branch cut. We cast the map F : A 7→ A1/2 as an empirical risk minimization problem over deep ReLU networks, establish a Rademacher-complexity generalization bound, and show that the principal square-root map is qualitatively approximable by ReLU networks on the compact spectral band An,α,β. Two hybrid algorithms are proposed in which a trained network warm-starts Newton or Denman-Beavers refinement; their iteration savings follow directly from the quadratic convergence of Newton's method without new convergence proofs.