A mathematical framework for improved weight initialization of neural networks using Lagrange multipliers